Learning from Prototypes: Contrastive Learning with Prior-Aware Multi-Label Chest X-ray Classification.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiology Informatics

Background:

• Multi-label Chest X-ray classification is hindered by data imperfections, including co-occurring pathologies, long-tailed distributions, and visually similar diseases.
• Existing methods struggle to effectively disentangle and learn discriminative representations for individual pathologies within complex CXR data.

Purpose of the Study:

• To propose a novel framework for Chest X-ray classification that synergizes medical prior knowledge with prototype-driven contrastive learning.
• To enable disentangled and discriminative per-pathology representation learning for improved diagnostic accuracy.

Main Methods:

• Integration of a co-occurrence modulated Label Graph Attention (LGA) module using LLM prior knowledge and co-occurrence patterns.
• Employment of a Label-Aware Decoupling (LAD) decoder to isolate pathology-specific features and mitigate class dominance.
• Introduction of an Adaptive Prototype Contrastive Learning (APCL) mechanism to enhance discriminability of visually similar pathologies.

Main Results:

• Achieved state-of-the-art mean AUCs of 0.834 on NIH ChestX-ray14 and 0.840 on CheXpert datasets.
• Demonstrated exceptional zero-shot and few-shot generalization capabilities on the MIMIC-CXR dataset.
• Validated the framework's robustness and potential for real-world clinical deployment.

Conclusions:

• The proposed framework effectively addresses challenges in multi-label CXR classification by leveraging medical priors and advanced contrastive learning.
• The method shows significant promise for improving the accuracy and generalization of automated radiological diagnosis systems.
• The developed approach offers a robust solution for clinical deployment, enhancing diagnostic capabilities in medical imaging.