DL-based Highly-Accelerated Hepatobiliary-Phase Gadoxetate-Enhanced MRI ; Preliminary Experience

Technical Summary: DL-based Highly-Accelerated Hepatobiliary-Phase Gadoxetate-Enhanced MRI

Problem Statement
Gadoxetate-enhanced MRI (Gd-EOB-MRI) is critical for liver tumor detection and characterization, particularly during the hepatobiliary phase (HBP). High-resolution, fat-suppressed T1-weighted imaging (FS-T1WI) is typically required within a single breath-hold to minimize motion artifacts. However, increasing spatial resolution inherently reduces the signal-to-noise ratio (SNR) and prolongs acquisition time, thereby increasing the risk of motion artifacts. While existing deep learning reconstruction (DLR) techniques like AIR Recon DL (ARDL) improve image quality by denoising and sharpening, they often rely on parallel imaging or compressed sensing for acceleration. These traditional fast imaging methods can introduce blurring and aliasing artifacts when acceleration factors (AF) are pushed too high. There is a need for a reconstruction method capable of handling highly undersampled k-space data directly to achieve faster, high-resolution HBP imaging without compromising diagnostic quality.

Methodology
This retrospective, single-center study evaluated the feasibility and diagnostic performance of "Sonic DL" (DLS), a next-generation DLR method that reconstructs images directly from highly undersampled k-space data without relying on parallel imaging or compressed sensing. The study analyzed 40 patients who underwent Gd-EOB-MRI on a 3.0-T system (Discovery MR750w).

The researchers compared five acquisition protocols:

1. Conv-LAVA: Conventional LAVA using ARDL, 1.5 mm isovoxel, acceleration factor (AF) 6, acquisition time 17s.
2. DLS-LAVA (4 variants): Protocols utilizing Sonic DL with varying resolutions (1.2 mm and 1.5 mm isovoxel) and acceleration factors (AF 10 and AF 12).

Evaluation Metrics:

• Qualitative: Two blinded, board-certified radiologists scored image quality on a 5-point scale across five categories: overall quality, sharpness, fat suppression, graininess, and aliasing artifacts.
• Quantitative: Liver-to-spleen signal-intensity ratios (L/S) were measured to assess contrast preservation.
• Statistical Analysis: Paired t-tests and Wilcoxon tests with Benjamini–Hochberg correction were used to compare protocols.

Key Contributions and Results
The study demonstrated that Sonic DL enables significantly accelerated HBP acquisitions while maintaining or improving image quality compared to conventional methods.

• Artifact Reduction: All DLS-LAVA protocols showed significantly higher sharpness scores and markedly reduced aliasing artifacts compared to Conv-LAVA (p < 0.01).
• Image Quality: Overall image quality was comparable across all protocols, with a slight tendency favoring the DLS-LAVA 1.5 mm protocols (AF 10 and 12).
• Contrast and Fat Suppression: Fat suppression was best preserved in the 1.5 mm resolution DLS protocols. The L/S ratio in the DLS-LAVA 1.5 mm AF 10 protocol was statistically comparable to Conv-LAVA (p = 0.32), whereas other DLS variants showed slightly lower ratios.
• Artifacts Specific to DLS: A "fine grainy texture," distinct from random noise, was occasionally observed in DLS-LAVA images, particularly at AF 12. The authors note this lacked reproducibility and appeared suggestive of reconstruction artifacts but did not interfere with diagnosis.
• Optimal Protocol: The DLS-LAVA 1.5 mm AF 10 protocol emerged as the optimal balance, offering a 12-second breath-hold acquisition (faster than the 17-second Conv-LAVA) with superior sharpness, reduced aliasing, and preserved contrast.

Significance and Claims
The paper claims that Sonic DL represents a significant advancement for HBP imaging by allowing high-resolution (1.5 mm isovoxel) acquisition within a short 12-second breath-hold. The authors posit that this technology provides a superior trade-off between speed, sharpness, and contrast compared to conventional ARDL-based LAVA.

Specifically, the study concludes that:

1. Sonic DL successfully reconstructs images from highly undersampled k-space data, minimizing the aliasing artifacts typically associated with high acceleration factors in traditional parallel imaging.
2. The 1.5 mm AF 10 protocol offers the best clinical compromise, potentially replacing conventional LAVA for HBP imaging, which is particularly beneficial for elderly or dyspneic patients who struggle with longer breath-holds.
3. While a distinct grainy texture is a noted artifact of the DLS reconstruction process, its diagnostic impact is minimal.

The authors maintain a modest tone, acknowledging limitations such as the retrospective, single-center design, the small sample size (n=40), and the lack of lesion delineation assessment or histopathologic confirmation. They state that further prospective, multicenter validation is required to confirm the clinical utility and lesion detectability of this approach.