2732

Artificial Sparsity Enhanced Deep Learning BUDA Imaging Enables Rapid and Distortion-corrected High-Resolution 3D-EPI and QSM
Zhifeng Chen1,2, Jin Jin3, Richard Mcintyre2, Kieran O'Brien3, Daniel Stäb4,5, Meng Law6,7, and Zhaolin Chen1,2
1Department of Data Science and AI, Faculty of IT, Monash University, Clayton, Australia, 2Monash Biomedical Imaging, Monash University, Clayton, Australia, 3Siemens Healthcare Pty Ltd, Brisbane, Australia, 4Siemens Healthcare Pty Ltd, Melbourne, Australia, Melbourne, Australia, 5Department of Radiology, The Royal Melbourne Hospital and The University of Melbourne, Parkville, Australia, 6Department of Neuroscience, Faculty of Medicine, Monash University, Melbourne, Australia, 7Department of Radiology, Alfred Health, Melbourne, Australia

Synopsis

Keywords: Image Reconstruction, Quantitative Susceptibility mapping, EPI, distortion correction

Motivation: MR susceptibility mapping serves as a highly valuable tool in various neuroscientific and clinical applications.

Goal(s): This innovative approach is designed to facilitate fast and robust high-resolution whole-brain imaging and quantitative susceptibility mapping (QSM).

Approach: In this study, our primary objective was to create a distortion-free 3D-EPI with blip-up/down acquisition (BUDA), incorporating controlled aliasing in parallel imaging (CAIPI) sampling, and applying artificial sparsity enhanced deep learning image reconstruction.

Results: Our developed technique holds the potential to produce distortion-free high-resolution whole-brain quantitative susceptibility mapping in just 12s at 3T and 9s at 7T, achieving an impressive resolution of 1 mm isotropic.

Impact: The proposed 3D-BUDA, incorporating a 2D CAIPIRINHA acquisition sequence with artificial sparsity-enhanced self-supervised deep learning reconstruction, demonstrated its ability to deliver rapid, distortion-free, high-resolution, whole-brain T2*-weighted imaging and QSM.

Purpose

MR susceptibility mapping has garnered increased attention recently due to its burgeoning applications in the fields of neuroscience and clinical practice1,2. 3D Echo Planar Imaging (3D-EPI) has shown promise in acquiring T2*w data suitable for susceptibility mapping3,4. Notwithstanding its rapid data acquisition capabilities, image distortion and voxel pile-up, remain open challenges in EPI5,6 and numerous technical advancements have been pursued to enhance geometric fidelity6–9.
In this study, we propose an innovative approach that combines 3D Blip-Up/Down Acquisition (BUDA)7,9 with Controlled Aliasing in Parallel Imaging (CAIPI)10 sampling and an artificial sparsity11–13 enhanced joint deep learning14 (ARTS-DL) reconstruction to yield high-resolution, distortion-corrected images.

