Bladder Semantic Segmentation

Abstract

Obtaining information about the shape and volume of the bladder plays a significant role in determining the pathologies of this organ. To collect the relevant data, the first thing to do is to separate the bladder from the background on the ultrasound image. The article is devoted to automation this process using an algorithm based on the Unet architecture with a pretrained imagenet encoder (encoder – ResNet50). The article gives a comparative analysis of some well-known methods in the literature that improve the accuracy of the proposed algorithm. The quality of the basic architecture has been improved by more than 4 percent on the PR AUC metric (from 84.49% to 89.62%) in the series of experiments with the help of automatic annotation of previously unmarked data. In addition, there are two important results showing practical effectiveness of using the data from another medical task (which raised the accuracy to 88.50%) and using time-dependent sequence of frames inside the video (raised the quality to 88.19%).