JMIR Publications

ABSTRACT

Background:

Comparisons of decoder modules for automated polyp segmentation are complicated by differences in encoder backbones, training protocols, and data splits. This study evaluates decoder modules under matched conditions and investigates a decoder combining semantic feature decoupling with boundary-guided attention.

Objective:

To evaluate decoder modules for colorectal polyp segmentation under matched conditions and develop a semantic decoupling and boundary-guided attention framework.

Methods:

Ten decoder modules were compared with a plain convolutional baseline using a fixed Twins-PCPVT-Small encoder and matched training settings. We developed ConBPoly, combining a Semantic Information Decoupling Module with Contrastive Boundary-Guided Attention, and evaluated architectural components and auxiliary loss configurations. Training used 1,450 images from Kvasir-SEG and CVC-ClinicDB; evaluation used 738 images across four test sets. Analyses included per-image paired Wilcoxon tests, bootstrap confidence intervals, and three-seed validation of the proposed configurations.

Results:

Three of ten modules exceeded the heuristic Kvasir-SEG screening threshold of Dice > 0.9142. ConBPoly-min achieved a Kvasir-SEG Dice of 0.9246 in the primary run and 0.9203 ± 0.0032 across three controlled training seeds. Its primary-run Kvasir-SEG Dice exceeded those of SSFormer (0.9197) and Polyp-PVT (0.9190). However, Polyp-PVT achieved a higher four-dataset mean Dice (0.8596 versus 0.8490). Removing auxiliary objectives from the original ConBPoly loss configuration increased Kvasir-SEG Dice by 0.0065, while the benefits of loss simplification varied across datasets.

Conclusions:

Under the evaluated backbone and training protocol, decoder module benefits were dataset-dependent. ConBPoly combined competitive Kvasir-SEG segmentation with a simplified training objective, but its cross-domain performance remained below Polyp-PVT. These findings support controlled evaluation of decoder design and further investigation of semantic decoupling and boundary-guided refinement. Validation across additional encoders and clinical video data is needed.