Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Skin-R1: Toward Trustworthy Clinical Reasoning for Dermatological Diagnosis

Zehao Liu, Wejieying Ren, Jipeng Zhang, Tianxiang Zhao, Jingxi Zhu, Xiaoting Li, Vasant G. Honavar

TL;DR

Skin-R1 presents a three-stage dermatology vision-language framework that combines textbook-grounded reasoning with reinforcement learning to overcome data heterogeneity, lack of grounded reasoning supervision, and limited scalability. It introduces SkinRationale to capture hierarchy-aware and DDx-informed reasoning trajectories, followed by supervised fine-tuning and a GRPO-based RL stage that transfers grounded reasoning to large, sparse datasets. The approach yields superior diagnostic accuracy and robustness across multiple dermatology datasets and shows strong generalization on out-of-distribution data and DDx/hierarchical reasoning tasks. This work advances trustworthy AI in dermatology by grounding decisions in expert-like reasoning while maintaining scalability to diverse, real-world data. The results suggest practical impact for triage and diagnostics in settings with limited access to dermatology specialists.

Abstract

The emergence of vision-language models (VLMs) has opened new possibilities for clinical reasoning and has shown promising performance in dermatological diagnosis. However, their trustworthiness and clinical utility are often limited by three major factors: (1) Data heterogeneity, where diverse datasets lack consistent diagnostic labels and clinical concept annotations; (2) Absence of grounded diagnostic rationales, leading to a scarcity of reliable reasoning supervision; and (3) Limited scalability and generalization, as models trained on small, densely annotated datasets struggle to transfer nuanced reasoning to large, sparsely-annotated ones. To address these limitations, we propose SkinR1, a novel dermatological VLM that combines deep, textbook-based reasoning with the broad generalization capabilities of reinforcement learning (RL). SkinR1 systematically resolves the key challenges through a unified, end-to-end framework. First, we design a textbook-based reasoning generator that synthesizes high-fidelity, hierarchy-aware, and differential-diagnosis (DDx)-informed trajectories, providing reliable expert-level supervision. Second, we leverage the constructed trajectories for supervised fine-tuning (SFT) empowering the model with grounded reasoning ability. Third, we develop a novel RL paradigm that, by incorporating the hierarchical structure of diseases, effectively transfers these grounded reasoning patterns to large-scale, sparse data. Extensive experiments on multiple dermatology datasets demonstrate that SkinR1 achieves superior diagnostic accuracy. The ablation study demonstrates the importance of the reasoning foundation instilled by SFT.

Skin-R1: Toward Trustworthy Clinical Reasoning for Dermatological Diagnosis

TL;DR

Skin-R1 presents a three-stage dermatology vision-language framework that combines textbook-grounded reasoning with reinforcement learning to overcome data heterogeneity, lack of grounded reasoning supervision, and limited scalability. It introduces SkinRationale to capture hierarchy-aware and DDx-informed reasoning trajectories, followed by supervised fine-tuning and a GRPO-based RL stage that transfers grounded reasoning to large, sparse datasets. The approach yields superior diagnostic accuracy and robustness across multiple dermatology datasets and shows strong generalization on out-of-distribution data and DDx/hierarchical reasoning tasks. This work advances trustworthy AI in dermatology by grounding decisions in expert-like reasoning while maintaining scalability to diverse, real-world data. The results suggest practical impact for triage and diagnostics in settings with limited access to dermatology specialists.

Abstract

The emergence of vision-language models (VLMs) has opened new possibilities for clinical reasoning and has shown promising performance in dermatological diagnosis. However, their trustworthiness and clinical utility are often limited by three major factors: (1) Data heterogeneity, where diverse datasets lack consistent diagnostic labels and clinical concept annotations; (2) Absence of grounded diagnostic rationales, leading to a scarcity of reliable reasoning supervision; and (3) Limited scalability and generalization, as models trained on small, densely annotated datasets struggle to transfer nuanced reasoning to large, sparsely-annotated ones. To address these limitations, we propose SkinR1, a novel dermatological VLM that combines deep, textbook-based reasoning with the broad generalization capabilities of reinforcement learning (RL). SkinR1 systematically resolves the key challenges through a unified, end-to-end framework. First, we design a textbook-based reasoning generator that synthesizes high-fidelity, hierarchy-aware, and differential-diagnosis (DDx)-informed trajectories, providing reliable expert-level supervision. Second, we leverage the constructed trajectories for supervised fine-tuning (SFT) empowering the model with grounded reasoning ability. Third, we develop a novel RL paradigm that, by incorporating the hierarchical structure of diseases, effectively transfers these grounded reasoning patterns to large-scale, sparse data. Extensive experiments on multiple dermatology datasets demonstrate that SkinR1 achieves superior diagnostic accuracy. The ablation study demonstrates the importance of the reasoning foundation instilled by SFT.

Paper Structure

This paper contains 42 sections, 10 equations, 3 figures, 8 tables, 4 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Medical Vision Large Language Models.
  4. RL-Driven Emergence of Reasoning Abilities on Med-VLMs.
  5. Methods
  6. Training Data for Grounded Reasoning
  7. Extracting Textbook Grounded Exemplar
  8. Constructing Reasoning Trajectories
  9. Hierarchical diagnosis completion.
  10. Differential diagnosis reasoning enrichment.
  11. Enhancing Grounded Medical Reasoning
  12. Supervised Fine-Tuning on SkinRationale.
  13. Reinforcement Learning via GRPO
  14. Reward function.
  15. Format Compliance Reward ($R_{\text{format}}$).
  16. ...and 27 more sections

Figures (3)

  • Figure 1: Top: Overview of the proposed training framework, illustrating (1) the synthesis of the SkinRationale dataset, (2) the SFT stage performed on SkinRationale, and (3) the GRPO-based RL stage. Bottom: Two representative cases comparing diagnostic responses from Skin-R1 and baseline models. The response from Skin-R1 is concise, accurate, and aligned with expert diagnostic reasoning. Red text highlights incorrect or hallucinated content; Green text denotes hierarchical-aware reasoning; and content enclosed within the blue box indicates the DDx comparison derived from prior differential diagnostic reasoning.
  • Figure 2: Training reward curve comparison of Skin-R1 with or without SFT stage.
  • Figure 3: Comparison of Skin-R1 with baseline models on DDx and hierarchical diagnosis test sets.