CoChat
Sign Up
Cited by 1,129Medical Image Segmentation TechniquesBrain Tumor Detection and ClassificationAdvanced Neural Network Applications

Deep learning for brain MRI segmentation: state of the art and future directions

Open access
Akkus, Zeynettin, Galimzianova, Alfiia, Hoogi, Assaf, Rubin, Daniel L., Erickson, Bradley J.·2017·Journal of Digital Imaging
DOI 10.1007/s10278-017-9983-4PMC PMC5537095
Open this paper in CoChat — free accountOpen PDF

Abstract

Quantitative analysis of brain MRI is routine for many neurological diseases and conditions and relies on accurate segmentation of structures of interest. Deep learning-based segmentation approaches for brain MRI are gaining interest due to their self-learning and generalization ability over large amounts of data. As the deep learning architectures are becoming more mature, they gradually outperform previous state-of-the-art classical machine learning algorithms. This review aims to provide an overview of current deep learning-based segmentation approaches for quantitative brain MRI. First we review the current deep learning architectures used for segmentation of anatomical brain structures and brain lesions. Next, the performance, speed, and properties of deep learning approaches are summarized and discussed. Finally, we provide a critical assessment of the current state and identify likely future developments and trends.

AI snapshot

Written from the paper’s full text.

TL;DR

This review surveys deep learning approaches for brain MRI segmentation, covering architectures for normal anatomical structures and brain lesions. It reports that CNN-based methods increasingly outperform classical machine learning and highlights remaining challenges in data size, generalization, preprocessing, and supervised ground-truth requirements.

Key findings

  • Deep learning architectures for brain MRI segmentation are grouped into patch-wise, semantic-wise, and cascaded styles; CNNs are the most commonly applied.
  • CNN approaches outperformed classical methods in reviewed studies: Zhang et al. reported overall DSC 85.03% ± 2.27% (CNN) vs. 76.95% ± 3.55% (SVM) and 83.15% ± 2.52% (RF) for infant brain tissue segmentation; Nie et al. improved DSC to 85.5% (CSF), 87.3% (GM), and 88.7% (WM) vs. 83.5%, 85.2%, and 86.4% by Zhang et al.
  • On BRATS, Havaei et al. reported DSC 0.88 (complete), 0.79 (core), and 0.73 (enhancing) and faster testing (3 vs. 100 min) than the BRATS 2013 winner; Pereira et al. reported DSC 0.88, 0.83, and 0.77 for complete, core, and enhancing regions; Kamnitsas et al. reported DSC 0.9, 0.75, and 0.73 on BRATS-2015.
  • Public challenges/datasets such as BRATS, ISLES, MSSEG, NeoBrainS12, and MRBrainS provide benchmarks; the MRBrainS winner achieved DSC 86.15% for gray matter, 89.46% for white matter, and 84.25% for cerebrospinal fluid segmentation.

Methods

  • Literature review using Google Scholar and NLM PubMed; included peer-reviewed journal publications and conference proceedings applying deep learning to brain MRI segmentation, plus relevant Arxiv preprints; conference proceedings with a follow-up journal publication were included only in final publication form.
  • Papers were divided into works on normal structures and brain lesions, then subdivided by architecture style: patch-wise, semantic-wise, or cascaded.
  • The review summarizes training/validation/evaluation methods, preprocessing steps, CNN architecture styles, and reported performance; it is a narrative survey, not a primary study or meta-analysis.

Limitations

  • The paper states that medical imaging datasets are relatively small compared with large-scale datasets such as ImageNet, limiting deep learning’s full potential; most methods are supervised and require manual ground-truth labels, which is tedious.
  • The paper states that developing a generic method robust to variations across institutions, scanners, acquisition protocols, and pathology remains challenging, and that transfer learning from ImageNet may not be the best choice for medical images.
  • Evident from the text: performance comparisons are difficult because algorithms are evaluated on different datasets and metrics; the review also contains an inconsistency for Dou et al. cerebral microbleed detection sensitivity (table lists 98.29%, text states 93.2%).

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1007/s10278-017-9983-4

Cites

66 of 66 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 46 cited sources checked · 0 supported · 0 flagged

  • Findingcitation

    Zhang et al. CNN approach outperforms prior methods and classical ML algorithms using SVM and RF (overall DSC 85.03±2.27 CNN vs 76.95±3.55 SVM, 83.15±2.52 RF).

