Real-Time Polyp Detection, Localization and Segmentation in Colonoscopy Using Deep Learning

Computer-aided detection, localisation, and segmentation methods can helpimprove colonoscopy procedures. Even though many methods have been built totackle automatic detection and segmentation of polyps, benchmarking ofstate-of-the-art methods still remains an open problem. This is due to theincreasing number of researched computer vision methods that can be applied topolyp datasets. Benchmarking of novel methods can provide a direction to thedevelopment of automated polyp detection and segmentation tasks. Furthermore,it ensures that the produced results in the community are reproducible andprovide a fair comparison of developed methods. In this paper, we benchmarkseveral recent state-of-the-art methods using Kvasir-SEG, an open-accessdataset of colonoscopy images for polyp detection, localisation, andsegmentation evaluating both method accuracy and speed. Whilst, most methods inliterature have competitive performance over accuracy, we show that theproposed ColonSegNet achieved a better trade-off between an average precisionof 0.8000 and mean IoU of 0.8100, and the fastest speed of 180 frames persecond for the detection and localisation task. Likewise, the proposedColonSegNet achieved a competitive dice coefficient of 0.8206 and the bestaverage speed of 182.38 frames per second for the segmentation task. Ourcomprehensive comparison with various state-of-the-art methods reveals theimportance of benchmarking the deep learning methods for automated real-timepolyp identification and delineations that can potentially transform currentclinical practices and minimise miss-detection rates.