4643

SmartBlade: AI-based reconstruction for motion robust abdominal PROPELLER imaging
Alexander Selivanov1, Holger Eggers1, Jakob Meineke1, Max-Heinrich Laves1, and Mariya Doneva1
1Philips Research, Hamburg, Germany

Synopsis

Keywords: Motion Correction, Motion Correction, PROPELLER, free breathing, registration, body MRI

Motivation: The motion correction in PROPELLER is typically limited to rigid body motion and only the averaging effect is exploited for motion artifact reduction for abdominal scans.

Goal(s): Our goal was to reduce motion artifacts and improve image quality in abdominal MRI.

Approach: We proposed an AI framework for reconstructing high-resolution motion-free image from a free-breathing abdominal PROPELLER scan. The AI model was trained on synthetic data and tested on abdominal T2W TSE scans.

Results: The proposed AI reconstruction outperforms the conventional PROPELLER reconstruction and PROPELLER with non-rigid motion correction in terms of residual motion artifacts and general image quality.

Impact: The proposed AI-based reconstruction allows obtaining motion-free images with high-spatial resolution from PROPELLER MRI scans, which facilitates abnormality detection.

Introduction

PROPELLER is a well-established technique for motion correction in MRI1. In PROPELLER, k-space data is acquired in blades that are rotated over time, and low-resolution data collected in each blade is utilized for motion estimation. The motion correction in PROPELLER is typically limited to rigid body motion. If more complex motion is expected to occur, as in abdominal imaging, the motion correction is often switched off and only the averaging effect is exploited for motion artifact reduction.

Recently, an AI-based approach (BladeNet)2 was proposed for time-resolved paediatric bowel wall imaging, showing promising results. However, BladeNet is based on bSSFP scans and requires full-resolution motion-free Cartesian data acquired in different orientations for training. Acquiring such target data is often not feasible, since PROPELLER is typically used in cases when the fully-sampled Cartesian scan would have motion artifacts.

In this work, we proposed an AI-based approach for motion correction in abdominal PROPELLER imaging that was trained on synthetic data, aiming to improve the generalization to contrasts where acquiring full-resolution target images is not possible. The proposed approach is demonstrated with abdominal T2W TSE scans in healthy volunteers and compared to the conventional PROPELLER reconstruction as well as PROPELLER with non-rigid motion correction.

Methods

The proposed deep learning reconstruction framework, called SmartBlade, generates high-resolution motion-free images from a free-breathing abdominal PROPELLER scan and is agnostic to different contrasts.

Data simulation. For better generalization, training was based on simulated data. Simulations of the PROPELLER sampling process and motion were performed on natural images and included the following steps:
  1. Crop a square patch of specified size
  2. Simulate several breathing motion states by performing affine transformations, including expansion and shrinkage with realistic motion amplitudes expected in a free-breathing MRI scan
  3. Apply local distortions to account for more complex types of motion, such as peristalsis
  4. Simulate PROPELLER sampling by projecting the images, simulated for each motion state, onto different blades using a NUFFT
  5. Reconstruct single-blade images using zero-filling and an adjoint NUFFT
  6. Normalize the data for more stable training
For the computation of the target image, one of the simulated motion states (3) was randomly selected and the corresponding image was projected to the fully-sampled PROPELLER trajectory. The target image was computed by gridding reconstruction of the simulated k-space data.

Training. We reduced the number of training parameters, by using 2D instead of a 3D ResUNet3,4 and treating the blade dimension as channels instead of as the third dimension for the input. A combination of L1 and SSIM loss was used.

Data acquisition. Abdominal T2-weighted TSE PROPELLER scans were performed in 2 healthy volunteers on a 1.5T Ingenia Scanner (Philips, Best, The Netherlands) with IRB approval and informed consent obtained. The following scan parameters were used: TR=4.2s, TE=142ms, flip angle=90°, TSE factor=64, SENSE factor=2.0, FOV=450×450mm2, voxel size=0.7×0.7×5mm3, #blades=7. The scans were repeated once under free-breathing conditions, once with respiratory triggering, and once with combined triggering and guided breath-holds. The triggering aims to minimize the impact of respiratory motion but leads to a significant scan time increase.