    “They showed that their CNN approach outperforms prior methods and classical machine learning algorithms using support vector machine (SVM) and random forest (RF) classifiers (overall DSC performance 85.03% ∓ 2.27% (CNN) vs. 76.95% ∓ 3.55% (SVM), 83.15% ∓ 2.52% (RF)).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [27] Zhang W et al.: Deep convolutional neural networks for multi-modality isointense infant brain image segmentation. Neuroimage 108:214–224, 2015 source

      Not checked in the free run.

  • Findingnone given

    The ISLES 2016 winner algorithm for subacute ischemic stroke lesion segmentation achieved DSC 0.59 ± 0.31 and HD 37.88 ± 30.06.

    “The performance of the winner algorithm on this dataset for subacute ischemic stroke lesion segmentation currently is 0.59 ± 0.31 (Dice similarity coefficient, DSC) and 37.88 ± 30.06 (Hausdorff Distance, HD).”
    Quote in the paperNot a substantial claim
  • Findingcitation

    Nie et al. semantic-wise FCNs obtained improved results compared to Zhang et al. (overall DSC 85.5 CSF, 87.3 GM, 88.7 WM vs 83.5, 85.2, 86.4).

    “They obtained improved results compared to [27]. Their overall DSC were 85.5% (CSF), 87.3% (GM), and 88.7% (WM) vs. 83.5% (CSF), 85.2 (GM), and 86.4 (WM) by [27].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [30] Nie D, Dong N, Li W, Yaozong G, Dinggang S: “Fully convolutional networks for multi-modality isointense infant brain image segmentation,” in 2016 I.E. 13th International Symposium on Biomedical Imaging (ISBI), 2016. source

      Not checked in the free run.

    • cites [27] Zhang W et al.: Deep convolutional neural networks for multi-modality isointense infant brain image segmentation. Neuroimage 108:214–224, 2015 source

      Not checked in the free run.

  • Premisecitation

    Applying normalization prior to deep learning based segmentation improves results in another work.

    “however, another work shows improvement when applying normalization prior to deep learning based segmentation procedure [22].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [22] Pereira S, Pinto A, Alves V, Silva CA: “Brain Tumor Segmentation using Convolutional Neural Networks in MRI Images,” IEEE Trans. Med. Imaging, Mar. 2016. source

      Not checked in the free run.

  • Interpretationreasoning

    Measures such as accuracy and specificity tend to be avoided in lesion segmentation because they do not discriminate between different segmentation outputs when the lesion is considerably smaller than background.

    “Measures such as accuracy and specificity tend to be avoided in the lesion segmentation context since these measures do not discriminate between different segmentation outputs when the object (lesion) is considerably smaller than the background (normal-appearing brain tissue).”
    Quote in the paperNot a substantial claim
  • May combine two claims. 16
  • Classified as probably not a substantial claim. 17
  • The quote only partly states this claim. 1

  • May combine two claims.

    Findingcitation

    Manual segmentation of brain tumor images has intra-expert variabilities of 20 ± 15% and inter-expert variabilities of 28 ± 12%.

    “In their work, Mazzara et al. [11] reported intra-expert variabilities of 20 ± 15% and inter-experts variabilities of 28 ± 12% for manual segmentations of brain tumor images.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [11] G. P. Mazzara, R. P. Velthuizen, J. L. Pearlman, H. M. Greenberg, and H. Wagner, “Brain tumor target volume determination for radiation treatment planning through automated MRI segmentation,” Int. J. Radiat. Oncol. Biol. Phys., vol. 59, no. 1, pp. 300–312, 2004. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    Bao et al. reported overall DSC 82.2 and 85% for IBSR and LPBA40.

    “They reported overall DSC of 82.2 and 85% for IBSR and LPBA40, respectively.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [41] Bao S, Siqi B, Chung ACS: “Multi-scale structured CNN with label consistency for brain MR image segmentation,” Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, pp. 1–5, 2016. source

      Not checked in the free run.

  • May combine two claims.

    Findingnone given

    The MRBrainS current winner algorithm achieved DSC 86.15% for gray matter, 89.46% for white matter, and 84.25% for cerebrospinal fluid segmentation.