Method & Experiments

Novel acquisition: Figure 1(a) depicts the schematic of the proposed 3D-BUDA acquisition, which is achieved by modifying a CAIPIRINHA-enabled research 3D-EPI sequence15 to include blip-up and blip-down phase encoding (yellow and green blips in GPE, respectively).
Novel reconstruction: Conventional BUDA image reconstruction can be formulated as 7,9,$$\tilde{I}=argmin_{I_{ud}}\|EI_{ud}-d_{ud}\|_2^2+\lambda\|{H}(I_{ud})\|_*$$
where $$$E$$$ is the encoding matrix, which contains the sampling mask, the discrete Fourier transform operator, the distortion correction operator with field-map estimated using Topup (http://fsl.fmrib.ox.ac.uk/fsl), and coil sensitivity map. $$$I_{ud}$$$ is the to-be-restored distortion-corrected image pairs, which is formed by concatenating blip-up image $$$I_{u}$$$ and blip-down image $$$I_{d}$$$. $$$d_{ud}$$$ is the acquired data. $$${H}(I_{ud})$$$ represents the Hankel low-rank matrix.
Parallel imaging is akin to a convolution operator in k-space and known to perform better on images with smaller image support, a concept referred to as "artificial sparsity" 11–13. Given Hankel low-rank reconstructions7–9 also involve convolutional operations, this work aims to investigate the hypothesis that introducing artificial sparsity in BUDA can enhance image reconstruction.
We designed a deep learning network to test this hypothesis, as illustrated in Figure 1(b). The new framework incorporates an artificial sparsity module within the deep learning network. The Hankel low-rank BUDA reconstruction is expressed as convolutional neural network models in deep learning14. The gradient of the new objective function can be reformulated as follows: $$\begin{aligned}\nabla\tilde{I}_{new}&=(E^HEI_{ud}-E^Hd_{ud})+\lambda S^H(\mathfrak{c}_{\theta_{2}}(ReLU\big(\mathfrak{c}_{\theta_{1}}(\ {SI}_{ud} )\big)))\end{aligned}$$
where $$${c_{\theta}}$$$ represents the convolution neural network and $$$S$$$ represents the SENSE operator. The nonlinear ReLU activation function, introduced into the new objective function compared to conventional neural network, enhances the reconstruction accuracy. Let $$$I=\mathbf{h}_\theta(\mathbf{x})$$$ be the output of the following loss function with parameter $$$\theta$$$.
$$argmin_{\theta}\|EI_{ud}-d_{ud}\|_1+\beta\||I_u|-|I_d|\|_1+\mu R(I)$$
Here $$$E$$$ also contains the artificial sparsity operator. The first term facilitates untrained reconstruction. $$$\||I_u|-|I_d|\|_1$$$ is the L1 norm of the pixel-wise magnitude difference between the blip-up and down images, ensuring robust sparsity through the consistency of magnitude information. $$$R(I)$$$ is the sparsity regularization term.
Data acquisition: In vivo multi-shot 3D-BUDA experiments were performed on two 3T and one 7T MRI systems (MAGNETOM Skyra, Trio and 7T, Siemens Healthcare, Erlangen, Germany), using dedicated 32-channel head coils and the three protocols summarized in Fig.1c.
All computations were implemented in MATLAB and Python (https://www.python.org/) with PyTorch.
Data Analysis: For reconstruction evaluation, we employed RMSE and SSIM.For QSM, tissue phase was estimated through Laplacian unwrapping 16 and V-SHARP 17,18 background removal. QSM was estimated using a nonlinear dipole inversion 19.

Results & Discussion

Figure 2 displays various reconstruction results obtained from Protocol 1. The proposed ARTS-DL-BUDA exhibits superior reconstruction accuracy compared to conventional BUDA (RMSE: 4.55% < 5.85%), and both outperform conventional SENSE 20 (RMSE: 9.03%). These results are consistent with the SSIM evaluation (0.9837 > 0.9721 > 0.9406). Furthermore, the difference map reveals that the contour of the SENSE reconstruction differs from the distortion-corrected reference, indicating the effectiveness of the BUDA reconstruction in distortion correction.
Figures 3 and 4 present a comparison of different approaches using 3T (Protocol 2) and 7T (Protocol 3) data, respectively. The effectiveness of BUDA in distortion correction is evident in both BUDA reconstructions, as indicated by the prominent yellow, blue, and orange arrows. A comparison between conventional BUDA and the proposed ARTS-DL-BUDA highlights that the enhanced deep learning reconstruction, incorporating artificial sparsity, results in a cleaner reconstruction with reduced noise, as indicated by the slender arrows in Figure 3.
Figure 5 displays QSM results from ARTS-DL-BUDA (Protocols 2 and 3). Notably, the ARTS-DL-BUDA scheme effectively addresses shot-to-shot phase variations arising from physiological differences. While 7T imaging offers improved SNR due to the higher field strength, it also introduces larger field inhomogeneity and heightened physiological variations, which can have an adverse impact on the quality of QSM.

