Longitudinal structural Magnetic Resonance Imaging (sMRI) is pivotal for capturing dynamic brain atrophy patterns associated with the progression of Alzheimer's Disease (AD). However, effectively leveraging longitudinal data remains challenging due to the difficulty in modeling complex inter-temporal correlations and the presence of inherent noise and redundancy. To address these issues, this paper proposes a novel Longitudinal Feature Disentanglement Network for AD diagnosis and progression prediction. Specifically, we design a lightweight 3D convolutional encoder with a weight-sharing strategy to extract spatial representations from dual-time-point images. A Variational Autoencoder (VAE)-based mechanism is then introduced to disentangle these representations into stable features that encapsulate disease-specific identity and marginal features that account for noise and irrelevant variations. To further enhance the disentanglement capability, we incorporate a cross-time-point contrastive learning mechanism to enforce the temporal consistency of stable features, alongside a longitudinal feature reconstruction module that ensures the preservation of comprehensive information by recombining decoupled features. Extensive experiments conducted on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset demonstrate that our proposed method outperforms state-of-the-art approaches in both AD diagnosis and MCI-to-AD conversion prediction tasks. Furthermore, visualization analysis validates the model's effectiveness in isolating discriminative brain regions and interpreting the decision-making process.
Dai et al. (Thu,) studied this question.