SVF-net: A Dual-encoder Network with Adaptive Gated Fusion for Colorectal Polyp Segmentation

Automatic colorectal polyp segmentation in retrospective still colonoscopy images remains challenging for small, low-contrast, partially occluded, or irregular lesions. We propose SVF-Net, a heterogeneous dual-encoder network combining a frozen SAM2 Hiera-L image trunk with a trainable MaxViT-RMLP-Small specialist branch. At four feature scales, the two streams are projected to a common width, spatially aligned, and reconciled by complementary channel-spatial gating before progressive decoding. Across five public benchmarks, three-run mean mDice is 0.922 on CVC-ClinicDB, 0.898 on Kvasir-SEG, 0.805 on CVC-ColonDB, 0.770 on ETIS-LaribPolypDB, and 0.896 on CVC-300. Matched projected-concatenation and fixed-weight ( $G=0.5$ ) controls retain the surrounding fusion path while removing input-dependent weighting; across ten matched training runs, adaptive gating gives higher mean mDice than both controls on all five datasets. Dataset-wide gate diagnostics on all 380 CVC-ColonDB and 196 ETIS cases indicate non-collapsed, scale- and region-dependent activations, while targeted one-factor sensitivity checks cover the MaxViT learning-rate multiplier, gate initialization, and auxiliary-loss weight. Matched per-case analyses further show that cross-domain gains are heterogeneous across lesion size, brightness, and an image-derived specular-reflection proxy. SVF-Net requires 92.8 ms per image (10.8 FPS) and 8.5 GB peak GPU memory on an NVIDIA RTX 3090 under the reported profiling protocol. The study evaluates retrospective static images only and does not establish real-time colonoscopy performance, prospective clinical effectiveness, or clinician benefit.