MLB-Net: A Multi-level Lesion-Aware and Boundary-Enhanced Network for Polyp Segmentation

Accurate automatic polyp segmentation is crucial for the early diagnosis and treatment of colorectal cancer. However, the task remains challenging due to large variations in polyp size, shape, and number, especially in cases involving very small or multiple polyps. To overcome the limitations of existing approaches in cross-level feature interaction and boundary detail modeling, we propose MLB-Net, a Multi-level Lesion-aware and Boundary-enhanced Network for polyp segmentation. MLB-Net first adopts a pyramid vision Transformer to extract multi-scale features, followed by a Multi-level Position and Boundary Fusion (MPDF) module that integrates spatial and boundary information across all feature levels. A Selective Step Feature Aggregation (SSFA) module is then introduced to adaptively filter and fuse adjacent hierarchical features, while a Multi-level Detail Injection (MDI) module is employed to refine structural details and enhance boundary representation in the segmentation map. Extensive experiments on five public datasets demonstrate that MLB-Net consistently achieves superior performance over state-of-the-art methods. In particular, it obtains mDice scores of 92.6% and 94.5% on Kvasir-SEG and CVC-ClinicDB, respectively. Both quantitative and qualitative results confirm its effectiveness in segmenting polyps with diverse appearances, including challenging cases with small-scale or multiple instances. These findings suggest that MLB-Net can substantially improve segmentation accuracy and holds strong potential for clinical application in computer-aided colorectal cancer diagnosis. The source code will be released at https://github.com/cloneiq/MLBNet .