BDKD-Net: Boundary-Probability Knowledge Distillation for Compact Polyp Segmentation

Accurate polyp segmentation in colonoscopy supports early detection of colorectal cancer, but compact models under a student-only inference budget tend to lose boundary fidelity. BDKD-Net is a 3.72M-parameter compact student trained under a composite knowledge-distillation loss that combines a response signal, a boundary-probability signal restricted to a static teacher-derived boundary band, and an auxiliary detail-feature alignment; the teacher is used only during training and discarded at inference. In a uniform five-seed re-run on a locked Kvasir-SEG and CVC-ClinicDB development split, the same-architecture scratch student reaches Dev Dice 0.9210 ± 0.0029 and boundary F1 at 3-pixel tolerance 0.7726 ± 0.0057, while full BDKD-Net reaches 0.9300 ± 0.0029 and 0.8008 ± 0.0094. Boundary-probability distillation is the load-bearing signal: it has the highest mean Dev BF1t3 among the single KD signals and carries the boundary gain at no inference cost, with the full model reaching four-external Dice 0.8225 ± 0.0071 at 1.76 GFLOPs and 140.9 FPS. Among the directly reproduced baselines, the higher-Dice Polyp-PVT (0.8417 ± 0.0091) needs 6.8× the parameters and 5.7× the compute at roughly half the frame rate. BDKD-Net thus delivers a compact, boundary-faithful student that keeps most of the accuracy of much larger models at a fraction of their inference cost.