CNN Architecture-Based Image Retrieval of Colonoscopy Polyp Frames

Manual interpretation and retrieval of colorectal polyps is a time-consuming and laborious task even for specialized medical experts. An automated system can help in information retrieval and timely treatment of polyps. This work comprises of a colonoscopy polyp image retrieval and detection pipeline through the proposed Convolutional Neural Network (CNN) architecture. A binary classification of polyps versus non-polyps has been carried out to retrieve information about polyps in the colonoscopic frames. To check the efficacy of the architecture, test set evaluation, feature mapping, and per epoch analysis of achieved loss and accuracy values have been done. An improved Jaccard index of 83.18% and specificity up to 94.50% have been reported for 33,000 polyp and non-polyp frames generated using publicly available colonoscopic databases. Results infer a maximum of 206 correctly detected polyps out of 215 polyp image frames. The developed architecture has also been compared with state-of-the-art work in this field.