Multi-scale 3D Convolutional LSTM for Longitudinal Alzheimer's Disease Identification

The clinical symptoms of Alzheimer's Disease (AD) are progressive decline in memory and cognitive abilities, and persistent brain atrophy, leading to progressive worsening of symptoms over time. Longitudinal data provides information on the temporal evolution of neuroimaging and biomarkers, achieving better classification performance than single-time-point data models. Therefore, we propose a novel model for the classification of AD and its early state, Mild Cognitive Impairment (MCI), which combines data from multiple time points. First, preprocessed Structural Magnetic Resonance Imaging (sMRI) data is the input of a 3D Convolutional Neural Network (3D CNN) for feature extraction, which can preserve changes in feature space information. Subsequently, the 3D shape features are input into 3D Convolutional LSTM (3D CLSTM) to learn the longitudinal features. Finally, a multi-scale feature fusion module (MSFF) is applied to integrate longitudinal features at different scales. Experimental results show that our method outperforms other state-of-the-art methods.