Localization and Semantic Segmentation of Polyp in an Effort of Early Diagnosis of Colorectal Cancer from Wireless Capsule Endoscopy Images

Cancer is characterized by the fast growth of aberrant cells that affect adjacent tissues. Colorectal cancer could be diagnosed in the early stage by identifying the predecessor polyp that are initially innocuous. Endoscopy aids in the identification of polyp in real time monitoring. Polyps are obscured by the mucosa that surrounds it in the lumen of the colon, making visual differentiation from the mucosa difficult for the physicians thereby increasing the miss rates. Recognizing the colorectal polyps is challenging as it varies widely in characteristics representing its features. With the emergence of deep learning techniques especially the convolutional neural network an effort was made to detect and segment the colorectal polyps. U-net architecture with the capability to learn deep features from the images is proposed in our paper for the semantic segmentation where polyps in the colon images are localized. Polyp images from Kvasir dataset obtained using wireless capsule endoscopy that provides the advantage of viewing the entire gastrointestinal tract was used in this paper. The deep learning framework was evaluated for assessing its performance and was optimized. The predicted results produced an accuracy, precision, sensitivity (recall), IoU and f1 score of 93.14%, 98.08%, 95.55%, 95.75% and 98.01% respectively.