Ultra-fast MRI for brain-age prediction in a real-world cognitive disorders clinic

Alzheimer's trials and memory-clinic workflows require frequent structural MRI, but standard 3D-T1 MPRAGE can be burdensome and motion-prone. The Wave-CAIPI sequence offers major time savings, yet it is unclear whether these ultra-fast scans can be used to derive dementia-related biomarkers from models that have been trained on standard scans.

We acquired paired scans from the standard and Wave-CAIPI MPRAGE protocols in 147 patients from a cognitive disorders clinic and generated measures of the brain's biological age. We applied six public brain-age pipelines (brainageR, DeepBrainNet, PyBrainAge, ENIGMA, pyment, MCCQR-MLP) to assess variability across software packages. We evaluated accuracy, interchangeability, cross-protocol agreement and clinical discrimination (subjective memory complaints versus neurodegenerative disorders), and tested effects of acquisition, diagnosis, and its interaction in a mixed-effects model.

Cross-protocol agreement was excellent across brain-age pipelines (intraclass correlation coefficient: ICC ≳ 0.90). Clinical discrimination was comparable between protocols, with effect sizes varying modestly by model-protocol combinations. Small, model-specific offsets and significant acquisition-by-diagnosis interactions were seen for some pipelines, consistent with a calibratable protocol effect; test-retest reliability was high and quality control measures were similar across protocols.

The ultra-fast Wave-CAIPI protocol could generate robust brain-age estimates in memory clinic patients, while markedly reducing scan time. When mixing ultra-fast and standard scans, a harmonization or transfer learning step is advisable to remove model-dependent offsets.