Automatic endometrial segmentation in ultrasound images using deep learning

Abstract

Endometrial segmentation plays a vital role in the computerized evaluation of uterine ultrasonic images. Accurate segmentation of endometrial regions may improve the accuracy and efficiency of diagnosis. Recent studies have been focused on the employment of deep learning in medical image segmentation. In this study, we compared six models, including five convolutional neural networks with different network architectures (UNet, Segnet) and backbones (Resnet50, Vanilla CNN, VGG16) for the segmentation of endometrium, and one model called deep dual-resolution networks (DDRNets). The training and test datasets were composed of 840 and 210 images from 302 and 68 cases, respectively. Through validation, DRRNets demonstrated the best performance for endometrial segmentation with an average Dice coefficient (DSC) of 0.895.