Self-supervised learning for medical image classification: a systematic review and implementation guidelines
Open accessAbstract
Advancements in deep learning and computer vision provide promising solutions for medical image analysis, potentially improving healthcare and patient outcomes. However, the prevailing paradigm of training deep learning models requires large quantities of labeled training data, which is both time-consuming and cost-prohibitive to curate for medical images. Self-supervised learning has the potential to make significant contributions to the development of robust medical imaging models through its ability to learn useful insights from copious medical datasets without labels. In this review, we provide consistent descriptions of different self-supervised learning strategies and compose a systematic review of papers published between 2012 and 2022 on PubMed, Scopus, and ArXiv that applied self-supervised learning to medical imaging classification. We screened a total of 412 relevant studies and included 79 papers for data extraction and analysis. With this comprehensive effort, we synthesize the collective knowledge of prior work and provide implementation guidelines for future researchers interested in applying self-supervised learning to their development of medical imaging classification models.
AI snapshot
Written from the paper’s full text.
TL;DR
This systematic review synthesizes 79 studies applying self-supervised learning (SSL) to medical image classification and finds that SSL pre-training generally improves downstream performance over purely supervised or non-SSL baselines. It also provides terminology, implementation guidelines, and highlights that optimal SSL strategies remain uncertain due to heterogeneous studies and limited direct comparisons.
Key findings
- From 412 identified studies, 148 full-text screened, and 79 included; most SSL models improved performance, with relative increases of 0.216–32.6% AUROC, 0.440–29.2% accuracy, and 0.137–14.3% F1 over the same architecture without SSL pre-training, while five studies reported slight decreases of 0.980–4.51%.
- Contrastive learning was the most common strategy (44/79), with SimCLR, MoCo, and BYOL used in 13, 8, and 3 papers, respectively; 36 studies comparing contrastive SSL to supervised pre-training reported an average improvement of 6.35%.
- Innate relationship SSL was used in 15/79 studies; clinically relevant pre-text tasks achieved greater average improvement over supervised methods than transformation-based tasks (13.7% vs. 5.03%). Generative SSL appeared in 3/79, self-prediction in 6/79, and combined strategies in 11/79; all 8 combined-strategy studies that compared with purely supervised approaches reported improvement (0.140–8.29%).
- End-to-end fine-tuning tended to outperform training a classifier on extracted features; of 11 studies with quantitative comparisons, 8 favored end-to-end fine-tuning. Radiology was the most represented clinical domain (47/79), followed by pathology (12/79).
Methods
- PRISMA-based systematic review searching PubMed, Scopus, and ArXiv for papers published between January 2012 and May 2021, using search terms combining “self-supervised learning” and “medical imaging.”
- Inclusion required English, peer-reviewed original research applying SSL to a downstream clinical medical image classification task; exclusions included semi-supervised learning, any manual labels during self-supervision, regression- or segmentation-only downstream tasks, and non-human medical imaging.
- Screening used Covidence by three independent researchers, with disagreements resolved by discussion and a fourth arbitrator if needed. Data extraction recorded SSL strategy, modality, clinical domain, task, framework, fine-tuning strategy, metrics, SSL and supervised performance, and relative differences; AUROC was prioritized, then accuracy, F1, and sensitivity.
Limitations
- Stated by the paper: publication bias may overestimate SSL benefits; the post-2012 search excluded pre-deep-learning SSL work; studies could not be aggregated or statistically compared because of different modalities, metrics, and objectives; categorizing SSL strategies introduced subjectivity; and the review was limited to classification, excluding segmentation, regression, and registration.
- Evident from the text: only 33/79 studies compared with another SSL strategy, only 18/79 used natural-image pre-trained weights, and only 13 compared feature extraction with end-to-end fine-tuning (two without numerical results), indicating limited baselines and ablation experiments.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
0 of 71 cited sources checked · 0 supported · 0 flagged
- Findingown data
79 studies were included for systematic review and data extraction.
“We screened a total of 412 relevant studies and included 79 papers for data extraction and analysis.”
- Also stated in: § Results · plain
“A total of 79 studies fulfilled our eligibility criteria and were included for systematic review and data extraction.”
- Also stated in: § Results · plain
“Out of 412 studies identified from PubMed, Scopus, and ArXiv, 79 studies were included in the systematic review.”
Quote in the paperSubstantial claim - Also stated in: § Results · plain
- Findingcitation
Nine of the innate relationship studies designed their pre-text task based on different image transformations, including rotation prediction, horizontal flip prediction, reordering shuffled slices, and patch order prediction.
“Nine of these studies designed their innate relationship pre-text task based on different image transformations, including rotation prediction44–47, horizontal flip prediction48, reordering shuffled slices49, and patch order prediction46,50–52.”
Quote in the paperNot a substantial claim0 of 6 sources checked- cites [44] Vats, A., Pedersen, M. & Mohammed, A. A Preliminary Analysis of Self-Supervision for Wireless Capsule Endoscopy. In 2021 9th European Workshop on Visual Information Processing (EUVIP) 1–6 (2021). source
Not checked in the free run.
- cites [45] Ewen, N. & Khan, N. Online Unsupervised Learning For Domain Shift In Covid-19 CT Scan Datasets. In 2021 IEEE International Conference on Autonomous Systems (ICAS) 1–5 (2021). source
Not checked in the free run.
- cites [46] Zhu, Y. Self-supervised Learning for Small Shot COVID-19 Classification. In 2021 3rd International Conference on Information Technology and Computer Communications 36–40 (2021). source
Not checked in the free run.
- cites [47] Long, Y. Pneumonia Identification with Self-supervised Learning and Transfer Learning. In Application of Intelligent Systems in Multi-modal Information Analytics 627–635 (2021). source
Not checked in the free run.
- cites [48] Ewen, N. & Khan, N. Targeted Self Supervision For Classification On A Small Covid-19 Ct Scan Dataset. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI) 1481–1485 (2021). source
Not checked in the free run.
- cites [49] Jiao, J., Droste, R., Drukker, L., Papageorghiou, A. T. & Alison Noble, J. Self-supervised Representation Learning for Ultrasound Video. In 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI) (2020). source
Not checked in the free run.
- Findingown data
The remaining six innate relationship studies designed clinically relevant pre-text tasks by exploiting the unique properties of medical images.
“The remaining six studies designed clinically relevant pre-text tasks by exploiting the unique properties of medical images.”
Quote in the paperNot a substantial claim - Findingown data
Innate relationship was used in 15 out of 79 included studies.
“Innate relationship was used in 15 out of 79 studies (Table 1).”
Quote in the paperNot a substantial claim - Findingown data
The systematic search identified 412 unique studies.
“We screened a total of 412 relevant studies and included 79 papers for data extraction and analysis.”
- Also stated in: § Results · plain
“A total of 412 unique studies were identified through our systematic search.”
- Also stated in: § Results · plain
“Out of 412 studies identified from PubMed, Scopus, and ArXiv, 79 studies were included in the systematic review.”
Quote in the paperNot a substantial claim - Also stated in: § Results · plain
- Findingown data
After removing duplicates and excluding studies based on title and abstract, 148 studies remained for full-text screening.
“After removing duplicates and excluding studies based on title and abstract using our study selection criteria (see Methods), 148 studies remained for full-text screening.”
Quote in the paperNot a substantial claim - Findingown data
On average, clinically relevant pre-text tasks achieved greater improvements in performance over transformation-based pre-text tasks when compared to purely supervised methods (13.7% vs. 5.03%).
“On average, clinically relevant pre-text tasks achieved greater improvements in performance over transformation-based pre-text tasks, when compared to purely supervised methods (13.7% vs. 5.03%).”
Quote in the paperSubstantial claim - Findingown data
Generative SSL was used in 3 out of 79 included studies.
“Generative SSL was used in 3 out of 79 studies (Table 1).”
Quote in the paperNot a substantial claim - Findingown data
The second most prevalent clinical domain was pathology (12/79).
“The second most prevalent clinical domain was pathology (12/79).”
Quote in the paperNot a substantial claim - Findingown data
Only 33 studies reported comparisons with another self-supervised learning strategy.
“Even though 60 out of 79 studies compared their results with purely supervised baselines, only 33 studies reported comparisons with another self-supervised learning strategy.”
Quote in the paperSubstantial claim - Findingown data
Over half of the combined-approach studies (6/11) combined contrastive with generative approaches.
“Over half of these studies (6/11) combined contrastive with generative approaches.”
Quote in the paperNot a substantial claim - Findingown data
Self-prediction was used in six out of all included studies.
“Self-prediction was used in six out of all included studies (Table 1).”
Quote in the paperNot a substantial claim - Findingown data
Two of the 13 studies comparing extracted features and end-to-end fine-tuning did not report numerical results.
“Lastly, merely 13 studies compared performance between classification on extracted features to end-to-end fine-tuning, two of which did not report numerical results.”
Quote in the paperNot a substantial claim - Findingown data
The majority of self-supervised pre-trained models led to a relative increase in AUROC of 0.216-32.6% over the same model architecture without SSL pre-training, including both ImageNet and random initialization.
“While five studies reported a slight decrease in performance (0.980–4.51%), the majority of self-supervised pre-trained models led to a relative increased performances of 0.216–32.6% AUROC, 0.440–29.2% accuracy, and 0.137–14.3% F1 score over the same model architecture without SSL pre-training, including both ImageNet and random initialization (Fig. 4e).”
Quote in the paperSubstantial claim - Findingcitation
The innate relationship tasks in combined studies range from augmentation prediction and patch positioning prediction, rotation prediction, and ultrasound video to speech correspondence prediction.
“The innate relationship tasks range from augmentation prediction and patch positioning prediction69, rotation prediction70, and ultrasound video to speech correspondence prediction53.”
Quote in the paperNot a substantial claim0 of 3 sources checked- cites [69] Tian, Y. et al. Constrained Contrastive Distribution Learning for Unsupervised Anomaly Detection and Localisation in Medical Images. Preprint at https://arxiv.org/abs/2103.03423 (2021). source
Not checked in the free run.
- cites [70] Li, X. et al. Rotation-Oriented Collaborative Self-Supervised Learning for Retinal Disease Diagnosis. IEEE Trans. Med. Imaging 40, 2284–2294 (2021). source
Not checked in the free run.
- cites [53] Jiao, J. et al. Self-Supervised Contrastive Video-Speech Representation Learning for Ultrasound. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2020 534–543 (2020). source
Not checked in the free run.
- Findingown data
The majority of self-supervised pre-trained models led to a relative increase in F1 score of 0.137-14.3% over the same model architecture without SSL pre-training, including both ImageNet and random initialization.
“While five studies reported a slight decrease in performance (0.980–4.51%), the majority of self-supervised pre-trained models led to a relative increased performances of 0.216–32.6% AUROC, 0.440–29.2% accuracy, and 0.137–14.3% F1 score over the same model architecture without SSL pre-training, including both ImageNet and random initialization (Fig. 4e).”
Quote in the paperSubstantial claim - Findingown data
The vast majority of studies explored the clinical domain of radiology (47/79).
“The vast majority of the studies explored the clinical domain of radiology (47/79), of which 27 were focused on investigating abnormalities on chest imaging such as pneumonia, COVID-19, pleural effusion and pulmonary embolism (see Table 1).”
Quote in the paperNot a substantial claim - Findingown data
Of the 11 studies that quantitatively reported performance, eight found end-to-end fine-tuning to outperform training a new classifier on extracted features.
“Of the 11 studies that quantitatively reported performance, eight found end-to-end fine-tuning to outperform training a new classifier on extracted features (Fig. 4d).”
Quote in the paperSubstantial claim - Findingown data
A majority of the included studies lacked strong baselines and ablation experiments.
“A majority of the included studies lacked strong baselines and ablation experiments.”
Quote in the paperSubstantial claim - Findingown data
A combination of contrastive and innate relationships was used in three studies.
“A combination of contrastive and innate relationships was used in three studies.”
Quote in the paperNot a substantial claim - Findingown data
Only 18 out of 79 studies indicated use of natural image pre-trained weights, either supervised or self-supervised, to initiate their model for subsequent in-domain self-supervised pre-training.
“Furthermore, only 18 out of 79 studies indicated use of natural image pre-trained weights, either supervised or self-supervised, to initiate their model for subsequent in-domain self-supervised pre-training.”
Quote in the paperSubstantial claim - Findingown data
The majority of studies remaining after full-text screening (44/79) used contrastive learning as their self-supervised pre-training strategy.
“The majority of the studies that remained after our full-text screening (44/79) used contrastive learning as their self-supervised pre-training strategy (Table 1).”
Quote in the paperNot a substantial claim - Findingown data
Eleven studies combined different self-supervised learning strategies to pre-train their medical imaging models.
“Eleven studies found creative ways to combine different self-supervised learning strategies to pre-train their medical imaging models (Table 1).”
Quote in the paperNot a substantial claim - Findingown data
The types of medical images utilized for model development and the downstream classification tasks encompassed a wide range of medical domains and applications.
“The types of medical images utilized for model development as well as the downstream classification task encompassed a wide range of medical domains and applications (Fig. 4a, b).”
Quote in the paperNot a substantial claim - Findingown data
Five studies reported a slight decrease in performance (0.980-4.51%).
“While five studies reported a slight decrease in performance (0.980–4.51%), the majority of self-supervised pre-trained models led to a relative increased performances of 0.216–32.6% AUROC, 0.440–29.2% accuracy, and 0.137–14.3% F1 score over the same model architecture without SSL pre-training, including both ImageNet and random initialization (Fig. 4e).”
Quote in the paperNot a substantial claim - Findingown data
Of the radiology studies, 27 focused on investigating abnormalities on chest imaging such as pneumonia, COVID-19, pleural effusion and pulmonary embolism.
“The vast majority of the studies explored the clinical domain of radiology (47/79), of which 27 were focused on investigating abnormalities on chest imaging such as pneumonia, COVID-19, pleural effusion and pulmonary embolism (see Table 1).”
Quote in the paperNot a substantial claim - Findingown data
60 out of 79 studies compared their results with purely supervised baselines.
“Even though 60 out of 79 studies compared their results with purely supervised baselines, only 33 studies reported comparisons with another self-supervised learning strategy.”
Quote in the paperSubstantial claim - Findingown data
13 studies compared performance between classification on extracted features and end-to-end fine-tuning.
“Lastly, merely 13 studies compared performance between classification on extracted features to end-to-end fine-tuning, two of which did not report numerical results.”
Quote in the paperSubstantial claim - Findingown data
Of the 33 studies, 26 compared with a self-supervised category that differs from their best performing model.
“Of the 33 studies, 26 compared with a self-supervised category that differs from their best performing model.”
Quote in the paperSubstantial claim - Findingown data
With the exception of Ke et al.'s work, all studies that combined contrastive with generative strategies utilized an autoencoder as their generative model.
“With the exception of Ke et al.’s work68, which uses a CycleGAN for histopathology slide stain normalization, all studies utilized an autoencoder as their generative model when combined with contrastive strategies.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [68] Ke, J., Shen, Y., Liang, X. & Shen, D. Contrastive Learning Based Stain Normalization Across Multiple Tumor in Histopathology. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2021 571–580 (Springer International Publishing, 2021). source
Not checked in the free run.
- Findingown data
Based on the aggregated results, end-to-end fine-tuning generally leads to better performance for medical images.
“Based on our aggregated results, we found that end-to-end fine-tuning generally leads to better performance for medical images (Fig. 4c).”
Quote in the paperSubstantial claim - Findingcitation
Some papers leveraged medical domain priors to create specialized strategies for creating positive pairs.
“Some papers leveraged medical domain priors to create specialized strategies for creating positive pairs.”
Quote in the paperNot a substantial claim - Findingown data
The majority of self-supervised pre-trained models led to a relative increase in accuracy of 0.440-29.2% over the same model architecture without SSL pre-training, including both ImageNet and random initialization.
“While five studies reported a slight decrease in performance (0.980–4.51%), the majority of self-supervised pre-trained models led to a relative increased performances of 0.216–32.6% AUROC, 0.440–29.2% accuracy, and 0.137–14.3% F1 score over the same model architecture without SSL pre-training, including both ImageNet and random initialization (Fig. 4e).”
Quote in the paperSubstantial claim - Findingown data
Self-supervised pre-training generally improves performance for medical imaging classification tasks over purely supervised methods.
“We found that self-supervised pre-training generally improves performance for medical imaging classification tasks over purely supervised methods.”
Quote in the paperSubstantial claim - Premisecitation
Work in computer vision has already demonstrated performance for natural images that is superior to that achieved by supervised models, especially in label-scarce scenarios.
“Such work in computer vision has already demonstrated performance for natural images that is superior to that achieved by supervised models, especially in label-scarce scenarios20.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [20] Chen, T., Kornblith, S., Norouzi, M. & Hinton, G. A Simple Framework for Contrastive Learning of Visual Representations. Proc. 37th Int. Conf. Mach. Learn., PMLR 119, 1597–1607 (2020).
Not checked in the free run.
- Premisecitation
Most medical imaging models underperform in their ability to generalize to external institutions or when repurposed for other tasks.
“most medical imaging models underperform in their ability to generalize to external institutions or when repurposed for other tasks16.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [16] Zech, J. R. et al. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: A cross-sectional study. PLoS Med. 15, e1002683 (2018). source
Not checked in the free run.
- Premisecitation
An average radiologist now needs to interpret one image every 3-4 s to keep up with clinical workloads.
