Real time Automatic Polyp Detection in White light Endoscopy videos using a combination of YOLO and DeepSORT

Abstract

Colorectal cancer is a form of cancer that has its incidence all over the globe. It starts in the colon as a group of cells growing on the inner wall; these are called polyps. Not all polyps are cancerous, but they should be identified and removed. The detection of polyps during colonoscopy may be susceptible to human errors. Missed polyps due to human errors can lead to colorectal cancer. Advancements in the field of artificial intelligence brought revolutionary changes in several fields. A computerized algorithm that guides doctors can be a better option for reducing human error. For this purpose we have implemented a tracking by detection model which helps doctors during screening process. For detection of polyps we have trained our detection algorithm using YOLO-v4. For training we have used 1705 polyp images taken from various databases. For tracking polyps we have implemented DeepSORT algorithm. To evaluate the model, we have tested it on 2 colonoscopy videos acquired from hospitals. Performance of the model on these two videos is evaluated by computing two metrics Multiple Object Tracking Accuracy(MOTA) and Multiple Object Tracking Precision(MOTP). Our model is able to track polyps and promising results were obtained.