Foundation models for brain imaging: A systematic review.

Authors

Abstract

Foundation models (FMs), large neural networks pretrained on extensive and diverse datasets, have revolutionized artificial intelligence and demonstrated significant promise in medical imaging by enabling robust performance with limited labeled data. Although numerous surveys have reviewed the application of FMs in healthcare, brain imaging remains underrepresented, despite its critical role in the diagnosis and treatment of neurological diseases using modalities such as magnetic resonance imaging (MRI), computed tomography (CT), and positron emission tomography (PET). To address this gap, we present the first comprehensive and curated review of FMs for brain imaging. We systematically analyze 161 brain imaging datasets and 143 FMs up to Jan, 2026, providing insights into key design choices, training paradigms, and optimizations driving recent advances. Our review highlights that the race for larger models has stabilized in 2026 towards more efficient models. FMs for brain imaging heavily rely on MRI (92%) and CT (57%) inputs, while PET imaging remains vastly underexplored (supported by only 15% of models). Our study also demonstrates architectural vulnerabilities caused by homogenization and lack of diversity, with Vision Transformers utilized in 48% of visual encoders, and models predominantly built by patching pre-existing natural image backbones like SAM (19%), and CLIP (12%) rather than utilizing native domain-specific 3D medical imaging innovations. For each of the eight tasks of the study the systematic review identifies the best models and discusses their innovations. Our study also uncovers critical gaps in the tasks, pathologies and clinical validation. We demonstrate that the literature is disproportionately skewed toward brain cancer research (37% of models) and neurodegenerative diseases (24%), and discuss the potential causes and remedies. Similarly, tasks are heavily weighted toward anomaly classification (44%) and segmentation (32%), leaving areas like mental health and image synthesis underrepresented. Besides, most models rely exclusively on traditional machine learning metrics (e.g., DICE or SSIM) rather than medically relevant measures, and only seven out of the 143 models incorporated human expert evaluations to verify real-world utility. Our systematic review concludes by outlining future research directions to advance FMs in brain imaging and actionable recommendations to build better FMs and to evaluate and deploy them in clinical and research settings.

Tags

Topics