Swin-MSEG-Pro: Uncertainty-aware boundary-focused swin transformer with deep supervision for reliable gastrointestinal polyp segmentation
Year: 2026
Type: article
Abstract: Automatic segmentation of gastrointestinal polyps from colonoscopy images is clinically important because computer-aided systems must identify suspicious lesions and delineate their extent under challenging visual conditions. Variations in polyp size, shape, texture, and contrast, together with weak boundaries, illumination changes, artefacts, and cross-dataset distribution shift, complicate this task. This study introduces Swin-MSEG-Pro, a reliability-focused segmentation framework that integrates a SwinV2-based encoder with multi-scale feature learning, deep supervision, evidential uncertainty learning, and boundary-focused supervision. The model was developed using Kvasir-SEG with 800 training, 100 validation, and 100 held-out test images; ETIS-LaribPolypDB and CVC-ClinicDB were reserved exclusively for external testing. Across three independent random seeds, Swin-MSEG-Pro achieved a validation Dice of 0.9289 ± 0.0064 and mean per-image Dice scores of 0.9153 ± 0.0024, 0.7174 ± 0.0016, and 0.8527 ± 0.0020 on the held-out Kvasir-SEG, ETIS-LaribPolypDB, and CVC-ClinicDB test sets, respectively. Evidential uncertainty remained informative for selective risk ranking, with TTA object-level RC-AURC values of 0.0487 ± 0.0045, 0.1027 ± 0.0195, and 0.0769 ± 0.0034 on the corresponding datasets. Component-wise ablation showed the clearest degradation from removing the boundary mechanism on ETIS-LaribPolypDB, where Dice decreased by 0.0188. Swin-MSEG-Pro therefore provides reproducible uncertainty-aware polyp segmentation and useful risk information under cross-dataset shift, although performance declines on more strongly shifted external data.
Source: Array
Language: English
Subfield: Oncology
Field: Medicine
Domain: Health Sciences
Open Access status: gold
Estimated APC paid: $1,680
OpenAlex ID: W7214101682