MFS-VQ-VAE: Synthesizing 7T-like MR images from 1.5T MR images with Prompt Attention in VQ-VAE

Magnetic resonance imaging (MRI) images exhibit significant quality differences at varying magnetic field strengths, such as 1.5T, 3T, and 7T. Higher magnetic field strengths generally provide better tissue contrast and richer structural details, thereby enhancing diagnostic accuracy. However, the high cost and limited accessibility of 7T MRI scanners restrict their clinical adoption, while 1.5T and 3T MRI systems are more commonly used in routine medical practice. To bridge the quality gap between low-field (1.5T/3T) and high-field (7T) MRI images, this study proposes a dynamic MRI synthesis framework that integrates a vector-quantized variational autoencoder (VQ-VAE) with a shared latent space to enable cross-field image synthesis. The proposed model extracts shared feature representations from MRI images acquired at different magnetic field strengths and, through a proposed prompt attention module combined with a field strength code, maps these features to the desired field strength, thereby achieving single-step synthesis across arbitrary field strengths. Experimental results demonstrate that the proposed method outperforms existing approaches across multiple key quantitative metrics, showing particularly strong performance in synthesis tasks between 1.5T and 3T, as well as between 3T and 7T. Furthermore, the framework enables high-quality one-step synthesis from 1.5T to 7T, effectively mitigating the cumulative errors commonly observed in multi-step synthesis methods.