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

TL;DR

This paper introduces DiagR1, a vision-language model trained with reinforcement learning that improves gastrointestinal pathology diagnosis by enhancing reasoning, reducing errors, and increasing clinical relevance through a specialized dataset and prompt strategy.

Contribution

The paper presents a large-scale gastrointestinal pathology dataset and a novel prompt argumentation strategy, combined with reinforcement learning, to improve diagnostic reasoning and output quality.

Findings

Outperforms existing models in clinical relevance by 18.7%

Achieves 32.4% better structural completeness

Reduces diagnostic errors by 41.2%

Abstract

Multimodal large models have shown great potential in automating pathology image analysis. However, current multimodal models for gastrointestinal pathology are constrained by both data quality and reasoning transparency: pervasive noise and incomplete annotations in public datasets predispose vision language models to factual hallucinations when generating diagnostic text, while the absence of explicit intermediate reasoning chains renders the outputs difficult to audit and thus less trustworthy in clinical practice. To address these issues, we construct a large scale gastrointestinal pathology dataset containing both microscopic descriptions and diagnostic conclusions, and propose a prompt argumentation strategy that incorporates lesion classification and anatomical site information. This design guides the model to better capture image specific features and maintain semantic consistency in generation. Furthermore, we employ a post training pipeline that combines supervised fine tuning with Group Relative Policy Optimization (GRPO) to improve reasoning quality and output structure. Experimental results on real world pathology report generation tasks demonstrate that our approach significantly outperforms state of the art open source and proprietary baselines in terms of generation quality, structural completeness, and clinical relevance. Our solution outperforms state of the art models with 18.7% higher clinical relevance, 32.4% improved structural completeness, and 41.2% fewer diagnostic errors, demonstrating superior accuracy and clinical utility compared to existing solutions.