CoChat
Sign Up
Cited by 207Multimodal Machine Learning ApplicationsArtificial Intelligence in Healthcare and EducationCOVID-19 diagnosis using AI

The DOI matches its Crossref record; 33 of the 33 references it cites were found in a registry or an index.

Knowledge-enhanced visual-language pre-training on chest radiology images

Open access
Zhang, Xiaoman, Wu, Chaoyi, Zhang, Ya, Xie, Weidi, Wang, Yanfeng·2023·Nature Communications
DOI 10.1038/s41467-023-40260-7PMC PMC10382552
Open this paper in CoChat — free accountOpen PDF

Abstract

While multi-modal foundation models pre-trained on large-scale data have been successful in natural language understanding and vision recognition, their use in medical domains is still limited due to the fine-grained nature of medical tasks and the high demand for domain knowledge. To address this challenge, we propose an approach called Knowledge-enhanced Auto Diagnosis (KAD) which leverages existing medical domain knowledge to guide vision-language pre-training using paired chest X-rays and radiology reports. We evaluate KAD on four external X-ray datasets and demonstrate that its zero-shot performance is not only comparable to that of fully supervised models but also superior to the average of three expert radiologists for three (out of five) pathologies with statistical significance. Moreover, when few-shot annotation is available, KAD outperforms all existing approaches in fine-tuning settings, demonstrating its potential for application in different clinical scenarios.

Overview

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1038/s41467-023-40260-7

Cites

33 of 33 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1 Overview of the KAD workflow. a Knowledge base used for training the knowledge encoder. It contains two parts, a knowledge graph consisting of concept-relation-concept triplets and a concept info list consisting of concept-definition pairs. b The knowledge encoder is trained to learn textual representations by maximizing similarities between positive pairs. c We first extract the clinical entities and relations from the radiology reports; this can be achieved by heuristic rules, using an off-the-shelf reports information extraction toolbox (Entity Extraction), or ChatGPT, then we employ the pre-trained knowledge encoder to perform image-text contrastive learning with paired chest X-rays and extracted entities and optimize a Disease Query Network (DQN) for classification. d During the inference stage, we simply encode the disease name as a query input, and DQN will output the probability that the pathology is present in the input image.
Fig. 1 Overview of the KAD workflow. a Knowledge base used for training the knowledge encoder. It contains two parts, a knowledge graph consisting of concept-relation-concept triplets and a concept info list consisting of concept-definition pairs. b The knowledge encoder is trained to learn textual representations by maximizing similarities between positive pairs. c We first extract the clinical entities and relations from the radiology reports; this can be achieved by heuristic rules, using an off-the-shelf reports information extraction toolbox (Entity Extraction), or ChatGPT, then we employ the pre-trained knowledge encoder to perform image-text contrastive learning with paired chest X-rays and extracted entities and optimize a Disease Query Network (DQN) for classification. d During the inference stage, we simply encode the disease name as a query input, and DQN will output the probability that the pathology is present in the input image.CC BY · PubMed Central
Fig. 2 Comparison of KAD with SOTA medical image-text pre-training models under zero-shot setting on radiographic findings or diagnoses in the PadChest dataset. We evaluate model on the human-annotated subset of the PadChest dataset (n = 39,053 chest X-rays), and mean AUC and 95% CI of KAD are shown for each radiographic finding or diagnosis (n > 50). a Results of seen classes. Note that CheXNet is a supervised model trained on the PadChest dataset. b Results of unseen classes. KAD achieves an AUC of at least 0.900 on 31 classes and at least 0.700 on 111 classes out of 177 unseen classes in the PadChest test dataset. Top 50 classes where (n > 50) in the test dataset (n = 39,053) are shown in the figure.
Fig. 2 Comparison of KAD with SOTA medical image-text pre-training models under zero-shot setting on radiographic findings or diagnoses in the PadChest dataset. We evaluate model on the human-annotated subset of the PadChest dataset (n = 39,053 chest X-rays), and mean AUC and 95% CI of KAD are shown for each radiographic finding or diagnosis (n > 50). a Results of seen classes. Note that CheXNet is a supervised model trained on the PadChest dataset. b Results of unseen classes. KAD achieves an AUC of at least 0.900 on 31 classes and at least 0.700 on 111 classes out of 177 unseen classes in the PadChest test dataset. Top 50 classes where (n > 50) in the test dataset (n = 39,053) are shown in the figure.CC BY · PubMed Central
Fig. 3 Comparison of proposed KAD with SOTA self-supervised baseline models and medical image-text pre-training models on ChestXray14 with different ratio of labeled data used for fine-tuning. AUC, F1 score are reported, and the metrics refer to the macro average on all the diseases. Note that for fairness, all baselines use the same backbone as the basic image encoder (that is, ResNet50). The percentages refer to the percentage of labels used in the training data.
Fig. 3 Comparison of proposed KAD with SOTA self-supervised baseline models and medical image-text pre-training models on ChestXray14 with different ratio of labeled data used for fine-tuning. AUC, F1 score are reported, and the metrics refer to the macro average on all the diseases. Note that for fairness, all baselines use the same backbone as the basic image encoder (that is, ResNet50). The percentages refer to the percentage of labels used in the training data.CC BY · PubMed Central
Fig. 4 Comparisons of proposed KAD with SOTA medical image-text pre-training models and three board-certified radiologists on five competition pathologies in the CheXpert test dataset (n = 500). Note that all models are directly evaluated on the CheXpert dataset under zero-shot setting. The AUC, F1 scores and MCC of five pathologies are shown in the plots, where the average and 95% CI are shown. Details in Supplementary Table 3.
Fig. 4 Comparisons of proposed KAD with SOTA medical image-text pre-training models and three board-certified radiologists on five competition pathologies in the CheXpert test dataset (n = 500). Note that all models are directly evaluated on the CheXpert dataset under zero-shot setting. The AUC, F1 scores and MCC of five pathologies are shown in the plots, where the average and 95% CI are shown. Details in Supplementary Table 3.CC BY · PubMed Central
Fig. 5 Comparisons of proposed KAD with SOTA medical image-text pre-training models on ChestX-Det10 dataset. AUC scores are shown for the zero-shot classification task in (a), and Pointing game scores are shown for the zero-shot grounding task in (b). We use the best results as the maximum value for each category in the radar chart, and 0.5 as the minimal value for (a), and 0 as the minimal value for (b). Details in Supplementary Table 4.
Fig. 5 Comparisons of proposed KAD with SOTA medical image-text pre-training models on ChestX-Det10 dataset. AUC scores are shown for the zero-shot classification task in (a), and Pointing game scores are shown for the zero-shot grounding task in (b). We use the best results as the maximum value for each category in the radar chart, and 0.5 as the minimal value for (a), and 0 as the minimal value for (b). Details in Supplementary Table 4.CC BY · PubMed Central
Fig. 6 Zero-shot visualization of randomly chosen samples from ChestX-Det10, we present both the original image (left) and attention maps generated from KAD, KAD-512, and KAD-1024. In the original images, red boxes denote lesion areas annotated by radiologists. In the attention maps, the red to blue spectrum is plotted on the original image, with red representing high-attention regions and blue representing low attention.
Fig. 6 Zero-shot visualization of randomly chosen samples from ChestX-Det10, we present both the original image (left) and attention maps generated from KAD, KAD-512, and KAD-1024. In the original images, red boxes denote lesion areas annotated by radiologists. In the attention maps, the red to blue spectrum is plotted on the original image, with red representing high-attention regions and blue representing low attention.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

