Unsupervised Analysis of Alzheimer's Disease Signatures using 3D Deformable Autoencoders

TL;DR

The paper tackles the challenge of detecting and localizing Alzheimer's disease (AD)–related brain atrophy using MRI in an unsupervised setting. It introduces MORPHADE, a 3D deformable autoencoder that employs dual deformation fields (constrained and unconstrained) and adversarial training to produce refined residual and folding maps, which are fused into an anomaly map for localization and severity assessment. Key innovations include the use of deformation fields within an unsupervised framework and the combination of $m_{residual}=|x-x_{recon}|$ with $m_{foldings}=\max(0,-\det(J_{\boldsymbol{\Phi}}))$ to highlight atrophy regions, as well as a training protocol that optimizes for reconstruction fidelity while preserving anatomical variability. On ADNI data, MORPHADE achieves $AUROC=0.80$ for AD detection, surpassing several supervised and unsupervised baselines, and shows spatial concordance with medial temporal atrophy (MTA) scores, supporting its potential for non-invasive disease monitoring. The authors also provide public code at $github.com/ci-ber/MORPHADE$ for broader adoption and validation.

Abstract

With the increasing incidence of neurodegenerative diseases such as Alzheimer's Disease (AD), there is a need for further research that enhances detection and monitoring of the diseases. We present MORPHADE (Morphological Autoencoders for Alzheimer's Disease Detection), a novel unsupervised learning approach which uses deformations to allow the analysis of 3D T1-weighted brain images. To the best of our knowledge, this is the first use of deformations with deep unsupervised learning to not only detect, but also localize and assess the severity of structural changes in the brain due to AD. We obtain markedly higher anomaly scores in clinically important areas of the brain in subjects with AD compared to healthy controls, showcasing that our method is able to effectively locate AD-related atrophy. We additionally observe a visual correlation between the severity of atrophy highlighted in our anomaly maps and medial temporal lobe atrophy scores evaluated by a clinical expert. Finally, our method achieves an AUROC of 0.80 in detecting AD, out-performing several supervised and unsupervised baselines. We believe our framework shows promise as a tool towards improved understanding, monitoring and detection of AD. To support further research and application, we have made our code publicly available at github.com/ci-ber/MORPHADE.

Figures (5)

• Figure 1: Our approach, MORPHADE, integrates a dual-deformation strategy with a 3D autoencoder and adversarial training. The constrained deformer refines the reconstruction to generate a residual map with reduced false positives, while the unconstrained deformer is used to produce a folding map that highlights anomalies. The residual and folding maps together produce an anomaly map that allows the localization and assessment of the severity of atrophy.
• Figure 2: a) During training, a high value of $\beta=10$ constrains the deformer, promoting the AE to learn to produce less blurry reconstructions. b) At inference, a lower value of $\beta=0.01$ is used to generate folding maps (here shown overlayed on the input brain) that enhance the identification of anomalies.
• Figure 3: Anomaly scores for subcortical brain regions for Alzheimer's Disease (AD) and Healthy Control (HC) samples, showcasing markedly higher scores for AD samples in the hippocampus and amygdala, consistent with clinical literature. alzheimer_affect
• Figure 4: Anomaly maps for AD patients alongside their corresponding medial temporal lobe atrophy (MTA) scores, demonstrating consistent alignment with AD-related structural changes and clinical MTA assessments.
• Figure 5: A comparison of the performance of MORPHADE ($\beta=10$) with adversarial AEs for a subject with AD (left) and a healthy control subject (right). The morphological adjustments facilitated by MORPHADE enhance reconstruction fidelity, yielding higher Structure Similarity Index (SSIM) values for our method's morphed reconstructions compared to those of the adversarial AE. The residual maps also demonstrate fewer reconstruction errors for the healthy subject, while highlighting atrophy for the subject with AD.