Anatomy-R1: Enhancing Anatomy Reasoning in Multimodal Large Language Models via Anatomical Similarity Curriculum and Group Diversity Augmentation

TL;DR

This paper tackles the instability and data-inefficiency of reinforcement-learning-based reasoning in medical multimodal language models for anatomical recognition. It introduces Anatomical Similarity Curriculum Learning (ASC-Learning) to progressively raise task difficulty by intra-instance distractor similarity, and Group Diversity Question Augmentation (GDQA) to diversify prompts and visual inputs, thus expanding the reasoning search space. Through experiments on SGG-VQA Anatomy and OmniMedVQA MI using Qwen-based VL models, the authors demonstrate that ASC-Learning and GDQA improve training stability and final performance, with larger gains on bigger models. The work shows that addressing intra-group diversity and curriculum design is crucial for reliable clinical reasoning in low-data medical domains. The proposed methods enhance robustness and can facilitate safer deployment of medical visual-language systems in practice.

Abstract

Multimodal Large Language Models (MLLMs) have achieved impressive progress in natural image reasoning, yet their potential in medical imaging remains underexplored, especially in clinical anatomical surgical images. Anatomy understanding tasks demand precise understanding and clinically coherent answers, which are difficult to achieve due to the complexity of medical data and the scarcity of high-quality expert annotations. These challenges limit the effectiveness of conventional Supervised Fine-Tuning (SFT) strategies. While recent work has demonstrated that Group Relative Policy Optimization (GRPO) can enhance reasoning in MLLMs without relying on large amounts of data, we find two weaknesses that hinder GRPO's reasoning performance in anatomy recognition: 1) knowledge cannot be effectively shared between different anatomical structures, resulting in uneven information gain and preventing the model from converging, and 2) the model quickly converges to a single reasoning path, suppressing the exploration of diverse strategies. To overcome these challenges, we propose two novel methods. First, we implement a progressive learning strategy called Anatomical Similarity Curriculum Learning by controlling question difficulty via the similarity of answer choices, enabling the model to master complex problems incrementally. Second, we utilize question augmentation referred to as Group Diversity Question Augmentation to expand the model's search space for difficult queries, mitigating the tendency to produce uniform responses. Comprehensive experiments on the SGG-VQA and OmniMedVQA benchmarks show our method achieves a significant improvement across the two benchmarks, demonstrating its effectiveness in enhancing the medical reasoning capabilities of MLLMs. The code can be found in https://github.com/tomato996/Anatomy-R1

Figures (6)

• Figure 1: Results of the baseline GRPO model trained without a curriculum. The MLLMs can learn simple anatomies but fail on more challenging ones. Specifically, the knowledge acquired from those simpler structures does not transfer well to the more difficult anatomies.
• Figure 2: Model responses after several training steps with the direct application of GRPO. The uniformity of the responses indicates that the model's reasoning paths have become restricted. This consequently causes the gradient in this group to become zero, preventing effective updates.
• Figure 3: The process of Anatomical Similarity Curriculum Learning. We progressively increase task complexity with aligned reward mechanisms. This enables the model to better acquire knowledge across anatomies of varying difficulty, thereby improving its performance.
• Figure 4: The framework of Group Diversity Question Augmentation, where $R_i$ is the MLLM's response to the i-th prompt, and $\overline{R}_G$ is the group-average response. We expand the question space through semantically equivalent transformations and subsequently explore diverse reasoning trajectories arising from different question variants.
• Figure 5: The reward and S(q) curves of the Qwen-2.5-VL-7B during training on the SGG-VQA dataset.