Self-supervised deep learning enables reliable 20-fold accelerated fMRI mapping

A multi-slice, self-supervised, physics-driven deep-learning method enabled reliable functional human brain mapping with 20-fold accelerated functional MRI acquisitions, according to a study evaluating both retrospective and prospective acceleration. The reconstructed images substantially improved on conventional approaches and supported attainable spatiotemporal resolutions at an acceleration level previously described as unachievable for functional mapping. Functional MRI often relies on accelerated acquisitions to obtain broad brain coverage and high spatiotemporal resolution, particularly when imaging the whole brain.

Greater acceleration, however, can leave residual aliasing and amplify noise in reconstructed data, creating a barrier as researchers seek increasingly demanding imaging performance. Physics-driven deep learning has shown promise in the wider MRI reconstruction community at high acceleration rates, where conventional methods can struggle to maintain output quality, but its application to fMRI presents additional challenges. One of those challenges is the need for fully sampled reference data when training supervised reconstruction models.

Such references are not readily available for high-resolution fMRI studies, limiting the practical use of conventional supervised deep learning. The reported results indicate that this approach reconstructed highly accelerated fMRI data while preserving functional maps that remained reliable at a 20-fold acceleration factor. The work assessed reconstruction under 20-fold retrospective and prospective accelerations, finding that the physics-driven deep-learning method substantially outperformed conventional reconstructions in the tested setting.

The findings therefore connect reconstruction performance directly to the practical objective of functional human brain mapping and indicate that high acceleration need not preclude usable maps in the evaluated acquisitions. The evidence comes from an accepted, open-access research article in Communications Engineering that was shared early for faster access and remains subject to further edits before the final version of record. The reported performance is specific to the evaluated retrospective and prospective 20-fold acceleration conditions; it does not establish that the same reliability will occur across every MRI system, acquisition protocol or application.

Within those limits, the study identifies self-supervised, physics-driven reconstruction as a possible route to more efficient high-resolution fMRI acquisitions that conventional approaches cannot readily achieve.

Subject of Research: Self-supervised deep-learning reconstruction for highly accelerated functional human brain mapping

Article Title: Overcoming MRI data acceleration barriers in functional human brain mapping with physics-driven deep learning

Article References: Demirel, Ö. B., Vizioli, L., Yaman, B., Moeller, S., Yacoub, E., Uğurbil, K., & Akçakaya, M. (2026). Overcoming MRI data acceleration barriers in functional human brain mapping with physics-driven deep learning. Communications Engineering. https://doi.org/10.1038/s44172-026-00784-4

Image Credits: AI Generated

DOI: 10.1038/s44172-026-00784-4

Keywords: functional MRI, brain mapping, deep learning, self-supervised learning, MRI reconstruction, data acceleration, spatiotemporal resolution, neural imaging, noise amplification, Overcoming, data, acceleration