The Data Behind the Model: Gaps and Opportunities for Foundation Models in Brain Imaging

Salah Ghamizi, Georgia Kanli, Magali Perquin, Olivier Keunen

Abstract: Foundation Models (FMs) have revolutionized machine learning in medical imaging, yet their application to brain imaging remains limited and fragmented. Despite the availability of diverse and extensive neuroimaging datasets, most FM research has focused narrowly on a handful of tasks, mainly tumor classification and segmentation, while neglecting prevalent neurological disorders such as ADHD and early-stage Parkinson’s disease. In this work, we present the largest and most comprehensive atlas of brain imaging datasets to date, comprising 151 datasets and over 541k volumetric imaging studies across a wide range of modalities and pathologies. Our meta-analysis of 86 brain imaging FMs reveals a disproportionate reliance on structural MRI and a small set of popular datasets, along with critical blind spots in both disease coverage and imaging modalities. We identify systemic challenges, including inconsistent model evaluation protocols, heterogeneous data formats, and limited availability. All of which hinder reproducibility, scalability, and clinical translation. Our publicly available atlases pave the way for more robust, scalable, and clinically meaningful FMs in brain imaging.

External IDs:doi:10.1007/978-3-032-07845-2_11