Foundation models for brain imaging: A systematic review.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Neuroscience

Background:

• Foundation models (FMs) are revolutionizing AI and medical imaging.
• Brain imaging is critical for neurological disease diagnosis but underrepresented in FM research.
• Existing surveys lack comprehensive analysis of FMs specifically for brain imaging.

Purpose of the Study:

• To provide the first comprehensive review of FMs for brain imaging.
• To analyze datasets, models, design choices, and training paradigms.
• To identify trends, vulnerabilities, and gaps in current FM research for brain imaging.

Main Methods:

• Systematic analysis of 161 brain imaging datasets and 143 FMs (up to Jan 2026).
• Evaluation of model architectures, training strategies, and task-specific performance.
• Identification of research gaps in tasks, pathologies, and clinical validation.

Main Results:

• FM development has shifted towards efficiency over size.
• MRI and CT dominate FM inputs; PET imaging is underexplored.
• Homogenization in architectures (e.g., Vision Transformers) and reliance on natural image backbones (e.g., SAM, CLIP) create vulnerabilities.
• Research is skewed towards brain cancer and neurodegenerative diseases, with a focus on anomaly classification and segmentation.
• Limited use of medical metrics and human expert evaluation.

Conclusions:

• Current FMs for brain imaging exhibit architectural vulnerabilities and lack diversity.
• Significant gaps exist in task coverage, pathology representation, and clinical validation.
• Future research should focus on developing domain-specific innovations, diverse datasets, and robust clinical evaluation methods.
• Recommendations are provided for building, evaluating, and deploying FMs in clinical and research settings.