Deep learning-based estimation of multi-tracer PET phenotypes from structural MRI for A/T/N-related analysis and progression prediction in Alzheimer's disease.
Authors
Affiliations (4)
Affiliations (4)
- Department of Nuclear Medicine, General Hospital of Northern Theater Command, Shenyang, Liaoning 110016, China; College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, Liaoning 110167, China.
- Department of Nuclear Medicine, General Hospital of Northern Theater Command, Shenyang, Liaoning 110016, China.
- Department of Nuclear Medicine, General Hospital of Northern Theater Command, Shenyang, Liaoning 110016, China; College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, Liaoning 110167, China. Electronic address: [email protected].
- Department of Nuclear Medicine, General Hospital of Northern Theater Command, Shenyang, Liaoning 110016, China; College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, Liaoning 110167, China. Electronic address: [email protected].
Abstract
Early identification and staging of Alzheimer's disease (AD) require integrated assessment of amyloid deposition, tau pathology, and neurodegeneration. PET provides spatially resolved in vivo imaging biomarkers, but its routine use is constrained by cost, tracer availability, and radiation exposure, limiting access to complete A/T/N imaging profiles. We propose a conditional flow matching-based framework that estimates SUV-space PET-derived phenotypes related to <sup>18</sup>F-FDG, <sup>18</sup>F-florbetapir and <sup>18</sup>F-flortaucipir from a single T1-weighted MRI scan. To address limitations in SUV-level quantitative fidelity, single-tracer applicability, and external generalization, the framework combines a PET prior network with a time-conditioned velocity network and uses multi-constraint losses to preserve SUV accuracy, regional uptake consistency, and anatomical boundaries. The model was developed using 1,978 paired ADNI MRI-PET records from 991 participants and evaluated in a predefined ADNI test set and an independent hospital PET/MR cohort of 514 participants. In ADNI, voxel-level MAEs were 0.012, 0.034, and 0.021 SUV, with representative ROI Pearson correlations of 0.978, 0.842, and 0.781. In the hospital cohort, corresponding MAEs were 0.014, 0.037, and 0.024 SUV, with ROI correlations of 0.952, 0.803, and 0.834. In OASIS, generated PET-derived features improved CDR stratification and future cognitive decline prediction, increasing AUCs from 0.662 to 0.739 and from 0.569 to 0.643. These findings suggest that the method can provide PET-derived phenotypes related to A/T/N analysis from structural MRI when PET is unavailable, but should not be interpreted as a replacement for molecular PET assessment.