基于多头自注意力机制与U-Net的脑卒中DWI图像梗死病灶分割方法研究  

Infarction lesion segmentation method for stroke DWI images based on multi-head self-attention mechanism and U-Net

作　　者：陆小妍 张华伟[2] 郭静丽[1] 郭群[1] 蒋红兵[2,3] 张媛 LU Xiao-yan;ZHANG Hua-wei;GUO Jing-li;GUO Qun;JIANG Hong-bing;ZHANG Yuan(Department of Radiology,Nanjing First Hospital,Nanjing Medical University,Nanjing 210006,China;Department of Medical Equipment,Nanjing First Hospital,Nanjing Medical University,Nanjing 210006,China;Nanjing Emergency Medical Center,Nanjing 210003,China)

机构地区：[1]南京医科大学附属南京医院(南京市第一医院)医学影像科,南京210006 [2]南京医科大学附属南京医院(南京市第一医院)临床医学工程处,南京210006 [3]南京市急救中心,南京210003

出　　处：《医疗卫生装备》2024年第12期9-13,共5页Chinese Medical Equipment Journal

摘　　要：目的:为解决脑卒中弥散加权成像(diffusion weighted imaging,DWI)图像中梗死病灶边界模糊、病灶区域小、分割难度大的问题,提出一种基于多头自注意力机制(multi-head self-attention,MHSA)与U-Net的分割方法。方法:以U-Net作为基础分割模型,在其编码器最后一次卷积操作后加入MHSA模块,建立MHSA-UNet分割模型。为了验证MHSA-UNet分割模型的有效性,在自建的数据集上进行训练和验证,并与U-Net模型、Attention U-Net模型对脑卒中DWI图像中梗死病灶的分割效果进行比较。结果:MHSA-UNet分割模型的Dice相似系数、交并比、95%豪斯多夫距离分别为0.790、0.571、9.982,在总体上优于U-Net模型、Attention U-Net模型。结论:提出的方法能有效分割脑卒中DWI图像中梗死病灶,可以辅助临床医生进行疾病诊断。Objective To propose a segmentation method based on multi-head self-attention(MHSA)mechanism and U-net for solving the problems of stroke diffusion weighted imaging(DWI)images in infarction lesion segmentation due to blurred boundary and small area.Methods U-Net was used as the basic segmentation model,and the MHSA module was added after the last convolution operation of its encoder to build a MHSA-UNet segmentation model.The MHSA-UNet segmentation model had its effectiveness verified by being trained and validated on a self-constructed dataset and compared with the U-Net model and the Attention U-Net model for the segmentation of infarction lesions in DWI images of stroke.Results The MHSA-UNet segmentation model behaved generally better than U-Net model and Attention U-Net model,whose Dice similarity coefficient,intersection over union and 95%Hausdorff distance of the MHSA-UNet model were 0.790,0.571 and 9.982,respectively.Conclusion The proposed method segments infarction lesions in stroke DWI images effectively,and can assist clinicians in disease diagnosis.

关 键 词：多头自注意力机制 U-Net 脑卒中 缺血性脑卒中 梗死病灶 DWI图像 病灶分割 

分 类 号：R318[医药卫生—生物医学工程]