Identification of Key Areas in Histology Images for Detection of Collagenous Colitis: A Deep Learning Approach

Abstract

Collagenous colitis (CC) is an inflammatory disease of the large bowel that causes chronic watery diarrhoea, abdominal pain, faecal incontinence, nightly defecation, and weight loss, resulting in significantly impaired quality of life. The diagnosis of CC is rather challenging as it can only be diagnosed upon histological examination of colonic biopsies taken from normal or near normal appearing mucosa. Current routine histological interpretation of biopsies involves subjective evaluation leading to inter-rater variability discrepancies in diagnosis and treatment plan. The aim of this study was to develop an algorithm for robust segmentation of light microscopy images of histological specimen slides identifying key areas containing important diagnostic features of CC. Images of histological specimens from 10 patients (~60 images per patient) were pre-segment-ed into superpixels using a simple linear iterative clustering algorithm. The areas containing candidate diagnostic features for identification of CC, in particular, the thickened subepithelial collagen layer – the essen-tial diagnostic feature, were marked by the experts. The feed – forward neural network containing three hidden layers with ten neurons in each was trained to identify the superpixels containing sought diagnostic fea-tures. The model was tested on 250 images from 5 patients not used for training and showed accuracy of 0.807, sensitivity – 0.801 and specificity – 0.813. The shown neural network’s ability to segment histology images could be used for assisted diagnostic process emphasising areas with candidate key features for identification of collagenous colitis.