Comparative Analysis of ResNet Architecture Enhanced with Self-Attention for Colorectal Cancer Detection

Colorectal cancer is a malignancy that arises in the colon or rectum, which is the key component of the large intestine and terminal part of the digestive system, and typically originates from benign cell clusters called polyps. Over time, these polyps may progress into cancer. Given the prevalence of this condition, early detection is crucial for achieving favourable treatment outcomes. Conventional diagnostic approaches, such as colonoscopy, often fail to identify early-stage lesions due to their subtle presentation. However, advancements in machine learning, particularly deep learning, offer promising avenues for improving diagnostic accuracy. This study provides a detailed evaluation of the ResNet deep learning model, specifically focusing on its performance in distinguishing classes within the CRC-5000 dataset. This dataset comprises 5,000 colonoscopy images categorized into eight distinct classes: Tumor, Stroma, Complex, Lympho, Debris, Mucosa, Adipose, and Empty. The research compares the performance of ResNet architectures (ResNet-50, ResNet-101, and ResNet-152) in two configurations: with and without the addition of Self-Attention layers. The results are significant, showing that ResNet-50 with Self-Attention achieved an accuracy of 92.13%, ResNet-101 with Self-Attention achieved an accuracy of 92.94%, and ResNet-152 with Self-Attention achieved an accuracy of 92.74%, while without Self-Attention ResNet-50, ResNet-101, and ResNet-152 demonstrated a lower accuracy of 89.51%. These findings highlight the considerable potential of enhancing ResNet with Self-Attention mechanisms to improve the early detection and treatment of colorectal cancer. Furthermore, this approach could significantly streamline clinical workflows, providing pathologists with robust tools for early diagnosis.