    “The performance (DSC) of the current winner algorithm on this dataset is 86.15% for gray matter, 89.46% for white matter, and 84.25% for cerebrospinal fluid segmentation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingcitation

    De Brebisson et al. achieved competitive results (DSC 72.5±16.3) in MICCAI 2012 challenge as first CNN approach.

    “They achieved competitive results (DSC = 72.5% ∓ 16.3%) in MICCAI 2012 challenge on multi-atlas labeling as the first CNN approach applied to the task.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [40] A. de Brebisson, M. Giovanni: “Deep neural networks for anatomical brain segmentation,” in 2015 I.E. Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2015. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    CNNs nearly halved the error rates of previously best computing approaches on ImageNet dataset of about a million images and 1000 classes.

    “Applied on a dataset of about a million images that included 1000 different classes, CNNs nearly halved the error rates of the previously best computing approaches [9].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [9] Krizhevsky A, Sutskever I, Hinton GE: ImageNet classification with deep convolutional neural networks. In: Pereira F, Burges CJC, Bottou L, Weinberger KQ Eds. Advances in neural information processing systems 25. USA: Curran Associates, Inc., 2012, pp. 1097–1105

      Not checked in the free run.

  • May combine two claims.

    Interpretationcitation

    CNNs adopted for segmentation of brain tissues avoid explicit definition of spatial and intensity features and provide better performance than classical approaches.

    “In recent years, CNNs have been adopted for segmentation of brain tissues, which avoid the explicit definition of spatial and intensity features and provide better performance than classical approaches, as we describe next”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Traditional machine learning algorithms do not generalize well.

  • May combine two claims.

    Interpretationcitation

    Deep learning can help avoid limitations of classical machine learning algorithms, and its self-learning of features may enable identification of new useful imaging features for quantitative analysis of brain MRI.

    “An emerging machine learning technique referred to as deep learning [1], can help avoid limitations of classical machine learning algorithms, and its self-learning of features may enable identification of new useful imaging features for quantitative analysis of brain MRI.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [1] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, 2015. source

      Not checked in the free run.

    • Rests on an assumption: Traditional machine learning algorithms do not generalize well.
    • Rests on an assumption: Automated segmentation of brain structures and detection of abnormalities remain an unsolved problem due to normal anatomical variations, variations in acquisition settings and MRI scanners, image acquisition imperfections, and variations in the appearance of pathology.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Random sampling of patches over a dataset potentially results in faster convergence compared to full image training in semantic-wise architectures.

    “On the other hand, random sampling of patches over a dataset potentially results in faster convergence (LeCun et al. 1998) compared to full image training in semantic-wise architectures.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites LeCun et al. 1998

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Premisereasoning

    It is important to have intensities of input brain MR images in a reference scale and normalized for each modality.

    “For example, it is important to have intensities of input brain MR images in a reference scale and normalized for each modality.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning

    This allowed training of deep learning algorithms with millions of images and provided robustness to variations in images.

    “This allowed training of deep learning algorithms with millions of images and provided robustness to variations in images.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: The rapid increase in GPU processing power has enabled the development of state-of-the-art deep learning algorithms.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Pereira et al. [22] were ranked as second place in Brats 2015 data.

    “They were also ranked as second place in Brats 2015 data.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [22] Pereira S, Pinto A, Alves V, Silva CA: “Brain Tumor Segmentation using Convolutional Neural Networks in MRI Images,” IEEE Trans. Med. Imaging, Mar. 2016. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisereasoning

    Proper network parameter initialization is important to maintain the gradient flow through network and to achieve convergence.

    “Another important step is proper network parameter initialization in the neural network to maintain the gradient flow through network and to achieve convergence.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Premisereasoning

    Image preprocessing plays a key role in learning.

    “Image preprocessing plays a key role in learning.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    Measures of clinical relevance are commonly incorporated, including correlation analysis of total lesion load or count and volume or volume change correlation.

    “In addition, measures of clinical relevance are also commonly incorporated. These include such measures as correlation analysis of total lesion load or count as detected by automated and manual segmentation and volume or volume change correlation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Brosch et al. [31] reported performance comparable to the state of the art methods and superior to the publicly available MS segmentation methods.

