[1]王彩莲,郑文斌.基于图卷积与Transformer的缺失模态脑肿瘤分割[J].计算机技术与发展,2025,(07):173-181.[doi:10.20165/j.cnki.ISSN1673-629X.2025.0064]
　WANG Cai-lian,ZHENG Wen-bin.Missing Modality Brain Tumor Segmentation Based on Graph Convolution and Transformer[J].,2025,(07):173-181.[doi:10.20165/j.cnki.ISSN1673-629X.2025.0064]
点击复制

基于图卷积与Transformer的缺失模态脑肿瘤分割()

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:
期数:
2025年07期
页码:
173-181
栏目:
新型计算应用系统
出版日期:
2025-07-10

文章信息/Info

Title:
Missing Modality Brain Tumor Segmentation Based on Graph Convolution and Transformer
文章编号:
1673-629X(2025)07-0173-09
作者:
王彩莲; 郑文斌
成都信息工程大学 软件工程学院,四川 成都 610225
Author(s):
WANG Cai-lian; ZHENG Wen-bin
School of Software Engineering,Chengdu University of Information Technology,Chengdu 610225,China
关键词:
脑肿瘤分割; 缺失模态; 三维适配器; 动态图卷积; Transformer
Keywords:
brain tumor segmentation; missing modality; three-dimensional adapter; dynamic graph convolution; Transformer
分类号:
TP391.41;R739.41
DOI:
10.20165/j.cnki.ISSN1673-629X.2025.0064
摘要:
多模态磁共振成像(MRI)在脑肿瘤分割中具有重要应用,但模态缺失问题在临床中普遍存在,导致肿瘤区域细节丢失及模态间关联性遭到破坏等问题,显著降低分割性能。 为此,该文提出一种基于图卷积与 Transformer 融合的缺失模态脑肿瘤分割模型。 模型首先采用三维卷积与 MBConv 结合的编码器模块提取局部特征,并利用三维适配器捕获脑肿瘤的空间信息,增强模态内特征表达能力。 随后,设计了动态图卷积-Transformer 融合模块:其中,动态图卷积用于生成模态相关性矩阵,捕获模态间局部依赖关系,并结合掩码机制有效处理模态缺失;同时,Transformer 模块则通过自注意力机制建模模态与肿瘤区域的全局相关性与上下文信息,从而显著提升特征融合鲁棒性。 在 BraTS2018 和 BraTS2020 数据集上的实验结果表明,该模型在完整肿瘤、肿瘤核心及增强肿瘤区域的平均 Dice 系数较M2FTrans 分别提升 1. 36 百分点、1. 28 百分点和 1. 39 百分点,验证了其在模态缺失场景下的优越性。
Abstract:
Multimodal Magnetic Resonance Imaging (MRI) has significant applications in brain tumor segmentation. However,the issue of missing modalities is common in clinical practice,leading to the loss of tumor region details and the disruption of inter-modality corre-lations,which significantly reduces segmentation performance. To address this,we propose a missing modality brain tumor segmentation model based on the fusion of Graph Convolutional Networks (GCN) and Transformer. The model first employs an encoder module com-bining 3D convolution and MBConv to extract local features,and utilizes a 3D adapter to capture spatial information of brain tumors,en-hancing the intra-modality feature representation capability. Subsequently,a dynamic graph convolution-Transformer fusion module is designed,where the dynamic graph convolution generates a modality correlation matrix to capture inter-modality local dependencies and effectively handles missing modalities using a masking mechanism. Meanwhile,the Transformer module models the global correlations and contextual information between modalities and tumor regions through a self-attention mechanism,thereby significantly enhancing the robustness of feature fusion. Experimental results on the BraTS2018 and BraTS2020 datasets demonstrate that the proposed model achieves an average Dice coefficient improvement of 1. 36 percentage points,1. 28 percentage points,and 1. 39 percentage points for the whole tumor,tumor core,and enhancing tumor regions,respectively,compared to M2FTrans,validating its superiority in scenarios with missing modalities.
更新日期/Last Update: 2025-07-10