Cleaning assessment in endoscopic esophageal images using U-Net and a classification model

摘要

The presence of water bubbles, foam, mucus or residual food in the gastrointestinal tract is a common problem during Upper gastrointestinal endoscopy (UGIE). This condition could cause difficulties to detect lesions as well as perform interventions if required for medical doctors. In every medical center all over the world, analyzing endoscopic images with bubbles is becoming a vital practical issue. In this paper, we propose an automatic scheme to segment regions of water bubbles thanks to the recent advantages of a deep neural network - U-net. Based on the segmentation results, we construct a classification model to evaluate the cleaning level of the current examined images. The classification model utilizes features which directly extracted from the segmented areas such as the number of the water bubbles, their concentrating distribution or scattering in the current view. The proposed method is evaluated on a testing dataset validated by the endoscopists. Accuracy of the quality assessment by the proposed techniques achieves the rate of 90%. Therefore, the proposed method presents a feasible tool for automatically eliminating of unclean UGIE images.