Comprehensive Study on Semantic Segmentation of Ovarian Tumors from Ultrasound Images

Abstract

Ovarian cancer is one of the most mortal diseases in women. It is commonly detected by medical experts while observing ultrasound images. With the increasing advance in artificial intelligence, deep learning in particular, many medical image analysis have been applied to improve the efficiency in diagnosis and support for training young doctors. This paper introduces a comprehensive study on the segmentation of ovarian tumors from ultrasound images. We investigate state-of-the-art segmentation models such as PSPNet, U-net, DANet, Deeplabv3, and PSANet and evaluate them on a recently published dataset MMOTU. Different from the original works on the MMOTU dataset that just provided binary segmentation, we generate also the label of 8 tumor categories for each pixel. By doing so, it provides not only the size and shape of tumors but also the diseases. Experimental results show that DANet gives the highest accuracy of 71.65% in average. Overall, Chocolate cyst can be accurately segmented with IoU of 96.33% by PSANet.