NLCG-Net: A Model-Based Zero-Shot Learning Framework for Undersampled Quantitative MRI Reconstruction

TL;DR

This work addresses biases and error propagation in the conventional two-step qMRI pipeline by introducing NLCG-Net, a model-based zero-shot framework that uses nonlinear conjugate gradient optimization with a scan-specific U-Net regularizer to directly estimate $T_1$ and $T_2$ maps from undersampled k-space. The method formulates qMRI reconstruction as a data-consistent optimization with a mono-exponential signal model for T1/T2 and solves it via an unrolled architecture where data- consistency steps are complemented by neural regularization. It demonstrates superior T2 mapping fidelity at high accelerations and comparable T1 performance without external training data, highlighting its potential for high-speed, bias-resistant quantitative MRI. The approach enables flexible, high-fidelity qMRI reconstruction in clinical and research settings where large-scale training data are unavailable.

Abstract

Typical quantitative MRI (qMRI) methods estimate parameter maps in a two-step pipeline that first reconstructs images from undersampled k-space data and then performs model fitting, which is prone to biases and error propagation. We propose NLCG-Net, a model-based nonlinear conjugate gradient (NLCG) framework for joint T2/T1 estimation that incorporates a U-Net regularizer trained in a scan-specific, zero-shot fashion. The method directly estimates qMRI maps from undersampled k-space using mono-exponential signal modeling with scan-specific neural network regularization, enabling high-fidelity T1 and T2 mapping. Experimental results on T2 and T1 mapping demonstrate that NLCG-Net improves estimation quality over subspace reconstruction at high acceleration factors.

Figures (5)

• Figure 1: $\mathbf{PFCM}$ forward operators. To derive the k-space expression, $\vec{x}$ is firstly converted to signal intensity following T2/T1 quality, then multiplied with coil sensitivity maps to form each coil image. After that, Fast Fourier Transform $\mathbf{F}$ is performed to transform data into k-space, then finally applying mask $\mathbf{P}$ to reproduce downsample process.
• Figure 2: a) NLCG-Net framework. The model performs 800 NLCG iterations first to initialize, then passes to Unroll blocks. A NLCG data consistency layer and U-Net are deployed in each block to perform iterative optimization. b) Self-supervised training strategy for NLCG-Net. Acquired data are divided into training and validation sets by masking. c) NLCG-Net deploys a light U-Net model with fewer convolutional layers and jointly regularizes all desired maps.
• Figure 3: T2 mapping reconstruction results under acceleration rate $R = 4$.
• Figure 4: T2 mapping reconstruction results under acceleration rate $R = 6$.
• Figure 5: T1 mapping reconstruction results under acceleration rate $R = 4$.