Current Vision-Language Models (VLMs) in computational pathology often function as "black boxes," providing direct diagnosis without verifiable evidence. PathReasoner-R1 bridges this gap by introducing structured Chain-of-Thought (CoT) reasoning into pathology, transforming models from simple classifiers into transparent clinical reasoners.
- PathReasoner Dataset**: The first large-scale Whole-Slide Image (WSI) reasoning dataset. It contains 15K+ high-quality instruction samples where findings and clinical reasoning are explicitly aligned with diagnoses.
- Knowledge-Guided Generation**: Unlike traditional unverified distillation, we use Medical Knowledge Graphs to generate rigorous, structured pathological reasoning trajectories.
- PathReasoner-R1 Framework**: A two-stage training framework that combines dual-modality SFT augmentation with knowledge-aware reinforcement learning (RL).
- Knowledge-Aware Reward Function**: Incorporates a specialized Entity Reward mechanism strictly aligned with knowledge graphs to optimize for logical consistency and eliminate hallucinations.
- State-of-the-Art Performance**: Achieves SOTA results across multiple benchmarks and image scales, providing clinically grounded and transparent decision-making.
We leverage medical knowledge graphs to convert raw pathology findings into structured reasoning paths. This ensures that every diagnosis is backed by a verifiable chain of evidence.
- Phase 1: Dual-Modality SFT Augmentation — trajectory masking on text and entity-drop-guided pseudo-bag masking on WSI features, teaching the model to reason reliably under partial evidence.
- Phase 2: Knowledge-Aware RL — a multi-granular reward (format + semantic + entity) whose Entity Reward grounds each reasoning step in KG-verified medical concepts.
PathReasoner provides a leap in data quality for CPath:
- Scalability: Over 15,000 instruction-following reasoning pairs.
- Granularity: Covers multiple image scales.
- Rigor: Every sample is aligned with a medical knowledge graph to ensure clinical validity.
- Available on Hugging Face: jshhhh/PathReasoner (Coming Soon)
To run a quick inference with PathReasoner-R1, use the code in ./src/demo.py
- PrimeKG (Medical KG): https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/IXA7BM
- PathoGraph (Pathology KG): https://github.com/Peiliang/PathoML
- MedResearch-R1 （Trajectory masking）: https://github.com/AQ-MedAI/MedResearcher-R1
- MedReason（Entity extraction）: https://github.com/UCSC-VLAA/MedReason
This project is licensed under the Apache 2.0 License. Disclaimer: This model is for research purposes only. Always consult with a qualified pathologist for clinical decisions.