“Studies have shown that an average radiologist now needs to interpret one image every 3–4 s to keep up with clinical workloads3–5.”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [3] McDonald, R. J. et al. The effects of changes in utilization and technological advancements of cross-sectional imaging on radiologist workload. Acad. Radiol. 22, 1191–1198 (2015). source
Not checked in the free run.
- cites [4] Hosny, A., Parmar, C., Quackenbush, J., Schwartz, L. H. & Aerts, H. J. W. L. Artificial intelligence in radiology. Nat. Rev. Cancer 18, 500–510 (2018). source
Not checked in the free run.
- cites [5] Dan Lantsman, C. et al. Trend in radiologist workload compared to number of admissions in the emergency department. Eur. J. Radiol. 149, 110195 (2022). source
Not checked in the free run.
- Premisecitation
Deep learning models have demonstrated diagnostic accuracy comparable to that of human experts in narrow clinical tasks for several medical domains and imaging modalities.
“deep learning models have demonstrated diagnostic accuracy comparable to that of human experts in narrow clinical tasks for several medical domains and imaging modalities, including chest and extremity X-rays8–10, computed tomography (CT)11, magnetic resonance imaging (MRI)12, whole slide images (WSI)13,14, and dermatology images15.”
Quote in the paperSubstantial claim0 of 5 sources checked- cites [8] Dunnmon, J. A. et al. Assessment of Convolutional Neural Networks for Automated Classification of Chest Radiographs. Radiology 290, 537–544 (2019). source
Not checked in the free run.
- cites [9] Rajpurkar, P. et al. Deep learning for chest radiograph diagnosis: A retrospective comparison of the CheXNeXt algorithm to practicing radiologists. PLoS Med. 15, e1002686 (2018). source
Not checked in the free run.
- cites [10] Larson, D. B. et al. Performance of a Deep-Learning Neural Network Model in Assessing Skeletal Maturity on Pediatric Hand Radiographs. Radiology 287, 313–322 (2018). source
Not checked in the free run.
- cites [11] Park, A. et al. Deep Learning–Assisted Diagnosis of Cerebral Aneurysms Using the HeadXNet Model. JAMA Netw. Open 2, e195600–e195600 (2019). source
Not checked in the free run.
- cites [12] Bien, N. et al. Deep-learning-assisted diagnosis for knee magnetic resonance imaging: Development and retrospective validation of MRNet. PLoS Med. 15, e1002699 (2018). source
Not checked in the free run.
- Premisecitation
Supervised learning creates specialist models that can perform well only on the tasks they were trained to do.
“This creates specialist models that can perform well only on the tasks they were trained to do17.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [17] LeCun, Y. & Misra, I. Self-supervised learning: The dark matter of intelligence. Meta AI https://ai.facebook.com/blog/self-supervised-learning-the-dark-matter-of-intelligence/ (2021). (Accessed: 17th February 2023).
Not checked in the free run.
- Premisecitation
Azizi et al. have shown that SSL pre-trained models using natural images tend to outperform purely supervised pre-trained models for medical image classification.
“Azizi et al. have shown that SSL pre-trained models using natural images tend to outperform purely supervised pre-trained models93 for medical image classification, and continuing self-supervised pre-training with in-domain medical images leads to the best results.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [93] Azizi, S., Mustafa, B., Ryan, F. & Beaver, Z. Big self-supervised models advance medical image classification. Proc. Estonian Acad. Sci. Biol. Ecol.
Not checked in the free run.
- Interpretationreasoning
Many abnormalities are characterized by very subtle and localized visual cues, which can become ambiguous or obscured by augmentations.
“Many abnormalities are characterized by very subtle and localized visual cues, which can become ambiguous or obscured by augmentations (Fig. 5c).”
Quote in the paperSubstantial claim - Interpretationcitation
The optimal strategy may be dependent on factors including the specific medical imaging domain, the size and complexity of the dataset, and the type of classification task.
“Furthermore, the optimal strategy may be dependent on a number of factors including the specific medical imaging domain, the size and complexity of the dataset, and the type of classification task91,92.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [91] Kotar, K., Ilharco, G. & Schmidt, L. Contrasting contrastive self-supervised representation learning pipelines. Proc. Estonian Acad. Sci. Biol. Ecol.
Not checked in the free run.
- cites [92] Cole, E., Yang, X., Wilber, K., Aodha, O. M. & Belongie, S. When Does Contrastive Visual Representation Learning Work? Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) 2022, 14755–14764 (2022).
Not checked in the free run.
- Premisecitation
Truong et al. have demonstrated the effectiveness of combining representations from multiple self-supervised methods to improve performance for three different medical imaging modalities.
“Truong et al. have demonstrated the effectiveness of combining representations from multiple self-supervised methods to improve performance for three different medical imaging modalities96.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [96] Truong, T. et al. 158 54–74 (2021). source
Not checked in the free run.
- Premisecitation
Azizi et al. found that using generic and large-scale supervised pre-trained models, such as BigTransfer, can benefit subsequent domain-specific self-supervised pre-training and ultimately improve model performance and robustness for different medical imaging modalities.
“More recently, Azizi et al. found that using generic and large-scale supervised pre-trained models, such as BigTransfer94, can also benefit subsequent domain-specific self-supervised pre-training, and ultimately improve model performance and robustness for different medical imaging modalities95.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [94] Kolesnikov, A. et al. Big Transfer (BiT): General Visual Representation Learning. Computer Vision–ECCV 2020: 16th European Conference, Part V 16. (2020). source
Not checked in the free run.
- cites [95] Azizi, S. et al. Robust and Efficient Medical Imaging with Self-Supervision. Preprint at https://arxiv.org/abs/2205.09723 (2022).
Not checked in the free run.
- Interpretationreasoning
The assumption made for natural images is not necessarily valid for medical images.
“However, the assumption made for natural images is not necessarily valid for medical images for two reasons.”
Quote in the paperSubstantial claim - Interpretationown data
Definitive conclusions on the optimal strategy for medical images cannot be made since only a subset of studies made comparisons between different types of self-supervised learning strategies.
“Definitive conclusions on the optimal strategy for medical images cannot be made since only a subset of studies made comparisons between different types of self-supervised learning strategies.”
Quote in the paperSubstantial claim - Interpretationnone given
Medical images have high inter-class visual similarities due to the standardized protocols for medical image acquisition and the homogeneous nature of human anatomy.
“First, medical images have high inter-class visual similarities due to the standardized protocols for medical image acquisition and the homogeneous nature of human anatomy.”
Quote in the paperNot a substantial claim - Interpretationreasoning
Due to the lack of ablation studies from current work, it cannot be determined whether fine-tuning only later layers of the model could lead to better performance relative to end-to-end fine-tuning.
“However, due to the lack of ablation studies from current work, we cannot determine whether fine-tuning only later layers of the model could lead to better performance relative to end-to-end fine-tuning.”
Quote in the paperSubstantial claim - Premisecitation
Osin et al. and Hashimoto demonstrated that training a self-supervised framework could be beneficial to generate representative latent features of brain fMRIs before fine-tuning on image-level class labels.
“In this case both Osin et al. and Hashimoto demonstrated that training a self-supervised framework could be beneficial to generate representative latent features of brain fMRIs before fine-tuning on image-level class labels57,90.”
Quote in the paperNot a substantial claim0 of 2 sources checked- cites [57] Osin, J. et al. Learning Personal Representations from fMRI by Predicting Neurofeedback Performance. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2020 469–478 (2020). source
Not checked in the free run.
- cites [90] Hashimoto, Y., Ogata, Y., Honda, M. & Yamashita, Y. Deep Feature Extraction for Resting-State Functional MRI by Self-Supervised Learning and Application to Schizophrenia Diagnosis. Front. Neurosci. 15, 696853 (2021). source
Not checked in the free run.
- Rests on an assumption: Current limited knowledge and understanding of possible imaging features arising in psychiatric diseases constitutes a major clinical challenge to making local annotations such as bounding boxes or segmentations on brain scans.
- Interpretationreasoning
How translatable state-of-the-art natural image SSL results are to medical datasets is unclear, mainly due to the unique properties of medical images.
“While many self-supervised methods have led to state-of-the-art results when fine-tuned on natural image datasets, how translatable these results are to medical datasets is unclear, mainly due to the unique properties of medical images.”
Quote in the paperSubstantial claim - Premisenone given
Many contrastive-based strategies have been developed based on the assumption that the class-defining object is the main focus of an image, and thus variations caused by image transformations should not alter the image's semantic meanings.
“For instance, many contrastive-based strategies have been developed based on the assumption that the class-defining object is the main focus of an image, and thus variations caused by image transformations should not alter the image’s semantic meanings (Fig. 5).”
Quote in the paperNot a substantial claim - Premisecitation
Continuing self-supervised pre-training with in-domain medical images leads to the best results.
“Azizi et al. have shown that SSL pre-trained models using natural images tend to outperform purely supervised pre-trained models93 for medical image classification, and continuing self-supervised pre-training with in-domain medical images leads to the best results.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [93] Azizi, S., Mustafa, B., Ryan, F. & Beaver, Z. Big self-supervised models advance medical image classification. Proc. Estonian Acad. Sci. Biol. Ecol.
Not checked in the free run.
- Interpretationnone given
Within the medical imaging domain, the semantic meaning of interest is rarely an object such as the anatomical organ, but is rather the presence or absence of pathological abnormalities within that organ or tissue.
“Second, within the medical imaging domain, the semantic meaning of interest is rarely an object such as the anatomical organ, but is rather the presence or absence of pathological abnormalities within that organ or tissue.”
Quote in the paperNot a substantial claim - Premisecitation
Cornelissen et al. demonstrated that using representations from earlier layers can improve downstream classification of neoplasia in Barrett's Esophagus.
“Likewise, Cornelissen et al. demonstrated that using representations from earlier layers can improve downstream classification of neoplasia in Barrett’s Esophagus71.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [71] Cornelissen S. & van der Putten J. A.. Evaluating self-supervised learning methods for downstream classification of neoplasia In barrett’s esophagus. 2021 IEEE International Conference on Image Processing (ICIP), 66–70 (2021). source
Not checked in the free run.
- Interpretationown data
A comparison of different SSL strategies on the same downstream task suggests that a combined strategy tends to outperform other self-supervised categories.
“In Fig. 4c we show a comparison of different SSL strategies on the same downstream task, which suggests that a combined strategy tends to outperform other self-supervised categories.”
Quote in the paperNot a substantial claim - Interpretationown data
The graph of fine-tuning strategies tends to indicate that end-to-end fine-tuning leads to better performance regardless of the dataset size.
“In Fig. 4d we additionally plot the performance of the two main types of fine-tuning strategies on the same task, and the graph tends to indicate that end-to-end fine-tuning leads to better performance regardless of the dataset size.”
Quote in the paperNot a substantial claim - Premisecitation
Recent work has demonstrated the benefit of using learned visual word masks or spatially constrained crops to encourage representational invariance with semantically more similar regions of an image.
“Recent work has demonstrated the benefit of using learned visual word masks98,99 or spatially constrained crops100,101 to encourage representational invariance with semantically more similar regions of an image.”
Quote in the paperNot a substantial claim0 of 4 sources checked- cites [98] Shi, Y., Siddharth, N., Torr, P. H. S. & Kosiorek, A. R. Adversarial Masking for Self-Supervised Learning. In International Conference on Machine Learning, 20026–20040. (2022).
Not checked in the free run.
- cites [99] Li, G. et al. SemMAE: Semantic-Guided Masking for Learning Masked Autoencoders. Preprint at https://arxiv.org/abs/2206.10207 (2022).
Not checked in the free run.
- cites [100] Van Gansbeke, X. & Vandenhende, S. Revisiting contrastive methods for unsupervised learning of visual representations. Adv. Neural Inf. Process. Syst. 34, 16238–16250 (2021).
Not checked in the free run.
- cites [101] Peng, X., Wang, K., Zhu, Z. & Wang, M. Crafting better contrastive views for siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 16031–16040 (2022). source
Not checked in the free run.
- Study design and conductnone given
The study selection criteria only included literature for the task of medical image classification.
“Lastly, our study selection criteria only included literature for the task of medical image classification, which limits the scope of this review paper, since we recognize that self-supervised pre-trained models can also be fine-tuned for other important medical tasks, including segmentation, regression, and registration.”
Quote in the paperSubstantial claim - Interpretationcitation
The unique properties of medical images can be leveraged to design self-supervised learning methods more suitable for specific downstream tasks.
“The unique properties of medical images can be leveraged to design self-supervised learning methods more suitable for specific downstream tasks.”
Quote in the paperSubstantial claim - Premisecitation
Involving additional modalities during self-supervised learning has been shown to improve a model's performance when fine-tuned for downstream tasks.
“Involving additional modalities during self-supervised learning has also been shown to improve a model’s performance when fine-tuned for downstream tasks.”
Quote in the paperNot a substantial claim - Study design and conductreasoning
The review was unable to aggregate or statistically compare the effects of each self-supervised learning strategy on performance gain.
“Furthermore, we are unable to aggregate or statistically compare the effects of each self-supervised learning strategy on performance gain, since the included studies use different imaging modalities, report different performance metrics, and investigate different objectives.”
Quote in the paperSubstantial claim - Interpretationreasoning
The study selection criteria limit the scope of this review paper.
“Lastly, our study selection criteria only included literature for the task of medical image classification, which limits the scope of this review paper, since we recognize that self-supervised pre-trained models can also be fine-tuned for other important medical tasks, including segmentation, regression, and registration.”
Quote in the paperSubstantial claim - Interpretationreasoning
Multi-modal self-supervised learning strategies are limited to implementation within a healthcare system until more datasets with medical image and report pairs are publicly released.
“Therefore, these multi-modal self-supervised learning strategies are limited to implementation within a healthcare system until more datasets with medical image and report pairs are publicly released.”
Quote in the paperSubstantial claim- Rests on an assumption: Currently there are very few publicly available medical image datasets with corresponding radiology reports, largely due to the difficulties in preserving patient privacy.
- Premisecitation
Several studies have shown performance improvement when constructing positive pairs with slices from the same CT series, images from the same imaging study, images from the same patient, and images from patients with similar age.
“In fact, several studies have shown performance improvement when constructing positive pairs with slices from the same CT series102, images from the same imaging study103, images from the same patient93 and images from patients with similar age62.”
Quote in the paperSubstantial claim0 of 3 sources checked- cites [102] Dong, H. et al. Case Discrimination: Self-supervised Feature Learning for the Classification of Focal Liver Lesions. In Innovation in Medicine and Healthcare 241–249 (2021). source
Not checked in the free run.
- cites [103] Vu, Y. N. T. et al. MedAug: Contrastive learning leveraging patient metadata improves representations for chest X-ray interpretation. In Proceedings of Machine Learning Research 126, 1–14 (2021).
Not checked in the free run.
- cites [93] Azizi, S., Mustafa, B., Ryan, F. & Beaver, Z. Big self-supervised models advance medical image classification. Proc. Estonian Acad. Sci. Biol. Ecol.
Not checked in the free run.
- Interpretationown data
Results from this systematic review have revealed that SSL for medical image classification is a growing and promising field of research across multiple medical disciplines and imaging modalities.
“Results from this systematic review have revealed that SSL for medical image classification is a growing and promising field of research across multiple medical disciplines and imaging modalities.”
Quote in the paperSubstantial claim
- Classified as probably not a substantial claim. 21
- May combine two claims. 28
Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).
Findingnone givenThere is already a growing trend in recent medical imaging AI literature to leverage SSL.
“While there is already a growing trend in recent medical imaging AI literature to leverage SSL (Fig. 1)”
Quote in the paperNot a substantial claimMay combine two claims.
FindingcitationOf the three generative SSL studies, two reported comparisons with purely supervised models and showed relative improvements of 16.6% and 24.5% with self-supervised learning.
“Of the three studies, two reported comparisons with purely supervised models and showed relative improvements of 16.6%58 and 24.5%56 with self-supervised learning.”
Quote in the paperNot a substantial claim0 of 2 sources checked- cites [58] Zhao, Q., Liu, Z., Adeli, E. & Pohl, K. M. Longitudinal Self-Supervised Learning. Med. Image Anal. 71, (2021). source
Not checked in the free run.
- cites [56] Gamper, J. & Rajpoot, N. Multiple Instance Captioning: Learning Representations from Histopathology Textbooks and Articles. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 16544–16554 (2021). source
Not checked in the free run.
May combine two claims.
Findingown dataOut of all innate relationship based studies, ten compared performance to supervised pre-trained models and eight of them showed improvement.
“Out of all innate relationship based studies, ten compared performance to those of supervised pre-trained models and eight of them showed improvement in performance.”
Quote in the paperNot a substantial claimMay combine two claims.
FindingcitationThe pathology tasks were focused on histopathological classification, with the majority focused on colorectal cancer classification.
“The tasks were focused on histopathological classification, where the majority focused on colorectal cancer classification68,84–87.”
Quote in the paperNot a substantial claim0 of 5 sources checked- cites [68] Ke, J., Shen, Y., Liang, X. & Shen, D. Contrastive Learning Based Stain Normalization Across Multiple Tumor in Histopathology. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2021 571–580 (Springer International Publishing, 2021). source
Not checked in the free run.
- cites [84] Yang, P., Hong, Z., Yin, X., Zhu, C. & Jiang, R. Self-supervised Visual Representation Learning for Histopathological Images. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2021 47–57 (Springer International Publishing, 2021). source
Not checked in the free run.
- cites [85] Ciga, O., Xu, T. & Martel, A. L. Self supervised contrastive learning for digital histopathology. Machine Learning with Applications 7, (2022). source
Not checked in the free run.
