Enhancing Radiology Report Generation via Multi-Phased Supervision

Area of Science:

• Artificial Intelligence in Medical Imaging
• Natural Language Processing for Healthcare

Background:

• Large language models (LLMs) show promise in radiology report generation, improving style and fluency.
• Current LLMs struggle with clinical accuracy due to supervision methods that don't prioritize critical clinical phrases.

Purpose of the Study:

• To develop a novel multi-phased supervision method for LLM-based radiology report generation.
• To enhance both the clinical accuracy and language fluency of generated radiology reports.

Main Methods:

• Proposed a curriculum learning-inspired multi-phased supervision approach.
• Phase 1: Disease label supervision for disease identification.
• Phase 2: Entity-relation triple supervision for clinical finding description.
• Phase 3: Whole-report supervision for final report generation adaptation.

Main Results:

• The multi-phased supervision method achieved state-of-the-art performance.
• Demonstrated significant improvements in both clinical accuracy and language fluency.
• The training process design is crucial for effective radiology report generation.

Conclusions:

• The proposed multi-phased supervision method effectively addresses limitations in current LLM radiology report generation.
• This approach enhances the clinical utility and readability of AI-generated radiology reports.
• Highlights the importance of structured training methodologies in medical AI.