A Cautionary Tale of Self-Supervised Learning for Imaging Biomarkers: Alzheimer's Disease Case Study

TL;DR

This study addresses the need for sensitive, biologically grounded MRI biomarkers for Alzheimer's disease and questions the universal utility of generic self-supervised learning. It introduces Residual Noise Contrastive Estimation (R-NCE), a domain-informed SSL framework that fuses auxiliary FreeSurfer features with augmentation-based learning via a patch-based 3D CNN and transformer aggregator. Across age, disease, conversion, and amyloid-prediction tasks, R-NCE outperforms traditional FreeSurfer features and standard SSL methods in several clinically relevant metrics, and BAG analyses reveal meaningful genetic and cellular associations (e.g., MAPT, IRAG1; astrocyte/oligodendrocyte enrichment) with substantial heritability. The findings demonstrate that integrating structured domain knowledge into SSL yields biomarkers that are not only predictive but also biologically interpretable, with potential for broader neurodegenerative applications and multimodal extensions.

Abstract

Discovery of sensitive and biologically grounded biomarkers is essential for early detection and monitoring of Alzheimer's disease (AD). Structural MRI is widely available but typically relies on hand-crafted features such as cortical thickness or volume. We ask whether self-supervised learning (SSL) can uncover more powerful biomarkers from the same data. Existing SSL methods underperform FreeSurfer-derived features in disease classification, conversion prediction, and amyloid status prediction. We introduce Residual Noise Contrastive Estimation (R-NCE), a new SSL framework that integrates auxiliary FreeSurfer features while maximizing additional augmentation-invariant information. R-NCE outperforms traditional features and existing SSL methods across multiple benchmarks, including AD conversion prediction. To assess biological relevance, we derive Brain Age Gap (BAG) measures and perform genome-wide association studies. R-NCE-BAG shows high heritability and associations with MAPT and IRAG1, with enrichment in astrocytes and oligodendrocytes, indicating sensitivity to neurodegenerative and cerebrovascular processes.

Figures (7)

• Figure 1: Ideal SSL Biomarker. a) Alzheimer's Disease (AD) progression curves Jack2013TrackingBiomarkers showing the sensitivity of the biomarkers to characterize disease progression. The x-axis shows relative time of biomarker measurements; the y-axis shows the degree of biomarker abnormality. Ideally, MRI-based SSL biomarkers should detect decline earlier than classical imaging biomarkers (e.g., FreeSurfer). Unlike different modalities (e.g. PET), this increased sensitivity would not come at an additional acquisition cost. b) Biomarker information content (blue circle) with different training strategies. Generic SSL (top) ignores FreeSurfer (FS) and therefore does not learn all of the FS information. An ideal SSL biomarker (bottom) learns all FS information and additional relevant information. c) Proposed Method. Auxiliary (FreeSurfer) features are learned (and reconstructed). Next, the information in the residual is maximized via contrastive learning The full representation contains both auxiliary feature information and additional information not captured by auxiliary features. d) AD and amyloid classification performance. The top baseline SSL method significantly underperforms FreeSurfer, while our approach improves on FreeSurfer, indicating a more sensitive AD biomarker.
• Figure 2: Biomarkers Performance Results. Evaluations showing the sensitivity of various imaging biomarkers are shown for a variety of tasks including disease classification, AD conversion prediction, amyloid status classification, and age regression. Error bars are computed using the standard deviation of 1000 bootstrap samples.
• Figure 3: Phenotypic and Genetic Correlation across BAGs
• Figure 4: Neuropathology outcomes (X-axis) and BAG scores. P-values are shown for each BAG + pathology combination, where * indicates statistical significance ($p < 0.05$ after false discovery rate correction)
• Figure 5: Cell type enchrichment for R-NCE-BAG, FS-BAG, and SimCLR-BAG genes using scRNA-seq from lateral geniculate (LGN) and medial temporal gyrus (MTG). Significantly enriched cell types (using Bonferroni correction) are shown in red.