Coordinate-Based Neural Representation Enabling Zero-Shot Learning for 3D Multiparametric Quantitative MRI

TL;DR

The proposed SUMMIT, an innovative imaging methodology that includes data acquisition and an unsupervised reconstruction for simultaneous multiparametric qMRI, introduces a novel zero-shot learning paradigm for multiparametric imaging applicable to various medical imaging modalities.

Abstract

Quantitative magnetic resonance imaging (qMRI) offers tissue-specific physical parameters with significant potential for neuroscience research and clinical practice. However, lengthy scan times for 3D multiparametric qMRI acquisition limit its clinical utility. Here, we propose SUMMIT, an innovative imaging methodology that includes data acquisition and an unsupervised reconstruction for simultaneous multiparametric qMRI. SUMMIT first encodes multiple important quantitative properties into highly undersampled k-space. It further leverages implicit neural representation incorporated with a dedicated physics model to reconstruct the desired multiparametric maps without needing external training datasets. SUMMIT delivers co-registered T1, T2, T2*, and quantitative susceptibility mapping. Extensive simulations and phantom imaging demonstrate SUMMIT's high accuracy. Additionally, the proposed unsupervised approach for qMRI reconstruction also introduces a novel zero-shot learning paradigm for multiparametric imaging applicable to various medical imaging modalities.

Figures (5)

• Figure 1: Overview of SUMMIT. (a) Diagram of the MRI sequence for data acquisition. (b) The reconstruction process of SUMMIT. The 3D coordinates $\gamma$ are fed into the encoding module $H_\phi$ and $MLP_\theta$ to simultaneously estimate the multiparametric quantitative maps and sensitivity maps. The predicted multiple-dimensional k-space $\hat{S}_{k}$ is generated through the Bloch equation and the physical forward model based on the estimated maps. The weights in the MLP and encoding module are optimized by minimizing the data consistency (DC) loss and weighted nuclear norm minimization (WNNM) loss. (c) The denoising procedure of WNNM module. First, similar patches are extracted through block matching. These patches are then vectorized and concatenated along one dimension to construct a low-rank matrix. Finally, singular value decomposition (SVD) is performed to obtain a series of singular values $\sigma$. The WNNM loss is a weighted sum of the singular values.
• Figure 2: Comparison between SUMMIT and LRT on the 4$\times$ retrospective simulation. (a) The ground truth maps used for simulation. (b-c) The reconstructed quantitative maps and corresponding errors of SUMMIT and LRT.
• Figure 3: The performance variation of SUMMIT and LRT on the 4$\times$ retrospective simulation at different SNRs. SUMMIT shows lower NRMSE on $T_1$, $T_2$, and $T_2^*$ maps compared with LRT.
• Figure 4: Comparison between SUMMIT and LRT on the 8$\times$ retrospective simulation. (a) The ground truth maps used for simulation. (b-c) The reconstructed quantitative maps and corresponding errors of SUMMIT and LRT.
• Figure 5: Quantitative evaluation on a standard phantom. (a) SUMMIT shows good image quality on the $T_1$, $T_2$, and $T_2^*$ maps, and agrees well with the references (Ref). Those maps from LRT show non-uniformity on the phantom due to the potential reconstruction errors. (b) The correlation analysis indicates SUMMIT correlates well with the reference sequences, as denoted by the coefficient of determination and slope approaching 1. The red line represents the linear regression fitting and the black dashed line corresponds to y=x.