基于图卷积的自适应特征融合MRI脑肿瘤分割方法
2025-08-09张野张睦卿袁学刚牛大田
中图分类号:TP391.41;R739.41 文献标识码:A DOI:10.7535/hbkd.2025yx04005
Graph convolution-based adaptive feature fusion method for MRI brain tumor segmentation
ZHANG Ye¹,ZHANG Muqing²,YUAN Xuegang1,NIU Datian1 (1.School of Science,Dalian Minzu University,Dalian,Liaoning1166oo,China; 2.School of Computer Science and Enginering,Dalian Minzu University,Dalian,Liaoning l166oo,China)
Abstract:Toaddress theisues ofinsuffcient global information capture and inadequatedeep semantic information fusion in the U-Netmodel for MRIbrain tumor segmentation,anovel braintumor segmentation network,ASGU-Net was proposed.
The algorithm wasbasedon 3D U-Net,incorporating agraph convolution inferencemodule tocaptureadditional long-range contextualfeatures.Aditionally,dynamicsnakeconvolution(DSConv)was introduced inthe encoder-decodertobetter accommodate the varied shapes of tumors,enhancing edge feature extractionand improving segmentation acuracy. Furthermore,anadaptivespatial featurefusion(ASFF)modulewas introduced in the decoder toenhance the feature fusion efect byintegrating semantic informationcaptured by multipleencoderblocks.Theevaluationonthepubliclyavailable BraT 2019—2021 datasets shows that the Dice values for whole tumor,tumor core and enhanced tumor are 90.70%/90.70% 1 91.00% , 84.90%/84.00%/88.80% and 77.30%/77.40%/82.50% ,respectively. The experimental results demonstrate the effectiveessofASGU-NetinthebraintumorsegmentationtaskASGU-Netcaneectiveladdressstheissuesofiadequateglobal informationcaptureand feature fusion,providing effective reference for high-precisionautomatedbrain tumorsegmentation.
KeyWords:computer neuralnetwork;brain tumor segmentation;3D U-Net;graph convolution inferencebotteneck layer; dynamic snake convolution;adaptive spatial feature fusion
脑肿瘤是脑部异常细胞形成的肿块,严重威胁人类的健康与生命。精确的脑肿瘤自动化分割在临床诊断、治疗方案制定以及术后监测中起着至关重要的作用。因此,开发脑肿瘤高精度自动化分割方法已经成为医疗图像分析领域的重要研究方向[]。
随着深度学习的迅猛发展,众多与医疗图像分割相关的研究取得了显著进展,基于卷积神经网络的方法在医疗图像分割领域得到了广泛应用。文献[2]提出编解码器构成的U-Net网络成为脑肿瘤分割任务的主流架构,但横向连接操作融合低级特征图与高级特征图时会出现局部信息不连续和空间信息减少等问题。文献[3]提出在编码器特征和解码器特征间采用加性注意力机制来进一步提升网络的分割性能。文献_4-5分别在三维U-Net中集成残差模块和跨阶段模块,增强网络提取特征信息的能力。但纯卷积结构不擅长捕获全局语义信息。Transformer模型近年来在捕获全局特征上展现出巨大的潜力,引起了研究者的关注。研究者们提出的基于 Transformer 模型的网络有 PMTrans[6]、MultiTrans[7]、SwinTransformer[8]等。尽管 Transformer具有学习长距离依赖关系的能力,但其高计算复杂性对模型效率和内存占用带来了重大挑战,因此,研究人员尝试引入图卷积神经网络。文献[9]提出了一种基于图卷积神经网络的脑肿瘤分割方法,利用图推理模块获得邻居节点信息以提升脑肿瘤分割精度。文献[10]则创新性地提出了一种双图推理单元的方法,通过将空间推理模块和信道推理模块并行处理,以关注图像中丰富的空间信息和上下文依赖关系。……
