阿尔茨海默病辅助诊断的多模态数据融合轻量级网络

Lightweight multimodal data fusion network for auxiliary diagnosis of Alzheimer’s disease

单模态阿尔茨海默病辅助诊断方法缺少专业标注的影像数据,特征提取不稳定且要求高计算能力,为此融合核磁成像、正电子发射断层扫描影像数据和精神认知评分数据,提出多模态轻量级阿尔茨海默病辅助诊断网络(LightMoDAD).在影像特征提取模块中,去冗余卷积以提取局部特征,引入全局滤波用于提取全局特征,通过配准并相加实现多模态影像特征融合.在文本特征提取模块中,由可分离深度卷积提取精神认知评分数据特征与多模态影像特征融合,通过迁移学习增强特征判别性.采用多层感知器识别复杂的模式和特征,提高所提网络的分类准确率.在ADNI数据库中开展有效性验证实验,LightMoDAD的分类准确率、敏感性和特异性分别为0.980、0.985和0.975.实验结果表明,所提网络有助于提高医生诊断效率,具有移动端部署潜力.

The auxiliary diagnostic methods for Alzheimer’s disease(AD)using single-modal data suffer from a lack of professionally annotated imaging data,unstable feature extraction and high requirements for computing power.A multimodal lightweight Alzheimer’s disease auxiliary diagnostic network(LightMoDAD)was proposed,and magnetic resonance imaging(MRI),positron emission tomography(PET)imaging data,and psychometric score data were utilized in the network.In the image feature extraction module,spatial and channel reconstruction convolution was employed to extract local features,global filtering to extract global features,and feature registration and addition to achieve multimodal feature fusion.In the text feature extraction module,the features extracted by separable convolution from psychometric score data were integrated with multimodal image feature,and transfer learning was used to improve feature discrimination.A multi-layer perceptron was applied to recognize complex patterns and features,and the classification accuracy of the proposed network was improved.Experiments were conducted on the ADNI dataset,and LightMoDAD’s accuracy,sensitivity and specificity were 0.980,0.985 and 0.975,respectively.Experimental results show that the proposed network enhances physicians’diagnosis efficiency and holds potential for mobile platform deployment.