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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Accurate polyp segmentation is crucial for early colorectal cancer detection.
• Current multi-scale methods show high accuracy but lack reliability in complex clinical settings.
• Existing uncertainty-aware methods inadequately leverage multi-scale information.

Purpose of the Study:

• To develop a trusted polyp segmentation (TPS) framework that enhances reliability in polyp detection.
• To integrate uncertainty estimation and segmentation optimization at the evidence level.
• To improve the performance of polyp segmentation in challenging clinical scenarios.

Main Methods:

• Proposed an end-to-end trusted polyp segmentation (TPS) framework using a multi-scale evidence ensemble.
• Unified uncertainty estimation and segmentation optimization at the evidence level via Dempster-Shafer theory.
• Employed a Dirichlet distribution for uncertainty estimation at each scale and an uncertainty-embedded attention module for refinement, alongside an asymmetric Dice loss.

Main Results:

• Achieved state-of-the-art performance on multiple public datasets, including CVC-ClinicDB and Kvasir.
• Demonstrated significant performance improvements on challenging datasets like CVC-ColonDB and ETIS.
• Produced more calibrated uncertainty estimates, enhancing reliability in colonoscopy images.

Conclusions:

• Introduced a novel paradigm for integrating uncertainty estimation and segmentation optimization at the evidence level.
• The TPS framework provides more precise and trustworthy polyp segmentation predictions.
• Offers strong potential for safer and more reliable assistance in clinical colonoscopy procedures.