Real-time Colonoscopy Image Segmentation Based on Ensemble Knowledge Distillation

Colonoscopy is an important means of detecting various intestinal diseases such as bleeding, polyps, Merck diverticula, and ulcers. The sooner these diseases are detected, the better the patient's recovery. But colonoscopy is a demanding process, often leads to the high rate of misdiagnosis by experts, professional physicians and nurses and costs a lot of time. Therefore, robot-assisted colonoscopy is considered as an important method to solve this problem. In recent years, many automated deep learning models for colonoscopy have been proposed. However, these models are usually large and time-consuming, and cannot meet actual needs. Besides, due to the disconnection of data between hospitals, the strength of medical resources between different departments in different hospitals is different, so the general multi-classification model cannot fit the characteristics of such data distribution. Therefore, in this article, we ensemble multiple binary classification models (each model detects one disease) and extracted a compression model using knowledge distillation technology, which can simultaneously detect polyps, Merkel diverticula, ulcers and bleeding from colonoscopy. We tested the performance of our model on public and real data sets and found that the model can achieve acceptable results and help doctors make decisions in practice.