An explainable framework for the relationship between dementia and glucose metabolism patterns

TL;DR

This work introduces a semi-supervised variational autoencoder with a similarity regularization that ties a subset of latent variables to dementia-related measures, enabling interpretable extraction of disease patterns from high-dimensional neuroimaging data. By guiding the first latent dimension $z_0$ to align with cognitive decline (e.g., ADAS13) and leveraging voxel-wise GLMs on decoded reconstructions, the framework reveals metabolic reductions in the Default Mode and Central Executive Networks and hippocampal regions consistent with Alzheimer's pathology. The approach also disentangles confounds such as brain size and acquisition variability within the latent space through systematic latent traversals, and demonstrates strong AD vs HC discrimination via bootstrap-validated classification. Overall, the method offers a flexible, explainable tool for studying disease progression in neuroimaging with adaptable supervision and robust interpretability.

Abstract

High-dimensional neuroimaging data presents challenges for assessing neurodegenerative diseases due to complex non-linear relationships. Variational Autoencoders (VAEs) can encode scans into lower-dimensional latent spaces capturing disease-relevant features. We propose a semi-supervised VAE framework with a flexible similarity regularization term that aligns selected latent variables with clinical or biomarker measures of dementia progression. This allows adapting the similarity metric and supervised variables to specific goals or available data. We demonstrate the approach using PET scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI), guiding the first latent dimension to align with a cognitive score. Using this supervised latent variable, we generate average reconstructions across levels of cognitive impairment. Voxel-wise GLM analysis reveals reduced metabolism in key regions, mainly the hippocampus, and within major Resting State Networks, particularly the Default Mode and Central Executive Networks. The remaining latent variables encode affine transformations and intensity variations, capturing confounds such as inter-subject variability and site effects. Our framework effectively extracts disease-related patterns aligned with established Alzheimer's biomarkers, offering an interpretable and adaptable tool for studying neurodegenerative progression.

Figures (10)

• Figure 1: Architecture of the vae model. Convolutional layers extract feature from high dimensional volumes into feature maps, where are hierarchically compressed into lower representations.
• Figure 2: Phase diagram of latent dimensionality vs. $\beta$. The color code shows the mean euclidean distance to the mean $\mathcal{D_{\mu}}$ of the latent variables. Large values of $\beta$ cause the model to collapse to the mean, whereas small values of latent dimension are not sufficient to capture the variability of the data.
• Figure 3: Phase diagram of KL divergence vs. Pearson regularization. The color code shows the convergence of the Pearson correlation. Large values of $\alpha$ lead to non-informative regularization of the latent space, whereas small values do not produce patterns correlated with dementia. The latent dimensionality is constant at $8$.
• Figure 4: Odd columns: input brain slices. Even columns: VAE reconstructions. Green boxes highlight input-reconstruction pairs. Color represents intensity; reconstructions show lower intensity but preserve relevant structure. The kl regularization counterweights the reconstruction, producing lower quality scans than the input.
• Figure 5: Relationship between the value of the latent variable $z_0$ (test set) and the adas13 score. $r = 0.790$, $p \ll 0.001$.