CORE-Seg: Reasoning-Driven Segmentation for Complex Lesions via Reinforcement Learning

Yuxin Xie¹, Yuming Chen¹, Yishan Yang¹, Yi Zhou¹∗, Tao Zhou², Zhen Zhao³, Jiacheng Liu³, Huazhu Fu⁴

¹School of Computer Science and Engineering, Southeast University, Nanjing, China    ²School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China    ³Zhongda Hospital, Southeast University, Nanjing, China    ⁴Institute of High-Performance Computing, A*STAR, Singapore
∗Corresponding author: Yi Zhou (yizhou.szcn@gmail.com)

Abstract

Medical image segmentation is undergoing a paradigm shift from conventional visual pattern matching to cognitive reasoning-driven 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 large-scale Chain-of-Thought (CoT) benchmark specifically curated to demand reasoning-driven understanding for complex lesion segmentation. To accomplish this task, we propose CORE-Seg, an end-to-end framework integrating reasoning with segmentation through a latent interaction mechanism. 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%, less than half that of general MLLMs.

Dataset

To address the lack of reasoning-oriented complex lesion segmentation data, we construct ComLesion-14K, the first large-scale Chain-of-Thought (CoT) driven multimodal benchmark for complex lesion segmentation, curated from 26 public authorized datasets with a three-stage construction pipeline. The dataset contains 13678 complex lesion samples covering 8 imaging modalities (MRI, CT, X-Ray, Ultrasound, OCT, etc.), 9 anatomical regions (Head, Abdomen, Lung, Eye, etc.) and 31 disease categories, focusing on clinical scenarios where traditional segmentation models fail.

The construction process includes: (1) Difficulty-Aware Filtering: Using U-Net to model segmentation error distribution (power-law function) and Kneedle algorithm to select hard samples with high morphological variability and low contrast; (2) CoT and VQA Generation: Constructing structured tuples (image, mask, query, reasoning, answer) via GPT-4o with a two-step prompting strategy, simulating clinician's diagnostic reasoning and stripping coordinate values to avoid shortcut learning; (3) Automated Quality Assurance: Using Qwen2.5VL-Max to score samples based on normal anatomy description, lesion characterization and reasoning logic (weighted 0.3:0.3:0.4), regenerating samples with scores below 0.8 to ensure quality.

Method

CORE-Seg is an end-to-end complex-lesion-centric reasoning segmentation framework that unifies the coherence of end-to-end models with the explicit reasoning of reinforcement learning, eliminating box-based error propagation and reward sparsity in traditional methods. The framework consists of three core components and a two-stage progressive training pipeline, with a custom adaptive dual-granularity reward mechanism for reinforcement learning optimization.

Core Components: (1) Multimodal Reasoning Module: Based on Qwen2.5-VL-3B, generates structured reasoning (/) and answer () with a special token as the semantic anchor for segmentation; (2) Semantic-Guided Prompt Adapter: Projects the token's hidden state from MLLM's textual space to SAM's visual feature space via ResMLP and cross-attention, serving as a global semantic descriptor for multi-lesion segmentation; (3) Segmentation Module: Based on MedSAM 2, fuses image embeddings and semantic prompts to generate pixel-level segmentation masks.

Two-Stage Training: (1) CoT-Based Semantic Alignment (SFT): Applies LoRA to MLLM and fine-tunes the adapter/mask decoder with a composite loss (text cross-entropy + DiceCE) to align linguistic reasoning and visual localization; (2) RL-Based Reasoning Exploration (GRPO): Uses Group Relative Policy Optimization to enhance generalization, with a composite reward function (format reward + bipartite matching reward + dual-granularity mask reward) to guide policy update and refine segmentation boundaries.

Experiment

Extensive experiments are conducted on the ComLesion-14K dataset, with evaluations on in-distribution (ID) complex lesion segmentation and out-of-distribution (OOD) generalization (TNSCUI2020, ISPY, CVC-ClinicDB). We compare CORE-Seg with three categories of SOTA models: General MLLMs (Qwen2.5-VL-72B, InternVL3-8B), Medical-Specific MLLMs (HuatuoGPT-7B, MedGemma-4B) and Grounding-Specific MLLMs (SegZero-7B, LISA-3B), with evaluation metrics including mean Dice (mDice), mean IoU (mIoU) and Failure Rate (Dice=0 or invalid format).

Key Results: (1) CORE-Seg achieves SOTA performance with 37.06% mDice and 27.79% mIoU on ID data, outperforming the second-best LISA-3B by 14.89% in mDice, and reduces the failure rate to 18.42% (less than half of general MLLMs); (2) Superior cross-modality and cross-anatomy performance: Outperforms baselines in 6 of 9 anatomical regions, and achieves robust results on high-noise modalities (Ultrasound/OCT) and mainstream radiology modalities (MRI/CT); (3) Strong OOD generalization: Outperforms SFT-only baselines by ~10% mDice on OOD datasets, with 48.20% mDice on the most challenging CVC-ClinicDB; (4) Ablation studies verify the effectiveness of each module: Semantic-Guided Prompt Adapter is essential for cross-modal alignment, dual-granularity mask reward solves reward sparsity, and token outperforms whole answer as semantic probe for RL training.

Qualitative Comparison