TL;DR
This paper investigates how different k-space subsampling strategies affect deep learning-based MRI reconstruction. They found that using radial subsampling, rather than the conventional rectilinear approach, leads to significantly more accurate and faithful MRI images when reconstructed by Recurrent Inference Machines (RIMs). This suggests radial subsampling could be a more effective strategy for accelerating MRI scans with deep learning.
What Problem Does This Paper Address?
Magnetic Resonance Imaging (MRI) is a powerful diagnostic tool, but its long acquisition times limit its use for dynamic imaging and can cause patient discomfort. Accelerating MRI scans is crucial, and two primary methods are Parallel Imaging (PI) and Compressed Sensing (CS). CS, in particular, relies on acquiring fewer k-space measurements and then reconstructing the image. The standard method for subsampling k-space in clinical settings is rectilinear (Cartesian) sampling. However, this often leads to aliased or blurred reconstructions, which can be mistaken for pathology. While deep learning (DL) has shown great promise in improving these reconstructions, the choice of subsampling strategy itself might be a bottleneck. This paper explores whether using a different subsampling pattern, specifically radial subsampling, can lead to better DL-based reconstructions compared to the traditional rectilinear approach.
Methods
The authors employed Recurrent Inference Machines (RIMs), a deep learning architecture known for its effectiveness in solving inverse problems like MRI reconstruction. They trained two separate RIM models: one on data subsampled using a rectilinear pattern and another on data subsampled using a radial pattern. Both models were trained retrospectively, meaning they applied subsampling masks to already fully sampled k-space data to simulate accelerated acquisitions. The RIM architecture is designed to iteratively refine an initial reconstruction over a series of time-steps, mimicking a gradient descent process. They also incorporated a U-Net to improve coil sensitivity map estimation during training. The reconstruction quality was evaluated using standard metrics like PSNR, SSIM, and VIF.
Datasets
The study utilized the Calgary-Campinas public brain multi-coil MRI dataset. This dataset consists of 67 three-dimensional T1-weighted, gradient-recalled echo, isotropic MRI volumes. The data was acquired on a Cartesian grid and comprised fully sampled k-spaces from twelve receiver coils. For their experiments, they randomly split these volumes into training, validation, and test sets.
Key Insights and Ideas
- Radial Subsampling Advantage: The core finding is that radial subsampling, when used with DL reconstruction, yields superior results compared to rectilinear subsampling. This is a significant insight because rectilinear sampling is the current clinical standard for CS-based acceleration.
- Denser Sampling Near Center: The paper hypothesizes that radial masks allow for more incoherent sampling and sample denser near the k-space center. This denser sampling of low-frequency information is crucial for accurate image reconstruction.
- Improved Fidelity: Radial subsampling not only leads to higher quantitative scores but also produces reconstructions that are qualitatively more faithful to the ground truth, with fewer visible artifacts.
- Paving the Way for New Strategies: By demonstrating the effectiveness of radial subsampling with RIMs, the paper encourages the exploration of radial patterns in other DL-based MRI reconstruction approaches, potentially leading to more efficient and accurate accelerated MRI.
Key Results
Quantitative evaluation showed that the Radial-RIM consistently outperformed the Rectilinear-RIM across both 5-fold and 10-fold acceleration factors. For instance, at 10-fold acceleration, Radial-RIM achieved a PSNR of 31.144 ± 0.655 and an SSIM of 0.899 ± 0.009, compared to Rectilinear-RIM's PSNR of 24.969 ± 0.755 and SSIM of 0.757 ± 0.021. Qualitatively, the reconstructions from Radial-RIM were visually much cleaner and more similar to the ground truth, with artifacts being significantly less pronounced, especially at higher acceleration rates. In some cases, the Radial-RIM reconstructions at 10-fold acceleration were even better than Rectilinear-RIM's at 5-fold acceleration.
Limitations and Future Work
The authors acknowledge that their experiments used retrospectively applied radial subsampling masks on data acquired with a Cartesian trajectory. They note that prospective radial subsampling patterns typically acquire data on a non-Cartesian grid, which would require regridding or NUFFT. The influence of this difference between retrospective and prospective radial sampling needs further investigation. Future work could explore the impact of view ordering in prospective radial sampling and investigate how these findings can be integrated into other DL architectures for accelerated MRI.
Summary updated