MCR-VQGAN: A Scalable and Cost-Effective Tau PET Synthesis Approach for Alzheimer's Disease Imaging

TL;DR

This work tackles the high cost, limited availability, and radiation exposure of tau PET imaging in Alzheimer’s disease by introducing MCR-VQGAN, a scalable MRI-to-tau PET synthesis framework. By integrating multi-scale convolutions, ResNet blocks, and CBAM into a VQGAN backbone, the model achieves superior reconstruction quality (lower MSE, higher PSNR, higher SSIM) than established GAN baselines. A downstream AD classification task shows that synthetic tau PET preserves diagnostically relevant features with accuracy comparable to real images, supporting clinical utility as a surrogate imaging modality. The approach promises broader accessibility and throughput for tau-pathology imaging in AD research and care, with future work aiming at larger datasets and multimodal extensions.

Abstract

Tau positron emission tomography (PET) is a critical diagnostic modality for Alzheimer's disease (AD) because it visualizes and quantifies neurofibrillary tangles, a hallmark of AD pathology. However, its widespread clinical adoption is hindered by significant challenges, such as radiation exposure, limited availability, high clinical workload, and substantial financial costs. To overcome these limitations, we propose Multi-scale CBAM Residual Vector Quantized Generative Adversarial Network (MCR-VQGAN) to synthesize high-fidelity tau PET images from structural T1-weighted MRI scans. MCR-VQGAN improves standard VQGAN by integrating three key architectural enhancements: multi-scale convolutions, ResNet blocks, and Convolutional Block Attention Modules (CBAM). Using 222 paired structural T1-weighted MRI and tau PET scans from Alzheimer's Disease Neuroimaging Initiative (ADNI), we trained and compared MCR-VQGAN with cGAN, WGAN-GP, CycleGAN, and VQGAN. Our proposed model achieved superior image synthesis performance across all metrics: MSE of 0.0056 +/- 0.0061, PSNR of 24.39 +/- 4.49 dB, and SSIM of 0.9000 +/- 0.0453. To assess the clinical utility of the synthetic images, we trained and evaluated a CNN-based AD classifier. The classifier achieved comparable accuracy when tested on real (63.64%) and synthetic (65.91%) images. This result indicates that our synthesis process successfully preserves diagnostically relevant features without significant information loss. Our results demonstrate that MCR-VQGAN can offer a reliable and scalable surrogate for conventional tau PET imaging, potentially improving the accessibility and scalability of tau imaging biomarkers for AD research and clinical workflows.

Figures (4)

• Figure 1: Overview of the proposed MCR-VQGAN framework. (a) The end-to-end synthesis pipeline, showing the encoding path (encoder and ResNet blocks), the vector quantization module, the decoding path (decoder and ResNet blocks), and the discriminator. (b) The multi-scale convolutional block used in the encoding path. (c) The standard convolutional block used in the decoding path. (d) The Convolutional Block Attention Module (CBAM) woo2018cbam architecture.
• Figure 2: Visual results of an ablation study on MCR-VQGAN architectural components for T1-weighted MR-to-tau PET synthesis. For two randomly selected axial slices, the figure compares the baseline VQGAN against progressively enhanced versions of the model (VQGAN+MC, VQGAN+MC+RB, and the full MCR-VQGAN). Each row displays the full image, a zoomed-in patch from the indicated region, and the corresponding absolute difference map.
• Figure 3: Qualitative comparison of five generative models for T1-weighted MR-to-tau PET synthesis. Results from two randomly selected axial slices are shown. For each slice, rows display the full image, a zoomed-in patch from the indicated boxed region, and the corresponding absolute difference map.
• Figure 4: MRI–to–tau PET synthesis across diagnostic groups (CN, MCI, AD). Columns: T1‐weighted MRI, ground‐truth tau PET, MCR-VQGAN output, and absolute difference map.