首页|融合CNN与Transformer的MRI脑肿瘤图像分割

融合CNN与Transformer的MRI脑肿瘤图像分割

扫码查看
原文链接
  • 万方数据
  • 维普
为解决卷积神经网络(convolutional neural network,CNN)在学习全局上下文信息和边缘细节方面受到很大限制的问题,提出一种同时学习局语义信息和局部空间细节的级联神经网络用于脑肿瘤医学图像分割.首先将输入体素分别送入CNN和Transformer分支,在编码阶段结束后,采用一种双分支融合模块将 2 个分支学习到的特征有效地结合起来以实现全局信息与局部信息的融合.双分支融合模块利用哈达玛积对双分支特征之间的细粒度交互进行建模,同时使用多重注意力机制充分提取特征图通道和空间信息并抑制无效的噪声信息.在BraTS竞赛官网评估了本文方法,在BraTS2019 验证集上增强型肿瘤区、全肿瘤区和肿瘤核心区的Dice分数分别为 77.92%,89.20%和 81.20%.相较于其他先进的三维医学图像分割方法,本文方法表现出了更好的分割性能,为临床医生做出准确的脑肿瘤细胞评估和治疗方案提供了可靠依据.
MRI brain tumor image segmentation by fusing CNN and Transformer
This study presents a cascaded neural network that learns both global semantic information and local spatial details for medical image segmentation of brain tumors,solving the problem that convolutional neural networks(CNN)are greatly restricted in learning global contextual information and edge details.First,the input voxels are fed into the CNN and Transformer branches separately.After the encoding phase,a two-branch fusion module is used to effectively combine the features learned in both branches to achieve the fusion of global and local information.The two-branch fu-sion module uses Hadamard products to model the fine-grained interactions between the two-branch features,while us-ing multiple attention mechanisms to fully extract the feature map channels and spatial information and suppress the in-valid noise information.The method of this paper has been evaluated on the BraTS competition website,with Dice scores of 77.92%,89.20%and 81.20%for the enhanced tumor region,full tumor region and tumor core region on the BraTS2019 validation set,respectively.Compared with other advanced 3D medical image segmentation methods,this method shows better segmentation performance,which provides a reliable basis for clinicians to make accurate brain tu-mor cell assessment and treatment plans.

medical image segmentationbrain tumorcascaded neural networkconvolutional neural networksTrans-formerfeature fusionmultiple attentionresidual learning

刘万军、姜岚、曲海成、王晓娜、崔衡

展开 >

辽宁工程技术大学 软件学院,辽宁 葫芦岛 125105

医学图像分割 脑肿瘤 级联神经网络 卷积神经网络 Transformer 特征融合 多重注意力 残差学习

辽宁省高等学校基本科研项目辽宁工程技术大学学科创新团队项目

LJKMZ20220699LNTU20T-D-23

2024

10.11992/tis.202301016

智能系统学报
中国人工智能学会　哈尔滨工程大学

智能系统学报

CSTPCD北大核心
影响因子：0.672
ISSN：1673-4785
年,卷(期)：2024.19(4)
  • 1