A diffusion model converts a routine brain MRI into a metabolism map resembling an expensive PET scan and improves dementia diagnostic accuracy in a blinded physician reading
A diffusion model converts a routine brain MRI into a metabolism map resembling an expensive PET scan and improves dementia diagnostic accuracy in a blinded physician reading
On September 17, a group at the Technical University of Munich (TUM) presented DB-SUiT, a diffusion model that constructs a PET-like map of brain metabolism on the unfolded cortical surface from a standard MRI. In a blinded reading, the synthetic PET achieved 85.5% diagnostic accuracy for dementia, compared with 75.8% for conventional MRI and 95.2% for real PET. The model was tested on a cohort with a dementia subtype it had never seen and maintained accuracy without any site-specific fine-tuning.
FDG-PET uses fluorodeoxyglucose, a radioactive sugar, to detect early declines in glucose consumption by the brain: the most sensitive marker of dementia, appearing years before the tissue atrophy that MRI can reveal. PET, however, is expensive, requires a radioactive injection, and is not available everywhere. MRI is safe, inexpensive, and nearly ubiquitous. The question is how much of PET's diagnostic value can be extracted from a single MRI scan.
The TUM team posted a preprint of DB-SUiT on arXiv. Previous attempts, including an earlier model from the same laboratory, generated synthetic PET in the volumetric space of the brain, a cube of voxels (volumetric pixels). The cortex is heavily folded, and nearly all metabolic changes in dementia occur on its surface: in volumetric space this geometry is blurred. DB-SUiT is the first model to translate MRI to PET on the unfolded cortical surface, a mesh that FreeSurfer constructs from the MRI in the same way regardless of the hospital. Voxel intensity, by contrast, depends on the scanner's acquisition protocol. This change of coordinate system is what enables cross-site generalization. The convolutional component accounts for the spherical geometry of the cortex, while a transformer links distant cortical regions, which matters because dementia-related changes are distributed across the entire brain. The training procedure is called a "diffusion bridge": the path runs from the patient's MRI to the corresponding PET, rather than from random noise as in standard diffusion models.
On the open ADNI dataset (an archive of MRI and PET scans from elderly subjects) and on TUM's own cohort, DB-SUiT's synthetic maps reproduced real PET scans more accurately than competing methods across all six cortical lobes.
The decisive test came from two board-certified physicians with 14 years of experience: one read the synthetic PET blindly, the other read the MRI of the same cohort, which included cases of frontotemporal dementia. In the three-class task (normal, Alzheimer's disease, frontotemporal dementia) the synthetic PET yielded 85.5% accuracy versus 75.8% for MRI and 95.2% for real PET. In the simpler binary task (dementia present or absent) the synthetic PET reduced diagnostic errors by 58% compared with MRI, according to the authors. The laboratory's earlier volumetric model required site-specific fine-tuning for this kind of cross-site transfer; the new model managed without any adaptation. Cross-site generalization is a well-known bottleneck for such models: a similar system that predicted diagnosis and cognitive scores from MRI lost accuracy at a new hospital because of differences in clinical data across sites.
The laboratory's head, Christian Wachinger, previously created FastSurfer, a program that since 2020 has been reconstructing cortical surfaces from MRI as a replacement for FreeSurfer. The choice of the cortical surface as the coordinate system for PET synthesis continues this line of work.
The study has not undergone peer review, the synthetic PET falls short of real PET, and the clinical evaluation involved 62 patients from a single center. The authors themselves write:
DB-SUiT may serve as an auxiliary MRI-based assessment of cortical metabolism where PET is unavailable, complementing clinical and cognitive data.
An inexpensive MRI-based metabolic estimate like this could expand the number of hospitals where dementia can be caught at a stage when intervention is still able to slow the loss of memory and personality.