UniChest: Conquer-and-Divide Pre-Training for Multi-Source Chest X-Ray Classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Vision-Language Pre-training (VLP) shows promise for Chest X-ray (CXR) diagnosis.
• Current VLP models often focus on single datasets, limiting potential with multi-source CXR data.
• Heterogeneity across diverse CXR sources poses challenges for model generalization.

Purpose of the Study:

• To develop a novel pre-training framework, UniChest, for effective multi-source CXR analysis.
• To leverage the benefits of diverse CXR datasets while mitigating issues from data heterogeneity.
• To enhance the generalization capabilities of VLP models in medical imaging.

Main Methods:

• Designed a "Conquer-and-Divide" pre-training framework (UniChest).
• The "Conquer" stage captures common patterns across multiple CXR sources.
• The "Divide" stage uses query networks (experts) to learn personalized patterns.

Main Results:

• UniChest demonstrated superior performance across various benchmarks, including ChestX-ray14, CheXpert, and others.
• The framework effectively balances learning commonalities and individualities from heterogeneous CXR data.
• Experimental results validate the effectiveness of the proposed approach over existing baselines.

Conclusions:

• UniChest offers a robust solution for Vision-Language Pre-training on multi-source Chest X-ray datasets.
• The "Conquer-and-Divide" strategy successfully addresses data heterogeneity challenges.
• The study provides valuable pre-trained models and code for advancing medical image analysis.