Introducing a deep neural network structure with practical implementation capability for polyp detection in endoscopic and colonoscopyvideos

In recent years, deep learning has gained much attention in computer assisted minimally invasivesurgery. Researches have shown that the use of deep learning models in colonoscopy can bedivided into four main categories: analysis of surgical images, analysis of surgical operations,evaluation of surgical skills, and surgical automation. Analysis of surgical images can be one ofthe main solutions for early detection of gastrointestinal lesions and taking appropriate actions fortreatment of cancer, which deep learning has shown exceptional performance in this area. Recentstudies show that the high accuracy of lesion detection by these models significantly improves colonoscopyefficiency. Therefore, in this study, a simple and accurate structure of deep neural networksfor polyp detection is introduce. This model can be implemented with low-cost hardwareand provides high-precision polyp detection in real-time. For this purpose, due to the shortage oflabeled colonoscopy images, transfer learning was implemented to extract appropriate featuresfrom the input images. In addition, multi-task learning with two goals of classifying the imagesand detecting the bounding boxes of existing polyps in the images. Considering the appropriateweight for each task in the total cost function is crucial in achieving to the best results. Due tothe lack of datasets with non-polyp images and the need for them to evaluate the performance ofthe proposed structure on both polyp and non-polyp images, data collection was carried out. Theproposed deep neural network structure was trained on KVASIR-SEG and CVC-CLINIC datasetsand tested using cross-validation. Experimental results verify that the proposed structure classifiesthe images into polyp and non-polyp ones with ۱۰۰% accuracy. Moreover, it detects the boundingboxes of the polyps with an accuracy rate of ۸۶%, and processing time of ۰.۰۱ seconds.