基于改进U-Net网络的多囊卵巢图像分割方法  

Image Segmentation of Polycystic Ovary Based on Improved U-Net Network

作　　者：巫笠平 段晓鹏 马玉良[1,2] 张建海 Wu Liping;Duan Xiaopeng;Ma Yuliang;Zhang Jianhai(Institute of Intelligent Control and Robotics,Hangzhou Dianzi University,Hangzhou 310018,China;Key Laboratory of Brain Machine Collaborative Intelligence of Zhejiang Province,Hangzhou 310018,China;School of Computer,Hangzhou Dianzi University,Hangzhou 310018,China)

机构地区：[1]杭州电子科技大学智能控制与机器人研究所,杭州310018 [2]浙江省脑机协同智能重点实验室,杭州310018 [3]杭州电子科技大学计算机学院,杭州310018

出　　处：《中国生物医学工程学报》2022年第6期663-671,共9页Chinese Journal of Biomedical Engineering

基　　金：国家自然科学基金(62071161);浙江省重点研发计划项目(2020C04009)。

摘　　要：多嚢卵巢综合征(PCOS)是一种严重危害女性健康的疾病。针对PCOS超声图像存在目标区域对比度低,背景噪声多等问题,提出一种基于改进U-Net网络的多囊卵巢图像分割方法。首先,对PCOS图像进行预处理以减少斑点噪声和阴影的影响;然后,通过八度卷积模块降低冗余低频特征图并进行特征融合,并采用分层残差跳连模块弥补U-Net编码器与解码器之间的语义鸿沟;接着,使用PCOS超声图像数据集进行实验;最后,使用包含2594张皮肤病变图像的公开数据集ISIC2018进行验证实验。所提方法在PCOS超声图像数据集上达到了88.42%分割精度,相比于U-Net提升了4.24%;并在ISIC2018数据集上实现了97.5%的分割精度。实验结果表明,所提方法不仅在多囊卵巢囊泡的分割上有所提升,在鲁棒性方面也有较好的表现,在其他医学图像分割领域有一定的参考价值。Polycystic ovary syndrome(PCOS)is a disease that seriously endangers women′s health.Aiming to solve the problem of low contrast in the targeted area and high background noise in PCOS ultrasound images,a segmentation method of polycystic ovary images based on improved U-Net network was proposed in this paper.Firstly,the PCOS images were preprocessed to reduce the influence of speckle noise and shadows;then,redundant low-frequency feature maps were reduced by octave convolution module and feature fusion is performed;then,the hierarchical residual skip connection module was used to compensate for U-Net semantic gap between encoder and decoder;secondly,experiments were performed using PCOS ultrasound image dataset;finally,validation experiments were performed using ISIC2018,a public dataset containing 2594 skin lesion images.The proposed method achieved a segmentation accuracy of 88.42%on the PCOS ultrasound image dataset,which was 4.24%higher than that of U-Net;and achieved a segmentation accuracy of 97.5%on the ISIC2018 dataset.The experimental results showed that the proposed method not only improved the segmentation of polycystic ovarian vesicles,but also had better performance in terms of robustness,which could also be referred to other medical image segmentation fields.

关 键 词：图像分割 超声图像 U-Net 八度卷积 残差结构 

分 类 号：R318[医药卫生—生物医学工程]