Estimating SV2A PET-Derived Synaptic Density from Quantitative MRI: 3D U-Net Reaches Pearson Correlation of 0.8838 in Gray Matter

The study shows that non-invasive quantitative MRI parameters alone can reconstruct synaptic density images closely resembling SV2A PET, reaching a Pearson correlation of 0.8838 and SSIM of 0.9832 in gray matter. Previously, SV2A synaptic density imaging relied on PET radioligands such as [18F]UCB-H, which carry radiation exposure and scalability limits; this work shifts the input to qMRI maps (R1, R2*, MTsat, PD) and reports voxel-level agreement metrics. Based on two combined datasets totaling n = 74, with subject-level splits of 70% training, 15% validation, and 15% test to prevent data leakage, evaluated with MSE, SSIM, PSNR, and Pearson correlation, reporting means and standard deviations.

On the same task, deep learning models outperformed classical machine learning models, with 3D U-Net architectures performing most consistently. The study systematically compared SVR, PLS, Elastic Net, and Random Forests against U-Net, ResUNet++, and Pix2Pix-like conditional GANs, testing 2D, 2.5D, and 3D input configurations and incorporating the Schaefer atlas as a structural prior, yielding an empirical ranking across architectures and input dimensionality. The best classical model was Elastic Net (R² = 0.50, RMSE = 0.448, MAE = 0.331), while deep learning improved prediction accuracy; training curves showed validation loss plateauing after approximately 50 epochs, indicating overfitting.

The study proposes and tests an end-to-end pipeline from qMRI to PET-like synaptic density images, including ComBat cross-dataset harmonization, MNI spatial normalization, subject-wise z-scoring of qMRI, and cerebellar-uptake normalization of PET followed by global z-scoring. The pipeline combines handling of multi-site/multi-acquisition variability with a voxel-wise prediction task, offering an actionable technical path for replication and extension on independent datasets. The methods description covers harmonization, spatial normalization, standardization, and evaluation, and states that data and code are available upon reasonable request to Mohamed Ali Bahri.

The authors interpret the results as evidence that qMRI parameters indirectly encode microstructural information related to synaptic organization, supporting the development of non-invasive, scalable alternatives to PET. This interpretation elevates imaging prediction performance into hypothesis-level support for an association between qMRI microstructural parameters and synaptic density, rather than merely reporting fit metrics. The authors also note the limited dataset size and variability across acquisition settings, and explicitly state that results should be interpreted in that light, with generalization still requiring more diverse training data and evaluation on independent datasets.

The results target research settings with paired multimodal qMRI and SV2A PET data, applicable to synaptic density estimation in gray matter within mixed cohorts of Alzheimer's disease, subjective cognitive decline, and healthy controls; their value lies in showing that qMRI can serve as a non-invasive surrogate signal source for PET and in providing a starting point for validation on independent datasets and for improving regularization and generalization strategies. The authors note that future work should also address PET normalization issues, including reducing reliance on cerebellar reference regions or mitigating associated biases.

The dataset is limited in size and varies across acquisition settings, and training curves show validation loss plateauing after approximately 50 epochs, indicating overfitting, so whether current performance generalizes to new sites, scanners, or independent cohorts remains an open question; PET normalization relying on a cerebellar reference region may introduce bias, which the authors also list as unresolved; additionally, the relationship between gray-matter evaluation results and non-gray-matter or whole-brain performance, and the specific meaning of metric distribution differences across clinical groups, still require further analysis to clarify.