Automated Colon Cancer Analysis in histopathological Images Using Deep Learning Techniques

Abstract

Colorectal cancer (CRC) is the most commonly occurring malignancy of the digestive system that has a significantly high mortality rate. A timely precise diagnosis can save millions of lives and lead to a better society worldwide. Hence, the presented thesis depicts clinically comparable computerized deep learning-based models beneficial in reducing time, abating unintentional human error, and boosting the precision of pathologists. These models are powered by enhancing feature map learning of convolutional layers, incorporating attention learning in the skip connections/feedback path, and infusion of multi-scalar feature maps in deep learning-based network architectures. The Hematoxylin and Eosin (HandE)-stained histopathology images are complex and inconsistent in nature. Hence, first the pre-processing step is applied to generate a real-life clinical analogous training dataset for all the proposed methodologies. newlineThe pixel-level segmentation of one of the important Region-of-Interest objects i.e., glandular morphology is done using the modified U-Net architecture incorporated with the described amendments. These segmented glandular morphometrics assist pathologists in accurate quantification of cancer spread. newlineThe image-level classification is done by infusing structural enhancement in the conventional convolutional neural network (CNN) through multi-level factorized convolutional and attention (elemental, channel-wise, and scale-wise) learning. It leads to the effectual capturing of distinguished feature maps to quantify the degree of cancer grades and tissue structure analysis done by the same network. newlineFurthermore, for the complete multitasking model, the amalgamation of U-Net and CNN architecture is done by integrating hybrid convolutional and attention learning. This will assist pathologists in devising precise colon cancer diagnosis by providing glandular morphometric and cancer grade quantification information. newlineThe proposed approaches are implemented using five benchmark datasets (GlaS, CRAG, LC-25000, Ka