Speaker
Description
Introduction & Methods:
Alzheimer's disease (AD) progression involves the accumulation and spread of tau pathology, measurable via tau PET imaging. Conventional analyses using regional standardized uptake value ratios (SUVRs) obscure fine-grained spatial heterogeneity and early tau propagation. We developed an Implicit Neural Representation (INR) model to reconstruct voxel-level tau distributions from regional measures. Using multimodal neuroimaging data from 61 ADNI participants, AV1451 Tau-PET scans were parcellated into 86 regions via the Desikan–Killiany atlas, and MRI-derived atrophy measures were obtained using FreeSurfer. The INR—a 16-layer fully connected network —took spatial coordinates, regional SUVRs, and z-scored atrophy as inputs to predict voxelwise SUVR.
Results:
The model achieved a mean per-subject correlation of R = 0.891 ± 0.111 and a pooled correlation of R = 0.959 (p < 0.01), with unbiased, normally distributed residuals. Excluding MRI atrophy reduced performance to R = 0.831, underscoring its importance. Region-wise accuracy averaged R = 0.936, with highest fidelity in AD-critical regions including the entorhinal cortex (R = 0.952), left thalamus (R = 0.974), and inferior parietal cortex (R = 0.958).
Conclusion:
Our study bridges coarse regional summaries and high-resolution voxel data, offering a scalable, biologically grounded tool for early AD detection, mechanistic modeling, and individualized tau pathology assessment.
