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.
作者
沈鑫
魏利胜
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)
出处
《智能系统学报》
CSCD
北大核心
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)。
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
image segmentation
dermoscopic
convolutional neural network
attention residual block-UNet(ARBUNet)
attention mechanism
convolutional block attention module(CBAM)
deep learning
residual network
