Addressing priority challenges in the detection and assessment of colorectal polyps from capsule endoscopy and colonoscopy in colorectal cancer screening using machine learning.

Background: Colorectal capsule endoscopy (CCE) is a potentially valuable patient-friendly technique for colorectal cancer screening in large populations. Before it can be widely applied, significant research priorities need to be addressed. We present two innovative data science algorithms which can considerably improve acquisition and analysis of relevant data on colorectal polyps obtained from capsule endoscopy. Material and methods: A fully paired study was performed (2015–2016), where 255 participants from the Danish national screening program had CCE, colonoscopy, and histopathology of all detected polyps. We developed: (1) a new algorithm to match CCE and colonoscopy polyps, based on objective measures of similarity between polyps, and (2) a deep convolutional neural network (CNN) for autonomous detection and localization of colorectal polyps in colon capsule endoscopy. Results and conclusion: Unlike previous matching methods, our matching algorithm is able to objectively quantify the similarity between CCE and colonoscopy polyps based on their size, morphology and location, and provides a one-to-one unequivocal match between CCE and colonoscopy polyps. Compared to previous methods, the autonomous detection algorithm showed unprecedented high accuracy (96.4%), sensitivity (97.1%) and specificity (93.3%), calculated in respect to the number of polyps detected by trained nurses and gastroenterologists after visualizing frame-by-frame the CCE videos. [ABSTRACT FROM AUTHOR]