基于注意力机制无监督心脏超声序列图像配准  

Unsupervised image registration of cardiac ultrasound sequences based on channel attention

作　　者：兰其斌[1] 黄立勤[2] LAN Qibin;HUANG Liqin(Concord University College,Fujian Normal University,Fuzhou,Fujian 350117,China;College of Physics and Information Engineering,Fuzhou University,Fuzhou,Fujian 350108,China)

机构地区：[1]福建师范大学协和学院,福建福州350117 [2]福州大学物理与信息工程学院,福建福州350108

出　　处：《福州大学学报（自然科学版）》2023年第1期41-48,共8页Journal of Fuzhou University(Natural Science Edition)

基　　金：福建省自然科学基金资助项目(2021J01547);福建师范大学协和学院院级科研资助项目(KY20180201)。

摘　　要：针对基于传统非刚性医学图像配准的心脏超声序列图像配准方法缺乏自动性及配准速度慢、准确率较低的问题,将基于深度学习的医学图像配准算法应用于心脏超声序列图像配准,通过引入通道注意力机制,构建由注意力机制模块、Unet卷积神经网络模块及空间转换模块STN构成的配准模型.实验选取不同的相似性损失函数和平滑损失函数,对比VoxelMorph配准模型,相关配准性能指标都有不同程度的改进,DICE指标提升0.42%,MI指标提升2.5%,SSIM提升3.7%,NRMSE减小9%,表明配准模型的有效性.从配准效果及配准时间分析,配准模型基本可以满足心脏超声序列图像配准的实时性需求,具有一定的临床应用价值.Aiming at the problems of lack of automaticity,slow registration speed and low accuracy of cardiac ultrasound sequence image registration method based on traditional non-rigid medical image registration.In this paper,deep learning-based medical image registration algorithm is applied to cardiac ultrasound image registration.By introducing the channel attention mechanism,the registration model was constructed consisting of attention mechanism module,Unet convolution neural network module and space conversion module STN.Different similarity loss function and smoothing loss function were selected in the experiment.Compare with the registration models of VoxelMorph,the relevant registration performance indexes have been improved to varying degrees.DICE index increased by 0.42%,MI index increased by 2.5%,SSIM increased by 3.7%,and NRMSE decreased by 9%,indicating the effectiveness of the registration model.From the analysis of registration effect and registration time,the registration model can basically meet the real-time requirements of cardiac ultrasound sequence image registration,and has certain clinical application value.

关 键 词：医学图像配准 心脏超声序列图像 深度学习 通道注意力 Unet卷积神经网络 

分 类 号：R445.1[医药卫生—影像医学与核医学] TP391.41[医药卫生—诊断学]