Segmentation of Ultrasound Brachial Plexus Based on U-Net

Brachial plexus block anesthesia (PNB) is one of the anesthesia methods commonly used by anesthesiologists in surgical operations. Anesthesiologists use ultrasonic equipment to detect and locate brachial plexus nerves in ultrasonic imaging, and then perform local anesthesia on the nerves. Therefore, it is particularly important to accurately locate and segment brachial plexus nerves in ultrasonic images. For the basic task of ultrasonic brachial plexus nerve segmentation, this paper takes deep learning neural network U-NET as the basic structure and improves it from different perspectives. At the same time, the recurrent neural network (RNN), which is applied to solve natural language processing and sequence problems, is added to the image segmentation. It enhances the ability of obtaining the context information of the original U-NET. Finally, in order to solve the imbalance of positive and negative samples of the brachial plexus in ultrasound, the auxiliary loss, used to judge the presence of the brachial plexus, was introduced in this paper to improve the accuracy of the model. In the ultrasound nerve segmentation data set published by Kaggle, the improved U-NET method in this paper is 0.06 higher than the original method, and the DICE coefficient got 0.6490.