Related Experiment Video
Updated: Mar 21, 2026
Quantitative Magnetic Resonance Imaging of Skeletal Muscle Disease
Published on: December 18, 2016
A physics-driven neural network with parameter embedding for generating quantitative MR maps from weighted images
Lingjing Chen1,2, Chengxiu Zhang1,2, Yinqiao Yi1,2
1Shanghai Key Laboratory of Magnetic Resonance, School of Physics and Electronic Science, East China Normal University, Shanghai, China.
Background:
Traditional quantitative MRI (qMRI) techniques require acquisition of multiple weighted images to create a single quantitative mapping, which prolongs the scan time and limits their clinical applications. Deep learning (DL) has emerged as a promising solution for synthesizing quantitative maps from conventional weighted MRI images. However, existing DL methods often overlook the underlying physical principles inherent in MR signals, which inevitably compromises the performance and generalizability of the model.
Purpose:
To develop a deep learning-based approach that integrates MRI sequence parameters to improve the accuracy and generalizability of quantitative image synthesis from clinical weighted MRI.
Methods:
We proposed a physics-driven neural network that embeds MRI sequence parameters-repetition time (TR), echo time (TE), and inversion time (TI) - directly into the model via parameter embedding. This design enables the network to learn the underlying physical principles of MRI signal formation. The model takes conventional T1-weighted, T2-weighted, and T2-fluid-attenuated inversion recovery (T2-FLAIR) images as input and synthesizes T1, T2, and proton density (PD) quantitative maps. The model was trained on healthy brain MR images and evaluated on both internal and external test datasets.
Results:
The proposed method consistently achieved the best performance across all evaluation metrics compared with conventional deep learning methods (pGAN and U-Net). On the internal test set, the model achieved mean percentage errors (MPE) below 6% for T1, 10% for T2, and 5% for PD, with corresponding global voxel-wise mean absolute errors (MAE) of approximately 60 ms for T1, 10 ms for T2, and 30 ms for PD. Notably, the proposed model accurately generated quantitative maps for previously unseen pathological regions, highlighting its superior generalization capability.
Conclusion:
Incorporating MRI sequence parameters via parameter embedding allows the neural network to better learn the physical characteristics of MR signals, significantly enhancing the performance and reliability of quantitative MRI synthesis. This method shows great potential for accelerating qMRI and improving its clinical utility.
