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

2025-07-24

AI Summary

To address factual hallucinations and opaque reasoning in multimodal models for gastrointestinal (GI) pathology image diagnosis, this work introduces the first large-scale GI pathology dataset annotated with explicit clinical reasoning chains. We propose a “prompt-argumentation” strategy that jointly optimizes lesion classification and anatomical localization, and design a Grouped Relative Policy Optimization (GRPO) framework integrating vision-language modeling with structured prompt engineering. Built upon supervised fine-tuning, GRPO enhances reasoning auditability via intra-group consistency optimization. Experiments on real-world pathology report generation demonstrate that our method achieves a 18.7% improvement in clinical relevance, a 32.4% increase in structural completeness, and a 41.2% reduction in diagnostic error rate over state-of-the-art baselines—significantly advancing model accuracy, trustworthiness, and clinical utility.

Technology Category

• Knowledge Representation and Reasoning: Diagnosis and Abductive Reasoning
• Natural Language Processing: Language Grounding & Multi-modal NLP
• Computer Vision: Multi-modal Vision

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

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.

Problem

• Improving diagnostic accuracy in gastrointestinal pathology via multimodal models
• Reducing factual hallucinations in pathology report generation
• Enhancing reasoning transparency and clinical trustworthiness in AI diagnostics

Innovation

• Reinforcement learning for pathology diagnosis
• Prompt argumentation with lesion classification
• Post-training pipeline with GRPO optimization