Introduction

Routine pediatric abdominal protocols utilize fat suppressed T1 images for assessment of contrast enhancement characteristics of solid organs and focal lesions, with Dixon style acquisitions commonly used to acquire these T1 images^1,2. One such sequence is 3D Liver Acquisition with Volume Acceleration – Flex (LAVA-Flex), which is a 3D fast spoiled gradient-recalled echo sequence that uses a dual echo acquisition and a 2-point DIXON reconstruction¹. This sequence is acquired both pre- and post- intravenous contrast administration and is very useful in identifying pathology across the whole abdomen, including lesions demonstrating enhancing elements, blood and fat. Additionally, the short acquisition time of LAVA-Flex is advantageous for patients with limited breath hold capabilities, such as the pediatric population. To help further improve the image quality and signal-to-noise ratio, we applied a deep learning (DL) reconstruction aimed at improving SNR and quality of LAVA-Flex images. Here we present our comparison of DL and non-DL reconstructed 3D LAVA-Flex abdominal scans in the pediatric population at both 1.5T and 3.0T.

Methods

19 pediatric patients (2-17 y.o.; mean 10.2 y.o.) underwent 3D LAVA-Flex scans on either a 1.5T Signa Artist (15 patients) or 3T Premier XT (4 patients) MRI system (GE HealthCare, Waukesha, WI, USA) as part of a routine clinical abdominal protocol. Example protocols included general free breathing abdomen, enterography/fistula, and sacrum. The specific parameters varied across the two field strengths and specific abdominal protocol used, with example parameters listed in Table 1. The resulting raw data was reconstructed with a conventional reconstruction (non-DL) and DL reconstruction with 75% noise reduction. The DL algorithm was based on a Convolutional Neural Network (AIR™ Recon DL, GE Healthcare)³ trained to reduce noise, remove Gibbs ringing, and increase sharpness in 3D Flex data.
Two fellowship-trained abdominal radiologists independently evaluated the reconstructed images for diagnostic quality on a 5-point Likert scale and indicated which of the two reconstructions they preferred.
In addition, noise characteristics were quantitatively analyzed for both reconstructions. Two metrics, total variation (TV) and number of edges, were chosen as representations of effective image denoising because lower TV correlates with less image noise and fewer number of edges correlates with increased sharpness through fewer detachments along each edge^4–6. The TV and total number of detected edges across the whole 3D volume were determined by calculating the gradients and edges across each image slice and summing across the whole 3D volume. For TV and edge analysis, the imgradient and edge functions (MATLAB 2022b, MathWorks Inc., Natick, MA), were utilized, respectively. The local TV and number of edges, as well as the signal-to-noise ratio (SNR), were calculated as the mean value within the liver of a representative slice.
Statistical analyses were performed using in Microsoft Excel with a p-value ≤ 0.05 considered statistically significant. Wilcoxon signed rank test, sign test, and weighted kappa were calculated on qualitative reads by the radiologists. Paired t-test was used on the quantitative calculations.

Results

Qualitatively, both radiologists preferred the DL reconstructed images due to reduced image noise and improved sharpness. In cases with focal lesions, such as a large hepatic lesion (Figure 1) and a terminal ilium stricture (Figure 2), there was better definition of the pathology. The image quality improvement was field strength independent with the radiologists preferring the DL in majority (13/15 for Reader 1 and 12/15 for Reader 2) of the 1.5T cases and in all four 3T cases. See Figure 3 for an example of a 3T case (images in Figures 1 and 2 were acquired at 1.5T).
Table 2 shows the statistical results. Both readers had higher scores for DL, with higher mean, median, and 25-75 percentile scores for the DL images. The sign test for each reader was also statistically significant in favor of DL (p < 0.01), with 17/19 and 16/19 cases preferring DL. The weighted kappa showed slight to fair agreement in image quality evaluation.
Figures 1-3 show the water, gradient, edge image for both DL and non-DL reconstruction from the three aforementioned examples. From quantitative standpoint, the SNR, TV, and number of edges were all significantly different (p < 0.001), with DL images having greater SNR, lower TV and fewer edges detected, both globally and locally.

Discussion and Conclusion

The study demonstrated that deep learning reconstruction can be an effective technique to improve image quality of routine 3D LAVA-Flex abdominal acquisitions in the pediatric population. Quantitative analysis showed improved edge detection, decreased noise, and consequently ~2-3x improved SNR after application of the DL reconstruction at both 1.5 and 3T. Future work will focus on further optimizing the trade-off between spatial resolution and breath-hold time due to the added SNR benefit offered by DL. The findings have the potential to improve the diagnostic accuracy and to help improve pediatric patient comfort by reducing scan times while preserving imaging quality.

Acknowledgements

No acknowledgement found.

References

1. Li et al. XH. Abdominal MRI at 3.0 T: LAVA‐flex compared with conventional fat suppression. Journal of Magnetic Resonance Imaging. 2014;40:58-66.

2. Yoon JH, Nickel MD, Peeters JM, Lee JM. Rapid Imaging: Recent Advances in Abdominal MRI for Reducing Acquisition Time and Its Clinical Applications. Korean Journal of Radiology. 2019;20(12):1597-1615.

3. Lebel RM. Performance characterization of a novel deep learning-based MR image reconstruction pipeline. August 2020. doi:10.48550/arXiv.2008.06559

4. Rudin LI, Osher S, Fatemi E. Nonlinear total variation based noise removal algorithms ☆. Physica D: Nonlinear Phenomena. 60(1-4).

5. Block KT, Uecker M, Frahm J. Suppression of MRI Truncation Artifacts Using Total Variation Constrained Data Extrapolation. International Journal of Biomedical Imaging. 2008;2008.

6. Ruslau MFV, Pratama RA, Asmal S. Edge detection in noisy images with different edge types. IOP Conf Series: Earth and Environmental Science. 2019;343.