SCA-Net: A Scale- and Contrast-Aware Network for Subtle and Low-Contrast Polyp Segmentation.

Accurate polyp segmentation is critical for early detection of colorectal cancer. However, existing methods often struggle with subtle polyps, weak boundaries, and poor cross-dataset generalization. To address these challenges, we propose SCA-Net, a scale- and contrast-aware network designed to improve semantic representation, scale adaptability, and boundary sensitivity within a unified encoder-decoder framework. Specifically, we propose a semantic module group (SMG) consisting of a cross-scale global aggregator (CSGA) and gated semantic injection (GSI) to enable effective cross-scale semantic aggregation and selective semantic propagation. To further handle large-scale variations of polyps, we design a size-adaptive dynamic router (SADR) that dynamically adjusts receptive fields according to object size. In addition, we propose a Laplacian-guided synergistic refiner (LGSR), which leverages Laplacian-derived high-frequency priors to enhance boundary-aware feature refinement. Across five benchmark datasets, SCA-Net remains competitive on seen datasets and shows clearer gains on challenging unseen benchmarks, particularly ETIS-LaribPolypDB, where it achieves 86.0% Dice with a PVTv2-B4 backbone.