皮肤癌病变组织分割的多通道并行U型网络  

Multi-Channel Parallel U-Shaped Network for Lesion Segmentation

作　　者：高钟宇 禹龙[2] 田生伟[1] 吴卫东[3] 张德志[3] GAO Zhongyu;YU Long;TIAN Shengwei;WU Weidong;ZHANG Dezhi(School of Software,Xinjiang University,Urumqi Xinjiang 830091,China;Network Center,Xinjiang University,Urumqi Xinjiang 830046,China;People’s Hospital of Xinjiang Uygur Autonomous Region,Urumqi Xinjiang 830000,China)

机构地区：[1]新疆大学软件学院,新疆乌鲁木齐830091 [2]新疆大学网络中心,新疆乌鲁木齐830046 [3]新疆维吾尔自治区人民医院,新疆乌鲁木齐830000

出　　处：《新疆大学学报（自然科学版）（中英文）》2022年第6期707-719,共13页Journal of Xinjiang University(Natural Science Edition in Chinese and English)

基　　金：新疆维吾尔自治区重点研究与发展项目(2021B03001-4)。

摘　　要：医学图像分割已经成为辅助诊断当中重要的一环.受困于单通道模型特征提取能力的限制,网络所能获取的信息总量有限,导致性能无法进一步提升.针对信息数量不足的问题,提出了一种多通道模型.与单通道模型相比,多通道模型提供了更多互补的特征信息,有助于更好地进行特征提取与数据表达.结果如下:(1)设计了动态卷积发散模块(DSC BM),用于构建多通道模型.(2)设计了动态卷积集束模块(DSC AM),用于融合多尺度特征.(3)使用动态卷积发散模块与动态卷积集束模块构建多通道并行U型网络(MCPU-Net).所提出的方法在国际公开数据集ISIC2017进行训练和评估,MCPU-Net的总体Acc指标为0.933,JI指标为0.772.Medical image segmentation has become an important part of assisted diagnosis.Trapped by the limitations of the single channel model feature extraction capability,the total amount of information that the network can acquire is limited,resulting in no further performance improvement.In response to the insufficient amount of information,this paper proposes a multi-channel model for processing unimodal data.Compared with the single-channel model,the multi-channel model brings more complementary feature information,which helps better data representation and feature extraction.The results are as follows:(1)The dynamic selective kernel module branch model(DSC BM)is designed to construct a multi-channel model.(2)The dynamic selective kernel module aggregation module(DSC AM)is designed for fusing multi-scale features.(3)A multi-channel parallel U-shape network(MCPU-Net)is constructed.The overall Acc metric of MCPU-Net was 0.933,and the J I metric was 0.772 by training and evaluation in the international public dataset ISIC2017.

关 键 词：多通道模型 皮肤病分割 深度学习 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]