Abstract:To address the challenges of significant anatomical variability, class imbalance, and fuzzy boundaries in 3D medical image segmentation, this study proposes a novel method based on label-aware encoding and group-specific feature enhancement (i.e., differentiated feature strengthening for different anatomical categories). A unified multi-module segmentation framework is constructed, which organically integrates dynamic label semantic modeling with class-specific feature optimization. First, in the feature extraction stage, a dynamic clustering and evidence accumulation mechanism is introduced. This mechanism computes stable class center embedding through iterative clustering, thereby achieving adaptive semantic label modeling and improving the feature representation of small or low-contrast organs. Furthermore, a group-specific feature enhancement and disentanglement module is designed. Leveraging semantic label embedding, this module generates spatial and channel-wise weight maps to explicitly separate and recalibrate features of different categories, significantly improving the model’s ability to discriminate fine-grained structures. The overall network employs a Transformer-based encoder for global context modeling and a progressive decoder to fuse multi-scale features for high-resolution structural recovery. Extensive experiments conducted on the public WORD and BTCV datasets demonstrate that the proposed method achieves superior performance in multi-organ segmentation tasks compared with existing mainstream models, showing significant advantages in terms of Dice similarity coefficient (DSC) and 95% Hausdorff distance (HD95).