基于改进UNet的皮肤镜图像分割

Image Segmentation of Dermoscopy Based on Improved UNet

医学图像的自动分割对于计算机辅助诊断具有重要意义。针对皮肤镜图像分割任务中,本文提出一种RA-UNet(Residual and Attentional-UNet)的分割方法。将原本的网络结构加深,在编码端卷积层引入残差机制减少信息丢失同时防止梯度消失或爆炸,每层采用两个残差特征提取模块充分提取学习特征;同时,每层引入改进的卷积注意力模块(Convolutional block attention block,CBAM)使模型更好地学习皮肤病理区域特征。在ISIC 2018数据集上对所提出的RA-UNet模型进行训练和测试,并与UNet和其他模型进行对比试验,实验结果中准确率(ACC)达到了93.82%,特异度(SP)达到了95.26%,灵敏度(SE)达到了90.78%,精准度(P)达到了90.04%,平均交并比(Miou)达到了86.89%,Dice相似指数(DSC)达到了0.900,整体优于其它模型。对于提高基于图像分析的皮肤病诊断具有一定的参考应用价值。

Automatic segmentation of medical images is of great significance to computer aided diagnosis.In this paper,a RA‑UNet(Residual and Attentional UNet)method is proposed for dermoscopic image segmentation.The original network structure is deepened,and the residual mechanism is introduced into the convolutional layer at the coding end to reduce information loss and prevent gradient disappearance or explosion.Two residual feature extraction modules are used in each layer to fully extract learning features.At the same time,an improved Convolutional block attention block(CBAM)was introduced in each layer to enable the model to better learn the features of skin pathological regions.The proposed RA‑UNet model was trained and tested on the ISIC 2018 dataset,and compared with UNet and other models,the accuracy(ACC)reached 93.82%,the specificity(SP)reached 95.26%,and the sensitivity(SE)reached 90.78%.The accuracy(P)reached 90.04%,the mean crossover ratio(Miou)reached 86.89%,and the Dice similarity index(DSC)reached 0.900,which were superior to other models.It has a certain reference value for improving the diagnosis of skin diseases based on image analysis.