面向阿尔茨海默症辅助诊断的多尺度域适应网络

PDF下载HTML

Title:

Multiscale domain adaptation network for the auxiliary diagnosis of Alzheimer’s disease

作者:

蔡鸿顺, 张琼敏, 龙颖

重庆理工大学 计算机科学与工程学院, 重庆 400054

Author(s):

CAI Hongshun, ZHANG Qiongmin, LONG Ying

College Of Computer Science And Engineering, Chongqing University of Technology, Chongqing 400054, China

关键词:

阿尔茨海默症; 磁共振影像; 域偏移; 多尺度信息; 域适应; 协同训练; 对抗学习; 距离度量

Keywords:

Alzheimer’s disease; magnetic resonance imaging; domain shift; multiscale information; domain adaptation; joint training; adversarial learning; distance metrics

分类号:

TP391

DOI:

10.11992/tis.202205050

摘要:

针对传统有监督学习忽略了磁共振影像数据（magnetic resonance imaging, MRI）由于个体差异和不同站点等原因导致的特征分布不一致这一域偏移问题，本文提出了一种多尺度域适应网络模型应用于阿尔茨海默症（Alzheimer disease，AD）的辅助诊断。首先在三维卷积神经网络中设计空洞空间金字塔模块进行特征的多尺度信息提取融合，并加入注意力一致性损失来保留域间转移的语义信息；然后协同训练两个域判别器和特征提取器进行对抗学习实现源域和目标域的特征对齐，并加入权重差异损失防止域判别器过拟合；最后，在对抗训练中引入基于最大密度差异的距离度量方法，增强两个域数据的特征对齐。实验结果表明，本文方法在面临域偏移的MRI数据上具有更好的识别精度和鲁棒性。

Abstract:

The paper proposes a multiscale domain adaptation network for the auxiliary diagnosis of Alzheimer’s disease. The network is designed to address the domain shift problem that traditional supervised learning ignores due to individual differences and different sites in magnetic resonance imaging (MRI) data. The network uses the Atrous Spatial Pyramid Pooling module in the three-dimensional convolutional neural network for feature extraction and fusion at multiple scales. Attention consistency loss is added to preserve the semantic information of interdomain transfer. Two domain discriminators and feature extractors are jointly trained in adversarial learning to achieve feature alignment of the source domain and target domain. Weight difference loss is added to prevent the domain discriminator from overfitting. A distance metric method based on the Maximum Density Divergence is introduced in the adversarial training to enhance the feature alignment of the two domain data. Experimental results reveal that this method exhibits superior recognition accuracy and robustness of MRI data facing the domain shift.