arXiv：cs.AI· Yuxin Xie, Yuming Chen, Yishan Yang, Yi Zhou, Tao Zhou, Zhen Zhao, Jiacheng Liu, Huazhu Fu·· 13 小时前
CORE-Seg：基于强化学习的推理驱动复杂病变分割
CORE-Seg: Reasoning-Driven Segmentation for Complex Lesions via Reinforcement Learning
AI 导读
研究者提出 CORE-Seg 端到端框架，通过语义引导提示适配器将推理与分割结合，并采用从 SFT 到 GRPO 的渐进训练策略与自适应双粒度奖励机制缓解奖励稀疏问题。该方法平均 Dice 达 37.06%，比第二优基线高 14.89%，失败率降至 18.42%。同时发布首个推理驱动复杂病变分割的 CoT 基准 ComLesion-14K。
正文
Abstract:Medical image segmentation is undergoing a paradigm shift from conventional visual pattern matching to cognitive reasoning analysis. Although Multimodal Large Language Models (MLLMs) have shown promise in integrating linguistic and visual knowledge, significant gaps remain: existing general MLLMs possess broad common sense but lack the specialized visual reasoning required for complex lesions, whereas traditional segmentation models excel at pixel-level segmentation but lack logical interpretability. In this paper, we introduce ComLesion-14K, the first diverse Chain-of-Thought (CoT) benchmark for reasoning-driven complex lesion segmentation. To accomplish this task, we propose CORE-Seg, an end-to-end framework integrating reasoning with segmentation through a Semantic-Guided Prompt Adapter. We design a progressive training strategy from SFT to GRPO, equipped with an adaptive dual-granularity reward mechanism to mitigate reward sparsity. Our Method achieves state-of-the-art results with a mean Dice of 37.06\% (14.89\% higher than the second-best baseline), while reducing the failure rate to 18.42\%. Project Page: this https URL
| Comments: | Accepted by Computational Visual Media |
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2603.05911 [cs.CV] |
| (or arXiv:2603.05911v2 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2603.05911 arXiv-issued DOI via DataCite |
Submission history
From: Yuxin Xie [view email]
[v1]
Fri, 6 Mar 2026 04:58:55 UTC (18,635 KB)
[v2]
Fri, 9 Oct 2026 01:26:23 UTC (7,296 KB)
来源：arXiv：cs.AI · arxiv.org