[Paper Review] Zero-DeepSub: Zero-Shot Deep Subspace Reconstruction for Rapid Multiparametric Quantitative MRI Using 3D-QALAS

Yohan Jun, Yamin Arefeen+8 more|PubMed|Jul 4, 2023

Advanced MRI Techniques and Applications86 references7 citations

TL;DR

This paper proposes Zero-DeepSub, a zero-shot deep learning method combined with low-rank subspace modeling to enable rapid, high-fidelity 3D multiparametric quantitative MRI using 3D-QALAS. By leveraging scan-specific deep learning within a subspace framework, it achieves whole-brain T1, T2, and proton density mapping at 1 mm isotropic resolution in under 2 minutes with 9-fold acceleration and reduced artifacts compared to conventional QALAS.

ABSTRACT

The proposed subspace QALAS along with Zero-DeepSub enabled high fidelity and rapid whole-brain multiparametric quantification and time-resolved imaging.

Motivation & Objective

• To develop a fast and accurate method for 3D multiparametric quantitative MRI using the 3D-QALAS sequence.
• To reduce reconstruction time and improve image fidelity in T1 and T2 mapping by exploiting low-rank subspace structure.
• To enable zero-shot deep learning-based reconstruction that generalizes across scan-specific contrasts without fine-tuning.
• To achieve high acceleration factors (up to 9×) while maintaining precision and reducing noise and artifacts in in vivo whole-brain imaging.

Proposed method

• A low-rank subspace model is applied to 3D-QALAS k-space data to exploit temporal and spatial correlations across time points.
• The method formulates T1 and T2 mapping as a low-rank subspace reconstruction problem, reducing the number of required encoding steps.
• A zero-shot deep learning model (Zero-DeepSub) is trained on a diverse set of scan contrasts and applied to new scans without fine-tuning.
• The deep learning component enhances reconstruction fidelity by learning scan-specific patterns while preserving the global subspace structure.
• The framework combines subspace modeling with deep neural networks to jointly optimize for accuracy and speed.
• The approach enables time-resolved imaging and whole-brain 1 mm isotropic T1, T2, and PD mapping with high temporal and spatial resolution.

Experimental results

Research questions

• RQ1 Can low-rank subspace modeling significantly accelerate 3D-QALAS while preserving T1 and T2 mapping accuracy?
• RQ2 How does zero-shot deep learning improve reconstruction fidelity in subspace-based 3D-QALAS without requiring fine-tuning on new scans?
• RQ3 To what extent can the proposed method achieve high acceleration factors (e.g., 9×) while minimizing noise, blurring, and artifacts in in vivo imaging?
• RQ4 How does the combination of subspace modeling and deep learning compare to conventional QALAS in terms of g-factor maps and precision?
• RQ5 Can the method achieve whole-brain 1 mm isotropic multiparametric mapping within 2 minutes of scan time?

Key findings

• Phantom experiments showed that subspace QALAS achieved good linearity with reference methods and reduced bias and improved precision, especially for T2 maps.
• In vivo results demonstrated that Zero-DeepSub reduced voxel blurring, noise, and artifacts compared to conventional QALAS at up to 9-fold acceleration.
• The method achieved whole-brain T1, T2, and proton density mapping at 1 mm isotropic resolution within 2 minutes of scan time.
• g-factor maps indicated improved robustness to acceleration, confirming better noise performance in high-undersampling regimes.
• The zero-shot deep learning component maintained high fidelity across diverse scan protocols without fine-tuning, demonstrating strong generalization.
• Compared to conventional QALAS, the proposed method showed superior precision and lower bias in both T1 and T2 mapping, particularly at high acceleration factors.