Experimental Deep Learning Object Detection in Real-time Colonoscopies

Abstract

Real-time automatic analysis of colonoscopies are important for learning and in practice. In our experiment of object detection we used a Mobilenet Deep Learning model retrained on a Jetson Xavier NX microsystem produced by NVIDIA. The training set of colonoscopy frames was build and annotated from our database of video colonoscopies. The retrained Mobilenet network model successfully detects 10 classes of objects during the playback of real-time video colonoscopies. Such a small and powerful microsystem can be improved to become a tool for students training.