Multi-parametric MRI synthesis for glioblastoma from quantitative MR fingerprinting: Quantitative synthetic neural...

Diagnosis and treatment of glioblastoma (GBM) rely on multiparametric MRI (mpMRI), but mpMRI is time-consuming and costly. Deep learning-based synthesis methods have been proposed to streamline acquisition; however, their generalizability is limited by variability in qualitative input contrasts across sites and scanners.

To overcome this limitation, we developed and evaluated a generalizable deep learning model that synthesizes mpMRI contrasts directly from quantitative magnetic resonance fingerprinting (MRF) maps in GBM patients. The proposed Quantitative Synthesis Network (QS-Net) employs a deeply supervised residual U-Net generator within an adversarial framework, combined with a two-stage training strategy to separate anatomical and pathological learning.

We collected MRF-derived T1 and T2 maps, along with conventional mpMRI sequences (T1w, T2w, T1-FLAIR, T2-FLAIR, and SWI), from 32 healthy volunteers and retrospectively from 18 GBM patient scans. The proposed QS-Net was initially trained on healthy volunteer data (20 scans for training, 12 for testing) to learn general anatomical features. Subsequently, it was fine-tuned using 9 GBM patient scans to adapt to pathological characteristics, with the remaining 9 patient scans reserved for independent testing. We compared the performance of QS-Net against three state of the art deep learning models: Res-Unet, conditional GAN, and Swin-Transformer, using both quantitative metrics (MAE, SSIM, and PSNR) and qualitative assessments. Additionally, we assessed the generalizability of the models by evaluating their external validation performance when trained with either conventional MRI or quantitative MRF inputs.

QS-Net outperformed the comparison models in synthesizing T1w, T2w, SWI, and T2-FLAIR images for GBM patients, achieving the best results across all quantitative metrics: MAE (1.18 ± 0.52, 1.01 ± 0.36, 1.05 ± 0.37, 1.45 ± 0.76), SSIM (0.934 ± 0.037, 0.939 ± 0.039, 0.934 ± 0.034, 0.926 ± 0.053), and PSNR (29.69 ± 3.21, 29.35 ± 2.29, 29.64 ± 2.58, 27.56 ± 3.39), respectively. Qualitative analysis demonstrated that QS-Net generated synthetic images with superior resemblance to ground truth, accurately delineating tumor boundaries and preserving intra-tumoral texture. Furthermore, the generalizability test revealed that models trained on standardized quantitative MRF input maps consistently outperformed models trained on vendor-specific qualitative MRI inputs across all architectures and metrics (p < 0.005).

We developed QS-Net, a deep learning model for high fidelity mpMRI synthesis from quantitative MRF maps, and demonstrated that this quantitative-input paradigm enables superior cross-vendor generalization over conventional qualitative MRI-based approaches.