[Paper Review] Disentangling brain heterogeneity via semi-supervised deep-learning and MRI: dimensional representations of Alzheimer's Disease

Zhijian Yang, Ilya M. Nasrallah+9 more|arXiv (Cornell University)|Feb 24, 2021

Dementia and Cognitive Impairment Research29 references4 citations

TL;DR

This paper introduces Smile-GAN, a semi-supervised deep-clustering framework that disentangles neuroanatomical heterogeneity in Alzheimer’s disease using MRI data. It identifies four distinct neurodegenerative patterns (P1–P4) and two longitudinal progression pathways, with pattern expression outperforming amyloid/tau biomarkers in predicting clinical progression.

ABSTRACT

Heterogeneity of brain diseases is a challenge for precision diagnosis/prognosis. We describe and validate Smile-GAN (SeMI-supervised cLustEring-Generative Adversarial Network), a novel semi-supervised deep-clustering method, which dissects neuroanatomical heterogeneity, enabling identification of disease subtypes via their imaging signatures relative to controls. When applied to MRIs (2 studies; 2,832 participants; 8,146 scans) including cognitively normal individuals and those with cognitive impairment and dementia, Smile-GAN identified 4 neurodegenerative patterns/axes: P1, normal anatomy and highest cognitive performance; P2, mild/diffuse atrophy and more prominent executive dysfunction; P3, focal medial temporal atrophy and relatively greater memory impairment; P4, advanced neurodegeneration. Further application to longitudinal data revealed two distinct progression pathways: P1$ ightarrow$P2$ ightarrow$P4 and P1$ ightarrow$P3$ ightarrow$P4. Baseline expression of these patterns predicted the pathway and rate of future neurodegeneration. Pattern expression offered better yet complementary performance in predicting clinical progression, compared to amyloid/tau. These deep-learning derived biomarkers offer promise for precision diagnostics and targeted clinical trial recruitment.

Motivation & Objective

• To address the challenge of brain disease heterogeneity in precision diagnosis and prognosis.
• To develop a method that identifies biologically meaningful subtypes of Alzheimer’s disease from structural MRI data.
• To leverage semi-supervised deep learning to discover imaging-based disease subtypes without requiring fully labeled data.
• To validate the clinical relevance of identified patterns using longitudinal data and prediction of neurodegeneration trajectories.
• To compare the predictive power of deep-learning-derived biomarkers against traditional amyloid/tau markers.

Proposed method

• Smile-GAN is a novel semi-supervised deep-clustering model combining clustering and generative adversarial networks (GANs) to learn disentangled representations from MRI scans.
• The method uses a contrastive loss to encourage feature disentanglement and cluster assignment consistency across views.
• It jointly optimizes a generator to synthesize realistic brain images and a discriminator to distinguish real from generated scans, while enforcing cluster structure.
• The model is trained on a combination of labeled controls and unlabeled individuals with cognitive impairment or dementia.
• Dimensional representations of neurodegeneration are derived from the learned latent space, with each dimension corresponding to a distinct neurodegenerative pattern.
• Progression pathways are inferred by tracking the evolution of pattern expression across time points in longitudinal MRI data.

Experimental results

Research questions

• RQ1 Can a semi-supervised deep-learning framework identify biologically meaningful subtypes of Alzheimer’s disease from structural MRI without full annotation?
• RQ2 What are the distinct neurodegenerative patterns (axes of atrophy) that emerge from unsupervised disentanglement of brain MRI data?
• RQ3 How do these patterns relate to cognitive performance and clinical phenotypes such as memory or executive function deficits?
• RQ4 Can the expression of these patterns predict future neurodegeneration trajectories and rates in longitudinal data?
• RQ5 How does the predictive performance of these deep-learning-derived biomarkers compare to established amyloid/tau biomarkers?

Key findings

• Smile-GAN identified four distinct neurodegenerative patterns: P1 (normal anatomy, best cognition), P2 (mild/diffuse atrophy, executive dysfunction), P3 (focal medial temporal atrophy, memory impairment), and P4 (advanced neurodegeneration).
• Longitudinal analysis revealed two primary progression pathways: P1 → P2 → P4 and P1 → P3 → P4, with baseline pattern expression predicting future trajectory.
• The expression of these patterns predicted clinical progression with higher accuracy than amyloid or tau biomarkers, offering complementary predictive power.
• The method successfully disentangled brain heterogeneity using only partial labels, demonstrating robustness in low-supervision settings.
• The dimensional representations derived from MRI scans captured clinically relevant neuroanatomical and cognitive phenotypes across diverse cognitive states.
• The approach enables precision diagnostics and targeted recruitment for clinical trials by identifying biologically distinct Alzheimer’s subtypes.