This project studies cross-field magnetic resonance image synthesis from 1.5T, 3T, and 7T MRI. The model aligns anatomical information in a shared vector-quantized latent space and uses target field prompts to generate controllable high-field-like images in a single step.
The model aligns 1.5T, 3T, and 7T anatomical structure inside one vector-quantized representation.
Target field codes guide the decoder so the same backbone can produce 3T or 7T-like outputs on demand.
The restored figures below show why direct synthesis is more stable than multi-step field conversion.
7T MRI provides richer tissue contrast and finer structural detail, but scanners are less available and more expensive than 1.5T or 3T systems. MFS-VQ-VAE reduces this gap by learning a shared vector-quantized latent representation across multiple magnetic field strengths.
The attached page centered on the idea of controllable synthesis with Prompt Attention: once the source image is encoded into a shared latent space, a target field-strength code guides the decoder toward the desired output domain. This lets the model directly produce 3T or 7T-like images without relying on multi-stage cascades.
The encoder first maps the source MRI into a continuous latent feature. Vector quantization then projects it into a shared codebook space, and the Prompt Attention Module conditions that representation with the target field code before decoding the final synthesized image.
Shared VQC latent space construction: joint 1.5T to 3T and 3T to 7T synthesis constrains multi-field latent vectors into a unified representation space.
MFS-VQ-VAE architecture: encoder, vector quantization, Prompt Attention Module, and decoder for target-field reconstruction.
Field-aware control. Prompt Attention introduces an explicit magnetic field-strength condition into generation, so the same latent representation can be decoded toward different target domains.
Cross-domain alignment. The VQ codebook encourages a shared anatomical basis while preserving enough flexibility for domain-specific appearance and contrast reconstruction.
The original attached page included interactive task switching for metrics. That interaction is preserved here while matching the site's Research page visual system.
| Task | SSIM | PSNR | LPIPS | Sharpness |
|---|---|---|---|---|
| MFS-VQ-VAE - 1.5T to 3T | 0.973 +/- 0.006 | 32.4 +/- 1.7 | 1.7 +/- 0.4 | - |
| MFS-VQ-VAE - 3T to 7T | 0.927 +/- 0.005 | 28.6 +/- 0.6 | 5.7 +/- 0.4 | 1.90 +/- 0.23 |
| Single-step 1.5T-like to 7T - Group T | 0.931 +/- 0.013 | 27.8 +/- 1.1 | 6.4 +/- 1.0 | 2.31 +/- 0.04 |
| External validation - 3T to 7T | 0.871 +/- 0.002 | 25.2 +/- 1.0 | 23.1 +/- 3.4 | - |
The table values are carried over from the attached standalone page and reorganized into the current site style.
Visual comparisons and error maps for 1.5T to 3T and 3T to 7T synthesis against representative baselines.
The restored visual set shows that the proposed model preserves anatomical detail more faithfully while reducing prediction error compared with representative baselines.
It also highlights why direct 1.5T-like to 7T synthesis is preferable to cascaded conversion, and how 7T-like outputs help suppress distortion and intensity inhomogeneity.
Single-step synthesis avoids cumulative degradation compared with multi-step 1.5T-like to 3T to 7T synthesis.
7T-like synthesis from 3T MRI can reduce artifacts, geometric distortion, and intensity inhomogeneity while retaining high-field contrast.
You can copy the current BibTeX entry below and update the venue field when the publication details are finalized.
@article{xie2025mfsvqvae,
title = {MFS-VQ-VAE: Synthesizing 7T-like MR Images from 1.5T MR Images with Prompt Attention in VQ-VAE},
author = {Xie, XingHe and Han, LuYi and Sun, Yue and Liu, MengZhu and Jiang, MingFeng and Chang, HingChiu and Tan, Tao},
journal = {IEEE Transactions and Journals},
year = {2025},
note = {Update venue information after publication}
}