MRI-Based Brain Age Estimation Using Supervised Contrastive Learning

Abstract

   Brain age estimation models aim to accurately assess a subject’s biological brain age based on neuroanatomical features. Various factors, including neurodegenerative diseases, can cause accelerated brain aging and measuring this phenomena could serve as a biomarker for clinical applications. While promising results have been achieved in previous works, there is no consensus on an optimal model for accurate prediction or clinical utility. This thesis proposes using supervised contrastive learning with Rank-N-Contrast (RNC) loss and Grad-RAM for explainability utilizing structural T1w MRI data. Results indicate that the supervised contrastive strategy significantly outperformed ResNet models, achieving a mean absolute error of 4.28 years and an R² of 0.93 with a limited dataset of aging subjects. Benchmark comparisons with state-of-the-art models demonstrated that the supervised contrastive approach achieved comparable performance in our test sample, particularly among older age groups. The Grad-RAM analysis revealed anatomically relevant regions associated with aging, with the more nuanced capabilities exhibited by the supervised contrastive learning approach. Analyses of disease populations revealed significantly higher brain age gaps in Alzheimer’s Disease (AD) patients, correlating strongly with ADAS-cog scores (r = 0.37, p = 0.0098), suggesting its potential as a biomarker for assessing AD severity. However, no significant correlation was found between brain age gap and UPDRSIII scores in Parkinson’s disease (PD) patients, suggesting that anatomical changes caused by PD may be obscured by natural aging or that T1w MRIs might not be optimal contrast for visualizing subcortical structures. The Grad-RAM focuses on regions known to be linked to AD and PD, but with little difference to the healthy population. Our study demonstrates the potential of supervised contrastive learning with an RNC loss in brain age prediction, highlighting its ability to outperform other models in smaller datasets.