    “They reported performance comparable to the state of the art methods and superior to the publicly available MS segmentation methods.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [31] T. Brosch et al., “Deep 3D convolutional encoder networks with shortcuts for multiscale feature integration applied to multiple sclerosis lesion segmentation,” IEEE Trans. Med. Imaging, vol. 35, no. 5, pp. 1229–1239, 2016. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The performance of deep learning methods depends highly on several key steps such as preprocessing, initialization, and post-processing.

    “The performance of the deep learning methods depends highly on several key steps such as preprocessing, initialization, and post-processing.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    Typically, methods for normal structure or tumor segmentation include voxel-wise metrics, such as DSC, TPR, PPV, and lesion surface metrics, such as HD and ASSD.

    “Typically, the methods for normal structure or tumor segmentation include voxel-wise metrics, such as DSC, true positive rate (TPR), positive predictive value (PPV), and lesion surface metrics, such as HD and average symmetric surface distance (ASSD).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Shin et al. [62] reported that they obtained best performance with transfer learning from pre-trained model on ImageNet dataset and fine-tuning on lymph node and interstitial lung disease rather than training from scratch.

    “Shin et al. [62] reported that they obtained best performance with transfer learning from pre-trained model on ImageNet dataset and fine-tuning on lymph node and interstitial lung disease rather than training from scratch.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [62] H.-C. Shin et al., “Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning,” IEEE Trans. Med. Imaging, vol. 35, no. 5, pp. 1285–1298, 2016. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Premisereasoning

    Preventing overfitting is critical to learn the true information in images and avoid overfitting to specific training examples.

    “Lastly, preventing overfitting is critical to learn the true information in images, and avoiding overfitting to specific training examples provided.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Semantic-wise architectures are more susceptible to class imbalance but this can be solved by weighting the classes in the loss function.

    “Semantic-wise architectures also are more susceptible to class imbalance but this can be solved by weighting the classes in the loss function [31].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [31] T. Brosch et al., “Deep 3D convolutional encoder networks with shortcuts for multiscale feature integration applied to multiple sclerosis lesion segmentation,” IEEE Trans. Med. Imaging, vol. 35, no. 5, pp. 1229–1239, 2016. source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Dou et al. [32] reported their method outperformed previous studies with low level descriptors and provided a high sensitivity of 93.2% for detecting cerebral microbleeds.

    “They reported their method outperformed previous studies with low level descriptors and provided a high sensitivity of 93.2% for detecting CM.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [32] Q. Dou et al., “Automatic detection of cerebral Microbleeds from MR images via 3D convolutional neural networks,” IEEE Trans. Med. Imaging, vol. 35, no. 5, pp. 1182–1195, 2016. source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Havaei et al. [26] reported that their CNN approach outperformed and was much faster at testing phase (3 vs. 100 min) than the winner of BRATS 2013 competition.

    “They reported that their CNN approach outperformed and was much faster at testing phase (3 vs. 100 min) than the winner of BRATS 2013 competition.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [26] Havaei M et al.: Brain tumor segmentation with deep neural networks. Med. Image Anal. 35:18–31, 2016 source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning

    Post-processing of the output of model is an important step to refine the segmentation results.

    “Post-processing of the output of model is also an important step to refine the segmentation results.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: There are always regions in images that overlap between classes (partial volume effect), which unavoidably leads to false positives or negatives.

  • May combine two claims.

    Premisecitation

    Pereira et al. [22] method showed the best performance on the Brats 2013 data, with DSC values of 0.88, 0.83, and 0.77 for complete, core, and enhancing regions, respectively.

    “Their method showed the best performance on the Brats 2013 data – DSC values of 0.88, 083, 0.77 for complete, core, and enhancing regions, respectively.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [22] Pereira S, Pinto A, Alves V, Silva CA: “Brain Tumor Segmentation using Convolutional Neural Networks in MRI Images,” IEEE Trans. Med. Imaging, Mar. 2016. source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Interpretationnone given

    Developing a generic deep learning approach that will work on datasets from different machines and institutions is challenging due to limited training and ground truth data, variations in image acquisition protocols, imperfections of each MRI scanner, and variations in appearance of healthy and pathological brain tissue.

