Assessment of a deep-learning system for colorectal cancer diagnosis using histopathology images

Abstract

Abstract Colorectal Cancer is the one of the most common forms of cancer hence, an early and accurate detection is crucial. Manual diagnosis is a tedious and time-consuming job which is prone to human errors as it involves visual examinations of pathological images. Therefore, it is imperative to use computer-aided detection (CAD) systems to interpret the medical images for a quicker and more accurate diagnosis. The traditional methods for diagnosis comprise extraction of features based on texture, pattern, illumination etc. from pathological images and then use these features in a Machine Learning model for binary classification i.e., cancerous, or non-cancerous. Deep-learning approaches like the Convolutional neural networks (CNNs) have proved to be very effective in classifying and predicting cancer from pathological images. In this study, we have assessed several CNN-based techniques for cancer diagnosis on digitized histopathology images. We have also compared the results of traditional methods for diagnosis with the deep-learning models. Moreover, we have proposed a new model by borrowing the idea from Xception architecture (Xception+), which outperforms the existing architectures. Furthermore, we have studied the effect of transfer learning technique by using models pre-trained on unrelated histopathology images.