Journal Article

Highly accelerated whole-brain T2 mapping using non-cartesian acquisition and model-based implicit neural representation reconstruction

Xiao, Tianyi, Liu, Bei, She, Huajun, Du, Yiping P.

2025

Overview

•Development of a novel T2-prepared stack-of-stars imaging sequence that can achieve whole-brain scan in 70 s for T2 mapping.•Introduction of a signal evolution model to account for incomplete magnetization recovery, enabling efficient data acquisition without compromising mapping accuracy.•Integration of implicit neural representation for direct, unsupervised reconstruction of T2 maps from undersampled k-space data, without requiring fully sampled training datasets.•Reformulation of the T2 mapping problem into a function optimization framework, guided by a physical model and leveraging the flexibility of neural networks for precise parametric quantification. To propose a technique for highly accelerated T2 mapping of the whole brain. A pulse sequence with T2 preparation and a highly undersampled golden-angle stack-of-stars trajectory is used for data acquisition. A multiresolution hash encoding implicit neural representation, embedded with a physical model of signal evolution, is employed for unsupervised reconstruction. By rotating the trajectory at different kz encodings and effective echo times (TEeff) during acquisition, undersampling is applied to three dimensions, i.e., (kx, ky, TEeff) domain. In the phantom experiment, T2 quantification using this T2-prepared stack-of-stars acquisition was compared with that using a Cartesian multi-echo spin-echo acquisition. In the human experiments, T2 quantification using the undersampled acquisition was validated in comparison with those using the fully sampled acquisition and multi-echo spin-echo. In phantom test, T2 quantification using the proposed reconstruction agreed well (slope = 0.999, R2 > 0.99) with that obtained by fitting the NUFFT-reconstructed results for the fully sampled T2-prepared stack-of-stars acquisition. In human experiments, T2 quantification using a 20-fold acceleration agreed well with that using fully sampled data (NRMSE = 0.0066) in retrospectively undersampled reconstruction. No significant difference in T2 values was found between our technique and the reference method in prospective experiments, and our technique showed good inter-scan reproducibility and intra-scan repeatability. Using the proposed technique, whole-brain T2 mapping can be acquired in 70 s with a 20-fold acceleration.

• Publisher: Elsevier Inc,Elsevier
• Subject: Implicit neural representation / Quantitative magnetic resonance imaging / Stack-of-stars / T2 mapping / Unsupervised deep learning