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

Abstract:Typical quantitative MRI (qMRI) methods estimate parameter maps after image reconstructing, which is prone to biases and error propagation. We propose a Nonlinear Conjugate Gradient (NLCG) optimizer for model-based T2/T1 estimation, which incorporates U-Net regularization trained in a scan-specific manner. This end-to-end method directly estimates qMRI maps from undersampled k-space data using mono-exponential signal modeling with zero-shot scan-specific neural network regularization to enable high fidelity T1 and T2 mapping. T2 and T1 mapping results demonstrate the ability of the proposed NLCG-Net to improve estimation quality compared to subspace reconstruction at high accelerations.