CAPB-Net: a lightweight color-adaptive boundary network for medical image segmentation

Abstract

Accurate RGB medical segmentation must combine high overlap, reliable boundaries, low computational cost, and robustness to color and illumination shifts. We present CAPB-Net, a lightweight color-adaptive boundary network that places an image-adaptive color polynomial (IACP) front-end before a compact U-shaped encoder decoder. IACP predicts a per-image residual color mapping in RGB space. Its identity initialization allows the model to learn color corrections only when they improve the segmentation loss. A pyramid-kernel depthwise embedding and selective large-kernel bottleneck then provide efficient local scale diversity and boundary-aware context. Under a fixed protocol, CAPB-Net achieves the highest Dice among the compared models on Kvasir-SEG and ISIC 2018. It reaches 0.8814 and 0.9065 Dice, respectively, with 0.706M parameters and 3.943G FLOPs. A Kvasir-SEG ablation confirms the coupled effect of IACP and color-jitter training and the contribution of selective large-kernel mixing. The results suggest that compact RGB medical segmentation benefits from explicit color adaptation and efficient local context rather than simply enlarging the backbone.