From Slices to Volumes: A Scalable Pipeline for Developing General-Purpose Brain MRI Foundation Models

Foundation models exhibit a remarkable capacity in extracting subtle features from brain MRI, demonstrating transformative potential for the precise diagnosis of brain diseases. Here, we introduce BrainMRIFM, a scalable pipeline for developing both slice and volume brain MRI foundation models that achieve computational efficiency alongside powerful representational capabilities. BrainMRIFM utilizes a novel slice-to-volume training paradigm: a slice model is initially pretrained for MRI slice representation, then its parameters are transferred to the corresponding volumetric model. The models were trained on an unprecedented dataset comprising 140,501 multi-modal MRI volumes from 42,297 subjects, aggregating data from 130 public datasets and 8 custom clinical centers. Our innovative in-house datasets contribute over 80% of the total patient cases for seven major brain diseases that lack open-source MRI data. Pretrained on 25 million high-quality MRI slices, the three slice foundation models demonstrated consistent performance improvements across all three downstream tasks, achieving a maximum accuracy improvement of 10.49% compared to ImageNet-initialized models. The SimMIM-SwinT volume MRI foundation model, building upon the high-performance slice foundation model, exhibited robust performance across seven brain tumor diagnostic tasks, with a maximum AUROC improvement of 12.19% compared to conventional ResNet50-based task-specific models. Additionally, attention maps and saliency visualizations confirmed the models capability to accurately localize pathological features. The BrainMRIFM pipeline and associated resources represent a significant advance toward developing brain MRI foundation models, with clear potential for extension to other neuroimaging applications.