Conclusion

The proposed ARTS-DL-BUDA combines CAIPI-sampling, BUDA acquisition with inverted polarity and joint artificial sparsity enhanced deep learning reconstruction to boost SNR and eliminate distortion simultaneously. This enables rapid, distortion-corrected whole-brain QSM at the resolution of 1×1×1 mm3 in 12s at 3T and 9s at 7T.

Acknowledgements

This work was supported by an early career researcher seed grant at Faculty of IT, Monash University, and ARC DP (DP210101863).

References

1. Langkammer C, Schweser F, Shmueli K, et al. Quantitative susceptibility mapping: Report from the 2016 reconstruction challenge. Magn Reson Med. 2018;79(3):1661-1673.

2. Bilgic B, Langkammer C, Marques JP, Meineke J, Milovic C, Schweser F. QSM reconstruction challenge 2.0: Design and report of results. Magn Reson Med. 2021;86(3):1241-1255.

3. Gao Y, Cloos M, Liu F, Crozier S, Pike GB, Sun H. Accelerating quantitative susceptibility and R2* mapping using incoherent undersampling and deep neural network reconstruction. Neuroimage. 2021;240.

4. Tourell M, Jin J, Stewart A, et al. Submillimeter, sub-minute quantitative susceptibility mapping using a multi-shot 3D EPI with 2D CAIPIRINHA acceleration. Proc Int Socitey Magn Reson Med. 2021; 787.

5. In MH, Posnansky O, Beall EB, Lowe MJ, Speck O. Distortion correction in EPI using an extended PSF method with a reversed phase gradient approach. PLoS One. 2015;10(2).

6. Zahneisen B, Aksoy M, Maclaren J, Wuerslin C, Bammer R. Extended hybrid-space SENSE for EPI: Off-resonance and eddy current corrected joint interleaved blip-up/down reconstruction. Neuroimage. 2017;153(March):97-108.

7. Sun K, Chen Z, Dan G, et al. Three-dimensional echo-shifted EPI with simultaneous blip-up and blip-down acquisitions for correcting geometric distortion. Magn Reson Med. 2023;2022(4390):2375-2387.

8. Liao C, Bilgic B, Tian Q, et al. Distortion-free, high-isotropic-resolution diffusion MRI with gSlider BUDA-EPI and multicoil dynamic B0 shimming. Magn Reson Med. 2021;86(2):791-803.

9. Chen Z, Liao C, Cao X, et al. 3D-EPI blip-up/down acquisition (BUDA) with CAIPI and joint Hankel structured low-rank reconstruction for rapid distortion-free high-resolution T2* mapping. Magn Reson Med. 2023;89(5):1961-1974.

10. Setsompop K, Gagoski BA, Polimeni JR, Witzel T, Wedeen VJ, Wald LL. Blipped-controlled aliasing in parallel imaging for simultaneous multislice echo planar imaging with reduced g-factor penalty. Magn Reson Med. 2012;67(5):1210-1224.

11. Chen Z, Xia L, Liu F, et al. An improved non-Cartesian partially parallel imaging by exploiting artificial sparsity. Magn Reson Med. 2017;78(1):271-279.

12. Chen Z, Kang L, Xia L, et al. Technical Note: Sequential combination of parallel imaging and dynamic artificial sparsity framework for rapid free-breathing golden-angle radial dynamic MRI: K-T ARTS-GROWL. Med Phys. 2018;45(1):202-213.

13. Zhang J, Chu Y, Ding W, et al. HF-SENSE: An improved partially parallel imaging using a high-pass filter. BMC Med Imaging. 2019;19(1):1-10.

