2023-05-10计算机应用,JournalofComputerApplications2023,43(5):1606-1611ISSN1001-9081CODENJYIIDUhttp://www.joca.cn基于自注意力连接UNet的磁共振成像去吉布斯伪影算法刘阳,陆志扬,王骏,施俊*(上海大学通信与信息工程学院,上海200444)(∗通信作者电子邮箱junshi@shu.edu.cn)摘要:为去除磁共振成像(MRI)中的吉布斯伪影,提出一种基于自蒸馏训练的自注意力连接UNet(SD-SacUNet)算法。为了缩小UNet框架中跳连接两端编码和解码特征之间的语义差距,帮助捕捉伪影的位置信息,将UNet编码端每个下采样层的输出特征分别输入各自的自注意力连接模块进行自注意力机制的运算,而后与解码特征进行融合,参与特征的重建;在网络解码端进行自蒸馏训练,通过建立深层与浅层特征之间的损失函数,使深层重建网络的特征信息可以用于指导浅层网络的训练,同时优化整个网络,提升图像重建水平。在公开的MRI数据集CC359上评估SD-SacUNet算法的性能,获得的峰值信噪比(PSNR)为30.26dB,结构相似性(SSIM)为0.9179;与GRACNN(Gibbs-RingingArtifactreductionusingConvolutionalNeuralNetwork)、SwinIR(ImageRestorationusingSwinTransformer)相比,SD-SacUNet的PSNR分别提高了0.77dB、0.14dB,SSIM分别提高了0.0183、0.0033。实验结果表明,SD-SacUNet算法提升了MRI去除吉布斯伪影的图像重建性能,具备潜在的应用价值。关键词:磁共振成像重建;深度学习;自蒸馏;Transformer;UNet;注意力机制中图分类号:TP391.41文献标志码:AGibbsartifactremovalalgorithmformagneticresonanceimagingbasedonself-attentionconnectionUNetLIUYang,LUZhiyang,WANGJun,SHIJun*(SchoolofCommunicationandInformationEngineering,ShanghaiUniversity,Shanghai200444,China)Abstract:ToremoveGibbsartifactsinMagneticResonanceImaging(MRI),aSelf-attentionconnectionUNetbasedonSelf-Distillationtraining(SD-SacUNet)algorithmwasproposed.InordertoreducethesemanticgapbetweentheencodinganddecodingfeaturesatbothendsoftheskipconnectionintheUNetframeworkandhelptocapturethelocationinformationofartifacts,theoutputfeaturesofeachdown-samplinglayerattheUNetencodingendwasinputtothecorrespondingself-attentionconnectionmoduleforthecalculationoftheself-attentionmechanism,thentheywerefusedwiththedecodingfeaturestopartic...