残差混合注意力与自适应特征融合的脑肿瘤分割

打开文本图片集

关键词：脑肿瘤分割；自适应特征融合；残差混合注意力；双动态卷积增强中图分类号：TP391.41 文献标志码：A 文章编号：1001-3695（2025）08-037-2525-07doi：10.19734/j.issn.1001-3695.2024.10.0432

Brin tumor segmenttion sed on residul mixed ttention nd dptive feture fusion

Wu Jinxua，b，WuYuna，bt （a.StateKeyLaboratoryofublicBigData，b.ColegeofomputerSience&Technology，GuzhouUniersty，Guiyang550ina）

Abstract：Theautomatedsegmentationofbrain tumorimagesiscrucialintheauxiliarydiagnosisandtreatmentof brain tumors.This paper proposed abrain tumor segmentation method RAC-Net，which combined aresidual mixedattention mechanismandadaptivefeaturefusion toaddress theproblemsofcomplexandvariablelesions inbrain tumorimages，as well as blurredboundariesbetweenlesionsandthebackgroud.FirstlyDDCEmoduleenhancedfeatureextractionflexibilityndimprovedthe adaptabilityofthe model.Then，RMA moduleextractedbothglobalandlocal features fromtheimage.Finally， AF2Mfused deepand shalowfeatures inthe decoding path，thereby enriching thefeature representation.Experiments on the public datasetsBraTS2019 and BraTS2O21andcros-dataset validation with BraTS2O23showthatRAC-Netoutperformsmost existingsegmentationmethodsacross varousmetrics，demonstrating itspotentialforasistingintediagosisofclinicalyrelevant brain tumor diseases.

Key words：brain tumor segmentation；adaptive feature fusion;residual mixed atention；dual dynamic convolution enhancement

0 引言

脑胶质瘤是一种源自脑内神经胶质细胞的肿瘤，是最常见的原发性颅脑恶性肿瘤，约占据所有原发性中枢神经系统肿瘤的 40%～60% 。（剩余19982字）