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?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Research works printed but not linked: 4

  1. Devlin, J., Chang, M.-W., Lee, K. & Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proc. 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 4171–4186 (ACL, 2019).
  2. Ma, C., Yang, Y., Wang, Y., Zhang, Y. & Xie, W. Open-vocabulary semantic segmentation with frozen vision-language models. British Machine Vision Conference (2022).
  3. Zhang, Y., Jiang, H., Miura, Y., Manning, C. D. & Langlotz, C. P. Contrastive learning of medical visual representations from paired images and text. PMLR 182, 2–25. Highest Starred Implementation: https://github.com/edreisMD/ConVIRT-pytorch (2022).
  4. Loshchilov, I. & Hutter, F. Decoupled weight decay regularization. In Proc. International Conference on Learning Representations (2019).

Cited by

Sorted by citations

See all 207

Related

See all

Metadata

Authors
Zhang, X., Wu, C., Zhang, Y., Xie, W., & Wang, Y.
Affiliations
Shanghai Jiao Tong University (CN)
Corresponding author
  • Xie, Weidi
  • Wang, Yanfeng
ORCID
Venue
Nature Communications
Publisher
Springer Science and Business Media LLC
Published
2023-07-28
Volume
14
Issue
1
Pages
4542-4542
ISSN
2041-1723 (ISSN-L)
DOI
10.1038/s41467-023-40260-7
PMC
PMC10382552
OpenAlex
W4385347692
Cited by
207
References
30
Citations by year
2026: 54 · 2025: 104 · 2024: 47 · 2023: 2
FWCI
18.75
Type
journal article
Language
en
Open access
Licence
Locations
  • Nature Communications (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • RePEc: Research Papers in Economics (submittedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
  • arXiv (Cornell University)
Funders
Science and Technology Commission of Shanghai Municipality (21DZ1100100, 22511106101, BP0719010, 18DZ2270700)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Multimodal Machine Learning Applications
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Health Sciences › Medicine › Radiology, Nuclear Medicine and Imaging › COVID-19 diagnosis using AI
Keywords
vision-language pre-training, chest radiography, radiology reports, medical foundation models, few-shot fine-tuning
MeSH terms
Radiologists, Humans, Language, Radiography, Radiology, Knowledge

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.

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