基于Mamba的轻量级多模态脑肿瘤MRI图像分割

打开文本图片集

关键词：多模态数据融合；脑肿瘤；轻量级架构；3DU-Net；Mamba；SE模块中图分类号：TN911.73-34；TP391.4 文献标识码：A 文章编号：1004-373X（2026）09-0032-06

Lightweight multi-modal brain tumor MRI image segmentation based on Mamba

Hou Xiangning1，²，Huang Xiaobin1，²，Xu Caocao1，²，Yao Jun1，2 （1.TheEngineeringand Technical Collegeof Chengdu Universityof Technology，Leshan 6140oo，China; 2.Southwestern InstituteofPhysics，Chengdu 61O2O0，China）

Abstract：The existing 3DU-Netmodel has high computationalcomplexityand pooraccuracywhen processing 3D medical imagedata，soitisdificulttomeetteneedsofclinicalreal-timediagnosis.Therefore，alightweighttechnologyacitecture basedon3DU-Netcombined with Mamba isproposed toreduce thecomputationalcomplexityand improve thesegmentation accuracyofthetraditionaldeplearningmodelsinmulti-modalbraintumorMRIimagesegmentation.Mamba'slinearcomplexity self-atention mechanismisusedtooptimizethe3DU-Netmodel toreduce thecomputationalcomplexityand improve the inferencespeedandsegmentationaccuracyThespecificmethodistointroduceLWMmoduleandattentionbridgemoduleonthe basisof 3D U-Netmodel.ParalelMambaarchitecture isadopted intheLWMmodule，whichreduces thenumberof model parametersbyreducing thenumberof input channels，greatlyreducesthecomputationalcomplexityand memoryconsumption whileextractingglobalinformation，andrealizeseficientsegmentationinthecaseoffewerparameters.Inadition，theLWM furtherimproves thesegmentationperformanceoffine-grained tumorcoresby introducing SEmodule.Theexperimentsonthe dataset BraTS220showthattheproposed methodsignificantlyimprovestheaccuracyof tumorsegmentation，and greatlyreduces the training and inference duration，so it shows great practical application potential.

Keywords：multi-modal data-fusion;brain tumor; lightweight framework; 3D U-Net; Mamba; SE module

0 引言

在临床实践中，精确分割多模态脑肿瘤MRI图像对于脑肿瘤的诊断和治疗决策至关重要。（剩余9211字）