Early Diagnosis of Alzheimer's Disease Using 3D Residual Attention Network Based on Hippocampal Multi-indices Feature Fusion

Alzheimer's disease (AD) is one of the most common causes of dementia in older individuals. Convergence evidence has confirmed that hippocampal atrophy is one of the most robust neuroimaging biomarkers of AD. However, most previous studies only independently consider the hippocampal volume or other morphological indicators, which cannot reflect the abnormal pattern of the hippocampus comprehensively and objectively. The primary aim of this study is to develop a classification model of AD based on a hippocampal multi-indices features fusion framework. The multi-indices features included 1) hippocampal gray volume block; 2) probability matrix obtained from the hippocampal segmentation; 3) hippocampal radiomics features. The 3D convolutional neural network based on the multi-indices feature fusion framework showed an ACC = 90.3% (AUC = 0.93) in classifying AD (N = 282) from NC (N = 603). The results suggested that the hippocampal multi-indices features are robust neuroimaging biomarkers in the early diagnosis of AD.