- cites [86] Li, J., Lin, T. & Xu, Y. SSLP: Spatial Guided Self-supervised Learning on Pathological Images. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2021 3–12 (2021). source
Not checked in the free run.
- cites [87] Srinidhi, C. L., Kim, S. W., Chen, F.-D. & Martel, A. L. Self-supervised driven consistency training for annotation efficient histopathology image analysis. Med. Image Anal. 75, (2021). source
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing).
Findingnone givenCombined strategies are typically the proposed method in the manuscripts.
“However, it is important to note that combined strategies are typically the proposed method in the manuscripts, and thus publication bias might have resulted in this trend.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataThirty-six studies compared contrastive SSL pre-training to supervised pre-training and reported an average improvement in performance of 6.35%.
“Thirty-six studies compared contrastive SSL pre-training to supervised pre-training and reported an average improvement in performance of 6.35%.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Findingnone givenThe pathology field is centered around medical imaging in the form of whole slide images.
“Similar to radiology, this field is centered around medical imaging in the form of whole slide images.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataAmong the SSL baselines, SimCLR was most frequently compared (16/26), followed by autoencoders (11/26) and MoCo (9/26).
“Among the SSL baselines, SimCLR was most frequently compared (16/26), followed by autoencoders (11/26) and MoCo (9/26).”
Quote in the paperNot a substantial claimMay combine two claims.
FindingcitationTwo of the five self-prediction studies compared their approach to models without self-supervised pre-training and reported slight relative improvements in performance of 1.12% and 0.690%.
“Two of the five studies compared their approach to models without self-supervised pre-training and reported slight relative improvements in performance of 1.12%67 and 0.690%63.”
Quote in the paperNot a substantial claim0 of 2 sources checked- cites [67] Park, J., Kwak, I.-Y. & Lim, C. A Deep Learning Model with Self-Supervised Learning and Attention Mechanism for COVID-19 Diagnosis Using Chest X-ray Images. Electronics 10, 1996 (2021). source
Not checked in the free run.
- cites [63] Liu, C. et al. TN-USMA Net: Triple normalization-based gastrointestinal stromal tumors classification on multicenter EUS images with ultrasound-specific pre-training and meta attention. Med. Phys. 48, 7199–7214 (2021). source
Not checked in the free run.
May combine two claims.
Findingown dataSelf-supervised learning strategies generate label-efficient representations, but the learning process typically requires a relatively large amount of unlabeled data.
“Furthermore, even though self-supervised learning strategies generate label-efficient representations, the learning process typically requires a relatively large amount of unlabeled data.”
Quote in the paperNot a substantial claimMay combine two claims.
FindingcitationThe remaining studies explored multiple other tasks and many focused on classification of malignant disorders such as breast cancer, skin cancer, and lung cancer.
“The remaining studies explored multiple other tasks and many focused on classification of malignant disorders such as breast cancer56,61,88, skin cancer89, and lung cancer59.”
Quote in the paperNot a substantial claim0 of 5 sources checked- cites [56] Gamper, J. & Rajpoot, N. Multiple Instance Captioning: Learning Representations from Histopathology Textbooks and Articles. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 16544–16554 (2021). source
Not checked in the free run.
- cites [61] Wang, X. et al. TransPath: Transformer-Based Self-supervised Learning for Histopathological Image Classification. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2021 186–195 (2021). source
Not checked in the free run.
- cites [88] Dehaene, O., Camara, A., Moindrot, O., de Lavergne, A. & Courtiol, P. Self-Supervision Closes the Gap Between Weak and Strong Supervision in Histology. in ML4H 2020 NeurIPS workshop (2020).
Not checked in the free run.
- cites [89] Liu, Q. et al. SimTriplet: Simple Triplet Representation Learning with a Single GPU. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2021. Part II 24, 102–112 (2021). source
Not checked in the free run.
- cites [59] Li, B., Li, Y. & Eliceiri, K. W. Dual-stream Multiple Instance Learning Network for Whole Slide Image Classification with Self-supervised Contrastive Learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. (2021). source
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing).
FindingcitationA particularly interesting medical task explored was classification of psychiatric diseases or psychiatric traits using fMRI.
“A particularly interesting medical task that was explored was classification of psychiatric diseases or psychiatric traits using fMRI57,62,90.”
Quote in the paperNot a substantial claim0 of 3 sources checked- cites [57] Osin, J. et al. Learning Personal Representations from fMRI by Predicting Neurofeedback Performance. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2020 469–478 (2020). source
Not checked in the free run.
- cites [62] Dufumier, B. et al. Contrastive Learning with Continuous Proxy Meta-data for 3D MRI Classification. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2021 58–68 (2021). source
Not checked in the free run.
- cites [90] Hashimoto, Y., Ogata, Y., Honda, M. & Yamashita, Y. Deep Feature Extraction for Resting-State Functional MRI by Self-Supervised Learning and Application to Schizophrenia Diagnosis. Front. Neurosci. 15, 696853 (2021). source
Not checked in the free run.
May combine two claims.
Findingown dataEight of the studies that combined different strategies compared self-supervised pre-training with purely supervised approaches, all of which reported performance improvement (0.140-8.29%).
“Eight of the studies that combined different strategies compared self-supervised pre-training with purely supervised approaches, all of which reported performance improvement (0.140–8.29%).”
Quote in the paperSubstantial claimMay combine two claims.
Findingown dataSimCLR, MoCo and BYOL were the three most used contrastive frameworks, applied in 13, 8, and 3 papers, respectively.
“SimCLR, MoCo and BYOL were the three most used frameworks, applied in 13, 8, and 3 papers, respectively.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingnone givenMasking and cropping operations can obscure the cavernoma and alter the semantic meaning of the image, as the MRI-scan no longer exhibits any abnormality.
“b and c Masking and cropping operations can obscure the cavernoma and alter the semantic meaning of the image, as the MRI-scan no longer exhibits any abnormality.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown dataThe included studies use different imaging modalities, report different performance metrics, and investigate different objectives.
“Furthermore, we are unable to aggregate or statistically compare the effects of each self-supervised learning strategy on performance gain, since the included studies use different imaging modalities, report different performance metrics, and investigate different objectives.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Findingnone givenMasking and cropping operations do not obscure the dog nor alter the semantic meaning of the image.
“d Image of a dog (bounded in red), (e) and (f) Masking and cropping operations do not obscure the dog nor alter the semantic meaning of the image.”
Quote in the paperNot a substantial claimMay combine two claims.
PremisecitationWith such an immense cognitive burden placed on radiologists, delays in diagnosis and diagnostic errors are unavoidable.
“With such an immense cognitive burden placed on radiologists, delays in diagnosis and diagnostic errors are unavoidable6,7.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [6] Alonso-Martínez, J. L., Sánchez, F. J. A. & Echezarreta, M. A. U. Delay and misdiagnosis in sub-massive and non-massive acute pulmonary embolism. Eur. J. Intern. Med. 21, 278–282 (2010). source
Not checked in the free run.
- cites [7] Hendriksen, J. M. T. et al. Clinical characteristics associated with diagnostic delay of pulmonary embolism in primary care: a retrospective observational study. BMJ Open 7, e012789 (2017). source
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
PremisecitationPre-trained large language models are capable of achieving state-of-the-art results for a wide range of NLP tasks, and have shown the ability to perform well on new tasks with only a fraction of the labeled data that traditional supervised learning techniques require.
“These pre-trained large language models18,19 are capable of achieving state-of-the-art results for a wide range of NLP tasks, and have shown the ability to perform well on new tasks with only a fraction of the labeled data that traditional supervised learning techniques require.”
Quote in the paperNot a substantial claim0 of 2 sources checked- cites [18] Devlin, J., Chang, M.-W., Lee, K. & Toutanova, K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. Preprint at https://arxiv.org/abs/1810.04805 (2018).
Not checked in the free run.
- cites [19] Brown, T., Mann, B. & Ryder, N. Language models are few-shot learners. Adv. Neural Inf. Process. Syst. (2020).
Not checked in the free run.
Probably not a substantial claim: may not pass S4 (based). May combine two claims.
InterpretationreasoningReducing the number of manual annotations required to train medical imaging models will significantly reduce both the cost and time required for model development, making automated systems more accessible to different specialties and hospitals, thereby reducing workload for radiologists and potentially improving patient care.
“Reducing the number of manual annotations required to train medical imaging models will significantly reduce both the cost and time required for model development, making automated systems more accessible to different specialties and hospitals, thereby reducing workload for radiologists and potentially improving patient care.”
Quote in the paperNot a substantial claim- Rests on an assumption: Annotating medical imaging datasets requires domain expertize, making large-scale annotations cost-prohibitive and time-consuming.
- Rests on an assumption: The current success of deep learning has been largely dominated by supervised learning frameworks, which typically require large-scale labeled datasets to achieve high performance.
- Rests on an assumption: The inability to generalize can be largely due to the process of supervised learning, which encourages the model to mainly learn features heavily correlated with specific labels rather than general features representative of the whole data distribution.
- Rests on an assumption: In a healthcare system where a myriad of opportunities and possibilities for automation exist, it is practically impossible to curate labeled datasets for all tasks, modalities, and outcomes for training supervised models.
Probably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductnone givenThe authors independently screened all records for eligibility.
“The authors independently screened all records for eligibility.”
Quote in the paperNot a substantial claimMay combine two claims.
PremisecitationAugmentations and transformations alter the semantic meaning of medical images but not natural images.
“Examples of augmentations and transformations that alter the semantic meaning of medical images144 but not natural images145.”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [144] Hellerhoff. File:Kavernom rechts parietal 59M - MR - 001.jpg. Wikimedia https://commons.wikimedia.org/wiki/File:Kavernom_rechts_parietal_59M_-_MR_-_001.jpg (2022).
Not checked in the free run.
- cites [145] Matio, H. File:Dog Breeds.jpg. Wikimedia Commons https://commons.wikimedia.org/wiki/File:Dog_Breeds.jpg (2019).
Not checked in the free run.
May combine two claims.
PremisecitationYan et al. have demonstrated the possibility of training self-supervised models with data from multiple healthcare centers without explicitly sharing data, and have shown improvement in robustness and performance over models trained using data from only one institution.
“Nevertheless, using federated learning, Yan et al. have demonstrated the possibility of training self-supervised models with data from multiple healthcare centers without the need for explicitly sharing data, and have shown improvement in robustness and performance over models trained using data from only one institution97.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [97] Yan, R. et al. Label-Efficient Self-Supervised Federated Learning for Tackling Data Heterogeneity in Medical Imaging. IEEE Transactions on Medical Imaging (2023). source
Not checked in the free run.
- Rests on an assumption: Curating large-scale medical image datasets from multiple institutions is often challenged by the difficulty of sharing patient data while preserving patient privacy.
May combine two claims.
PremisecitationFor natural image datasets, MoCo outperforms MAE by training a linear model on extracted features (linear probing), while MAE achieves better performance than MoCo as the number of fine-tune layers increases.
“For instance, for natural image datasets, MoCo outperforms MAE by training a linear model on extracted features (linear probing), while MAE achieves better performance than MoCo as the number of fine-tune layers increases41.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [41] He, K. et al. Masked Autoencoders Are Scalable Vision Learners. Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) 2022, 16000–16009 (2022).
Not checked in the free run.
May combine two claims.
InterpretationcitationRepresentations learned using certain SSL strategies can be relatively more linearly separable, while representations from other strategies can achieve better performance when more layers or the entire model are fine-tuned.
“It is worth noting that representations learned using certain SSL strategies can be relatively more linearly separable, while representations from other strategies can achieve better performance when more layers or the entire model are fine-tuned.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Interpretationnone givenThe random masking operation often utilized by self-prediction self-supervised learning methods may also alter a medical image's semantic meaning by removing image regions with diseases or abnormalities.
“The random masking operation often utilized by self-prediction self-supervised learning methods may also alter a medical image’s semantic meaning by removing image regions with diseases or abnormalities (Fig. 5b).”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing), S4 (based).
Interpretationnone givenFactors such as the degree of domain shift between SSL pre-training data and downstream task data could also affect the linear separability of the representations.
“Factors such as the degree of domain shift between SSL pre-training data and downstream task data could also affect the linear separability of the representations.”
Quote in the paperNot a substantial claimMay combine two claims.
PremisereasoningThese contrastive methods typically apply strong transformations to the original image and encourage the model to learn similar global representations for images with similar semantic meanings.
“Therefore, these methods typically apply strong transformations to the original image and encourage the model to learn similar global representations for images with similar semantic meanings.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
InterpretationreasoningThe unique attributes of medical images can also be utilized for creating relevant pre-text tasks.
“The unique attributes of medical images can also be utilized for creating relevant pre-text tasks.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductnone givenIn cases where consensus could not be achieved, a fourth arbitrating researcher was consulted.
“In cases where consensus could not be achieved a fourth arbitrating researcher was consulted (A.S.C.).”
Quote in the paperNot a substantial claimMay combine two claims.
Study design and conductcitationThe search was limited to papers published after 2012, which excluded papers that applied self-supervised learning prior to the era of deep learning for computer vision.
“We also limited our search to only consider papers published after 2012, which excluded papers that applied self-supervised learning prior to the era of deep learning for computer vision108.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [108] Russakovsky, O. et al. ImageNet Large Scale Visual Recognition Challenge. Int. J. Comput. Vis. 115, 211–252 (2015). source
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing).
PremisecitationRivail et al. proposed a self-supervised approach to model disease progression by estimating the time interval between pairs of optical coherence tomography (OCT) scans from the same patient.
“Rivail et al. proposed a self-supervised approach to model disease progression by estimating the time interval between pairs of optical coherence tomography (OCT) scans from the same patient104.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [104] Rivail, A. et al. Modeling Disease Progression in Retinal OCTs with Longitudinal Self-supervised Learning. in Predictive Intelligence in Medicine 44–52 (2019). source
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.
Interpretationnone givenPublication bias can be a considerable limitation due to disproportionately reported positive results in the literature, which can lead to overestimation of the benefits of self-supervised learning.
“For this review paper, publication bias can be a considerable limitation due to disproportionately reported positive results in the literature, which can lead to overestimation of the benefits of self-supervised learning.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing), S4 (based).
Interpretationnone givenSelf-supervised pre-trained models can also be fine-tuned for other important medical tasks, including segmentation, regression, and registration.
“Lastly, our study selection criteria only included literature for the task of medical image classification, which limits the scope of this review paper, since we recognize that self-supervised pre-trained models can also be fine-tuned for other important medical tasks, including segmentation, regression, and registration.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductcitationThis systematic review was conducted based on the PRISMA guidelines.
“This systematic review was conducted based on the PRISMA guidelines109.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [109] Page, M. J. et al. “The PRISMA 2020 statement: an updated guideline for reporting systematic reviews.” International journal of surgery. 88 (2021).
Not checked in the free run.
Classified as probably not a substantial claim. May combine two claims.
Study design and conductown dataStudy selection was performed by three independent researchers, and disagreements were resolved through discussion.
“Study selection was performed by three independent researchers (S.-C.H., A.P., and M.J.), and disagreements were resolved through discussion.”
Quote in the paperSubstantial claimMay combine two claims.
PremisecitationJiao et al. leveraged the correlation between fetal ultrasonography and the narrative speech of the sonographer to create a pre-text task for self-supervision, and subsequently used the learned representations for downstream standard plane classification on sonograms.
“Jiao et al. cleverly leveraged the correlation between fetal ultrasonography and the narrative speech of the sonographer to create a pre-text task for self-supervision, and subsequently used the learned representations for downstream standard plane classification on sonograms53.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [53] Jiao, J. et al. Self-Supervised Contrastive Video-Speech Representation Learning for Ultrasound. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2020 534–543 (2020). source
Not checked in the free run.
May combine two claims.
PremisecitationTaleb et al. proposed a multimodal contrastive learning strategy between retinal fundus images and genetics data and showed improvement in performance over single modality pre-training.
“For example, Taleb et al. proposed a multimodal contrastive learning strategy between retinal fundus images and genetics data and showed improvement in performance over single modality pre-training105.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [105] Taleb A., Kirchler M. & Monti R. ContIG: Self-supervised Multimodal Contrastive Learning for Medical Imaging with Genetics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 20908–20921 (2022). source
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductreasoningSubjectivity may have been introduced when categorizing the self-supervised learning strategy in each paper, especially for studies that implemented novel, unconventional, or a mixture of methods.
“In addition, subjectivity may have been introduced when categorizing the self-supervised learning strategy in each paper, especially for studies that implemented novel, unconventional, or a mixture of methods.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim.
Study design and conductnone givenThe search encompassed papers published between January 2012 and May 2021.
“The search encompassed papers published between January 2012 and May 2021.”
Quote in the paperSubstantial claimMay combine two claims.
PremisecitationSeveral studies have utilized medical reports to provide supervision signals during self-supervised learning and shown label efficiency for downstream tasks.
“Several studies have utilized these medical reports to provide supervision signals during self-supervised learning and shown label efficiency for downstream tasks60,106.”
Quote in the paperNot a substantial claim0 of 2 sources checked- cites [60] Ji, Z. et al. Improving Joint Learning of Chest X-Ray and Radiology Report by Word Region Alignment. In Machine Learning in Medical Imaging 110–119 (2021). source
Not checked in the free run.
- cites [106] Zhang, Y., Jiang, H., Miura, Y., Manning, C. D. & Langlotz, C. P. Contrastive Learning of Medical Visual Representations from Paired Images and Text. Preprint at https://arxiv.org/abs/2010.00747 (2020).
