DiagR1: A Vision-Language Model Trained via Reinforcement Learning for Digestive Pathology Diagnosis.

Area of Science:

• Pathology
• Artificial Intelligence
• Medical Imaging

Background:

• Multimodal large models show promise for pathology image analysis.
• Current models face challenges with data quality and reasoning transparency, leading to errors and lack of trust.

Purpose of the Study:

• To improve the accuracy and trustworthiness of AI-generated diagnostic reports in gastrointestinal pathology.
• To address data quality issues and enhance reasoning transparency in multimodal models.

Main Methods:

• Constructed a large-scale gastrointestinal pathology dataset with microscopic descriptions and diagnostic conclusions.
• Developed a prompt augmentation strategy incorporating lesion classification and anatomical site information.
• Employed a post-training pipeline with supervised fine-tuning and Group Relative Policy Optimization (GRPO).

Main Results:

• The proposed approach significantly outperforms state-of-the-art baselines in generation quality, structural completeness, and clinical relevance.
• Achieved 18.7% higher clinical relevance, 32.4% improved structural completeness, and 41.2% fewer diagnostic errors.
• Demonstrated superior accuracy and clinical utility in real-world pathology report generation tasks.

Conclusions:

• The developed method enhances multimodal AI for gastrointestinal pathology, yielding more reliable and clinically useful diagnostic reports.
• The approach effectively tackles data quality and reasoning transparency issues, paving the way for trustworthy AI in clinical practice.