Distinct subtypes of spatial brain metabolism patterns in Alzheimer’s disease identified by deep learning-based FDG PET clusters

Introduction: Alzheimer’s disease (AD) presents a broad spectrum of clinicopathologic profiles, despite common pathologic features including amyloid and tau deposition. When trying to find AD subtypes using brain imaging, there has been a difficulty in defining subtypes by excluding effect of disease progression. To overcome this limitation and understand distinct spatial patterns of neurodegeneration, we aimed to identify AD subtypes using deep learning-based clustering on FDG PET images. We also aimed to investigate clinicopathologic features of subtypes defined by spatial brain metabolism patterns.

Methods: A total of 3620 FDG brain PET images with AD, mild cognitive impairment (MCI), and cognitively normal controls (CN) at baseline and follow-up visits were obtained from Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. In order to identify representations of brain metabolism patterns different from disease progression in AD, a conditional variational autoencoder (cVAE) was used, followed by clustering using the encoded representations. FDG brain PET images with AD (n=838) and Clinical Demetria Rating Scale Sum of Boxes (CDR-SB) scores were used as inputs of cVAE model and the k-means algorithm was applied for the clustering. The trained deep learning model was also transferred to FDG brain PET image with MCI (n=1761) to identify differential trajectories in subtypes. Statistical parametric maps were generated to visualize spatial patterns of clusters, and clinical and biological characteristics were compared among the clusters. The conversion rate from MCI to AD was also compared among the subtypes.

Results: Four distinct subtypes were identified by deep learning-based FDG PET clusters: (i) S1 (angular), showing prominent hypometabolism in angular gyrus with diffuse cortical hypometabolism pattern; frequent in males; more amyloid; less tau; more hippocampal atrophy; cognitive decline in the earlier stage. (ii) S2 (occipital), showing prominent hypometabolism in occipital cortex with posterior-predominant hypometabolism pattern; younger age; more tau; less hippocampal atrophy; lower executive and visuospatial scores; faster conversion from MCI to AD. (iii) S3 (orbitofrontal), showing prominent hypometabolism in orbitofrontal cortex with anterior-predominant hypometabolism pattern; older age; less amyloid; more hippocampal atrophy; higher executive and visuospatial scores. (iv) S4 (minimal), showing minimal hypometabolism; frequent in females; less amyloid; more tau; less hippocampal atrophy; higher cognitive scores.

Conclusions: In conclusion, we could identify distinct subtypes in AD with different brain pathologies and clinical profiles. Also, our deep learning model was successfully transferred to MCI to predict the prognosis of subtypes for conversion from MCI to AD. Our results suggest that distinct AD subtypes on FDG PET may have implications for the individual outcomes and provide a clue to understanding a broad spectrum of AD in terms of pathophysiology.