A unified multi-task framework enables interpretable chest radiograph analysis.

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Radiology

Background:

• Current deep learning models for medical imaging often function as black boxes, limiting their application in clinical diagnosis.
• Clinical diagnosis is a multi-task process that requires trust and interpretability, which are often lacking in AI systems.

Purpose of the Study:

• To develop an interpretable multi-task transformer framework for chest X-ray analysis that emulates a radiologist's diagnostic workflow.
• To enhance the clinical utility and trustworthiness of AI in medical imaging.

Main Methods:

• The IMT-CXR framework utilizes a unified transformer architecture with medical-domain instruction tuning.
• It performs four sequential clinical tasks: disease classification, lesion localization, anatomical segmentation, and report generation.
• The framework provides traceable decision pathways from findings to conclusions.

Main Results:

• IMT-CXR demonstrated competitive performance across ten chest X-ray benchmarks.
• A blinded evaluation showed that 66% of AI-generated reports were rated as comparable to or better than original clinical reports by radiologists.
• The framework highlights significant translational potential for AI in clinical settings.

Conclusions:

• The proposed framework bridges the gap between AI technical performance and clinical utility by providing traceable diagnostic pathways.
• This work advances the development of trustworthy AI systems for medical imaging analysis.
• The interpretable nature of the AI enhances clinical adoption and trust.