Exploiting Multiple Color Representations to Improve Colon Cancer Detection in Whole Slide H&E Stains

Abstract

Currently, colon cancer diagnosis is based on manual assessment of tissue samples stained with hematoxylin and eosin (H&E). This is a high volume, time consuming, and subjective task which could be aided by automatic cancer detection. We propose an algorithm for automatic cancer detection within WSI H&E stains using a multi class colon tissue classifier based on features extracted from 5 different color representations. Approx. 32000 tissue patches were extracted for the classifier from manual annotations of 9 representative colon tissue types from 74 WSI H&E stains. Colon tissue classifiers based on gray level or color features were trained using leave-one-out forward selection. The best colon tissue classifier was based on color texture features obtaining an average tissue precision-recall (PR) area under the curve (AUC) of 0.886 and a cancer PR-AUC of 0.950 on 20 validation WSI H&E stains.