Trusted polyp segmentation via multi-scale evidence ensemble in colonoscopy images

Accurate and reliable polyp segmentation is essential for early colorectal cancer detection. Although recent methods employing multi-scale feature aggregation have achieved high accuracy, their reliability in challenging clinical scenarios remains insufficient. Existing uncertainty-aware approaches primarily estimate uncertainty only at the decision layer, failing to exploit multi-scale trustworthiness across the network.

We propose an end-to-end trusted polyp segmentation (TPS) framework based on a multi-scale evidence ensemble. TPS unifies uncertainty estimation and segmentation optimization at the evidence level via Dempster-Shafer theory. Evidence is parameterized with a Dirichlet distribution to explicitly estimate uncertainty at each prediction scale. An uncertainty-embedded attention module is developed to refine polyp delineation using fused multi-scale uncertainty maps. Furthermore, an asymmetric Dice loss is introduced to emphasize error-prone regions dynamically revealed by uncertainty estimates.

Experiments on four public datasets demonstrate consistent performance gains. TPS achieves the best results on CVC-ClinicDB (mDice = 93.32 ± 0.17) and Kvasir (mDice = 91.95 ± 0.34), and substantially improves performance on the challenging CVC-ColonDB and ETIS datasets, achieving the highest mDice across all benchmarks. In particular, TPS produces more calibrated uncertainty estimates, indicating improved reliability in challenging colonoscopy images.

TPS introduces a new paradigm that integrates uncertainty estimation and segmentation optimization at the evidence level. By providing more precise and trustworthy predictions, TPS offers strong potential for safer and more reliable assistance in clinical colonoscopy procedures.