Not checked in the free run.
- Rests on an assumption: Many medical imaging modalities have corresponding radiology reports that contain detailed descriptions of the medical conditions observed by radiologists.
May combine two claims.
PremisecitationHuang et al. demonstrated the model's ability to localize chest abnormalities on chest x-rays without segmentation labels and revealed the possibility of zero-shot learning by converting the classification classes into textual captions and framing the image classification task as measuring image-text similarity.
“By leveraging radiology reports, Huang et al. demonstrated the model’s ability to localize chest abnormalities on chest x-rays without segmentation labels and revealed the possibility of zero-shot learning by converting the classification classes into textual captions and framing the image classification task as measuring the image-text similarity107.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [107] Huang, S.-C., Shen, L., Lungren, M. P. & Yeung, S. GLoRIA: A multimodal global-local representation learning framework for label-efficient medical image recognition. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV) (2021). source
Not checked in the free run.
- Rests on an assumption: Many medical imaging modalities have corresponding radiology reports that contain detailed descriptions of the medical conditions observed by radiologists.
Classified as probably not a substantial claim.
Study design and conductnone givenA systematic literature search was implemented in three literature databases: PubMed, Scopus and ArXiv.
“A systematic literature search was implemented in three literature databases: PubMed, Scopus and ArXiv.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S2 (checkable), S4 (based).
Interpretationnone givenThe flexibility in creating self-supervised methods as well as the adaptability and transferability to multiple medical domains highlights the importance and utility of self-supervised techniques in clinical applications.
“Overall, the flexibility in creating self-supervised methods as well as the adaptability and transferability to multiple medical domains highlights the importance and utility of self-supervised techniques in clinical applications.”
Quote in the paperNot a substantial claim
The paper relies on this without support.
Annotating medical imaging datasets requires domain expertize, making large-scale annotations cost-prohibitive and time-consuming.
“However, the prevailing paradigm of training deep learning models requires large quantities of labeled training data, which is both time-consuming and cost-prohibitive to curate for medical images.”
“However, annotating medical imaging datasets requires domain expertize, making large-scale annotations cost-prohibitive and time-consuming, which fundamentally limits building effective medical imaging models across varying clinical use cases.”
The paper relies on this without support.
The current success of deep learning has been largely dominated by supervised learning frameworks, which typically require large-scale labeled datasets to achieve high performance.
“While deep learning provides promising solutions for improving medical image interpretation, the current success has been largely dominated by supervised learning frameworks, which typically require large-scale labeled datasets to achieve high performance.”
The paper relies on this without support.
The inability to generalize can be largely due to the process of supervised learning, which encourages the model to mainly learn features heavily correlated with specific labels rather than general features representative of the whole data distribution.
“The inability to generalize can be largely due to the process of supervised learning, which encourages the model to mainly learn features heavily correlated with specific labels rather than general features representative of the whole data distribution.”
The paper relies on this without support.
In a healthcare system where a myriad of opportunities and possibilities for automation exist, it is practically impossible to curate labeled datasets for all tasks, modalities, and outcomes for training supervised models.
“In a healthcare system where a myriad of opportunities and possibilities for automation exist, it is practically impossible to curate labeled datasets for all tasks, modalities, and outcomes for training supervised models.”
The paper relies on this without support.
SSL can create generalist models that can be fine-tuned for many downstream tasks without large-scale labeled datasets.
“Unlike supervised learning, SSL can create generalist models that can be fine-tuned for many downstream tasks without large-scale labeled datasets.”
The paper relies on this without support.
The most suitable strategies and best practices for medical images have not been sufficiently investigated.
“the most suitable strategies and best practices for medical images have not been sufficiently investigated.”
The paper relies on this without support.
Many clinical tasks are based on classification.
“We focus on medical image classification tasks because many clinical tasks are based on classification, and thus our research may be directly applicable to deep learning models for clinical workflows.”
The paper relies on this without support.
SSL is task agnostic, and the same SSL pre-trained model can be fine-tuned for different types of downstream tasks, including classification, segmentation, and object detection.
“It is worth emphasizing that SSL is task agnostic, and the same SSL pre-trained model can be fine-tuned for different types of downstream tasks, including classification, segmentation, and object detection.”
The paper relies on this without support.
Current limited knowledge and understanding of possible imaging features arising in psychiatric diseases constitutes a major clinical challenge to making local annotations such as bounding boxes or segmentations on brain scans.
“Current limited knowledge and understanding of possible imaging features arising in psychiatric diseases constitutes a major clinical challenge to making local annotations such as bounding boxes or segmentations on brain scans.”
The paper relies on this without support.
Curating large-scale medical image datasets from multiple institutions is often challenged by the difficulty of sharing patient data while preserving patient privacy.
“Curating large-scale medical image datasets from multiple institutions is often challenged by the difficulty of sharing patient data while preserving patient privacy.”
The paper relies on this without support.
Many medical imaging modalities have corresponding radiology reports that contain detailed descriptions of the medical conditions observed by radiologists.
“Furthermore, many medical imaging modalities have corresponding radiology reports that contain detailed descriptions of the medical conditions observed by radiologists.”
The paper relies on this without support.
Currently there are very few publicly available medical image datasets with corresponding radiology reports, largely due to the difficulties in preserving patient privacy.
“However, currently there are very few publicly available medical image datasets with corresponding radiology reports, largely due to the difficulties in preserving patient privacy.”
The paper relies on this without support.
Many models pre-trained in a self-supervised manner on large-scale natural image datasets are publicly available.
“Although self-supervised pre-training can be computationally demanding, many models pre-trained in a self-supervised manner on large-scale natural image datasets are publicly available and should be utilized.”
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The expanding literature on SSL in medical imaging focuses primarily on 2D images for diagnostic applications.
“The expanding literature on SSL in medical imaging13 focuses primarily on two-dimensional (2D) images (X-ray, whole slide images, dermatology images, fundus images and so on) for diagnostic applications.”
From Foundation Model for Cancer Imaging Biomarkers (2024) · cites this paper as [13]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Search end date conflicts with the stated review period
MajorThe abstract says the review covered papers through 2022, but Methods says the search ended in May 2021. Table 1 also includes a study dated 2022, so the stated search window cannot account for all included studies.
“papers published between 2012 and 2022 on PubMed, Scopus, and ArXiv”
“The search encompassed papers published between January 2012 and May 2021.”
Fine-tuning results are attributed to the wrong figure panel
MinorThe implementation guidelines cite Fig. 4c for the finding about end-to-end fine-tuning. The figure caption assigns fine-tuning comparisons to panel d; panel c compares SSL strategies.
“Based on our aggregated results, we found that end-to-end fine-tuning generally leads to better performance for medical images (Fig. 4c).”
“c Relative performance difference between different types of self-supervised learning strategies on the same task. d Performance comparison between end-to-end fine-tuning vs. training a classifier using extracted features from pre-trained self-supervised models.”
Self-prediction comparison count omits a study
MinorThe self-prediction section describes comparisons by two of five studies, although it earlier says six studies were included. Table 1 reports supervised and SSL performance for Tamkin as well as the Park and Liu comparisons cited in that section, making at least three comparisons.
“Two of the five studies compared their approach to models without self-supervised pre-training and reported slight relative improvements in performance of 1.12%67 and 0.690%63.”
“| Self-prediction | 2021 | Alex Tamkin143 | Chest X-ray | Radiology | Chest abnormality | – | SHED | Extract features from encoder -> Linear classifier | Accuracy | 0.745 | 0.681 | 0.094 |”
Cites
The PRISMA 2020 statement: An updated guideline for reporting systematic reviews
Page, M. J., McKenzie, J. E., Bossuyt, P. M., Boutron, I., Hoffmann, T. C., Mulrow, C. D., Shamseer, L., Tetzlaff, J. M., Akl, E. A., Brennan, S. E., et al. (2021) · BMJ · doi:10.1136/bmj.n71 · Cited by 108,283 · see citing works
ImageNet Large Scale Visual Recognition Challenge
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., et al. (2015) · International Journal of Computer Vision · doi:10.1007/s11263-015-0816-y · Cited by 39,380 · cited 1× · see citing works
“We also limited our search to only consider papers published after 2012, which excluded papers that applied self-supervised learning prior to the era of deep learning for computer vision 108 .”
BackgroundAuto-Encoding Variational Bayes
Kingma, D. P., & Welling, M. (2013) · arXiv · doi:10.48550/arxiv.1312.6114 · Cited by 20,067 · see citing works
Deep learning in neural networks: an overview
Schmidhuber, J. (2015) · Neural Networks · doi:10.1016/j.neunet.2014.09.003 · Cited by 18,570 · cited 1× · see citing works
“Generative models, popularized through the advent of traditional autoencoders 27 , variational autoencoders 28 and generative adversarial networks (GANs) 29 , are able to learn the distribution of training data, and thereby reconstruct the original input or create new synthetic data instances.”
BackgroundDermatologist-level classification of skin cancer with deep neural networks
Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H. M., & Thrun, S. (2017) · Nature · doi:10.1038/nature21056 · Cited by 14,362 · see citing works
Momentum Contrast for Unsupervised Visual Representation Learning
He, K., Fan, H., Wu, Y., Xie, S., & Girshick, R. (2020) · 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr42600.2020.00975 · Cited by 13,347 · cited 1× · see citing works
“To reduce the large batch size required by SimCLR to ensure enough informative negative samples, Momentum Contrast (MoCo) introduced a momentum encoded queue to keep negative samples 33 .”
MethodThe PRISMA 2020 Statement: An Updated Guideline for Reporting Systematic Reviews
Page, M. J., McKenzie, J. E., Bossuyt, P. M., Boutron, I., Hoffmann, T. C., Mulrow, C. D., Shamseer, L., Tetzlaff, J. M., Akl, E. A., Brennan, S. E., et al. (2021) · International Journal of Surgery · doi:10.1016/j.ijsu.2021.105906 · Cited by 11,748 · see citing works
Extracting and composing robust features with denoising autoencoders
Vincent, P., Larochelle, H., Bengio, Y., & Manzagol, P.-A. (2008) · Proceedings of the 25th international conference on Machine learning - ICML '08 · doi:10.1145/1390156.1390294 · Cited by 7,436 · cited 1× · see citing works
“Early work that leverages generative models for self-supervised learning rely on autoencoders, where an encoder converts inputs into latent representations and a decoder reconstructs the representation back to the original image 30 .”
BackgroundEmerging Properties in Self-Supervised Vision Transformers
Caron, M., Touvron, H., Misra, I., Jegou, H., Mairal, J., Bojanowski, P., & Joulin, A. (2021) · 2021 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv48922.2021.00951 · Cited by 5,638 · cited 1× · see citing works
“More recently, a subtype of contrastive self-supervised learning called instance discrimination, which includes methods such as DINO 34 , BYOL 35 and SimSiam 36 , further eliminates the need for negative samples.”
BackgroundArtificial intelligence in radiology
Hosny, A., Parmar, C., Quackenbush, J., Schwartz, L. H., & Aerts, H. J. W. L. (2018) · Nature Reviews Cancer · doi:10.1038/s41568-018-0016-5 · Cited by 3,803 · see citing works
Unsupervised Learning of Visual Representations by Solving Jigsaw Puzzles
Noroozi, M., & Favaro, P. (2016) · Lecture Notes in Computer Science · doi:10.1007/978-3-319-46466-4_5 · Cited by 2,775 · see citing works
Unsupervised Visual Representation Learning by Context Prediction
Doersch, C., Gupta, A., & Efros, A. A. (2015) · 2015 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2015.167 · Cited by 2,629 · cited 1× · see citing works
“Examples of innate relationship for visual inputs include predicting image rotation angle 24 , solving jigsaw puzzles of an image 25 , or determining the relative positions of image patches 26 .”
Unsupervised Learning of Visual Features by Contrasting Cluster Assignments
Caron, M., Misra, I., Mairal, J., Goyal, P., Bojanowski, P., & Joulin, A. (2020) · arXiv · doi:10.48550/arxiv.2006.09882 · Cited by 2,499 · cited 1× · see citing works
“Instead of contrastive augmented pairs from the same image, several studies have explored contrasting clustering assignments of augmented versions of the same image 37 – 39 .”
BackgroundUnsupervised Representation Learning by Predicting Image Rotations
Gidaris, S., Singh, P., & Komodakis, N. (2018) · arXiv · doi:10.48550/arxiv.1803.07728 · Cited by 2,477 · cited 1× · see citing works
“Examples of innate relationship for visual inputs include predicting image rotation angle 24 , solving jigsaw puzzles of an image 25 , or determining the relative positions of image patches 26 .”
Adversarial Feature LearningPreprint
Donahue, J., Krähenbühl, P., & Darrell, T. (2016) · International Conference on Learning Representations · doi:10.48550/arxiv.1605.09782 · Cited by 1,982 · see citing works
MIMIC-CXR, a de-identified publicly available database of chest radiographs with free-text reports
Johnson, A. E. W., Pollard, T. J., Berkowitz, S. J., Greenbaum, N. R., Lungren, M. P., Deng, C.-Y., Mark, R. G., & Horng, S. (2019) · Scientific Data · doi:10.1038/s41597-019-0322-0 · Cited by 1,816 · see citing works
Self-citationZech, J. R., Badgeley, M. A., Liu, M., Costa, A. B., Titano, J. J., & Oermann, E. K. (2018) · PLOS Medicine · doi:10.1371/journal.pmed.1002683 · Cited by 1,719 · cited 1× · see citing works
“Besides facing challenges with training data, most medical imaging models underperform in their ability to generalize to external institutions or when repurposed for other tasks 16 .”
BackgroundRajpurkar, P., Irvin, J., Ball, R. L., Zhu, K., Yang, B., Mehta, H., Duan, T., Ding, D., Bagul, A., Langlotz, C. P., et al. (2018) · PLOS Medicine · doi:10.1371/journal.pmed.1002686 · Cited by 1,449 · see citing works
Self-citationBEiT: BERT Pre-Training of Image Transformers
Bao, H., Dong, L., Piao, S., & Wei, F. (2021) · arXiv · doi:10.48550/arxiv.2106.08254 · Cited by 1,392 · see citing works
Big Transfer (BiT): General Visual Representation Learning
Kolesnikov, A., Beyer, L., Zhai, X., Puigcerver, J., Yung, J., Gelly, S., & Houlsby, N. (2020) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-58558-7_29 · Cited by 1,039 · cited 2× · see citing works
“More recently, Azizi et al. found that using generic and large-scale supervised pre-trained models, such as BigTransfer 94 , can also bene fi t subsequent domain-speci fi c self-supervised pre-training, and ultimately improve model performance and robustness for different medical imaging modalities 95 .”
Methodk-Nearest Neighbour Classifiers - A Tutorial
Cunningham, P., & Delany, S. J. (2021) · ACM Computing Surveys · doi:10.1145/3459665 · Cited by 1,033 · see citing works
Li, B., Li, Y., & Eliceiri, K. W. (2021) · 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr46437.2021.01409 · Cited by 957 · see citing works
Bien, N., Rajpurkar, P., Ball, R. L., Irvin, J., Park, A., Jones, E., Bereket, M., Patel, B. N., Yeom, K. W., Shpanskaya, K., et al. (2018) · PLOS Medicine · doi:10.1371/journal.pmed.1002699 · Cited by 734 · cited 11× · see citing works
“In this case both Osin et al. and Hashimoto demonstrated that training a self-supervised framework could be bene fi cial to generate representative latent features of brain fMRIs before fi ne-tuning on image-level class labels 57,90 .”
BackgroundInfluentialSelf-citationTrends in Use of Medical Imaging in US Health Care Systems and in Ontario, Canada, 2000-2016
Smith‐Bindman, R., Kwan, M. L., Marlow, E. C., Theis, M. K., Bolch, W. E., Cheng, S. Y., Bowles, E. J. A., Duncan, J. R., Greenlee, R. T., Kushi, L. H., et al. (2019) · JAMA · doi:10.1001/jama.2019.11456 · Cited by 711 · cited 1× · see citing works
“The importance of medical imaging is exempli fi ed by the consistent rate of growth in medical imaging utilization in modern healthcare 1,2 .”
BackgroundSelf-supervised learning in medicine and healthcare
Krishnan, R., Rajpurkar, P., & Topol, E. J. (2022) · Nature Biomedical Engineering · doi:10.1038/s41551-022-00914-1 · Cited by 605 · cited 1× · see citing works
“1), as well as a few narrative reviews 21,22 , the most suitable strategies and best practices for medical images have not been suf fi ciently investigated.”
BackgroundHuang, S.-C., Shen, L., Lungren, M. P., & Yeung, S. (2021) · 2021 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv48922.2021.00391 · Cited by 489 · see citing works
Self-citationContrastive Learning of Medical Visual Representations from Paired Images and TextPreprint
Zhang, Y., Jiang, H., Miura, Y., Manning, C. D., & Langlotz, C. P. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2010.00747 · Cited by 465 · see citing works
McDonald, R. J., Schwartz, K. M., Eckel, L. J., Diehn, F. E., Hunt, C. H., Bartholmai, B. J., Erickson, B. J., & Kallmes, D. F. (2015) · Academic Radiology · doi:10.1016/j.acra.2015.05.007 · Cited by 460 · see citing works
Larson, D. B., Chen, M. C., Lungren, M. P., Halabi, S. S., Stence, N. V., & Langlotz, C. P. (2018) · Radiology · doi:10.1148/radiol.2017170236 · Cited by 456 · cited 1× · see citing works
“…accuracy comparable to that of human experts in narrow clinical tasks for several medical domains and imaging modalities, including chest and extremity X-rays 8 – 10 , computed tomography (CT) 11 , magnetic resonance imaging (MRI) 12 , whole slide images (WSI) 13,14 , and dermatology images 15 .”
