Adaptive kernel fusion network with a boundary-aware decoder for generalizable colorectal polyp segmentation

초록

Accurate and generalizable polyp segmentation is a critical prerequisite for computer-aided colorectal cancer screening; however, it remains challenging due to small or low-contrast lesions, large appearance variability, and domain shift across clinical centers. To address these challenges, we propose the Adaptive Kernel Fusion network (AKF-Net) with a boundary-aware decoder for robust polyp segmentation. The AKF-Net adopts a pyramidstructured encoder to capture multi-scale features and incorporates three complementary context modeling modules at shallow, intermediate, and deep feature levels. The Context-Aware High Frequency Refinement (CAHFR) module enhances fine-grained edge and texture details, the Multi-Kernel Dynamic Routing Block (MKDRB) adaptively selects receptive fields to accommodate diverse polyp sizes and shapes, and the Top-Global Context (TGC) block improves robustness to illumination and contrast variation. In addition, a Region Hybrid Boundary Attention (RHBA) mechanism is integrated into a cross-scale decoder to model boundary cues and mitigate segmentation ambiguity explicitly. Extensive experiments on five public colonoscopy benchmarks demonstrate strong in-domain performance. In the single-run evaluation, AKF-Net achieves Dice scores of 94.0% and 92.3% on CVC-ClinicDB and Kvasir-SEG, respectively. Under cross-dataset evaluation, it achieves Dice scores of 84.5%, 81.5%, and 81.7% on BKAI-IGH, CVC-ColonDB, and ETIS, respectively. Additional evaluations on skin lesions, foot ulcers, and breast ultrasound lesion segmentation indicate the potential applicability of AKF-Net to non-colonoscopy medical image segmentation tasks, while more extensive validation across larger and more diverse datasets remains necessary. AKF-Net improves pixel-level segmentation by combining multi-scale context with boundary-aware fusion, supporting robust performance across challenging polyp datasets.

키워드