SLIM-Brain: A Data- and Training-Efficient Foundation Model for fMRI Data Analysis

TL;DR

SLIM-Brain tackles the data- and training-efficiency bottlenecks of fMRI foundation models by combining a lightweight temporal extractor with a selective 4D voxel-level encoder. Through a two-stage adaptive pipeline—ranking informative windows with a global MAE and then applying a 4D Hiera-JEPA encoder to top-$k$ segments—it preserves fine-grained spatial details while drastically reducing memory and compute. It achieves state-of-the-art results across seven downstream tasks, using only ~4k pretraining sessions and about 30% of the memory of competing voxel-level methods. The approach demonstrates strong data and model scaling, robust OOD transfer, and neurobiologically meaningful voxel-level representations aligned with Neurosynth maps, suggesting practical atlas-free voxel-level foundation modeling for fMRI.

Abstract

Foundation models are emerging as a powerful paradigm for fMRI analysis, but current approaches face a dual bottleneck of data- and training-efficiency. Atlas-based methods aggregate voxel signals into fixed regions of interest, reducing data dimensionality but discarding fine-grained spatial details, and requiring extremely large cohorts to train effectively as general-purpose foundation models. Atlas-free methods, on the other hand, operate directly on voxel-level information - preserving spatial fidelity but are prohibitively memory- and compute-intensive, making large-scale pre-training infeasible. We introduce SLIM-Brain (Sample-efficient, Low-memory fMRI Foundation Model for Human Brain), a new atlas-free foundation model that simultaneously improves both data- and training-efficiency. SLIM-Brain adopts a two-stage adaptive design: (i) a lightweight temporal extractor captures global context across full sequences and ranks data windows by saliency, and (ii) a 4D hierarchical encoder (Hiera-JEPA) learns fine-grained voxel-level representations only from the top-$k$ selected windows, while deleting about 70% masked patches. Extensive experiments across seven public benchmarks show that SLIM-Brain establishes new state-of-the-art performance on diverse tasks, while requiring only 4 thousand pre-training sessions and approximately 30% of GPU memory comparing to traditional voxel-level methods.

Figures (7)

• Figure 1: Performance & Pretraining size.Our method SLIM-Brain reaches 64.5% age-classification accuracy with only about 4 thousand sessions in pretraining.
• Figure 2: (a) ROI-based. Atlas parcellation coarsely downsamples space, introducing atlas bias and erasing voxel-level detail. (b) Volume sliding-window pipelines. Resolution is retained, but fixed window lengths (e.g., 40 frames) with simple averaging dilute transient events and miss cross-window dynamics. (c) Ours. A lightweight global pass ranks windows; the top small windows (e.g., 5 frames) are concatenated to a set (e.g., 40 frames) and encoded with a 4D encoder and fused with global features, yielding efficient multi-granularity representations with fine spatial semantics and long-range spatiotemporal structure.
• Figure 3: SLIM-Brain Pipeline. (a) Atlas-free 4D fMRI input at voxel resolution. (b) A lightweight ViT processes the full recording to produce robust global features. (c) Using the same masking mechanism, a cross-window masked-reconstruction score ranks temporal windows and selects informative segments. (d) Selected windows are routed to a voxel-level 4D Hiera encoder to extract fine-grained representations without any predefined atlas.
• Figure 4: Scaling study. Performance on AD vs. CN at varying amounts of pre-training data and model parameter sizes..
• Figure 5: Ablation on window selection (HCP sex). Classification accuracy (%) / F1-score (%) at different frame budgets.