Automated Polyp Segmentation in Wireless Capsule Endoscopy Images Using Modified U-Net Architecture to Diagnose Colorectal Cancer

The timely identification of colorectal cancer (CRC) through the characterization of polyp in wireless capsule endoscopy (WCE) images is crucial for mitigating the progressive nature of the disease. The miss-rates of the polyp are higher when the WCE images are examined by the clinician as the polyp may sometimes blend with the background pixels. Furthermore, the cognitive state of the clinician is crucial in accurately identifying polyps. The process of automatically segmenting polyps is greatly enhanced through the utilization of the Modified U-Net architecture, which builds upon the foundational U-Net framework by integrating residual blocks. We utilized the Kvasir dataset for training and validation of the Modified U-net and evaluated the performance metrics. The hyperparameters were tuned to achieve optimal performance. We achieved an accuracy of 95.43%, mIoU of 0.9298 and mF1 of 0.9642. The deployment of Modified U-net to localize the polyp in medical imaging has great promise, since it could improve the effectiveness of early CRC detection.