Automated Detection of Colorectal Polyp Utilizing Deep Learning Methods With Explainable AI

Detecting colorectal polyps promptly and accurately is crucial in preventing the progression of colorectal cancer. These polyps cause severe conditions in the colon or rectum, presenting a significant diagnostic challenge. Traditional manual detection through medical imaging is not only bulky and prone to errors but also incurs substantial costs, requiring expert endoscopist. Inefficient detection and treatment can lead to critical health complications. Addressing these issues, we extensively employed various configurations of the state-of-the-art YOLOv8 (n-nano, s-small, m-medium, l-large, and x-extra-large) models for effective polyp localization. Complementing this, we proposed a novel TR-SE-Net model for segmentation, integrating Squeeze-and-Excite Networks (SE-Net) to elevate performance and real-time processing capabilities. The Kvasir-SEG dataset is utilized for training and testing models, supplemented by external validation CVC-ClinicDB, PolypGen, ETIS-LaribPolypDB, EDD 2020, and BKAI-IGH to confirm their efficacy in processing unseen, real-time data. This study delves into the interpretability of these models using explainable AI (XAI), such as eigen visualization for localization and heatmap analysis for segmentation. This exploration provides deeper insights into the decision-making processes of the models, thereby enhancing their reliability. Notably, the YOLOv8m model showcased remarkable prediction speed (approximately 16.61 ms) and excelled in precision (0.946), recall (0.771), F1-score (0.85), mAP50 (0.886), and mAP50–95 (0.695), catering to diverse clinical scenarios. The TR-SE-Net demonstrated significant improvements in segmentation performances, including DSC (0.8754), F2-score (0.8786), precision (0.9027), recall (0.8879), accuracy (0.9647), competitive mIoU (0.7961), FPS (54), parameters (27.27 million), and flops (10.59 GMac). Furthermore, A graphical Computer Aided Diagnosis (CAD) system developed utilizing both models can substantially reduce the miss rate because segmentation will assist in polyp detection or vice versa if localization fails. Conclusively, integrating these advanced computer-aided methods substantially enhances colonoscopy procedures by mitigating the risks of colorectal cancer.