基于多头自注意力机制与U-Net的增强CT图像肾脏小肿瘤自动分割研究  

Study on the automatic segmentation of enhanced CT image of small kidney tumors based on MHSA mechanism and U-Net

作　　者：吴倩倩[1] 周蕾蕾[2] 陆小妍 蒋红兵[1,3] WU Qian-qian;ZHOU Lei-lei;LU Xiao-yan(Department of Clinical Medical Engineering,Nanjing First Hospital,Nanjing Medical University,Nanjing 210006,China;不详)

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

出　　处：《中国医学装备》2022年第2期27-31,共5页China Medical Equipment

基　　金：南京市医学科技发展课题(QRX11033)“南京市卫生青年人才培养工程”;南京医科大学科技发展基金一般项目(NMUB2019155)“基于卷积神经网络的肾脏肿瘤CT图像定位分割研究”。

摘　　要：目的:构建一种基于多头自注意力(MHSA)机制与U-Net网络的M-UNet优化分割模型,以提高增强CT图像中肾脏小肿瘤横断面最大直径≤3 cm的分割准确度。方法:选取医学图像计算和计算机辅助干预协会(MICCAI)的2019肾脏肿瘤分割挑战(KiTS19)数据集中64例最大层面肾脏肿瘤直径≤3 cm的数据,将其按7∶3划分为训练集与测试集,训练集进行五折交叉验证。建立基于MHSA机制与U-Net网络的M-UNet优化分割模型,对M-UNet和U-Net分别进行训练与测试,计算交并比(IOU)、Dice相似系数(Dice系数)和95%豪斯多夫距离(HD_95),对比M-Unet与U-Net二者对肾脏及肾脏肿瘤的分割精度。结果:M-UNet网络的IOU、Dice系数相较于U-Net分别提升3.19%和3.00%,HD_95下降41.63%。结合分割结果视觉图,M-UNet与U-Net相比,其对肾脏小肿瘤分割准确率更高,检测更为敏感。结论:M-UNet相对于传统的U-Net能够更准确分割增强CT图像中的肾脏小肿瘤,为临床对肾脏小肿瘤定位及诊断提供便利,有助于提升肾脏小肿瘤检出率。Objective:To propose one kind of optimized M-UNet segmentation model based on Multi-Head Self-Attention(MHSA)and U-Net network,so as to improve the accuracy of segmentation that the maximum diameter of transverse section of small kidney tumor was≤3 cm in enhancing computed tomography(CT)images.Methods:The data of 64 cases whose largest tumor diameter were≤3 cm in the data sets of kidney tumor segmentation challenge(KiTS19)of Medical Image Computing and Computer Assisted Intervention Society(MICCAI)were selected and were divided into training set and testing set as 7:3,and training set underwent five-fold cross validation.And then,M-UNet and U-Net were trained and tested respectively,and the intersection over union(IOU),Dice similarity coefficient and 95%hausdorff distance(HD_95)were calculated,and the segmentation accuracies of M-Unet and U-Net on kidney and renal tumor were compared.Results:Compared with the IOU,Dice similarity coefficient of U-Net,those of M-UNet were increased by 3.19%and 3.00%,respectively,and HD_95 was reduced by 41.63%.Combined with the visual images of the segmentation results,the segmentation accuracy of M-UNet was higher and the test of that was more sensitive than those of U-Net on small kidney tumors.Conclusion:Compared with conventional U-Net,M-UNet can more accurately segment the small kidney tumors in enhanced CT images,which provide convenience for clinical localization and diagnosis of small kidney tumors,and contribute to improve the detection rate of small kidney tumors.

关 键 词：肾脏 小肿瘤 图像分割 多头自注意力(MHSA) 机制 增强CT 

分 类 号：R737.11[医药卫生—肿瘤]