Transfer Learning for Optimal Adaptation of Normative Neuroimaging models in Alzheimer's Disease

201 

Abstract Submission 

Camille Elleaume^1,2, Bruno Hebling Vieira¹, Dorothea Floris¹, Nicolas Langer^1,2

¹University of Zurich, Zürich, Switzerland, ²Neuroscience Center Zürich (ZNZ), Zürich, Switzerland

Camille Elleaume  
University of Zurich|Neuroscience Center Zürich (ZNZ)
Zürich, Switzerland|Zürich, Switzerland

Bruno Hebling Vieira  
University of Zurich
Zürich, Switzerland

Dorothea Floris  
University of Zurich
Zürich, Switzerland

Nicolas Langer  
University of Zurich|Neuroscience Center Zürich (ZNZ)
Zürich, Switzerland|Zürich, Switzerland

Alzheimer's Disease (AD) is a neurodegenerative disorder impacting memory and cognition, with associated hippocampal atrophyⁱ. Disentangling healthy-aging related shrinkage and AD-related pathological atrophy is crucial for early disease detection and understanding. Building on this research objective, recent advances in neuroimaging methods, such as normative modelling, offer promising avenues. These methods establish normative trajectories using large-scale datasets, allowing the assessment of deviations in clinical populations²⁻⁵. Despite their growing use and benefits, the application of normative models to independent clinical populations presents methodological challenges, such as site-specific variations in MRI scanner parameters⁴˒⁶. Our study addresses these challenges by using transfer learning to align pretrained models with new datasets. We explore how sample size and scanner variations impact model adaptation in healthy controls (HC) and examine the influence of sample size on accurately representing AD neuroanatomical deviations, assessed through a classifier's performance in differentiating between HC and AD individuals.

Utilizing FreeSurfer to extract hippocampal volume from T1-weighted MRI scans, normative models for left and right hippocampal volumes were established in the UK Biobank dataset (N=42,747). These models were transferred to the AIBL dataset (N=462, 12% AD) using 80% of healthy controls (HC) as an adaptation set (N_adj=322). The remaining 20% of HC, along with participants diagnosed with AD (N_test=140, with 42% AD), were used for testing (Fig. 1A-B). Bayesian Linear Regression implemented in the PCN toolkit⁷ was employed for normative modeling, incorporating age, sex, and image acquisition site as covariates. Deviation from the normative model was quantified as Z-scores (Fig. 1A).
To transfer models to AIBL, we used bootstrapping to sub-sample the adaptation set with sample sizes ranging from 5 to 100 subjects per site. For each sample size bin, model adaptation was evaluated through evaluation parameters (MSLL, SMSE, EV, Rho) on HC in the test set from AIBL (Fig. 2.A). To further evaluate the impact of both sample size and scanner variability in the adaptation set, we calculated Z-score differences between the full adaptation set (N=322) and obtained Z-scores across different sample sizes and scanners.
To highlight the improved AD classification achieved with transfer learning of normative models, we compared the Receiver Operating Characteristic Area Under the Curve (ROC-AUC) obtained from each normative model with those derived from raw hippocampal volumes. This comparison was conducted using a Logistic Regression classifier on the complete test set.

The results indicated that transfer learning reached an optimal plateau, as determined by model evaluation parameters, at approximately 20 samples in the adaptation dataset (Fig. 2A). This was further substantiated by the significant decrease in Z-scores differences compared to the full model (Fig. 2B). The scanner with a distinct magnetic field strength (i.e. 1.5T) exhibited a significant differences in Z-score deviations, which indicates a bias in transfer learning across different magnetic field strength for small adaption sets (Fig. 2B).
The AD-classification confirms that 20 samples in the adaption set is sufficient to reach the performance asymptote, while 9 samples already surpass the classification performance achieved with raw hippocampal volumes (Fig. 2C).

Our study indicates that for the AIBL data, a minimum of 20 samples are necessary to adapt the UK-based normative models to a new site and correctly map clinical deviations in the hippocampus for AD. Going forward, we aim to validate these results using independent datasets to ensure generalizability and extend this work by including a comprehensive study of all brain regions. This work can aid future studies to economically use resources for efficient biomarker development in AD.

Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) ¹

Bayesian Modeling ²

Classification and Predictive Modeling

Subcortical Structures

Aging

Computational Neuroscience

Degenerative Disease

Machine Learning

Modeling

Statistical Methods

STRUCTURAL MRI

Sub-Cortical