MedReason-R1: Learning to Reason for CT Diagnosis with Reinforcement Learning and Local Zoom

TL;DR

This work targets the limited accuracy and interpretability of general Vision-Language Models on CT diagnosis by introducing the CT-RATE-VQA dataset with 84k QA pairs and MedReason-R1, a VLM designed for explicit coarse-to-fine reasoning via a local zoom-in ROI strategy. The model is trained in two stages: supervised fine-tuning of a medical-adapted VL model, followed by reinforcement learning with Group Relative Policy Optimization (GRPO) using a composite reward that enforces structured reasoning, category validity, and prediction correctness, expressed with $R = \alpha R_{format} + \beta R_{validity} + \gamma R_{correctness}$ and policy updates via $J(\theta) = \mathbb{E}[ \min(\rho^{(i)} \hat{A}^{(i)}, \text{clip}(\rho^{(i)}, 1-\delta, 1+\delta) \hat{A}^{(i)}) - \beta \mathrm{KL}[\pi_{\theta_{new}} || \pi_{ref}] ]$. Results show that MedReason-R1 achieves an average accuracy of $0.5218$ on CT-RATE-VQA, outperforming strong baselines and maintaining general-domain VQA performance, with ablations confirming the importance of zoom-in data augmentation and the composite reward design. The findings offer a practical data and training solution for medically interpretable CT diagnosis using vision-language models, enabling robust performance and reasoning without expensive manual annotations.

Abstract

General-purpose large Vision-Language Models (VLMs) demonstrate strong capabilities in generating detailed descriptions for natural images. However, their performance in the medical domain remains suboptimal, even for relatively straightforward tasks, primarily due to the lack of large-scale, high-quality, specialized medical imaging datasets and the neglect of the diagnostic process that progresses from coarse to fine-grained. To address the first issue, we construct the CT-RATE-VQA dataset, which has 84K QA pairs. For the second issue, we propose MedReason-R1, a medical VLM with explicit reasoning process for disease diagnosis. MedReason-R1 incorporates a novel strategy that embeds zoom-in disease region-of-interest areas into the image, highlighting the crucial role of both global localization and disease-specific details in enhancing the model's diagnostic performance. Furthermore, we introduce the GRPO reinforcement learning framework to MedReason-R1, which enables effective reasoning without relying on costly manual annotations. Compared to recent general-purpose and medical VLMs, MedReason-R1 achieves state-of-the-art performance in CT disease diagnosis while retaining generalization. The code, checkpoints, and dataset are available at: https://github.com/Leevan001/MedReason-R1

Figures (3)

• Figure 1: Framework and results of the proposed method. (a) Global-to-local reasoning enables lesion localization and fine-grained feature analysis. (b) Our dataset comprises 84K QA pairs, substantially larger than existing VQA-RAD and SLAKE. (c) GRPO pipeline integrates correctness, format, and validation rewards to optimize model outputs. (d) Comparative evaluation shows our model achieves superior or competitive performance against both general-purpose and medical vision–language models on chest CT abnormalities.
• Figure 2: Overview of Stage I: supervised fine-tuning. Only the MLP layers of Qwen2.5-VL are updated to align the model with the disease diagnosis task.
• Figure 3: Accuracy (left) and CoT reasoning–answer consistency (right) under different training and inference setups.