    “Developing a generic deep learning approach that will work on datasets from different machines and institutions is challenging due to limited training and ground truth data, variations and image acquisition protocols, imperfections of each MRI scanner, and variations in appearance of healthy and pathological brain tissue.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationnone given

    Current deep learning architectures are based on supervised learning and require generation of manual ground truth labels, which is tedious work on a large-scale data.

    “Moreover, current deep learning architectures are based on supervised learning and require generation of manual ground truth labels, which is tedious work on a large-scale data.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    Data augmentation helps increase the size of training examples and reduce overfitting by introducing random variations to the original data.

    “Data augmentation helps increase the size of training examples and reduce overfitting by introducing random variations to the original data.”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [9] Krizhevsky A, Sutskever I, Hinton GE: ImageNet classification with deep convolutional neural networks. In: Pereira F, Burges CJC, Bottou L, Weinberger KQ Eds. Advances in neural information processing systems 25. USA: Curran Associates, Inc., 2012, pp. 1097–1105

      Not checked in the free run.

    • cites [22] Pereira S, Pinto A, Alves V, Silva CA: “Brain Tumor Segmentation using Convolutional Neural Networks in MRI Images,” IEEE Trans. Med. Imaging, Mar. 2016. source

      Not checked in the free run.

    • cites [29] Akkus Z, et al.: “Predicting 1p19q Chromosomal Deletion of Low-Grade Gliomas from MR Images using Deep Learning,” arXiv [cs.CV], 2016.

      Not checked in the free run.

  • May combine two claims. The quote only partly states this claim.

    Interpretationreasoning

    Deep learning approaches tend to outperform previous state of the art classical machine learning algorithms and are becoming more mature.

    “Even though deep learning approaches have been applied to brain MRI only recently, they tend to outperform previous state of the art classical machine learning algorithms and are becoming more mature.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    Methods for multifocal brain lesions often also include lesion-wise metrics, such as LTPR and LPPV.

    “On the other hand, methods for multifocal brain lesions often also include lesion-wise metrics, such as lesion-wise true positive rate (LTPR) and lesion-wise positive predictive value (LPPV).”
    Quote in the paperNot a substantial claim
29 of 29

  • The paper relies on this without support.

    Manual segmentation is the gold standard for in vivo images.

    “Manual segmentation is the gold standard for in vivo images.”

  • The paper relies on this without support.

    Manual segmentation is expensive and tedious.

    “However, this requires outlining structures slice-by-slice, and is not only expensive and tedious, but also inaccurate due to human error.”

  • The paper relies on this without support.

    Traditional machine learning algorithms do not generalize well.

    “Furthermore, traditional machine learning algorithms do not generalize well.”

  • The paper relies on this without support.

    Automated segmentation of brain structures and detection of abnormalities remain an unsolved problem due to normal anatomical variations, variations in acquisition settings and MRI scanners, image acquisition imperfections, and variations in the appearance of pathology.

    “Despite a significant effort from the medical imaging research community, automated segmentation of the brain structures and detection of the abnormalities remain an unsolved problem due to normal anatomical variations in brain morphology, variations in acquisition settings and MRI scanners, image acquisition imperfections, and variations in the appearance of pathology.”

  • The paper relies on this without support.

    Deep learning techniques achieve impressive results and generalizability by training on large amount of data.

    “They achieve impressive results and generalizability by training on large amount of data.”

  • The paper relies on this without support.

    The rapid increase in GPU processing power has enabled the development of state-of-the-art deep learning algorithms.

    “The rapid increase in GPU processing power has enabled the development of state-of-the-art deep learning algorithms.”

  • The paper relies on this without support.

    CNNs are the most commonly applied to image segmentation and classification.

    “CNNs are the most commonly applied to image segmentation and classification.”

  • The paper relies on this without support.

    Ground truth is usually obtained with manual delineations of brain lesions or structures by experts for segmentation tasks.

    “Ground truth is usually obtained with manual delineations of brain lesions or structures by experts for segmentation tasks.”

  • The paper relies on this without support.

    For classification tasks of brain lesions, the ground truth is obtained with biopsy and pathological tests.

    “For classification tasks of brain lesions, the ground truth is obtained with biopsy and pathological tests.”

  • The paper relies on this without support.

    Most algorithms are evaluated on different sets of data and reported different similarity metrics, making it hard to compare performance.