BackgroundSelf-citationSELF-LABELLING VIA SIMULTANEOUS CLUSTERING AND REPRESENTATION LEARNINGPreprint
Asano, Y. M., Rupprecht, C., & Vedaldi, A. (2019) · Oxford University Research Archive (ORA) (University of Oxford) · doi:10.48550/arxiv.1911.05371 · Cited by 450 · see citing works
CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison
Irvin, J., Rajpurkar, P., Ko, M., Yu, Y., Ciurea-Ilcus, S., Chute, C., Marklund, H., Haghgoo, B., Ball, R., Shpanskaya, K., et al. (2019) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v33i01.3301590 · Cited by 392 · see citing works
Self-citationSelf supervised contrastive learning for digital histopathology
Ciga, O., Xu, T., & Martel, A. L. (2022) · Machine Learning with Applications · doi:10.1016/j.mlwa.2021.100198 · Cited by 294 · see citing works
Deep Learning–Assisted Diagnosis of Cerebral Aneurysms Using the HeadXNet Model
Park, A., Chute, C., Rajpurkar, P., Lou, J., Ball, R. L., Shpanskaya, K., Jabarkheel, R., Kim, L. H., McKenna, E., Tseng, J., et al. (2019) · JAMA Network Open · doi:10.1001/jamanetworkopen.2019.5600 · Cited by 258 · cited 1× · see citing works
“…accuracy comparable to that of human experts in narrow clinical tasks for several medical domains and imaging modalities, including chest and extremity X-rays 8 – 10 , computed tomography (CT) 11 , magnetic resonance imaging (MRI) 12 , whole slide images (WSI) 13,14 , and dermatology images 15 .”
BackgroundSelf-citationSelf-supervised learning methods and applications in medical imaging analysis: a survey
Shurrab, S., & Duwairi, R. (2022) · PeerJ Computer Science · doi:10.7717/peerj-cs.1045 · Cited by 258 · cited 1× · see citing works
“1), as well as a few narrative reviews 21,22 , the most suitable strategies and best practices for medical images have not been suf fi ciently investigated.”
BackgroundEsteva, A., Feng, J., van der Wal, D., Huang, S.-C., Simko, J. P., DeVries, S., Chen, E., Schaeffer, E. M., Morgan, T. M., Sun, Y., et al. (2022) · npj Digital Medicine · doi:10.1038/s41746-022-00613-w · Cited by 217 · cited 3× · see citing works
“In recent years, deep learning models have demonstrated diagnostic accuracy comparable to that of human experts in narrow clinical tasks for several medical domains and imaging modalities, including chest and extremity X-rays 8 – 10 , computed tomography (CT) 11 , magnetic resonance imaging (MRI) 12 , whole slide imag…”
BackgroundMethodAssessment of Convolutional Neural Networks for Automated Classification of Chest Radiographs
Dunnmon, J. A., Yi, D., Langlotz, C. P., Ré, C., Rubin, D. L., & Lungren, M. P. (2019) · Radiology · doi:10.1148/radiol.2018181422 · Cited by 206 · see citing works
Self-citationYan, R., Qu, L., Wei, Q., Huang, S.-C., Shen, L., Rubin, D. L., Xing, L., & Zhou, Y. (2023) · IEEE Transactions on Medical Imaging · doi:10.1109/tmi.2022.3233574 · Cited by 183 · see citing works
Self-citationTransPath: Transformer-Based Self-supervised Learning for Histopathological Image Classification
Wang, X., Yang, S., Zhang, J., Wang, M., Zhang, J., Huang, J., Yang, W., & Han, X. (2021) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-87237-3_18 · Cited by 178 · see citing works
Hossain, M. B., Iqbal, S., Islam, M. M., Akhtar, M. N., & Sarker, I. H. (2022) · Informatics in Medicine Unlocked · doi:10.1016/j.imu.2022.100916 · Cited by 161 · see citing works
Self-Supervised Feature Learning via Exploiting Multi-Modal Data for Retinal Disease Diagnosis
Li, X., Jia, M. J., Islam, M. T., Yu, L., & Xing, L. (2020) · IEEE Transactions on Medical Imaging · doi:10.1109/tmi.2020.3008871 · Cited by 150 · see citing works
Haghighi, F., Taher, M. R. H., Zhou, Z., Gotway, M. B., & Liang, J. (2021) · IEEE Transactions on Medical Imaging · doi:10.1109/tmi.2021.3060634 · Cited by 144 · see citing works
Crafting Better Contrastive Views for Siamese Representation Learning
Peng, X., Wang, K., Zhu, Z., Wang, M., & You, Y. (2022) · 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52688.2022.01556 · Cited by 122 · cited 1× · see citing works
“Recent work has demonstrated the bene fi t of using learned visual word masks 98,99 or spatially constrained crops 100,101 to encourage representational invariance with semantically more similar regions of an image.”
BackgroundSelf-supervised driven consistency training for annotation efficient histopathology image analysis
Srinidhi, C. L., Kim, S. W., Chen, F.-D., & Martel, A. L. (2022) · Medical Image Analysis · doi:10.1016/j.media.2021.102256 · Cited by 116 · see citing works
Rotation-Oriented Collaborative Self-Supervised Learning for Retinal Disease Diagnosis
Li, X., Hu, X., Qi, X., Yu, L., Zhao, W., Heng, P.-A., & Xing, L. (2021) · IEEE Transactions on Medical Imaging · doi:10.1109/tmi.2021.3075244 · Cited by 105 · cited 1× · see citing works
“Out of 412 studies identi fi ed from PubMed, Scopus, and ArXiv, 79 studies were included in the systematic review. prediction 70 , and ultrasound video to speech correspondence prediction 53 .”
BackgroundSelf-Supervised Pretraining Improves Self-Supervised Pretraining
Reed, C., Yue, X., Nrusimha, A., Ebrahimi, S., Vijaykumar, V., Mao, R., Li, B., Zhang, S., Guillory, D., Metzger, S. L., et al. (2022) · IEEE Winter Conference on Applications of Computer Vision · doi:10.1109/wacv51458.2022.00112 · Cited by 105 · see citing works
The RSNA Pulmonary Embolism CT Dataset
Colak, E., Kitamura, F., Hobbs, S., et al. (2021) · Radiology Artificial Intelligence · doi:10.1148/ryai.2021200254 · Cited by 104 · cited 2× · see citing works
“The choice of this domain is likely a combination of the availability of large-scale public chest datasets such as CheXpert 73 , RSPECT 74 , RadFusion 75 and MIMIC-CXR 76 , as well as the motivation to solve acute or emerging healthcare threats, which was the case during the coronavirus pandemic…”
BackgroundSelf-citationMulti-task contrastive learning for automatic CT and X-ray diagnosis of COVID-19
Li, J., Zhao, G., Tao, Y., Zhai, P., Chen, H., He, H., & Cai, T. (2021) · Pattern Recognition · doi:10.1016/j.patcog.2021.107848 · Cited by 100 · cited 2× · see citing works
“The majority of the studies that remained after our full-text screening (44/79) used contrastive learning as their self-supervised pre-training strategy (Table 1).”
BackgroundZhou, H.-Y., Yu, S., Bian, C., Hu, Y., Ma, K., & Zheng, Y. (2020) · Lecture notes in computer science · doi:10.1007/978-3-030-59710-8_39 · Cited by 100 · see citing works
Learning Representations by Predicting Bags of Visual Words
Gidaris, S., Bursuc, A., Komodakis, N., Pérez, P., & Cord, M. (2020) · IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings · doi:10.1109/cvpr42600.2020.00696 · Cited by 97 · see citing works
Trend in radiologist workload compared to number of admissions in the emergency department
Dan Lantsman, C., Barash, Y., Klang, E., Guranda, L., Konen, E., & Tau, N. (2022) · European Journal of Radiology · doi:10.1016/j.ejrad.2022.110195 · Cited by 85 · cited 1× · see citing works
“Studies have shown that an average radiologist now needs to interpret one image every 3 – 4 s to keep up with clinical workloads 3 – 5 .”
BackgroundAbbas, A., Abdelsamea, M. M., & Gaber, M. M. (2021) · IEEE Transactions on Neural Networks and Learning Systems · doi:10.1109/tnnls.2021.3082015 · Cited by 83 · see citing works
ContIG: Self-supervised Multimodal Contrastive Learning for Medical Imaging with Genetics
Taleb, A., Kirchler, M., Monti, R., & Lippert, C. (2022) · 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52688.2022.02024 · Cited by 83 · see citing works
Contrastive Learning with Continuous Proxy Meta-data for 3D MRI Classification
Dufumier, B., Gori, P., Victor, J., Grigis, A., Wessa, M., Brambilla, P., Favre, P., Polosan, M., McDonald, C., Piguet, C. M., et al. (2021) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-87196-3_6 · Cited by 82 · see citing works
Robust and Efficient Medical Imaging with Self-SupervisionPreprint
Azizi, S., Culp, L., Freyberg, J., Mustafa, B., Baur, S., Kornblith, S., Chen, T., MacWilliams, P., Mahdavi, S. S., Wulczyn, E., et al. (2022) · arXiv.org · doi:10.48550/arxiv.2205.09723 · Cited by 82 · see citing works
Burlina, P., Paul, W., Mathew, P., et al. (2020) · JAMA Ophthalmology · doi:10.1001/jamaophthalmol.2020.3269 · Cited by 77 · see citing works
Tian, Y., Pang, G., Liu, F., Chen, Y., Shin, S. H., Verjans, J. W., Singh, R., & Carneiro, G. (2021) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-87240-3_13 · Cited by 77 · cited 1× · see citing works
“The innate relationship tasks range from augmentation prediction and patch positioning prediction 69 , rotation Fig.”
BackgroundSelf-Supervised Deep Convolutional Neural Network for Chest X-Ray Classification
Gazda, M., Plavka, J., Gazda, J., & Drotár, P. (2021) · IEEE Access · doi:10.1109/access.2021.3125324 · Cited by 76 · see citing works
Multiple Instance Captioning: Learning Representations from Histopathology Textbooks and Articles
Gamper, J., & Rajpoot, N. (2021) · 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr46437.2021.01628 · Cited by 75 · see citing works
Delay and misdiagnosis in sub-massive and non-massive acute pulmonary embolism
Alonso-Martínez, J. L., Sánchez, F. J. A., & Echezarreta, M. A. U. (2010) · European Journal of Internal Medicine · doi:10.1016/j.ejim.2010.04.005 · Cited by 74 · see citing works
Hong, A. S., Levin, D. C., Parker, L., et al. (2019) · Radiology · doi:10.1148/radiol.2019191116 · Cited by 73 · cited 1× · see citing works
“The importance of medical imaging is exempli fi ed by the consistent rate of growth in medical imaging utilization in modern healthcare 1,2 .”
BackgroundLongitudinal self-supervised learning
Zhao, Q., Liu, Z., Adeli, E., & Pohl, K. M. (2021) · Medical Image Analysis · doi:10.1016/j.media.2021.102051 · Cited by 71 · cited 3× · see citing works
“Finally, Zhao et al. use a generative approach with an autoencoder with an additional constraint that explicitly associates brain age to the latent representations for longitudinally acquired brain MRIs 58 .”
BackgroundInfluentialFederated Contrastive Learning for Decentralized Unlabeled Medical Images
Dong, N., & Voiculescu, I. D. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-87199-4_36 · Cited by 64 · see citing works
Vu, Y. N. T., Wang, R., Balachandar, N., Liu, C., Ng, A. Y., & Rajpurkar, P. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2102.10663 · Cited by 63 · see citing works
Self-Supervised Representation Learning for Ultrasound Video
Jiao, J., Droste, R., Drukker, L., Papageorghiou, A. T., & Noble, J. A. (2020) · 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI) · doi:10.1109/isbi45749.2020.9098666 · Cited by 56 · see citing works
Self-supervised Visual Representation Learning for Histopathological Images
Yang, P., Hong, Z., Yin, X., Zhu, C., & Jiang, R. (2021) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-87196-3_5 · Cited by 56 · cited 1× · see citing works
“The tasks were focused on histopathological classi fi cation, where the majority focused on colorectal cancer classi fi cation 68,84 – 87 .”
MethodYellapragada, B., Hornauer, S., Snyder, K. R., Yu, S. X., & Yiu, G. C. (2021) · Ophthalmology Retina · doi:10.1016/j.oret.2021.06.010 · Cited by 54 · see citing works
Self-supervised Mean Teacher for Semi-supervised Chest X-Ray Classification
Liu, F., Tian, Y., Cordeiro, F. R., Belagiannis, V., Reid, I. D., & Carneiro, G. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-87589-3_44 · Cited by 52 · see citing works
Hendriksen, J. M. T., Koster-van Ree, M., Morgenstern, M. J., Oudega, R., Schutgens, R. E. G., Moons, K. G. M., & Geersing, G.-J. (2017) · BMJ Open · doi:10.1136/bmjopen-2016-012789 · Cited by 48 · cited 1× · see citing works
“With such an immense cognitive burden placed on radiologists, delays in diagnosis and diagnostic errors are unavoidable 6,7 .”
BackgroundSemMAE: Semantic-Guided Masking for Learning Masked AutoencodersPreprint
Li, G., Zheng, H., Liu, D., et al. (2022) · arXiv (Cornell University) · doi:10.48550/arxiv.2206.10207 · Cited by 48 · see citing works
SALAD: Self-supervised Aggregation Learning for Anomaly Detection on X-Rays
Bozorgtabar, B., Mahapatra, D., Vray, G., & Thiran, J. (2020) · Lecture notes in computer science · doi:10.1007/978-3-030-59710-8_46 · Cited by 42 · see citing works
Zhou, Y., Huang, S.-C., Fries, J., Youssef, A., Amrhein, T. J., Chang, M., Banerjee, I., Rubin, D. L., Xing, L., Shah, N. H., et al. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2111.11665 · Cited by 40 · see citing works
Self-citationSelf supervised learning improves dMMR/MSI detection from histology slides across multiple cancersPreprint
Saillard, C., Dehaene, O., Marchand, T., Moindrot, O., Kamoun, A., Schmauch, B., & Jégou, S. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2109.05819 · Cited by 39 · see citing works
Self-Supervised Contrastive Video-Speech Representation Learning for Ultrasound
Jiao, J., Cai, Y., Alsharid, M., Drukker, L., Papageorghiou, A. T., & Noble, J. A. (2020) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-59716-0_51 · Cited by 38 · cited 3× · see citing works
“…et al. pre-trained their models simultaneously with two innate relationship pre-text tasks (slice order prediction and geometric transformation prediction), and showed that a weight-sharing Siamese network out-performs a single disentanged model for combining the two pre-training objectives 53 .”
BackgroundMethodSimTriplet: Simple Triplet Representation Learning with a Single GPU
Liu, Q., Louis, P. C., Lu, Y., Jha, A., Zhao, M., Deng, R., Yao, T., Roland, J. T., Yang, H., Zhao, S., et al. (2021) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-87196-3_10 · Cited by 38 · see citing works
Improve Image-based Skin Cancer Diagnosis with Generative Self-Supervised Learning
Ren, Z., Guo, Y., Yu, S. X., & Whitney, D. C. (2021) · IEEE/ACM Conference on Connected Health: Applications, Systems and Engineering Technologies · doi:10.1109/chase52844.2021.00011 · Cited by 36 · see citing works
HistoTransfer: Understanding Transfer Learning for Histopathology
Sharmay, Y., Ehsany, L., Syed, S., & Brown, D. E. (2021) · IEEE-EMBS International Conference on Biomedical and Health Informatics (BHI ...) · doi:10.1109/bhi50953.2021.9508542 · Cited by 32 · see citing works
Ultrasound Image Representation Learning by Modeling Sonographer Visual Attention
Droste, R., Cai, Y., Sharma, H., Chatelain, P., Drukker, L., Papageorghiou, A. T., & Noble, J. A. (2019) · Lecture notes in computer science · doi:10.1007/978-3-030-20351-1_46 · Cited by 31 · cited 1× · see citing works
“For instance, Droste et al. utilized a gaze tracking dataset and pre-trained a model to predict sonographers ’ gazes on ultrasound video frames with gaze-point regression 54 .”
MethodTargeted Self Supervision For Classification On A Small Covid-19 Ct Scan Dataset
Ewen, N., & Khan, N. (2021) · 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI) · doi:10.1109/isbi48211.2021.9434047 · Cited by 29 · see citing works
Contrastive Learning Based Stain Normalization Across Multiple Tumor in Histopathology
Ke, J., Shen, Y., Liang, X., & Shen, D. (2021) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-87237-3_55 · Cited by 28 · cited 2× · see citing works
“With the exception of Ke et al. ’ s work 68 , which uses a CycleGAN for histopathology slide stain normalization, all studies utilized an autoencoder as their generative model when combined with contrastive strategies.”
