NeuroSTORM

Functional MRI (fMRI) measures brain activity through blood-oxygen-level-dependent (BOLD) signals, producing 4D data (three spatial dimensions over time) that is central to cognitive neuroscience and clinical neuroimaging. Yet most fMRI analysis pipelines remain bespoke: studies typically reduce the raw 4D volumes to handcrafted features such as region-of-interest time series or functional connectivity matrices, then train narrow, single-task models that transfer poorly across cohorts, scanners, and sites. This fragmentation contributes to well-documented reproducibility and generalization problems in brain imaging.

NeuroSTORM (Neuroimaging Foundation Model with Spatial-Temporal Optimized Representation Modeling) addresses this by learning generalizable representations directly from full 4D fMRI volumes, rather than from pre-extracted summaries. The model was developed by a multi-institution collaboration led by the Chinese University of Hong Kong with Massachusetts General Hospital, Yonsei University, the University of Sydney, Peking University, the University of Georgia, Lehigh University, and Emory University, first released as a preprint in June 2025 and published in Nature Biomedical Engineering in 2026.

Its central contribution is a scalable pretraining recipe paired with lightweight task adaptation, allowing a single pretrained backbone to be transferred efficiently to a wide range of downstream brain-imaging problems instead of training a new network for each study.

#Key Features

• Direct 4D volume modeling: NeuroSTORM operates on raw voxel-based 4D fMRI rather than ROI time series or connectivity matrices, preserving spatial and temporal structure that handcrafted features discard.
• Shifted-Window Mamba backbone: A Shifted-Window Mamba (SWM) design combines linear-time state-space sequence modeling with a shifted scanning strategy, making it computationally feasible to process long, high-dimensional 4D volumes.
• Large-scale, diverse pretraining: The model is pretrained on 28.65 million fMRI frames from more than 50,000 subjects across multiple centers and ages spanning 5 to 100 years.
• Lightweight task adaptation: In addition to full fine-tuning, NeuroSTORM supports task-specific prompt learning, enabling fast transfer to new tasks with modest compute and data.
• Broad task coverage: A single backbone supports five evaluation tasks — demographic prediction, phenotype prediction, disease diagnosis, re-identification (fMRI retrieval), and task-state classification.

#Technical Details

NeuroSTORM is a self-supervised foundation model built on a Shifted-Window Mamba backbone, released in several configurations (including Base and Large variants, as well as low-resolution and long-sequence variants tuned for different compute and input regimes). Pretraining uses self-supervised objectives — masked autoencoding together with contrastive learning — applied directly to 4D volumes, drawing on 28.65 million frames from over 50,000 subjects across multiple sites and a wide age range. Downstream evaluation spans large open cohorts such as UK Biobank, the Adolescent Brain Cognitive Development (ABCD) study, the Human Connectome Project, ABIDE, and ADHD-200, alongside multi-hospital clinical datasets covering 17 diagnoses. Across these benchmarks NeuroSTORM reports consistent improvements over prior fMRI methods, with strong results on disease diagnosis and cognitive phenotype prediction and demonstrated clinical utility on data from hospitals in the United States, South Korea, and Australia.

#Applications

NeuroSTORM is intended as a reusable backbone for both neuroscience research and clinical neuroimaging. Researchers can fine-tune or prompt-adapt the pretrained model for demographic and cognitive phenotype prediction, brain-state decoding, and subject re-identification, reducing the per-study engineering and labeling burden. Its demonstrated performance on multi-hospital cohorts across three countries points toward clinical decision-support applications such as psychiatric and neurological disease diagnosis from fMRI, where consistent, transferable representations across sites and scanners are particularly valuable.

#Impact

NeuroSTORM extends the foundation-model paradigm to 4D functional neuroimaging, a modality that has lagged behind protein, genomic, and natural-image domains in part because of the cost of modeling full spatiotemporal volumes. By showing that a single backbone pretrained on tens of millions of frames can transfer across demographic, phenotypic, diagnostic, retrieval, and state-classification tasks, it offers a concrete template for addressing the reproducibility and transferability challenges that have limited fMRI machine learning. The code is released under the Apache-2.0 license at the CUHK-AIM-Group repository, supporting both pretraining and fine-tuning workflows; broad reuse will depend on the continued availability and documentation of the pretrained checkpoints.