Supervision Meets Self-supervision: A Deep Multitask Network for Colorectal Cancer Histopathological Analysis

Abstract

Colorectal cancer is one of the most common types of cancer worldwide and the leading cause of death due to cancer. As such, an early detection and diagnosis is of paramount importance, which is, however, limited due to insufficient medical practitioners available for large-scale histopathological screening. This demands a reliable computer-aided framework that can automatically analyse histopathological slide images and assist pathologists in quick decision-making. To this end, we propose a novel deep learning framework that combines supervised learning with self-supervision for robust learning of histopathological features from colorectal tissue images. Specifically, our framework comprises a multitask training pipeline using deep metric learning that learns the embedding space using triplet loss, which is augmented using a self-supervised image reconstruction module that enhances learning of pixel-level texture features. The downstream classification is done by extracting features using the pre-trained encoder and feeding them into a support vector machine classifier. We perform qualitative and quantitative analysis on a publicly available colorectal cancer histopathology dataset, as well as compare the proposed framework against some state-of-the-art works, where the model is found to outperform several existing works in literature. The source codes of the proposed method can be found at https://github.com/soumitri2001/DMTL-CRCH.