Accessing Mesoscale Functional Organization from Standard Functional MRI Leveraging AI-Based Super-Resolution

Summary statement

Precision functional magnetic resonance imaging (fMRI) at the mesoscale is transforming our understanding of human brain functions by enabling the visualization of fine-grained functional organization in the human brain. Recent advances have shown that precision fMRI can reveal heterogeneous cortical architecture, resolve mesoscale processing streams underlying sensory perception, attention, and decision-making, identify functional alterations associated with neurological disorders, and help explain inter-individual variability in cognition. However, the widespread adoption of these approaches remains constrained by the limited availability of ultra-high-field MRI systems.

Here, we introduce f-LENNS (functional Localization Enhancement via Neural-Net Sharpening), an artificial intelligence (AI)-based framework that broadens access to mesoscale functional neuroimaging by enhancing the effective spatial resolution of conventional fMRI. The model is trained exclusively on domain-randomized synthetic MRI data and generalizes to diverse datasets without retraining : f-LENNS can super-resolve data obtained with different acquisition parameters and at different magnetic field strengths (3T and 7T), while minimizing hallucination artifacts. F-LENNS substantially improves the detectability and spatial fidelity of fine-scale functional organization defined by stimulus-evoked activity and accurately recovers mesoscale functional patterns from conventional-resolution fMRI. Importantly, the same model generalizes to individuals with atypical brain function, revealing disorder-related functional abnormalities without data- or disease-specific retraining.

Together, these findings demonstrate that domain randomization enables AI-based super-resolution models to overcome domain shifts arising from heterogeneous MRI hardware and acquisition protocols. Compared with conventional approaches, including interpolation-based upsampling, AFNI sharpening, and AI models trained on real functional and/or structural MRI data, f-LENNS consistently provides superior reconstruction of fine-scale functional organization. By enabling precision functional imaging in individual participants using widely available MRI systems, f-LENNS expands access to mesoscale brain mapping and represents an important step toward scalable, individualized functional neuroimaging and precision medicine.