Chest X-Ray Foundation Model With Global and Local Representations Integration

Area of Science:

• Medical Imaging AI
• Computer Vision in Healthcare
• Self-Supervised Learning

Background:

• Chest X-ray (CXR) is a primary diagnostic tool, but task-specific models face limitations in scope, data requirements, and generalizability.
• Current approaches often struggle with out-of-distribution datasets and require extensive labeled data.

Purpose of the Study:

• To introduce CheXFound, a self-supervised vision foundation model for learning robust and generalizable CXR representations.
• To enhance downstream task performance using a novel Global and Local Representations Integration (GLoRI) head.

Main Methods:

• Pretraining CheXFound on a large-scale dataset (CXR-987K) from 12 public sources.
• Developing the GLoRI head to integrate global image features with fine- and coarse-grained local disease features.
• Evaluating CheXFound on diverse downstream tasks, including multilabel classification, disease detection, risk estimation, and segmentation.

Main Results:

• CheXFound surpassed state-of-the-art models in classifying 40 disease findings on the CXR-LT 24 dataset.
• Demonstrated superior label efficiency on downstream tasks with limited training data.
• Showcased significant improvements on out-of-distribution datasets for various clinical applications.

Conclusions:

• CheXFound exhibits strong generalization capabilities for diverse downstream CXR analysis tasks.
• The model offers improved label efficiency, paving the way for broader clinical applications.
• Publicly available code facilitates future research and development in medical imaging AI.