Brain MRI: a network trained solely on synthetic images reconstructs the cortex from almost any clinical scan

The model works on MRI scans of varying resolution and contrast, and provides an uncertainty estimate that serves as a quality control measure, thereby enabling the analysis of large hospital databases.

The analysis of the surface of the cerebral cortex, which makes it possible to measure its thickness or curvature, is usually based on highly standardized research MRIs (T1 to 1 mm isotropic). The examinations performed routinely in hospitals, which are more varied, remain largely unexploited. On October 12, 2026, a team presented on arXiv, a neural network trained exclusively on synthetic images with random contrasts and resolutions, which allows it to process almost any type of MRI.

Uncertainty useful for quality control

On standard T1 MRIs, the model is on par with the best deep learning methods; on low-resolution and non-T1 examinations, it outperforms the only other method capable of this task. It also estimates, point by point, the uncertainty of the reconstructed surface: it monitors the overall quality of the examinations and highlights poorly reconstructed areas, which can be used as a quality control. Finally, it performs spherical registration to the fsaverage reference space with an accuracy comparable to existing tools.

Fast to inference, the tool could make it possible to analyze large clinical MRI databases that have already been built up and to look for disease signatures beyond the reach of conventional research studies.