This paper benchmarks several deep learning models for real-time polyp detection, localization, and segmentation in colonoscopy videos and introduces ColonSegNet, a novel lightweight architecture. ColonSegNet achieves competitive performance compared to state-of-the-art methods while being significantly faster, processing up to 182 frames per second.

The work of Debesh Jha is funded by the Research Council of Norway. Parts of the computational resources were used from research supported by the National Institute for Health Research (NIHR) Oxford BRC and the Wellcome Trust. Sharib Ali is supported by the NIHR Oxford Biomedical Research Centre. While potential COIs exist due to funding sources, the authors explicitly state that the views expressed are their own and not necessarily those of the funding bodies.

This paper presents a strong contribution to the field of automated polyp detection and segmentation by benchmarking various deep learning methods and introducing a novel architecture, ColonSegNet, that achieves real-time performance. The extensive experiments and comparison on the Kvasir-SEG dataset provide valuable insights. Despite the limitations of a retrospective study design and the potential for further hyperparameter optimization, the paper demonstrates the potential for significant improvements in colonoscopy procedures.