    “In general, most of the algorithms are evaluated on different sets of data and reported different similarity metrics. This makes it hard to compare the performance of different algorithms against each other.”

  • The paper relies on this without support.

    The MS lesion segmentation challenge dataset (MICCAI 2008) was one of the first publicly available datasets with ground truth; its test dataset is released without ground truth and evaluated only by organizers, which helps avoid overfitting and makes comparison more objective.

    “One of the first such datasets was released in the framework of an MS lesion segmentation challenge, which was held in conjunction with MICCAI 2008.”
    “The latter helps avoid overfitting of the methods and makes comparison more objective.”

  • The paper relies on this without support.

    Significance tests commonly accompany contributions that build on or compare to other methods, most often nonparametric tests such as Wilcoxon’s signed rank or Wilcoxon’s rank sum tests.

    “Significance tests commonly accompany contributions that build on or compare to other methods, most often nonparametric tests such as Wilcoxon’s signed rank of Wilcoxon’s rank sum tests are preferred.”

  • The paper relies on this without support.

    Reliable extraction of imaging biomarkers for brain lesions depends on prior accurate segmentation.

    “Reliable extraction of these biomarkers depends on prior accurate segmentation.”

  • The paper relies on this without support.

    Accurate segmentation of brain lesions remains a challenge despite significant effort.

    “Despite the significant effort in brain lesion segmentation and advanced imaging techniques, accurate segmentation of brain lesions remains a challenge.”

  • The paper relies on this without support.

    Brain image analysis is a great challenge to computer-aided techniques due to complex brain anatomy, variability of appearance, non-standardized MR scales, image acquisition imperfection, and presence of pathology.

    “Brain image analysis has been a great challenge to computer-aided techniques due to complex brain anatomy and variability of its appearance, non-standardized MR scales due to variability in imaging protocols, image acquisition imperfection, and presence of pathology.”

  • The paper relies on this without support.

    There are always regions in images that overlap between classes (partial volume effect), which unavoidably leads to false positives or negatives.

    “The goal of any learning method is to have a perfect classification, but there are always regions in images that overlap between classes, known as partial volume effect, which unavoidably leads to false positives or negatives.”
16 of 16

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?

  • Several review articles on deep learning in medical image analysis focus on the whole field of medical image analysis or on other single-imaging modalities such as MRI and microscopy.

    “Several review articles have been written to date on the application of deep learning to medical image analysis; these articles focus on either the whole field of medical image analysis [11][12][13][14][15] or other single-imaging modalities such as MRI [53] and microscopy [54].”

    From Deep Learning in Medical Ultrasound Analysis: A Review (2019) · cites this paper as [53]

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.

  • Conflicting sensitivity for cerebral microbleed detection

    Major

    Table 4 reports 98.29% sensitivity for Dou et al., while the discussion of the same study reports 93.2%.

1 of 1

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Akkus, Z., Galimzianova, A., Hoogi, A., Rubin, D. L., & Erickson, B. J.
Affiliations
  • Mayo Clinic (US)
  • Stanford Medicine (US)
Corresponding author
Erickson, Bradley J.
ORCID
Venue
Journal of Digital Imaging
Publisher
Springer Science and Business Media LLC
Published
2017-06-02
Volume
30
Issue
4
Pages
449-459
ISSN
  • 0897-1889 (ISSN-L)
  • 1618-727X
DOI
10.1007/s10278-017-9983-4
PMC
PMC5537095
OpenAlex
W2621028221
Cited by
1,129
References
66
Citations by year
2026: 60 · 2025: 111 · 2024: 110 · 2023: 162 · 2022: 171 · 2021: 166 · 2020: 154 · 2019: 138 · 2018: 52 · 2017: 3 · 2015: 1
FWCI
5.28
Type
journal article
Language
en
Open access
Licence
Locations
  • Journal of Imaging Informatics in Medicine (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
Funders
National Cancer Institute (U01CA160045, U01CA142555)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Medical Image Segmentation Techniques
  • Life Sciences › Neuroscience › Neurology › Brain Tumor Detection and Classification
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
Keywords
brain MRI segmentation, deep learning, brain lesion segmentation, segmentation speed, future research directions
MeSH terms
Machine Learning, Algorithms, Brain, Forecasting, Humans, Magnetic Resonance Imaging

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 2026-10-10.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-10.

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