基于Mamba-UNet架构的3D MRI脑肿瘤分割方法

打开文本图片集

关键词：深度学习；MRI脑肿瘤分割；多面体卷积；三维U-Net；Mamba 中图分类号：TP391.41 文献标志码：A 文章编号：1001-3695（2026）01-037-0305-08 doi：10.19734/j. issn.1001-3695.2025.03.0147

3D MRI brain tumor segmentation method based on Mamba-UNet architecture

Zhang Ye，Niu Datian† （SchoolofScience，DalianMinzuUniversity，DalianLiaoning116ooO，China）

Abstract：AcuratesegmentationofmultimodalMRIbrain tumorimages iscrucialforclinicaldiagnosisandprognosisssesmentofbraincancer.Toaddress thelimitationsofconvolutionalneuralnetworks incapturingglobalcontextualinformationand modeling long-rangedependencies，thispaper proposedanovel segmentationmodelnamedPolyhedron Conv-Tri-orientated Mamba（PhC-ToMamba）byintegrating Mamba witha U-Netarchitecture.Itembedded a Tri-orientated Mamba （ToM）moduleinthebotlenecklayertomodelhigh-dimensionalglobalfeaturesbycomputingandinteractingdependenciesalongthreedirections，therebyenhancing global featurerepresentationin3Dmedicalimages.Inadition，thispaperintroducedanovel Polyhedron Convolution（PhConv）intotheencodertoenlargethereceptivefieldandimproved theextractionofcritical target regions.These modules efectivelyenhanced global context awarenessand focusedatentiononkey tumorregions.Extensive experiments wereconductedonthe BraTS 2O21and MSDTaskO1_BrainTumordatasets.The proposed PhC-ToMambaachieves Dice scores of 95.05%/90. 46% ，94 ∴53%189.91% ，and 90. 74% /75. 91% for whole tumor，tumor core，and enhancing tumorsegmentation，respectively.Comparedwithstate-of-the-artmethods，PhC-ToMambademonstratessuperiorsegmentation accuracyand parametereficiency，providingarobust solutionforbrain tumorsegmentationandimproving diagnosticprecision.

Key words： deep learning；MRI brain tumor segmentation; polyhedron convolution;3D U-Net；Mamba

0 引言

脑肿瘤是一种具有高致死率和高发病率的疾病，根据肿瘤细胞的恶性程度，可分为高级别胶质瘤（HGG）和低级别胶质瘤（LGG），这两类肿瘤均可导致颅内压升高，严重威胁患者生命[1]。（剩余21515字）