PWT: Advancing Alzheimer’s PET Synthesis from MRI with Enhanced Pathological Feature Preservation

Alzheimer’s disease (AD) is a neurodegenerative condition that significantly impairs cognitive function. Early and accurate diagnosis is critical for effective intervention. While MRI provides detailed structural information, it lacks the functional insights that PET scans offer, particularly in detecting metabolic changes associated with AD. However, PET imaging is costly and involves radiation exposure, limiting its frequent use. In this work, we propose an enhanced framework, PWT (PASTA With TransUNet), to generate synthetic PET images from MRI data, integrating transformer-based attention mechanisms to improve the preservation of pathological features. By replacing the UNet architecture in the original PASTA model, our approach captures both local and global features more effectively, resulting in higher-quality synthetic PET images. Evaluation on a dataset of MRI and PET images shows that PWT with an embedding dimension of 512 outperforms the original PASTA model in metrics such as FID, SSIM, and PSNR, offering a promising tool for improving AD diagnosis while reducing the need for direct PET scans.