Adaptive Neighborhoods in Contrastive Regression Learning for Brain Age Prediction

Jakob Träuble, Lucy V Hiscox, Curtis Johnson, Carola-Bibiane Schönlieb, Gabriele S Kaminski Schierle, Angelica I Aviles-Rivero

Keywords: self-supervised learning, contrastive learning, regression, brain age prediction, neuroscience

TL;DR: We propose a dynamic contrastive learning method for brain age prediction using stiffness maps, adapting local sample neighborhoods during training to handle non-uniform data and outperform state-of-the-art models in neuroimaging.

Abstract: In neuroimaging, accurate brain age prediction is key to understanding brain aging and early neurodegenerative signs. Recent advancements in self-supervised learning, particularly contrastive learning, have shown robustness with complex datasets but struggle with non-uniformly distributed data common in medical imaging. We introduce a novel contrastive loss that dynamically adapts during training, focusing on localized sample neighborhoods. Additionally, we incorporate brain stiffness, a mechanical property sensitive to aging. Our approach outperforms state-of-the-art methods and opens new directions for brain aging research.

Submission Number: 27