基于特征融合与无参数注意力机制的MRI图像超分辨率重建方法

打开文本图片集

关键词：超分辨率重建；MRI图像；SwinIR；特征分离；无参数注意力机制中图分类号：TP391.4 文献标志码：A 文章编号：1001-3695（2026）04-035-1251-07doi：10. 19734/j. issn. 1001-3695.2025.06.0264

Super-resolution reconstruction method for MRI images based on feature separation and parameter-free attention mechanism

Zhao Hong， Zhang Runyan† （SchoolofComputerScienceandArtificial Intelligence，Lanzhou UniversityofTechnology，Lanzhou730o5o，China）

Abstract：ForMRI，wheresuper-resolutionreconstructionoftensufersfrominsuffcientdetail representationandhighcomputationalcost，this paper proposedafeature fusionand parameter-freeattention network （FFPAN）toaddress these issues.The network consistedofthreeparts：shallowfeature extraction，deepfeatureextractionandimagereconstruction.Thedeepfeature extractionincludedafeatureseparationblock（FSB），aparameter-freeatention（PA）module，andageneralizedselftention （GSA）module.TheFSB divided image features into high-frequency detailsandlow-frequencyglobal information.The PA modulecomputedimage featuresthroughconvolutions tocapture global dependencies between features.Unliketraditionalaention mechanisms，PA generatedthe atention map fromtheoutputofconvolutionlayers without leaing additional parameters， resultinginlowcomputationalcost.TheGSAmoduleefectivelyextractedglobalorhigh-frequencyinformationthroughselfatentioncombined with residual connections.Itconducted experimentsonthe public datasets BraTS21andFastMRI.Theresults show that，compared with SwinIR，the proposed method reduces the number of parameters by 48.3% while improving PSNRand SSIMbyO.15dBand0.0446，respectively.Inaddition，subjectiveevaluation indicates thatMRIimagesreconstructed by this network better preserve image details，demonstrating high clinical value.

KeyWords：super-resolutionreconstruction；MRI image；SwinIR；featureseparation；parameter-freeatentionmechanism

0 引言

磁共振成像（MRI）技术具有多参数成像能力、无电离辐射等优点，生成的图像具有高软组织对比度，能够为医生提供精准的病变识别和诊断依据。（剩余18143字）