MethodSelf-supervised Longitudinal Neighbourhood Embedding
Ouyang, J., Zhao, Q., Adeli, E., Sullivan, E. V., Pfefferbaum, A., Zaharchuk, G., & Pohl, K. M. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-87196-3_8 · Cited by 26 · see citing works
SSLP: Spatial Guided Self-supervised Learning on Pathological Images
Li, J., Lin, T., & Xu, Y. (2021) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-87196-3_1 · Cited by 26 · see citing works
Modeling Disease Progression in Retinal OCTs with Longitudinal Self-supervised Learning
Rivail, A., Schmidt-Erfurth, U., Vogl, W.-D., Waldstein, S. M., Riedl, S., Grechenig, C., Wu, Z., & Bogunovic, H. (2019) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-32281-6_5 · Cited by 25 · cited 1× · see citing works
“Rivail et al. proposed a self-supervised approach to model disease progression by estimating the time interval between pairs of optical coherence tomography (OCT) scans from the same patient 104 .”
BackgroundOn Self-Supervised Multimodal Representation Learning: An Application To Alzheimer’s Disease
Fedorov, A., Wu, L., Sylvain, T., Luck, M. M., DeRamus, T. P., Bleklov, D., Plis, S., & Calhoun, V. D. (2021) · IEEE International Symposium on Biomedical Imaging (ISBI) · doi:10.1109/isbi48211.2021.9434103 · Cited by 25 · see citing works
Park, J., Kwak, I.-Y., & Lim, C. (2021) · Electronics · doi:10.3390/electronics10161996 · Cited by 24 · cited 2× · see citing works
“Two of the fi ve studies compared their approach to models without self-supervised pre-training and reported slight relative improvements in performance of 1.12% 67 and 0.690% 63 .”
BackgroundResultInfluentialSelf-Taught Semi-Supervised Anomaly Detection On Upper Limb X-Rays
Spahr, A., Bozorgtabar, B., & Thiran, J. (2021) · IEEE International Symposium on Biomedical Imaging (ISBI) · doi:10.1109/isbi48211.2021.9433771 · Cited by 23 · see citing works
Context Matters: Graph-based Self-supervised Representation Learning for Medical Images
Sun, L., Yu, K., & Batmanghelich, K. N. (2021) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v35i6.16620 · Cited by 22 · cited 1× · see citing works
“…domain is likely a combination of the availability of large-scale public chest datasets such as CheXpert 73 , RSPECT 74 , RadFusion 75 and MIMIC-CXR 76 , as well as the motivation to solve acute or emerging healthcare threats, which was the case during the coronavirus pandemic 45,46,48,67,77 – 83 .”
BackgroundLiu, C., Qiao, M., Jiang, F., Guo, Y., Jin, Z., & Wang, Y. (2021) · Medical Physics · doi:10.1002/mp.15172 · Cited by 22 · see citing works
Imbalance-Aware Self-supervised Learning for 3D Radiomic Representations
Li, H., Xue, F.-F., Chaitanya, K., Luo, S., Ezhov, I., Wiestler, B., Zhang, J., & Menze, B. H. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-87196-3_4 · Cited by 21 · see citing works
Improving Joint Learning of Chest X-Ray and Radiology Report by Word Region Alignment
Ji, Z., Shaikh, M. A., Moukheiber, D., Srihari, S. N., Peng, Y., & Gao, M. (2021) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-87589-3_12 · Cited by 21 · cited 2× · see citing works
“In radiology, Ji et al. used multimodal contrastive learning by matching X-rays with corresponding radiology reports 60 .”
BackgroundHashimoto, Y., Ogata, Y., Honda, M., & Yamashita, Y. (2021) · Frontiers in Neuroscience · doi:10.3389/fnins.2021.696853 · Cited by 19 · cited 2× · see citing works
“A particularly interesting medical task that was explored was classi fi cation of psychiatric diseases or psychiatric traits using fMRI 57,62,90 .”
BackgroundMethodZhao, Z., & Yang, G. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-87196-3_24 · Cited by 17 · see citing works
DABS: A Domain-Agnostic Benchmark for Self-Supervised LearningPreprint
Tamkin, A., Liu, V., Lu, R., Fein, D., Schultz, C., & Goodman, N. D. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2111.12062 · Cited by 16 · see citing works
Self-supervised representation learning for detection of ACL tear injury in knee MR videos
Manna, S., Bhattacharya, S., & Pal, U. (2022) · Pattern Recognition Letters · doi:10.1016/j.patrec.2022.01.008 · Cited by 14 · cited 1× · see citing works
“Nine of these studies designed their innate relationship pre-text task based on different image transformations, including rotation prediction 44 – 47 , horizontal fl ip prediction 48 , reordering shuf fl ed slices 49 , and patch order prediction 46,50 – 52 .”
BackgroundSeeking an Optimal Approach for Computer-Aided Pulmonary Embolism Detection
Islam, N. U., Gehlot, S., Zhou, Z., Gotway, M. B., & Liang, J. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-87589-3_71 · Cited by 13 · see citing works
Case Discrimination: Self-supervised Feature Learning for the Classification of Focal Liver Lesions
Dong, H., Iwamoto, Y., Han, X., Lin, L., Hu, H., Cai, X., & Chen, Y.-W. (2021) · Smart Innovation, Systems and Technologies · doi:10.1007/978-981-16-3013-2_20 · Cited by 12 · see citing works
Context Matters: Graph-based Self-supervised Representation Learning for Medical Images.
Sun, L., Yu, K., & Batmanghelich, K. N. (2021) · PubMed · Cited by 10 · see citing works
Inter-regional High-Level Relation Learning from Functional Connectivity via Self-supervision
Jung, W., Heo, D.-W., Jeon, E., Lee, J.-I., & Suk, H. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-87196-3_27 · Cited by 10 · cited 1× · see citing works
“Jung et al. created a functional connectivity matrix between pairs of region-of-interest in rs-fMRI for each subject, and created a masked auto-encoder task by randomly masking out different rows and columns of the matrix for restoration 66 .”
MethodContext Matters: Graph-based Self-supervised Representation Learning for Medical ImagesPreprint
Sun, L., Yu, K., & Batmanghelich, K. N. (2020) · arXiv (Cornell University) · doi:10.48550/arxiv.2012.06457 · Cited by 10 · cited 1× · see citing works
“…domain is likely a combination of the availability of large-scale public chest datasets such as CheXpert 73 , RSPECT 74 , RadFusion 75 and MIMIC-CXR 76 , as well as the motivation to solve acute or emerging healthcare threats, which was the case during the coronavirus pandemic 45,46,48,67,77 – 83 .”
BackgroundDezaki, F. T., Ginsberg, T., Luong, C. L., et al. (2021) · IEEE International Symposium on Biomedical Imaging (ISBI) · doi:10.1109/isbi48211.2021.9433766 · Cited by 9 · see citing works
Human Embryo Classification Using Self-Supervised Learning
Wicaksono, R. S. H., Septiandri, A. A., & Jamal, A. (2021) · International Conference on Artificial Intelligence and Data Sciences (AiDAS) · doi:10.1109/aidas53897.2021.9574328 · Cited by 9 · see citing works
Real-World Multi-Domain Data Applications for Generalizations to Clinical Settings
Mojab, N., Noroozi, V., Yi, D., et al. (2020) · IEEE International Conference on Machine Learning and Applications (ICMLA) · doi:10.1109/icmla51294.2020.00112 · Cited by 8 · see citing works
Liver Fibrosis And NAS Scoring From CT Images Using Self-Supervised Learning And Texture Encoding
Jana, A., Qu, H. Y., Minacapelli, C. D., Catalano, C., Rustgi, V. K., & Metaxas, D. (2021) · IEEE International Symposium on Biomedical Imaging (ISBI) · doi:10.1109/isbi48211.2021.9433920 · Cited by 8 · see citing works
Triplet Self-Supervised Learning for Gastritis Detection with Scarce Annotations
Li, G., Togo, R., Ogawa, T., et al. (2021) · IEEE Global Conference on Consumer Electronics (GCCE) · doi:10.1109/gcce53005.2021.9622100 · Cited by 8 · see citing works
A Semi-Supervised Learning Method to Remedy the Lack of Labeled Data
Nguyen, N.-Q., & Le, T.-S. (2021) · International Conference on Advanced Computing and Applications (ACOMP) · doi:10.1109/acomp53746.2021.00017 · Cited by 8 · see citing works
Esteva, A., Feng, J., Huang, S.-C., van der Wal, D., Simko, J., DeVries, S., Chen, E., Schaeffer, E. M., Morgan, T. M., Monson, J. M., et al. (2022) · Journal of Clinical Oncology · doi:10.1200/jco.2022.40.6_suppl.222 · Cited by 7 · cited 1× · see citing works
“…accuracy comparable to that of human experts in narrow clinical tasks for several medical domains and imaging modalities, including chest and extremity X-rays 8 – 10 , computed tomography (CT) 11 , magnetic resonance imaging (MRI) 12 , whole slide images (WSI) 13,14 , and dermatology images 15 .”
BackgroundSelf-citationDeep learning models based on automatic labeling with application in echocardiography
Danu, M. D., Ciusdel, C. F., & Lucian Mihai, I. T. U. (2020) · International Conference on System Theory, Control and Computing (ICSTCC) · doi:10.1109/icstcc50638.2020.9259701 · Cited by 6 · see citing works
Hsieh, W., Lefort‐Besnard, J., Yang, H.-C., et al. (2020) · Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society · doi:10.1109/embc44109.2020.9175312 · Cited by 5 · see citing works
A Preliminary Analysis of Self-Supervision for Wireless Capsule Endoscopy
Vats, A., Pedersen, M., & Mohammed, A. (2021) · 2021 9th European Workshop on Visual Information Processing (EUVIP) · doi:10.1109/euvip50544.2021.9484012 · Cited by 5 · cited 2× · see citing works
“Nine of these studies designed their innate relationship pre-text task based on different image transformations, including rotation prediction 44 – 47 , horizontal fl ip prediction 48 , reordering shuf fl ed slices 49 , and patch order prediction 46,50 – 52 .”
BackgroundOnline Unsupervised Learning For Domain Shift In Covid-19 CT Scan Datasets
Ewen, N., & Khan, N. (2021) · 2021 IEEE International Conference on Autonomous Systems (ICAS) · doi:10.1109/icas49788.2021.9551146 · Cited by 5 · cited 1× · see citing works
“…domain is likely a combination of the availability of large-scale public chest datasets such as CheXpert 73 , RSPECT 74 , RadFusion 75 and MIMIC-CXR 76 , as well as the motivation to solve acute or emerging healthcare threats, which was the case during the coronavirus pandemic 45,46,48,67,77 – 83 .”
BackgroundZhong, H., Wu, J., Zhao, W., Xu, X., Hou, R., Zhao, L., Deng, Z., Zhang, M., & Zhao, J. (2021) · Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) · doi:10.1109/embc46164.2021.9630228 · Cited by 5 · see citing works
Tian, Y., Pang, G., Liu, F., et al. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2103.03423 · Cited by 5 · see citing works
SISE-PC: Semi-supervised Image Subsampling for Explainable Pathology Classification
Roychowdhury, S., Tang, K. S., Ashok, M., et al. (2021) · Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society · doi:10.1109/embc46164.2021.9630682 · Cited by 4 · see citing works
Zhang, S., Zou, B., Xu, B., Su, J., & Hu, H. (2021) · IEEE International Conference on Progress in Informatics and Computing · doi:10.1109/pic53636.2021.9687080 · Cited by 4 · see citing works
Learning Personal Representations from fMRI by Predicting Neurofeedback Performance
Osin, J., Wolf, L., Gurevitch, G., Keynan, J. N., Fruchtman-Steinbok, T., Or-Borichev, A., & Hendler, T. (2020) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-59728-3_46 · Cited by 3 · see citing works
Jian, G.-Z., Lin, G.-S., Wang, C.-M., et al. (2021) · International Conference on Graphics and Signal Processing (ICGSP) · doi:10.1145/3474906.3474912 · Cited by 3 · see citing works
Fedorov, A., Geenjaar, E., Wu, L., DeRamus, T. P., Calhoun, V. D., & Plis, S. (2021) · arXiv (Cornell University) · doi:10.48550/arxiv.2103.15914 · Cited by 3 · see citing works
Self Supervised Contrastive Learning on Multiple Breast Modalities Boosts Classification Performance
Perek, S., Amit, M., & Hexter, E. (2021) · Lecture notes in computer science · doi:10.1007/978-3-030-87602-9_11 · Cited by 3 · see citing works
Interpretive self-supervised pre-training
Manna, S., Bhattacharya, S., & Pal, U. (2021) · Indian Conference on Computer Vision, Graphics and Image Processing · doi:10.1145/3490035.3490273 · Cited by 3 · see citing works
Fast Mixing of Hard Negative Samples for Contrastive Learning and Use for COVID-19
Zhao, X. B., & Zhou, S. (2021) · International Conference on Big Data Technologies (ICBDT) · doi:10.1145/3490322.3490324 · Cited by 3 · see citing works
The Forgotten Underrepresented Minorities: A Call for Data Disaggregation
Truong, O. (2022) · Family Medicine · doi:10.22454/fammed.2022.582301 · Cited by 3 · see citing works
Self-supervised Learning for Small Shot COVID-19 Classification
Zhu, Y. (2021) · 2021 3rd International Conference on Information Technology and Computer Communications · doi:10.1145/3473465.3473472 · Cited by 2 · see citing works
Pneumonia Identification with Self-supervised Learning and Transfer Learning
Long, Y. (2021) · Advances in Intelligent Systems and Computing · doi:10.1007/978-3-030-74811-1_90 · Cited by 2 · see citing works
Cornelissen, S., van der Putten, J. A., Boers, T. G. W., Jukema, J. B., Fockens, K. N., Bergman, J. J. G. H. M., van der Sommen, F., & de With, P. H. N. (2021) · 2021 IEEE International Conference on Image Processing (ICIP) · doi:10.1109/icip42928.2021.9506121 · Cited by 2 · see citing works
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805
Flagged first, then by citations
6 of the 130 references couldn’t be matched to a known paper.
Research works printed but not linked: 21
- Brown, T., Mann, B. & Ryder, N. Language models are few-shot learners. Adv. Neural Inf. Process. Syst. (2020).
- Chen, T., Kornblith, S., Norouzi, M. & Hinton, G. A Simple Framework for Contrastive Learning of Visual Representations. Proc. 37th Int. Conf. Mach. Learn., PMLR 119, 1597–1607 (2020).
- Lilian Weng, J. W. K. Self-Supervised Learning: Self-Prediction and Contrastive Learning. Adv. Neural Inf. Proc. Syst. (2021).
- Goodfello w, I. J. et al. Generative Adversarial Networks. Preprint at https://community.unix.com/uploads/short-url/oXsmq2VZ9hc2X6hwPRXZRMGbV20.pdf (2014).
- Donahue J. & Simonyan K. Large scale adversarial representation learning. Adv. Neural Inf. Process. Syst. (2019).
- Grill J. B., Strub F., Altché F. & Tallec C. Bootstrap your own latent-a new approach to self-supervised learning. Adv. Neural Inf. Proc. Syst. 33, (2020).
- Chen, X. & He, K. Exploring Simple Siamese Representation Learning. Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, 15750–15758 (2021).
- Pathak, D., Krahenbuhl, P., Donahue, J., Darrell, T. & Efros, A. A. Context Encoders: Feature Learning by Inpainting. Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR) 2016, 2536–2544 (2016).
- He, K. et al. Masked Autoencoders Are Scalable Vision Learners. Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) 2022, 16000–16009 (2022).
- Vu, Y. N. T., Tsue, T., Su, J. & Singh, S. An improved mammography malignancy model with self-supervised learning. Med. Imaging 2021: Comput-Aided Diagn. 11597, 210–216 (2021).
- Hao, H., Didari, S., Woo, J. O., Moon, H. & Bangert, P. Highly Efficient Representation and Active Learning Framework for Imbalanced Data and its Application to COVID-19 X-Ray Classification. Conference on Neural Information Processing Systems (NeurIPS 2021) (2021).
- Dehaene, O., Camara, A., Moindrot, O., de Lavergne, A. & Courtiol, P. Self-Supervision Closes the Gap Between Weak and Strong Supervision in Histology. in ML4H 2020 NeurIPS workshop (2020).
- Kotar, K., Ilharco, G. & Schmidt, L. Contrasting contrastive self-supervised representation learning pipelines. Proc. Estonian Acad. Sci. Biol. Ecol.
- Cole, E., Yang, X., Wilber, K., Aodha, O. M. & Belongie, S. When Does Contrastive Visual Representation Learning Work? Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR) 2022, 14755–14764 (2022).
- Azizi, S., Mustafa, B., Ryan, F. & Beaver, Z. Big self-supervised models advance medical image classification. Proc. Estonian Acad. Sci. Biol. Ecol.
- Shi, Y., Siddharth, N., Torr, P. H. S. & Kosiorek, A. R. Adversarial Masking for Self-Supervised Learning. In International Conference on Machine Learning, 20026–20040. (2022).
- Van Gansbeke, X. & Vandenhende, S. Revisiting contrastive methods for unsupervised learning of visual representations. Adv. Neural Inf. Process. Syst. 34, 16238–16250 (2021).
- Sowrirajan, H., Yang, J., Ng, A. Y. & Rajpurkar, P. MoCo-CXR: MoCo Pre-training Improves Representation and Transferability of Chest X-ray Models. Proc. Mach. Learn. Res. 143, 727–743 (2021).
- Bao, W., Jin, Y., Huang, C. & Peng, W. CT Image Classification of Invasive Depth of Gastric Cancer based on 3D-DPN Structure. In The 11th International Workshop on Computer Science and Engineering (WCSE 2021) 115–121 (2021).
- Kaku, A., Upadhya, S. & Razavian, N. Intermediate layers matter in momentum contrastive self supervised learning. In 35th Conference on Neural Information Processing Systems (NeurIPS 2021). (2021).
