基于注意力残差U-Net的皮肤镜图像分割方法  

Dermoscope image segmentation method based on ARB-UNet

作　　者：沈鑫 魏利胜[2] SHEN Xin;WEI Lisheng(School of Electrical Engineering,Anhui Polytechnic University,Wuhu 241000,China;Anhui Key Laboratory of Electric Drive and Control,Anhui Polytechnic University,Wuhu 241000,China)

机构地区：[1]安徽工程大学电气工程学院,安徽芜湖241000 [2]安徽工程大学安徽省电气传动与控制重点实验室,安徽芜湖241000

出　　处：《智能系统学报》2023年第4期699-707,共9页CAAI Transactions on Intelligent Systems

基　　金：安徽省教育厅重大项目(KJ2020ZD39);安徽省检测技术与节能装置重点实验室开放基金项目(DTESD-2020A02)。

摘　　要：针对皮肤镜图像类内差异性、类间相似性、数据集不平衡等问题,本文提出了一种基于注意力残差U-Net(attention residual block-UNet,ARB-UNet)的皮肤镜图像分割方法。将卷积块注意力机制模块(convolutional block attention module,CBAM)引入到U-Net模型的“跳过连接”中;同时将CBAM模块集成到残差模块DRB(dilated residual networks)中得到注意力残差结构(attention residual block,ARB);且选取Focal Tversky Loss作为该模型的损失函数;在ISIC2016数据集上对所提ARB-UNet模型进行训练和测试,并与传统方法和UNet等经典方法进行了对比实验,实验结果中灵敏度(sensitivity,SE)达到了92.9%,特异性(specificity,SP)达到了94.1%,Dice相似指数(dice similarity cofficient,DSC)达到了92.1%,整体上均优于其他对比方法,从而验证了本文方法是有效的和可行的。Aiming at the problems of intra-class difference,inter-class similarity,and dataset imbalance in dermoscopic images,a dermoscopic image segmentation method based on attention residual block-UNet(ARB-UNet)is proposed.Firstly,the convolutional block attention module(CBAM)is introduced into the“skip connection”of U-Net model;at the same time,the CBAM module is integrated into the residual module DRB(dilated residual networks)to obtain the attention residual block(ARB);Focal Tversky loss is selected as the loss function of the model;Finally,the proposed ARB-UNet model is trained and tested on ISIC2016 data set,and compared with traditional methods and classical methods such as U-Net.The experimental results show that the sensitivity(SE),specificity(SP),and dice similarity index(DSC)have reached 92.9%,94.1%,and 92.1%,respectively,which are all better than other comparative methods in overall.Thus,the feasibility and effectiveness of the method in this paper are verified.

关 键 词：图像分割 皮肤镜 卷积神经网络 注意力残差U-Net 注意力机制 卷积块注意力机制模块 深度学习 残差网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]