Neuroanatomical heterogeneity in Alzheimer's disease (AD) hinders precision diagnosis and treatment, as distinct brain phenotypes may correspond to different disease subtypes. However, MRI-based subtype classifications are often confounded by co-occurring pathologies and non-AD factors, such as genetic predisposition and environmental influences, limiting their clinical interpretability. We propose 3D-DisAD, an unsupervised deep learning framework that disentangles AD-specific neuroanatomical variations from unrelated influences and clusters patients into subtypes with homogeneous brain phenotypes. The framework comprises two synergistic networks: (1) Contrastive Disentanglement Network, which separates AD-specific variations from those shared by AD patients and healthy controls; and (2) Transformation Generation Network, which refines these disease-specific variations by transforming healthy brain representations into realistic, pathology-consistent anatomies via diffusion-based generative modeling. Evaluated on four public datasets, 3D-DisAD reveals strong correlations between the disentangled AD-specific variations and diverse clinical and biological profiles, validating their relevance. Using these variations, we identify four AD subtypes with significant differences in biomarkers, cognitive trajectories, and genetic signatures, and uncover distinct longitudinal progression patterns that suggest potential windows for early intervention. By disentangling AD-specific variations, our method enables more precise patient stratification and personalized treatments, particularly in the early stage of AD. Code is available at: https://github.com/cnuzh/3D-DisAD.
Zhang et al. (Thu,) studied this question.