Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A foundation model for generalizable disease diagnosis in chest X-ray images

Lijian Xu, Ziyu Ni, Hao Sun, Hongsheng Li, Shaoting Zhang

TL;DR

CXRBase is a foundational model designed to learn versatile representations from unlabelled CXR images, facilitating efficient adaptation to various clinical tasks and improving model performance and alleviating the annotation workload of experts to enable broad clinical AI applications from chest imaging.

Abstract

Medical artificial intelligence (AI) is revolutionizing the interpretation of chest X-ray (CXR) images by providing robust tools for disease diagnosis. However, the effectiveness of these AI models is often limited by their reliance on large amounts of task-specific labeled data and their inability to generalize across diverse clinical settings. To address these challenges, we introduce CXRBase, a foundational model designed to learn versatile representations from unlabelled CXR images, facilitating efficient adaptation to various clinical tasks. CXRBase is initially trained on a substantial dataset of 1.04 million unlabelled CXR images using self-supervised learning methods. This approach allows the model to discern meaningful patterns without the need for explicit labels. After this initial phase, CXRBase is fine-tuned with labeled data to enhance its performance in disease detection, enabling accurate classification of chest diseases. CXRBase provides a generalizable solution to improve model performance and alleviate the annotation workload of experts to enable broad clinical AI applications from chest imaging.

A foundation model for generalizable disease diagnosis in chest X-ray images

TL;DR

CXRBase is a foundational model designed to learn versatile representations from unlabelled CXR images, facilitating efficient adaptation to various clinical tasks and improving model performance and alleviating the annotation workload of experts to enable broad clinical AI applications from chest imaging.

Abstract

Medical artificial intelligence (AI) is revolutionizing the interpretation of chest X-ray (CXR) images by providing robust tools for disease diagnosis. However, the effectiveness of these AI models is often limited by their reliance on large amounts of task-specific labeled data and their inability to generalize across diverse clinical settings. To address these challenges, we introduce CXRBase, a foundational model designed to learn versatile representations from unlabelled CXR images, facilitating efficient adaptation to various clinical tasks. CXRBase is initially trained on a substantial dataset of 1.04 million unlabelled CXR images using self-supervised learning methods. This approach allows the model to discern meaningful patterns without the need for explicit labels. After this initial phase, CXRBase is fine-tuned with labeled data to enhance its performance in disease detection, enabling accurate classification of chest diseases. CXRBase provides a generalizable solution to improve model performance and alleviate the annotation workload of experts to enable broad clinical AI applications from chest imaging.

Paper Structure

This paper contains 22 sections, 6 figures.

Table of Contents

  1. Introduction
  2. Results
  3. Overview
  4. Image Quality Assessment
  5. 3D Reconstruction
  6. Disease Diagnosis
  7. Prognosis and prediction
  8. SSL Strategies
  9. Model Interpretation
  10. Discussion
  11. Model achieves comparable results with expert radiologists
  12. Limitation and future work
  13. Material and Method
  14. Building dataset for customized instruction tuning
  15. Datasets for pre-training
  16. ...and 7 more sections

Figures (6)

  • Figure 1: Schematic of foundation model (i.e., CXRBase).参考nature文章图1，俩阶段
  • Figure 2: Model performance of Image Quality on 8 single-class Foreign Object Detection(a) and Position Assess (b). The datasets include three important disease categories: COVID-19, pneumonia, and tuberculosis. ACC and AUC are used as evaluation metrics. The MAE_DR model utilizes few-shot results, while other comparative models utilize full-data results.
  • Figure 3: Compariosn of different SSL strategies.
  • Figure 4: Model performance on 8 single-class disease datasets. The datasets include three important disease categories: COVID-19, pneumonia, and tuberculosis. ACC and AAUC are used as evaluation metrics. The MAE_DR model utilizes few-shot results, while other comparative models utilize full-data results.
  • Figure 5: Illustration of training pipeline of CXRBase. 参考nature 文字附图1
  • ...and 1 more figures