[Paper Review] Extreme MRI: Large-Scale Volumetric Dynamic Imaging from Continuous Non-Gated Acquisitions

Frank Ong, Xucheng Zhu+5 more|arXiv (Cornell University)|Sep 30, 2019

Advanced MRI Techniques and Applications45 references4 citations

TL;DR

This paper presents Extreme MRI, a framework for reconstructing large-scale volumetric dynamic MRI from continuous, non-gated, undersampled k-space data using 3D non-Cartesian trajectories. It combines multi-scale low-rank matrix factorization and stochastic optimization to enable high-resolution, motion-robust reconstruction without periodicity assumptions, revealing transient dynamics and sharper features in irregular motion scenarios.

ABSTRACT

Purpose: To develop a framework to reconstruct large-scale volumetric dynamic MRI from rapid continuous and non-gated acquisitions, with applications to pulmonary and dynamic contrast enhanced (DCE) imaging. Theory and Methods: The problem considered here requires recovering hundred-gigabytes of dynamic volumetric image data from a few gigabytes of k-space data, acquired continuously over several minutes. This reconstruction is vastly under-determined, heavily stressing computing resources as well as memory management and storage. To overcome these challenges, we leverage intrinsic three dimensional (3D) trajectories, such as 3D radial and 3D cones, with ordering that incoherently cover time and k-space over the entire acquisition. We then propose two innovations: (1) A compressed representation using multi-scale low rank matrix factorization that constrains the reconstruction problem, and reduces its memory footprint. (2) Stochastic optimization to reduce computation, improve memory locality, and minimize communications between threads and processors. We demonstrate the feasibility of the proposed method on DCE imaging acquired with a golden-angle ordered 3D cones trajectory and pulmonary imaging acquired with a bit-reversed ordered 3D radial trajectory. We compare it with "soft-gated" dynamic reconstruction for DCE and respiratory resolved reconstruction for pulmonary imaging. Results: The proposed technique shows transient dynamics that are not seen in gating based methods. When applied to datasets with irregular, or non-repetitive motions, the proposed method displays sharper image features. Conclusion: We demonstrated a method that can reconstruct massive 3D dynamic image series in the extreme undersampling and extreme computation setting.

Motivation & Objective

• Overcome the limitations of gating-based methods in volumetric dynamic MRI, which rely on periodic motion assumptions and lose transient dynamics.
• Address the extreme computational and memory challenges of reconstructing hundreds of gigabytes of dynamic volumetric data from a few gigabytes of undersampled k-space data.
• Enable high spatiotemporal resolution reconstruction in the absence of periodic motion, particularly for non-repetitive or irregular physiological dynamics such as coughing or irregular breathing.
• Develop a scalable reconstruction framework that is robust to bulk motion and motion irregularities, applicable to challenging clinical scenarios like pulmonary and DCE-MRI.
• Demonstrate feasibility of reconstructing dynamic series with sharp temporal features and preserved transient events not captured by conventional soft-gating or respiratory-resolved methods.

Proposed method

• Leverages intrinsic 3D non-Cartesian k-space trajectories—3D radial and 3D cones—with pseudo-random, incoherent ordering over time and k-space to improve sampling efficiency and reduce aliasing.
• Introduces a multi-scale low-rank (MSLR) matrix factorization to compress the dynamic image volume, reducing memory footprint and constraining the underdetermined reconstruction problem.
• Employs stochastic optimization to minimize computation, enhance memory locality, and reduce inter-processor communication, enabling scalable processing on large datasets.
• Uses a finite-difference temporal penalty on temporal bases to enforce smoothness in dynamic evolution, balancing temporal fidelity and artifact suppression.
• Applies a Burer-Monteiro factorization approach to the low-rank problem, enabling efficient non-convex optimization with theoretical guarantees under idealized incoherence conditions.
• Supports retrospective motion correction by enabling signal intensity curve computation across reconstructed time points, even in the presence of non-periodic motion.

Experimental results

Research questions

• RQ1 Can a reconstruction framework achieve high-resolution volumetric dynamic MRI from continuous, non-gated acquisitions without relying on periodic motion assumptions?
• RQ2 How can multi-scale low-rank modeling effectively reduce memory and computational demands in extreme undersampling scenarios?
• RQ3 To what extent does stochastic optimization improve scalability and memory efficiency in large-scale dynamic MRI reconstruction?
• RQ4 Can the proposed method recover transient physiological dynamics—such as coughing or irregular breathing—that are lost in conventional gating-based methods?
• RQ5 What are the limitations of low-rank modeling in handling large bulk motions, and how can they be mitigated in future work?

Key findings

• The proposed method successfully reconstructs massive 3D dynamic MRI series from continuous, non-gated acquisitions, achieving high spatiotemporal resolution in extreme undersampling regimes.
• For datasets with irregular breathing, the method reveals transient dynamics such as coughing that are invisible in soft-gated reconstructions, demonstrating superior motion robustness.
• Image quality is significantly improved in non-periodic motion scenarios, with sharper features and reduced blurring compared to conventional gating techniques.
• The method exhibits flickering artifacts in some reconstructions, primarily due to sensitivity of small blocks in the multi-scale low-rank model to noise and aliasing.
• Temporal resolution is lower than prescribed due to the finite-difference temporal penalty, and actual spatiotemporal resolution remains unquantified due to lack of ground truth.
• For large bulk motions, such as those in supporting video S7, image quality degrades and temporal blurring increases, indicating limitations in low-rank modeling for large displacements.