- Liu, Q. et al. SimTriplet: Simple Triplet Representation Learning with a Single GPU. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2021 102–112 (2021). · DOI · Paper page
Cited by
Sorted by citations
Advances in medical image analysis with vision Transformers: A comprehensive review
Azad, R., Kazerouni, A., Heidari, M., Aghdam, E. K., Molaei, A., Jia, Y., Jose, A., Roy, R., & Merhof, D. (2023) · Medical Image Analysis · doi:10.1016/j.media.2023.103000 · Cited by 523 · see citing works
Foundation Model for Cancer Imaging Biomarkers
Pai, S., Bontempi, D., Hadzic, I., Prudente, V., Sokač, M., Chaunzwa, T. L., Bernatz, S., Hosny, A., Mak, R. H., Birkbak, N. J., et al. (2024) · Nature Machine Intelligence · doi:10.1038/s42256-024-00807-9 · Cited by 199 · see citing works
“The expanding literature on SSL in medical imaging 13 focuses primarily on two-dimensional (2D) images (X-ray, whole slide images, dermatology images, fundus images and so on) for diagnostic applications.”
Machine learning in precision diabetes care and cardiovascular risk prediction
Oikonomou, E. K., & Khera, R. (2023) · Cardiovascular Diabetology · doi:10.1186/s12933-023-01985-3 · Cited by 190 · see citing works
“However, instead of focusing on tasks like clustering, SSL attempts to solve tasks traditionally addressed through supervised learning, such as classification and regression [26].”
BackgroundWaqas, A., Bui, M. M., Glassy, E. F., Naqa, I. E., Borkowski, P., Borkowski, A., & Rasool, G. (2023) · Laboratory Investigation · doi:10.1016/j.labinv.2023.100255 · Cited by 149 · see citing works
“[44, 72, 73, 74] neering algorithms that have significantly impacted pathology practices by aiding in diagnosing and characterizing various diseases.”
A Comprehensive Survey of Deep Learning Approaches in Image Processing
Τρίγκα, Μ., & Dritsas, E. (2025) · Sensors · doi:10.3390/s25020531 · Cited by 159 · see citing works
A GAN-Based Data Augmentation Method for Imbalanced Multi-Class Skin Lesion Classification
Su, Q., Hamed, H. N. A., Isa, M. A., Xue, H., & Dai, X. (2024) · IEEE Access · doi:10.1109/access.2024.3360215 · Cited by 131 · see citing works
“Additionally, the associated distortion magnitude parameter of geometric and color space transformations need to be carefully chosen, since heavily transformed images could alter the semantic meaning of the medical images [17], [24].”
BackgroundA vision–language foundation model for the generation of realistic chest X-ray images
Bluethgen, C., Chambon, P., Delbrouck, J.-B., van der Sluijs, R., Połacin, M., Zambrano Chaves, J. M., Abraham, T. M., Purohit, S., Langlotz, C. P., & Chaudhari, A. S. (2024) · Nature Biomedical Engineering · doi:10.1038/s41551-024-01246-y · Cited by 108 · see citing works
“SSL is a method for deriving meaningful representations from unla-belled data by pre-training models on auxiliary objectives instead of directly on a target downstream task 22 .”
Self-citationSilva-Rodríguez, J., Chakor, H., Kobbi, R., Dolz, J., & Ayed, I. B. (2024) · Medical Image Analysis · doi:10.1016/j.media.2024.103357 · Cited by 121 · see citing works
“Larger-scale pre-training, which leverages unlabeled data via self-supervision (Chen et al., 2022a; Huang et al., 2023), is a promising alternative.”
BackgroundFoundation Models in Radiology: What, How, Why, and Why Not
Paschali, M., Chen, Z., Blankemeier, L., Varma, M., Youssef, A., Blüthgen, C., Langlotz, C. P., Gatidis, S., & Chaudhari, A. S. (2025) · Radiology · doi:10.1148/radiol.240597 · Cited by 118 · see citing works
Self-citationGalić, I., Habijan, M., Leventić, H., & Romić, K. (2023) · Electronics · doi:10.3390/electronics12214411 · Cited by 108 · see citing works
Self-supervised learning for medical image analysis: a comprehensive review
Rani, V., Kumar, M., Gupta, A., Sachdeva, M., Mittal, A., & Kumar, K. (2024) · Evolving Systems · doi:10.1007/s12530-024-09581-w · Cited by 76 · see citing works
“• Object detection : Semi-supervised learning for object detection in medical images can aid in the identification and localization of diverse objects, including tumors, nodules, or lesions (Huang et al. 2023).”
InfluentialCXR-LLaVA: a multimodal large language model for interpreting chest X-ray images
Lee, S., Youn, J., Kim, H., Kim, M., & Yoon, S. H. (2025) · European Radiology · doi:10.1007/s00330-024-11339-6 · Cited by 76 · see citing works
Merlin: A Vision Language Foundation Model for 3D Computed TomographyPreprint
Blankemeier, L., Cohen, J., Kumar, A., et al. (2024) · Research Square · doi:10.21203/rs.3.rs-4546309/v1 · Cited by 58 · see citing works
“Prior approaches for 3D CT pretraining focus on image-only pretraining [67, 68, 69].”
Self-citationThe Future of AI and Informatics in Radiology: 10 Predictions
Langlotz, C. P. (2023) · Radiology · doi:10.1148/radiol.231114 · Cited by 56 · see citing works
Wolf, D., Payer, T., Lisson, C. S., Lisson, C. G., Beer, M., GOETZ, M. A., & Ropinski, T. (2023) · Scientific Reports · doi:10.1038/s41598-023-46433-0 · Cited by 64 · see citing works
“A promising solution to overcome these challenges is self-supervised learning (SSL) 8 .”
InfluentialConsequential Advancements of Self-Supervised Learning (SSL) in Deep Learning Contexts
Abdulrazzaq, M. M., Ramaha, N. T. A., Hameed, A. A., Salman, M., Yon, D., Fitriyani, N. L., Syafrudin, M., & Lee, S. (2024) · Mathematics · doi:10.3390/math12050758 · Cited by 52 · see citing works
“In their overview [20], the scholars identified a collection of some medical disciplines in which SSL models can be advantageous in conducting the classification process flexibly, which can be illustrated in Figure 12.”
BackgroundMethodInfluentialVanberlo, B., Hoey, J., & Wong, A. (2024) · BMC Medical Imaging · doi:10.1186/s12880-024-01253-0 · Cited by 59 · see citing works
“This survey distinguishes itself from prior works in that it includes more recent literature and its scope is limited to four radiological modalitiesUn-like [13], this survey includes applications other than classification.”
MethodMerlin: a computed tomography vision–language foundation model and dataset
Blankemeier, L., Kumar, A., Cohen, J., Liu, J., Liu, L., Veen, D. V., Gardezi, S. J. S., Yu, H., Paschali, M., Chen, Z., et al. (2026) · Nature · doi:10.1038/s41586-026-10181-8 · Cited by 124 · see citing works
Self-citationZambrano Chaves, J. M., Huang, S.-C., Xu, Y., Xu, H., Usuyama, N., Zhang, S., Wang, F., Xie, Y., Khademi, M., Yang, Z., et al. (2025) · Nature Communications · doi:10.1038/s41467-025-58344-x · Cited by 45 · see citing works
“Additionally, we propose CheXprompt, a factuality evaluation method based on GPT-4.”
Self-citationSelf-supervised learning framework application for medical image analysis: a review and summary
Zeng, X., Abdullah, N., & Sumari, P. (2024) · BioMedical Engineering OnLine · doi:10.1186/s12938-024-01299-9 · Cited by 49 · see citing works
Weakly Supervised Deep Learning in Radiology
Misera, L., Müller‐Franzes, G. A., Truhn, D., & Kather, J. N. (2024) · Radiology · doi:10.1148/radiol.232085 · Cited by 39 · see citing works
Radiology and multi-scale data integration for precision oncology
Paverd, H., Zormpas-Petridis, K., Clayton, H., Burge, S., & Crispin-Ortuzar, M. (2024) · npj Precision Oncology · doi:10.1038/s41698-024-00656-0 · Cited by 41 · see citing works
Benachour, Y., Flitti, F., & Khalid, H. M. (2025) · IEEE Access · doi:10.1109/access.2025.3542160 · Cited by 35 · see citing works
“These methods have been shown to be effective in not only learning useful representations from medical images but also in improving the accuracy of classifications, making it a valuable tool in enhancing healthcare and patient outcomes [24].”
Explainable AI reveals Clever Hans effects in unsupervised learning models
Kauffmann, J., Dippel, J., Ruff, L., Samek, W., Müller, K., & Montavon, G. (2025) · Nature Machine Intelligence · doi:10.1038/s42256-025-01000-2 · Cited by 37 · see citing works
“Further motivated by the need to accommodate the critically small number of COVID-19 instances in the aggregated dataset and to avoid overfitting, we choose to rely on the representations provided by unsupervised foundation models [15, 35–37].”
Mohammad‐Rahimi, H., Dianat, O., Abbasi, R., Zahedrozegar, S., Ashkan, A., Motamedian, S. R., Rohban, M. H., & Nosrat, A. (2023) · Journal of Endodontics · doi:10.1016/j.joen.2023.11.004 · Cited by 38 · see citing works
Linguraru, M. G., Bakas, S., Aboian, M., Chang, P., Flanders, A. E., Kalpathy–Cramer, J., Kitamura, F., Lungren, M. P., Mongan, J., Prevedello, L. M. S., et al. (2024) · Radiology Artificial Intelligence · doi:10.1148/ryai.240225 · Cited by 39 · see citing works
Self-citationOn the use of contrastive learning for standard-plane classification in fetal ultrasound imaging
Migliorelli, G., Fiorentino, M. C., Cosmo, M. D., Villani, F. P., Mancini, A., & Moccia, S. (2024) · Computers in Biology and Medicine · doi:10.1016/j.compbiomed.2024.108430 · Cited by 29 · see citing works
Wen, R., Yuan, H., Ni, D., Xiao, W., & Wu, Y. (2023) · IEEE Workshop/Winter Conference on Applications of Computer Vision · doi:10.1109/wacv57701.2024.00052 · Cited by 49 · see citing works
“In the medical field, several works have demonstrated that SSL can produce a pretrained model to advance supervised tasks [14], such as image classification [1, 14] and image segmentation [15, 29].”
BackgroundMedical image foundation models in assisting diagnosis of brain tumors: a pilot study
Chen, M., Zhang, M., Yin, L., Ma, L., Ding, R., Zheng, T., Yue, Q., Lui, S., & Sun, H. (2024) · European Radiology · doi:10.1007/s00330-024-10728-1 · Cited by 29 · see citing works
“The pretext tasks are usually designed to randomly transform the input data, subsequently tasking the model with either reconstructing the original input data or predicting the applied transformations [11].”
Sükei, E., Rumetshofer, E., Schmidinger, N., Mayr, A., Schmidt-Erfurth, U. M., Klambauer, G., & Bogunović, H. (2024) · Scientific Reports · doi:10.1038/s41598-024-78515-y · Cited by 29 · see citing works
Metaheuristics role in image processing and computer vision applications: a comprehensive review
Şahin, M. F., & Anka, F. (2025) · Cluster Computing · doi:10.1007/s10586-025-05610-8 · Cited by 31 · see citing works
“However, due to the presence of different organs in these dis - eases, imaging systems vary [81, 82 ].”
Evaluating Large Language Models in Ophthalmology: Systematic Review
Zhang, Z., Zhang, H., Pan, Z., Bi, Z., Wan, Y., Song, X., & Fan, X. (2025) · Journal of Medical Internet Research · doi:10.2196/76947 · Cited by 35 · see citing works
“…original or modified LLMs (fine-tuned, retrieval-augmented generation enhanced [34], or pipeline integrated) In line with recent AI-focused reviews [35], we used domains that are more appropriate for AI model evaluations rather than the traditional patient-centered population, intervention,…”
Deep learning in pediatric neuroimaging
Wang, J., Wang, J., Wang, S., et al. (2023) · Displays · doi:10.1016/j.displa.2023.102583 · Cited by 25 · see citing works
Beyond Supervised: The Rise of Self-Supervised Learning in Autonomous Systems
Taherdoost, H. (2024) · Information · doi:10.3390/info15080491 · Cited by 26 · see citing works
Nielsen, M., Wenderoth, L., Sentker, T., & Werner, R. (2023) · Bioengineering · doi:10.3390/bioengineering10080895 · Cited by 24 · see citing works
Guo, H., Somayajula, S. A., Hosseini, R., & Xie, P.-T. (2024) · Scientific Reports · doi:10.1038/s41598-024-53955-8 · Cited by 22 · see citing works
“This method has been particularly beneficial in medical imaging, where labeled data can be scarce and expensive to obtain 42,43 .”
MethodArtificial Intelligence Virtual Organoids (AIVOs)
Bai, L., & Su, J. (2025) · Bioactive Materials · doi:10.1016/j.bioactmat.2025.12.030 · Cited by 34 · see citing works
“Classical machine learning models, including supervised, unsupervised and representation learning methods [30], are used to classify organoid samples, cluster cell states, infer latent factors and link molecular profiles to phenotypic outputs.”
Shen, Y., Li, H., Sun, C., Ji, H., Zhang, D., Hu, K., Tang, Y., Chen, Y., Wei, Z., & Lv, J. (2024) · npj Digital Medicine · doi:10.1038/s41746-024-01014-x · Cited by 21 · see citing works
“There are prior works exploring contrastive learning in dermatological diagnosis, emphasizing its capability to extract consistent representations and enhance generalizability and diagnostic accuracy 16,17 .”
BackgroundSuper-resolution techniques for biomedical applications and challenges
Shin, M., Seo, M., Lee, K., & Yoon, K. (2024) · Biomedical Engineering Letters · doi:10.1007/s13534-024-00365-4 · Cited by 21 · see citing works
“Biomedical images often have unique characteristics such as lower contrast, higher noise levels, and more complex structures compared to natural images [31–33].”
Gomaa, A., Huang, Y., Stephan, P., Breininger, K., Frey, B., Doerfler, A., Schnell, O., Delev, D., Coras, R., Donaubauer, A.-J., et al. (2025) · Scientific Reports · doi:10.1038/s41598-025-02026-7 · Cited by 20 · see citing works
“The attention mechanism of transformers enables them to capture both global and local context information within the data, making them well-suited for tasks involving complex spatial relationships 18,20,21 .”
Overcoming barriers in the use of artificial intelligence in point of care ultrasound
Vega, R., Dehghan, M., Nagdev, A. D., Buchanan, B. M., Kapur, J., Jaremko, J. L., & Zonoobi, D. (2025) · npj Digital Medicine · doi:10.1038/s41746-025-01633-y · Cited by 26 · see citing works
“Similarly, the use of self-supervised learning, which leverages the use of large amountsofunlabeleddata,isanactiveareaofresearchthatpromisestolead tobetterpredictivemodels 71 .”
Shi, G., Lu, H., Hui, H., et al. (2025) · Medical Image Analysis · doi:10.1016/j.media.2024.103442 · Cited by 19 · see citing works
A Method‐Oriented Review of Explainable Artificial Intelligence for Neurological Medical Imaging
Peng, C., Li, L., & Peng, D. (2025) · Expert Systems · doi:10.1111/exsy.70111 · Cited by 20 · see citing works
Hu, Y., Lui, A., Goldstein, M., et al. (2024) · European Heart Journal Acute Cardiovascular Care · doi:10.1093/ehjacc/zuae037 · Cited by 18 · see citing works
Retinal disease diagnosis with unsupervised Grad-CAM guided contrastive learning
Zhao, Z., Chen, H., Wang, Y., Meng, D., Xie, Q., Yu, Q., & Wang, L. (2024) · Neurocomputing · doi:10.1016/j.neucom.2024.127816 · Cited by 18 · see citing works
Vaidya, S. S., Ali, R. H., Faiz, S., et al. (2025) · Algorithms · doi:10.3390/a18100624 · Cited by 17 · see citing works
A Review on Research and Application of AI-Based Image Analysis in the Field of Computer Vision
Wu, P., He, X., Dai, W., Zhou, J., Shang, Y., Fan, Y., & Hu, T. (2025) · IEEE Access · doi:10.1109/access.2025.3565300 · Cited by 19 · see citing works
“SSL in [26] reduces annotation dependency in multimodal medical classification, with end-to-end fine-tuning increasing Area Under the Receiver Operating Characteristic Curve (AUROC) by 0.113.”
InfluentialFoundation models for radiology—the position of the AI for Health Imaging (AI4HI) network
de Almeida, J. G., Alberich, L. C., Tsakou, G., Kostas Marias, A., Tsiknakis, M., Lekadir, K., Martí‐Bonmatí, L., & Papanikolaou, N. K. (2025) · Insights into Imaging · doi:10.1186/s13244-025-02056-9 · Cited by 21 · see citing works
Self-supervised approach for diabetic retinopathy severity detection using vision transformer
Ohri, K., Kumar, M., & Sukheja, D. (2024) · Progress in Artificial Intelligence · doi:10.1007/s13748-024-00325-0 · Cited by 14 · see citing works
“The problem is even more severe in medical field where data collection and its annotation are more challenging and expensive considering the data privacy issues and unavailability of expert annotators [9, 10].”