14. Kim TH, Garg P, Haldar JP. LORAKI: Autocalibrated Recurrent Neural Networks for Autoregressive MRI Reconstruction in k-Space. Published online 2019:1-24.

15. Jin J, Tourell M, Pascal S, et al. Segmented 3D EPI with CAIPIRINHA for Fast, High-Resolution T2*-Weighted Imaging. Proc Int Socitey Magn Reson Med. Published online 2021: 4190.

16. Li W, Wu B, Liu C. Quantitative susceptibility mapping of human brain reflects spatial variation in tissue composition. Neuroimage. 2011;55(4):1645-1656.

17. Wu B, Li W, Guidon A, Liu C. Whole brain susceptibility mapping using compressed sensing. Magn Reson Med. 2012;67(1):137-147.

18. Schweser F, Sommer K, Deistung A, Reichenbach JR. Quantitative susceptibility mapping for investigating subtle susceptibility variations in the human brain. Neuroimage. 2012;62(3):2083-2100.

19. Polak D, Chatnuntawech I, Yoon J, et al. Nonlinear dipole inversion (NDI) enables robust quantitative susceptibility mapping (QSM). NMR Biomed. 2020;33(12):1-13.

20. Samsonov AA, Kholmovski EG, Parker DL, Johnson CR. POCSENSE: POCS-based reconstruction for sensitivity encoded magnetic resonance imaging. Magn Reson Med. 2004;52(6):1397-1406.

Figures

FIG.1. 3D-BUDA multi-shot sequence schematic (a) and artificial sparsity enhanced deep learning image reconstruction (b). The protocols explored in this study were shown in (c). Blip up/down were implemented in 3D acquisition (see yellow and green blips in phase encoding gradient, respectively).

FIG.2. Comparison of reconstructed results and error maps of different reconstruction approaches. Reference image was generated by fully sampled 3D-BUDA image reconstruction. SENSE, conventional BUDA and artificial sparsity enhanced deep learning BUDA (ARTS-DL-BUDA) were displayed from the first column to third column. The corresponding difference maps were listed in the second row.

FIG.3. Comparison of different approaches (SENSE, conventional 3D-BUDA and ARTS-DL-BUDA) on image quality and distortion-correction effect for the same 2-shot BUDA dataset. First column: Blip-up EPI SENSE. Second column: Blip-down EPI SENSE. Third column: conventional BUDA. Last column: The ARTS-DL-BUDA image reconstruction. The total acquisition times of blip-up EPI, blip-down EPI, conventional 3D-BUDA and ARTS-DL-BUDA are 7 s, 7 s, 12 s, and 12 s, respectively. A 2-s FOV-matched FLASH low-resolution scan for coil sensitivity map is included in these experiments.

FIG.4. Comparison of different approaches (SENSE, conventional 3D-BUDA and ARTS-DL-BUDA) on image quality and distortion-correction effect for the same 2-shot BUDA at 7T. First column: Blip-up SENSE. Second column: Blip-down SENSE. Third column: conventional 3D-BUDA. Last column: The proposed ARTS-DL-BUDA image reconstruction. The total acquisition times of blip-up EPI, blip-down EPI, conventional 3D-BUDA and ARTS-DL-BUDA are 5.5 s, 5.5 s, 9 s, and 9 s, respectively. A 2-s FOV-matched FLASH low-resolution scan for coil sensitivity map is also counted in these acquisitions.

FIG.5. Tissue phase (left column) and quantitative susceptibility mapping (right column) results estimated using ARTS-DL-BUDA image reconstruction (Reduction factor: Rinplane = 4, Rz = 2, TA = 12s @ 3T; Reduction factor: Rinplane = 5, Rz = 2, TA = 9s @ 7T). The local tissue phase was obtained by Laplacian unwrapping and V-SHARP filtering (10 mm largest kernel size).

Proc. Intl. Soc. Mag. Reson. Med. 32 (2024)
2732
DOI: https://doi.org/10.58530/2024/2732