Deep Learning-Driven Subtype Classification of Colorectal Cancer Using Histopathology Images for Personalized Medicine

Abstract

Precise and timely diagnosis of colorectal cancer (CRC) is vital for enhancing patient outcomes. Histopathological examination of tissue samples remains the gold standard for CRC diagnosis, but it is a time-consuming and subjective approach tending to inter-observer variability. This study explores the use of deep learning, particularly ResNet architectures, for the automated classification of CRC using histopathology images. Our research focuses on evaluating different ResNet models (ResNet-18, ResNet-34, ResNet-50) to extract relevant visible characteristics. Additionally, we use Grad-CAM heatmaps to understand the focus areas of model, ensuring alignment with established diagnostic criteria. To address limited data availability, we examine data augmentation techniques to improve the adaptability of model. Our analysis indicates that ResNet-34 strikes a balance between model complexity and performance, demonstrating notable accuracies of 91.10%, 99.11%, and 100.00% for overall, top-2, and top-3 accuracy classifications, respectively, outperforming both shallower (ResNet-18) and deeper (ResNet-50) models. This indicates that a moderate depth is practical in capturing features of CRC images. Our findings have significant importance for the development of an interpretable AI-assisted diagnostic tool for CRC, with the potential to improve the efficiency and accuracy of pathologists. This approach aims to automate image analysis, offer insights into model decisions, and ultimately improve diagnostic consistency and patient care in oncology.