Update README.md

README.md

@@ -1,6 +1,5 @@
 # Med-R1
-
-Vision-language models (VLMs) have made significant progress in reasoning within natural scenes, yet their potential in medical imaging remains largely underexplored. Medical reasoning tasks, which require robust image analysis and the generation of well-justified answers, present unique challenges due to the inherent complexity of medical images. Transparency and trustworthiness are critical not only for gaining clinicians’ confidence but also for meeting stringent regulatory requirements. To address these challenges, we propose Med-R1, a novel framework that investigates whether reinforcement learning (RL) can enhance the generalizability and trustworthiness of VLMs in medical reasoning. Building on the recently introduced DeepSeek strategy, we adopt Group Relative Policy Optimization (GRPO) for RL, which encourages models to explore reasoning paths guided by reward signals. Unlike supervised fine-tuning (SFT), which often overfits to training data and struggles with generalization, RL enables models to develop more robust and diverse reasoning capabilities. We comprehensively evaluate Med-R1 across eight distinct medical imaging modalities: Computed Tomography (CT), Magnetic Resonance Imaging (MRI), Ultrasound, Dermoscopy, Fundus Photography, Optical Coherence Tomography (OCT), Microscopy Images, and X-ray Imaging. Compared to the base model, Qwen2-VL-2B, Med-R1 achieves a 29.94% improvement in average accuracy across these modalities and even outperforms Qwen2-VL-72B—a model with 36 times more parameters. To assess model’s generalization abilities, we further test on five different question types: modality recognition, anatomy identification, disease diagnosis, lesion grading, and biological attribute analysis. Med-R1 also demonstrates superior cross-task generalization, outperforming Qwen2-VL-2B by 32.06% and Qwen2-VL-72B in question-type generalization accuracy. These results highlight that RL not only enhances medical reasoning capabilities but also enables parameter-efficient models to exceed the performance of significantly larger counterparts. Combined with its interpretable reasoning outputs, Med-R1 establishes a new potential direction for developing generalizable, trustworthy, and clinically deployable medical VLMs.
+Med-R1 is a reinforcement learning (RL)-enhanced vision-language model (VLM) designed for medical reasoning across 8 imaging modalities (CT, MRI, Ultrasound, Dermoscopy, Fundus Photography, Optical Coherence Tomography (OCT), Microscopy, and X-ray) and 5 key tasks (modality recognition, anatomy identification, disease diagnosis, lesion grading, and biological attribute analysis). Using Group Relative Policy Optimization (GRPO), Med-R1 improves generalization and trustworthiness, surpassing Qwen2-VL-2B by 29.94% and even outperforming the much larger Qwen2-VL-72B. Our model checkpoints provide researchers with a powerful tool for advancing medical AI with RL-driven enhancements.
 
 ## Start with the Model