A Structured Review and Quantitative Profiling of Public Brain MRI Datasets for Foundation Model Development

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Data Science

Background:

• Foundation models for brain MRI require large, diverse, and consistent datasets.
• Systematic assessments of public brain MRI data for foundation model development are lacking.

Purpose of the Study:

• To systematically assess the scale, diversity, and consistency of publicly available brain MRI datasets.
• To evaluate the impact of preprocessing variability on data harmonization.
• To inform the development of generalizable brain MRI foundation models.

Main Methods:

• Analysis of 54 public brain MRI datasets (538,031 scans) at dataset and image levels.
• Quantification of image properties (voxel spacing, orientation, intensity) across 14 datasets.
• Evaluation of preprocessing steps (normalization, bias correction, registration, etc.) and their impact on data.
• Feature-space analysis using a 3D DenseNet121 to assess residual covariate shift.

Main Results:

• Significant imbalances exist between large healthy cohorts and smaller clinical populations in public MRI datasets.
• Substantial heterogeneity in image properties (voxel spacing, orientation, intensity) across datasets.
• Standardized preprocessing improves within-dataset consistency but leaves residual inter-dataset differences.
• Residual covariate shift observed even after harmonization, indicating limitations of harmonization alone.

Conclusions:

• Public brain MRI resources exhibit considerable variability, posing challenges for foundation model development.
• Harmonization strategies alone are insufficient to overcome inter-dataset bias.
• Preprocessing-aware and domain-adaptive approaches are essential for creating robust and generalizable brain MRI foundation models.