Reconstructions. The proposed SmartBlade approach was compared to:
  1. Conventional gridding reconstruction without motion correction5
  2. Reconstruction with non-rigid motion correction6 instead of the typically used rigid-body motion correction

Results

A comparison between the conventional PROPELLER reconstruction, the non-rigid motion-corrected PROPELLER reconstruction, and the proposed AI-based reconstruction for the free-breathing scans is shown in Figure 1. There are visible motion artifacts in the conventional reconstruction (Fig.1a, Fig.1d) that were partially reduced by the non-rigid motion correction (Fig1.b, Fig1.e). The proposed reconstruction (Fig.1c, Fig.1f) results in significantly reduced motion artifacts especially related to intestinal motion.

Figure 2 shows results for the triggered scans. Triggering (with or without breath-holding) helps to reduce the respiratory motion artifacts but does not improve the blurring or reduce artifacts due to intestinal motion. Also here, SmartBlade shows the least motion artifacts.

Figure 3 shows a comparison of a single-blade image from the target motion state (Fig.3a,c) and the proposed reconstruction (Fig3.b,d). SmartBlade shows no ringing and improved resolution compared to the single-blade image, while preserving the motion state.

Conclusion

We proposed a new AI-based approach for motion-robust abdominal imaging based on training with synthetic data and demonstrated its feasibility with T2W PROPELLER scans. The proposed approach shows significantly reduced motion artifacts compared to the conventional PROPELLER reconstruction and PROPELLER with non-rigid motion correction. Further validation is needed to assess the potential of SmartBlade for robust abdominal imaging.

Acknowledgements

No acknowledgement found.

References

  1. Pipe JG. Motion correction with PROPELLER MRI: application to head motion and free-breathing cardiac imaging. Magn Reson Med. 1999 Nov;42(5):963-9. doi: 10.1002/(sici)1522-2594(199911)42:5<963::aid-mrm17>3.0.co;2-l. PMID: 10542356.
  2. Shimron, Efrat, et al. "BladeNet: Rapid PROPELLER Acquisition and Reconstruction for High Spatio-Temporal Resolution Abdominal MRI." Proceedings of the 31st Annual International Society for Magnetic Resonance in Medicine, London, UK (2022): 7-12.
  3. Ronneberger O, Fischer P, Brox T (2015) U-net: convolutional networks for biomedical image segmentation. In: International conference on medical image computing and computer-assisted intervention. Springer, pp 234–241
  4. K. He, X. Zhang, S. Ren and J. Sun, "Deep Residual Learning for Image Recognition," 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 2016, pp. 770-778, doi: 10.1109/CVPR.2016.90.
  5. Pipe, J.G., Gibbs, W.N., Li, Z., Karis, J.P., Schar, M. and Zwart, N.R. (2014), Revised motion estimation algorithm for PROPELLER MRI. Magn. Reson. Med., 72: 430-437. https://doi.org/10.1002/mrm.24929
  6. Ashburner J. A fast diffeomorphic image registration algorithm. Neuroimage. 2007 Oct 15;38(1):95-113. doi: 10.1016/j.neuroimage.2007.07.007. Epub 2007 Jul 18. PMID: 17761438.

Figures

Figure 1. Comparison of different reconstructions for the free-breathing acquisition. a,d - conventional reconstruction without motion correction. b,e – non-rigid motion correction applied to each blade followed by conventional reconstruction. c,f – the proposed AI-based reconstruction.

Figure 2. Comparison of different reconstructions for the acquisitions with a-c - triggering and d-f - combined triggering and guided breath-holds. a,d - conventional reconstruction without motion correction. b,e – non-rigid motion correction applied to each blade followed by conventional reconstruction. c,f – the proposed AI-based reconstruction.

Figure 3. Comparison of a,c - single-blade images from the target motion state and b,d - the proposed AI reconstruction for free-breathing datasets.

Proc. Intl. Soc. Mag. Reson. Med. 32 (2024)
4643
DOI: https://doi.org/10.58530/2024/4643