PathReasoner-R1: Instilling Structured Reasoning into Pathology Vision-Language Model via Knowledge-Guided Policy Optimization

TL;DR

PathReasoner tackles the lack of transparent reasoning in pathology vision-language models by building PathReasoner, a large-scale WSI reasoning dataset grounded in medical knowledge graphs. It introduces PathReasoner-R1, a two-stage training pipeline combining trajectory-masked supervised fine-tuning and knowledge-guided reinforcement learning with an Entity Reward to promote clinically grounded reasoning. The knowledge graphs integrate PrimeKG and PathoGraph to connect micro-scale histology features to macro-scale diagnoses, enabling verifiable chain-of-thought reasoning across gigapixel WSIs. Across open-ended and external benchmarks, PathReasoner-R1 achieves state-of-the-art results while providing interpretable reasoning traces, advancing trustworthy AI in computational pathology. The dataset and code are publicly available at the project repository.

Abstract

Vision-Language Models (VLMs) are advancing computational pathology with superior visual understanding capabilities. However, current systems often reduce diagnosis to directly output conclusions without verifiable evidence-linked reasoning, which severely limits clinical trust and hinders expert error rectification. To address these barriers, we construct PathReasoner, the first large-scale dataset of whole-slide image (WSI) reasoning. Unlike previous work reliant on unverified distillation, we develop a rigorous knowledge-guided generation pipeline. By leveraging medical knowledge graphs, we explicitly align structured pathological findings and clinical reasoning with diagnoses, generating over 20K high-quality instructional samples. Based on the database, we propose PathReasoner-R1, which synergizes trajectory-masked supervised fine-tuning with reasoning-oriented reinforcement learning to instill structured chain-of-thought capabilities. To ensure medical rigor, we engineer a knowledge-aware multi-granular reward function incorporating an Entity Reward mechanism strictly aligned with knowledge graphs. This effectively guides the model to optimize for logical consistency rather than mere outcome matching, thereby enhancing robustness. Extensive experiments demonstrate that PathReasoner-R1 achieves state-of-the-art performance on both PathReasoner and public benchmarks across various image scales, equipping pathology models with transparent, clinically grounded reasoning capabilities. Dataset and code are available at https://github.com/cyclexfy/PathReasoner-R1.

Figures (24)

• Figure 1: Comparisons of mainstream vision-language models in computational pathology. Existing models like SlideChat and Qwen3-VL perform direct diagnosis, while Patho-R1 generates superficial reasoning. In contrast, our PathReasoner-R1 employs medically grounded step-by-step reasoning, explicitly linking visual evidence to the diagnosis. Text colors correspond to the bounding boxes, and underlines highlight the logical flow.
• Figure 2: Overview of the PathReasoner construction pipeline. The framework transforms unstructured reports into structured CoT annotations through three key stages: constructing a medical knowledge graph from public platforms, aligning entities extracted from WSI pathology reports with graph nodes, and generating explicit reasoning paths that logically link visual findings to the final diagnosis.
• Figure 3: Statistical overview of PathReasoner. (a) Data distribution across 10 cancer types and various anatomical sites. (b) Diversity of pathological concepts covered in the dataset.
• Figure 4: Overview of the PathReasoner-R1 framework. Building upon the PathReasoner dataset, the framework implements a two-stage post-training process for SlideChat: SFT-based reasoning activation and RL-based reasoning enhancement. This pipeline sequentially generates the initial policy model, PathReasoner-SFT-7B, and the final model, PathReasoner-R1-7B. The resulting model is optimized for open-ended VQA tasks, delivering well-organized outputs with superior reasoning capabilities.
• Figure 5: Qualitative comparison of different VLMs for pathology diagnosis. Red text indicates an incorrect diagnosis, while bold text indicates a correct diagnosis. The proposed PathReasoner‑R1 successfully captures pathological visual features (blue, orange, and green texts) and provides a comprehensive reasoning process for accurate slide‑level diagnosis. More samples are in Appendix \ref{['subsec:qualitative']}.