MCR-VQGAN generates synthetic tau PET from MRI with 0.93 SSIM, cutting costs

Tau PET imaging is critical for diagnosing Alzheimer's disease, but its use is limited by radiation exposure, high costs, and low availability. Researchers from Wake Forest University and collaborators propose MCR-VQGAN, a generative adversarial network that synthesizes high-fidelity tau PET images from routine structural T1-weighted MRI scans. The model builds on standard VQGAN architecture with three key enhancements: multi-scale convolutions to capture features at varying resolutions, ResNet blocks for deeper feature extraction, and Convolutional Block Attention Modules (CBAM) to focus on diagnostically relevant regions. Trained on 222 paired scans from the ADNI database, MCR-VQGAN outperformed cGAN, WGAN-GP, CycleGAN, and baseline VQGAN across all metrics.

MCR-VQGAN achieved MSE of 0.0056, PSNR of 30.65 dB, and SSIM of 0.9263. Critically, a CNN-based Alzheimer's classifier trained on real tau PET achieved 63.64% accuracy versus 65.91% on synthetic images, indicating that key diagnostic features are preserved. Regional SUVR-equivalent analysis across Braak-defined regions showed strong correlation (Pearson r = 0.78–0.88) and ICC values up to 0.838 in Braak V/VI. These results suggest MCR-VQGAN could serve as a surrogate for actual tau PET, dramatically lowering barriers to tau biomarker access and enabling broader Alzheimer's research and clinical screening without radiation or high costs.