Unlocking the Potential of Weakly Labeled Data: A Co-Evolutionary Learning Framework for Abnormality Detection and...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Chest X-ray (CXR) analysis involves abnormality detection and report generation, crucial for clinical diagnosis.
• Current methods often address these tasks in isolation, neglecting their inherent correlation.
• Integrating these tasks can lead to more comprehensive and accurate radiological assessments.

Purpose of the Study:

• To develop a unified framework, CoE-DG, for co-evolutionary abnormality detection and report generation in CXRs.
• To leverage both fully and weakly labeled data for mutual task promotion.
• To enhance the performance of both individual tasks through their synergistic interaction.

Main Methods:

• Proposed a co-evolutionary abnormality detection and report generation (CoE-DG) framework.
• Introduced bi-directional information interaction: Generator-Guided Information Propagation (GIP) and Detector-Guided Information Propagation (DIP).
• Implemented a semi-supervised approach using GIP for detection and DIP for report generation, enhanced by a Self-Adaptive Non-Maximum Suppression (SA-NMS) module.

Main Results:

• CoE-DG demonstrated superior performance compared to existing state-of-the-art models on two public CXR datasets.
• The framework effectively integrated abnormality detection and report generation, achieving mutual promotion.
• Experimental results validated the efficacy of GIP, DIP, and SA-NMS in improving task-specific and overall performance.

Conclusions:

• The CoE-DG framework offers a significant advancement in automated CXR analysis by jointly optimizing detection and generation.
• Co-evolutionary training and bi-directional information flow are effective strategies for improving performance in medical image analysis.
• This approach holds promise for enhancing clinical workflow and diagnostic accuracy in radiology.