Standard fMRI reaches mesoscale brain maps: f-LENNS super-resolves to 1 mm from synthetic-data training, no retraining across 3T and 7T
Synopsis
f-LENNS super-resolves conventional-resolution fMRI magnitude images to 1 mm isotropic, letting mesoscale functional organization be recovered without an ultra-high-field scanner. The model is trained only on domain-randomized synthetic MRI, with no adversarial or generative components, so it applies to experimental data across field strengths, sequences and resolutions without retraining. In visual cortex, color-, motion- and disparity-selective sites localized from super-resolved maps came closer to the original high-resolution acquisitions than interpolation upsampling and sharpening did.
Interpretation
A publicly available AI post-processing framework raises the effective spatial resolution of conventional fMRI to 1 mm isotropic, recovering mesoscale functional organization from ordinary scanner data. Previously such fine functional maps typically required ultra-high-field MRI systems plus extensive signal averaging, and ultra-high-field scanners are scarce while scan time is limited. In 7T data from eleven neurotypical participants, original 1 mm data were downsampled to 2 mm or 3 mm and then super-resolved; image fidelity metrics beat standard upsampling and AFNI sharpening (t(10)>11.18, p<10⁻³, corrected for multiple comparisons).
The model processes each fMRI volume independently and receives only real-valued magnitude images resampled to a common 1 mm grid, without complex-valued data, phase or raw k-space. Earlier AI super-resolution methods mostly trained on real functional or structural MRI, degraded when acquisition conditions changed, and were typically confined to a single site, protocol or stimulus paradigm. Training data were synthetic images generated from anatomical segmentations of 5,241 subjects across 10 public MRI datasets, with voxel sizes sampled between 2 and 4.5 mm so the network never learns the true input voxel size.
One pretrained model works without retraining on 3T and 7T data, different pulse sequences and different stimulus paradigms, and extends to people with amblyopia. On 2 mm input, f-LENNS matched a model trained on real data with a minimal domain gap (t(10)=2.82; p=0.07) and clearly outperformed a real-data model with a larger domain gap (t(10)=41.94; p<10⁻³). On the harder 3 mm input, the closely matched real-data model performed better (t(10)=6.78; p<10⁻³), indicating f-LENNS trades some condition-specific peak performance for generality across conditions.
Perspective
Validation centers on visual cortex: 7T data from eleven neurotypical participants, data from ten participants with amblyopia, and 3T data from six participants, with tasks localizing color-, motion- and disparity-selective sites. The model is currently developed and validated on data from anatomically intact brains, and the authors propose systematic future evaluation on brains with structural abnormalities.
The authors note that f-LENNS is not designed to fully resolve differences in functional map spatial organization arising from scanner platforms, pulse sequences (such as BOLD versus VASO) or field strength, and that a model tailored to each acquisition could reach a higher performance ceiling. Frame-wise processing is a deliberate tradeoff that forgoes potentially useful temporal information to stay compatible with standard fMRI analysis. Brain regions and cognitive domains beyond the visual system remain to be evaluated.