Avram, O., Durmus, B., Rakocz, N., Corradetti, G., An, U., Nittala, M. G., Terway, P., Rudas, Á., Chen, Z. J., Wakatsuki, Y., et al. (2024) · Nature Biomedical Engineering · doi:10.1038/s41551-024-01257-9 · Cited by 17 · see citing works
“…that SLIViT’s simplicity is one of its major strengths. https://doi.org/10.1038/s41551-024-01257-9 The utility of self-supervised pre-training has been validated in numerous biomedical-imaging learning tasks 7,19,20,71,78,79 ; however, its general translatability across domains remains unclear 22 .”
InfluentialA dynamic model using k-NN algorithm for predicting diabetes and breast cancer
Al-Khamees, H. A. A., Sani, N. S., Gifal, A. S., Liu, L. X. W., & Esa, M. I. (2025) · Computers in Biology and Medicine · doi:10.1016/j.compbiomed.2025.110276 · Cited by 15 · see citing works
Santurro, A., Simone, M. D., Choucha, A., Morena, D., Consalvo, F., Romano, D. G., Terrasi, P., Corrivetti, F., Scrofani, R., Narciso, N., et al. (2025) · International Journal of Molecular Sciences · doi:10.3390/ijms26167808 · Cited by 15 · see citing works
A survey on self-supervised learning: Recent advances and open problems
Zhang, J., Yang, L., Mohammadabadi, S. M. S., & Yan, F. (2025) · Neurocomputing · doi:10.1016/j.neucom.2025.131409 · Cited by 19 · see citing works
Deshmukh, M. T., Thakre, Y., Bisen, W. H., Sankar, K., Avthankar, A., Shahade, A. K., Bhaiyya, M. L., & Kulkarni, M. B. (2025) · Analytical Chemistry · doi:10.1021/acs.analchem.5c06356 · Cited by 18 · see citing works
Huang, J., Galal, G., Mukhin, V., Etemadi, M., & Tanna, A. P. (2024) · Journal of Glaucoma · doi:10.1097/ijg.0000000000002359 · Cited by 18 · see citing works
“Given the scale of the data set and the demonstrated ability of self-supervised learning algorithms to learn robust image features from noisy data, 17 image quality control was not performed beyond that done for initial entry into the database.”
BackgroundRousta, F., Esteki, A., Shalbaf, A., Sadeghi, A., Moghadam, P. K., & Voshagh, A. (2024) · Computer Methods and Programs in Biomedicine · doi:10.1016/j.cmpb.2024.108205 · Cited by 16 · see citing works
Exploring the Landscape of Ubiquitous In-home Health Monitoring: A Comprehensive Survey
Pourpanah, F., & Etemad, A. (2024) · ACM Transactions on Computing for Healthcare · doi:10.1145/3670854 · Cited by 19 · see citing works
“For a review on wearable technologies, please see [94, 184].”
BackgroundEvaluating artificial intelligence for medical imaging: a primer for clinicians
Keni, S. (2024) · British Journal of Hospital Medicine · doi:10.12968/hmed.2024.0312 · Cited by 15 · see citing works
Guo, B., Chen, Y., Lin, J., Huang, B., Bai, X., Gao, B., Gao, B., Gong, Q., & Bai, X. (2024) · Nature Communications · doi:10.1038/s41467-024-53550-5 · Cited by 15 · see citing works
Suo, X., Chen, M., Chen, L., Luo, C., Kemp, G. J., Lui, S., & Sun, H. (2025) · EBioMedicine · doi:10.1016/j.ebiom.2025.105773 · Cited by 15 · see citing works
“24,25 Self-supervised learning has been applied to research in natural language processing (NLP), computer vision, and medical data analysis, 24,26,27 where labelled data are particularly scarce or expensive.”
Sethia, K., Strakoš, P., Jaros, M., Kubíček, J., Roman, J., Penhaker, M., & Řı́ha, L. (2025) · Artificial Intelligence Review · doi:10.1007/s10462-025-11310-x · Cited by 19 · see citing works
Pantelidis, P., Dilaveris, P., Ruipérez-Campillo, S., Goliopoulou, A., Giannakodimos, A., Theofilis, P., LUCIA, R. D., Κατσαρού, Ο., Zisimos, K., Kalogeras, K. I., et al. (2025) · Biomedicines · doi:10.3390/biomedicines13051019 · Cited by 18 · see citing works
“Self-supervised learning has already shown promising results in certain areas of cardiovascular research, such as echocardiographic view classification [83,84] and anomaly detection in ECGs [85].”
Chen, X., Xu, H., Yu, S., Hu, W., Zhang, Z., Wang, X., Yuan, Y., Wang, M., Chen, L., Lin, X., et al. (2025) · Biology · doi:10.3390/biology14060651 · Cited by 18 · see citing works
“Unlike traditional supervised learning that relies on extensive labeled datasets (the model can only be trained with defined data) [93], RL emphasizes the autonomous learning of optimal behavioral trajectories via trial-and-error processes [94].”
Zakaria, T. M., Langi, A. Z. R., Nazaruddin, M. S., & Anshori, I. (2025) · IEEE Access · doi:10.1109/access.2025.3582805 · Cited by 16 · see citing works
“Supervised Learning: The AI model is trained using labelled data collected from individuals with and without attention disorders [109], [110], [111], [112], [113].”
Radio DINO: A foundation model for advanced radiomics and AI-driven medical imaging analysis
Zedda, L., Loddo, A., & Ruberto, C. D. (2025) · Computers in Biology and Medicine · doi:10.1016/j.compbiomed.2025.110583 · Cited by 17 · see citing works
Estimating chlorophyll content in tea leaves using spectral reflectance and deep learning methods
Tsuchiya, Y., Hirono, Y., & Sonobe, R. (2025) · Ecological Informatics · doi:10.1016/j.ecoinf.2025.103399 · Cited by 13 · see citing works
Medical multimodal large language models: A systematic review
Hu, Y., Xu, C., Lin, B., Yang, W., & Tang, Y. Y. (2025) · Intelligent Oncology · doi:10.1016/j.intonc.2025.09.005 · Cited by 14 · see citing works
Machine learning and transformers for thyroid carcinoma diagnosis
Habchi, Y., Kheddar, H., Himeur, Y., & Ghanem, M. C. (2025) · Journal of Visual Communication and Image Representation · doi:10.1016/j.jvcir.2025.104668 · Cited by 16 · see citing works
“For instance, [259] discusses various SSL strategies and their applications in medical imaging, emphasizing their potential to improve diagnostic performance.”
Don, A. K. K., & Khalil, I. (2024) · ACM Transactions on Quantum Computing · doi:10.1145/3660647 · Cited by 13 · see citing works
“Additionally, its effectiveness depends on the availability of a sufficiently large labeled dataset, which limits its practicality in situations where labeled data is severely insufficient [ 27 ].”
Li, K., Yang, J., Liang, W., Li, X. D., Zhang, C., Chen, L., Wu, C., Zhang, X., Xu, Z., Wang, Y., et al. (2024) · Medical Image Analysis · doi:10.1016/j.media.2024.103319 · Cited by 13 · see citing works
Multimodal Foundation Models for Medical Imaging - A Systematic Review and Implementation GuidelinesPreprint
Huang, S.-C., Jensen, M., Yeung, S. Y., Lungren, M. P., Poon, H., & Chaudhari, A. S. (2024) · medRxiv · doi:10.1101/2024.10.23.24316003 · Cited by 15 · see citing works
Self-citationRickard, D., Kabir, M. A., & Homaira, N. (2025) · Computer Methods and Programs in Biomedicine · doi:10.1016/j.cmpb.2025.108802 · Cited by 14 · see citing works
Espis, A., Marzi, C., & Diciotti, S. (2025) · Scientific Reports · doi:10.1038/s41598-025-99000-0 · Cited by 17 · see citing works
“In our study, we focused on self-contrastive learning because it has been empirically shown to be usually more performant than other SSL paradigms when dealing with classification tasks in the medical field 45,46 .”
Jomeiri, A., Navin, A. H., & Shamsi, M. (2025) · Computers in Biology and Medicine · doi:10.1016/j.compbiomed.2025.111065 · Cited by 16 · see citing works
Karagöz, M. A., Nalbantoğlu, Ö. U., Karaboğa, D., Akay, B., Baştürk, A., Ulutabanca, H., Doğan, S., COŞKUN, D., & DEMİR, O. (2024) · TURKISH JOURNAL OF ELECTRICAL ENGINEERING & COMPUTER SCIENCES · doi:10.55730/1300-0632.4076 · Cited by 12 · see citing works
Gholami, S., Scheppke, L., Kshirsagar, M., et al. (2024) · JAMA Ophthalmology · doi:10.1001/jamaophthalmol.2023.6454 · Cited by 12 · see citing works
A Systematic Review and Implementation Guidelines of Multimodal Foundation Models in Medical ImagingPreprint
Huang, S.-C., Jensen, M., Yeung, S. Y., Lungren, M. P., Poon, H., & Chaudhari, A. S. (2025) · Research Square · doi:10.21203/rs.3.rs-5537908/v1 · Cited by 17 · see citing works
Self-citationA review and systematic guide to counteracting medical data scarcity for AI applications
Gröger, F., Amruthalingam, L., Lionetti, S., Navarini, A. A., Ille, F., & Pouly, M. (2025) · Computer Methods and Programs in Biomedicine Update · doi:10.1016/j.cmpbup.2025.100220 · Cited by 13 · see citing works
Fu, H., Novak, A., Robert, D., Kumar, S., Tanamala, S., Oke, J. L., Bhatia, K., Shah, R., Romsauerova, A., Das, T., et al. (2024) · BMJ Open · doi:10.1136/bmjopen-2023-079824 · Cited by 10 · see citing works
Wang, Y., Ni, H., Zhou, J., Liu, L., Lin, J., Yin, M., Gao, J., Zhu, S., Yin, Q., Zhu, J., et al. (2024) · Journal of Imaging Informatics in Medicine · doi:10.1007/s10278-024-01123-9 · Cited by 12 · see citing works
“SSL is now a novel solution for the successful use of various effective deep learning models [6].”
BackgroundQi, L., Jiang, Z., Shi, W., Qu, F., & Feng, G. (2024) · Computers in Biology and Medicine · doi:10.1016/j.compbiomed.2024.108547 · Cited by 12 · see citing works
Zanini, L. G. K., Rubira‐Bullen, I. R. F., & Nunes, F. L. S. (2024) · Computers in Biology and Medicine · doi:10.1016/j.compbiomed.2024.109221 · Cited by 11 · see citing works
Li, Y., Wynne, J. F., Wu, Y., Qiu, R. L. J., Tian, S., Wang, T., Patel, P. R., Yu, D. S., & Yang, X. (2025) · Radiotherapy and Oncology · doi:10.1016/j.radonc.2025.110711 · Cited by 14 · see citing works
Ahmed, I., Ahmad, M., Chehri, A., & Jeon, G. (2025) · Information Fusion · doi:10.1016/j.inffus.2025.103527 · Cited by 10 · see citing works
Li, J., Liu, S., Aligayev, A., et al. (2025) · Chemical Engineering Journal · doi:10.1016/j.cej.2025.168343 · Cited by 10 · see citing works
“Advances in deep learning and computer vision provide promising solutions for medical image analysis [30].”
Shi, W., Yu, L., Tian, S., Yang, Q., Zhang, D., Yu, S., & Wu, W. (2025) · Engineering Applications of Artificial Intelligence · doi:10.1016/j.engappai.2025.113083 · Cited by 10 · see citing works
Cervical OCT image classification using contrastive masked autoencoders with Swin Transformer
Wang, Q., Xiong, Y., Zhu, H., et al. (2024) · Computerized Medical Imaging and Graphics · doi:10.1016/j.compmedimag.2024.102469 · Cited by 10 · see citing works
Ryu, S. Y., Choi, J. Y., & Yoo, T. K. (2025) · Medical & Biological Engineering & Computing · doi:10.1007/s11517-025-03353-7 · Cited by 10 · see citing works
MedMAE: A Self-Supervised Backbone for Medical Imaging Tasks
Gupta, A., Osman, I. I., Shehata, M., Braun, W. J., & Feldman, R. E. (2025) · Computation · doi:10.3390/computation13040088 · Cited by 12 · see citing works
Ma, D., Stocks, J., Rosen, H., Kantarci, K., Lockhart, S. N., Bateman, J. R., Craft, S., Gurcan, M. N., Popuri, K., Beg, M. F., et al. (2024) · Frontiers in Neuroscience · doi:10.3389/fnins.2024.1331677 · Cited by 9 · see citing works
“…inhibitors) are currently underway (Boxer et al., 2013; Tsai and Boxer, 2016; Mis et al., 2017; Logroscino et al., 2019; Panza et al., 2020; Huang et al., 2023), the significant degree of clinical, pathological and genetic heterogeneity observed in FTD hinders the development of sensitive…”
BackgroundMethodEfficient sparse spiking auto-encoder for reconstruction, denoising and classification
Walters, B., Kalatehbali, H. R., Cai, Z., et al. (2024) · Neuromorphic Computing and Engineering · doi:10.1088/2634-4386/ad5c97 · Cited by 9 · see citing works
“Self-supervised learning has emerged as a promising method, eliminating the laborious process of labeling datasets [1].”
Robust image representations with counterfactual contrastive learning
Roschewitz, M., Ribeiro, F. D. S., Xia, T., Khara, G. S., & Glocker, B. (2025) · Medical Image Analysis · doi:10.1016/j.media.2025.103668 · Cited by 15 · see citing works
“Recent work has successfully applied DINO-v2 pre-training to chest radiography, achieving state-of-the-art downstream performance across different tasks 16 .”
Wu, Y., Ramai, D., Smith, E. R., Mega, P. F., Qatomah, A., Spadaccini, M., Maida, M. F., & Papaefthymiou, A. (2024) · Cancers · doi:10.3390/cancers16244196 · Cited by 9 · see citing works
Zhang, T., Bi, Y., Liang, J., et al. (2025) · Swarm and Evolutionary Computation · doi:10.1016/j.swevo.2025.101875 · Cited by 9 · see citing works
Domain-invariant representation learning via SAM for blood cell classification
Li, Y., Cai, L., Lu, Y., Lin, C., Zhang, Y., Jiang, J., Dai, G., Zhang, B., Cao, J., Zhang, X., et al. (2025) · Pattern Recognition · doi:10.1016/j.patcog.2025.112000 · Cited by 10 · see citing works
Hussien, A., Elkhateb, A., Saeed, M., Elsabawy, N. M., Elnakeeb, A. E., & El-Rashidy, N. M. (2025) · Scientific Reports · doi:10.1038/s41598-025-15604-6 · Cited by 12 · see citing works
“Additionally, as in a few narrative reviews 29,30 , the most suitable strategies and best practices for medical images have not been sufficiently investigated.”
Improving Medical Image Classification in Noisy Labels Using only Self-supervised Pretraining
Khanal, B., Bhattarai, B., Khanal, B., & Linte, C. A. (2023) · Lecture notes in computer science · doi:10.1007/978-3-031-44992-5_8 · Cited by 9 · see citing works
A Fusion of Supervised Contrastive Learning and Variational Quantum Classifiers
Don, A. K. K., Khalil, I., & Atiquzzaman, M. (2024) · IEEE Transactions on Consumer Electronics · doi:10.1109/tce.2024.3351649 · Cited by 15 · see citing works
“Furthermore, the effectiveness of SCL relies heavily on access to a sufficiently large labeled dataset, limiting its practicality in scenarios with severely limited labeled data [20].”
BackgroundAnatomical foundation models for brain MRIs
Barbano, C. A., Brunello, M., Dufumier, B., & Grangetto, M. (2025) · Pattern Recognition Letters · doi:10.1016/j.patrec.2025.11.028 · Cited by 11 · see citing works
“Its applicability on neuroimage data is still unclear and very few studies have tackled this problem so far [17].”
Li, C., Yang, D., Yao, S., Wang, S., Wu, Y., Zhang, L., Li, Q., Cho, K. I. K., Seitz‐Holland, J., Ning, L., et al. (2025) · Computerized Medical Imaging and Graphics · doi:10.1016/j.compmedimag.2024.102489 · Cited by 8 · see citing works
- Authors
- Huang, S.-C., Pareek, A., Jensen, M., Lungren, M. P., Yeung, S., & Chaudhari, A. S.
- Affiliations
- Stanford University (US)
- Stanford Medicine (US)
- Corresponding author
- Huang, Shih-Cheng
- Venue
- npj Digital Medicine
- Published
- 2023-04-26
- Volume
- 6
- Issue
- 1
- Pages
- 74-74
- ISSN
- 2398-6352 (ISSN-L)
- PMC
- PMC10131505
- OpenAlex
- W4367055910
- Cited by
- 479
- References
- 130
- Citations by year
- 2026: 145 · 2025: 208 · 2024: 102 · 2023: 24
- FWCI
- 10.03
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Locations
- npj Digital Medicine (publishedVersion, cc-by)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- National Institutes of Health (75N92020C00021, 75N92020C00008, AR077604, EB002524)
- National Heart, Lung, and Blood Institute (75N92020C00008, 75N92020C00021, R01HL155410)
- National Institute of Arthritis and Musculoskeletal and Skin Diseases (R01AR077604, R01AR079431)
- National Institute of Biomedical Imaging and Bioengineering (P41EB027060, R01EB002524)
- U.S. National Library of Medicine (R01LM012966)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
- Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Medical Image Segmentation Techniques
- Keywords
- self-supervised learning, medical image classification, diagnostic imaging, unlabeled data, implementation guidelines, deep learning, limited labeled data
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-07.
Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-07.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
