3395
Corrupted Data Rejection Strategy in k-space Based Multi-shot Diffusion Reconstruction1China National Clinical Research Center for Neurological Diseases, Beijing Tiantan Hospital, Capital Medical University, Beijing, China, 2Department of Neurology, Beijing Tiantan Hospital, Capital Medical University, Beijing, China, 3Center for Biomedical Imaging Research, Department of Biomedical Engineering, School of Medicine, Tsinghua University, Beijing, China, 4Philips Healthcare, Beijing, China, 5Department of Radiology, University of Washington, Seattle, WA, United States
Synopsis
In diffusion imaging, bulk and physiological motion together with strong diffusion encoding gradient introduces extra image phase or data corruption in the diffusion-weighted images. Corrupted data identification/rejection procedure has not been integrated in the recently proposed k-space based multi-shot diffusion reconstruction pipelines. In this work, two corrupted data rejection strategies were proposed, compared and evaluated. Results show that using corrupted data identification and rejection after the CK-GRAPPA reconstruction is potentially a robust choice for multi-shot diffusion imaging reconstruction.
Purpose
In diffusion imaging, bulk and physiological motion together with diffusion encoding gradient introduces extra image phase or data corruption in the diffusion-weighted images. Recently, k-space based phase variation correction methods have been proposed using the virtual coil GRAPPA-like concept in multi-shot diffusion imaging1-4. However, the corrupted data identification/rejection is not integrated in these reconstruction pipelines. Unlike the single-shot scenario, the outlier exclusion in multi-shot acquisition is preferred to be implemented in the reconstruction of multiple shot data rather than in the post-processing of diffusion model fitting5. This is because data from multiple shots contribute to the reconstructed image, and by excluding certain corrupted shots can still result in image without artifact; while excluding the whole diffusion-weighted image during post-processing may waste the uncorrupted, usable data. In this work, two corrupted data rejection strategies in k-space based multi-shot diffusion reconstruction were proposed, compared and evaluated. This work aims to help researchers set up robust reconstruction pipeline for multi-shot diffusion imaging.Theory
In this work, diffusion weighted interleaved EPI acquisition with CK-GRAPPA (GRAPPA with a compact kernel2) reconstruction (Fig. 1A and Fig. 2A) is used for demonstration. For the original CK-GRAPPA implementation (Fig. 2A), each shot is recovered to a fully-sampled k-space using CK-GRAPPA interpolation from multi-shot and multi-channel data. When the data corruption occurs, the identification/rejection can be done before (method 1) or after (method 2) the k-space interpolation, and the corresponding pipeline is shown in Fig. 2B and Fig. 2C. When the corrupted data are excluded before the k-space reconstruction, the acquired data are no longer uniformly interleaved in k-space. In this work, we used flexible kernel GRAPPA reconstruction (FK-GRAPPA, Fig. 1B) for reconstruction under this circumstance. By skipping the corrupted k-space lines, the kernel shapes and weights will be flexible when the different shots are excluded.Methods
In order to have a good reference when comparing different reconstruction approach, a 3-shot data set was simulated from an acquired single-shot data set. The single-shot diffusion EPI data set was acquired on a healthy volunteer on a 3T scanner (Philips, Best, the Netherland) with a 32-channel head coil. The scan was repeated 80 times using diffusion directions along superior/inferior direction with b=0, 2000s/mm2. FOV=230×230mm2, matrix=64×63, TE/TR=97/2000ms. For each slice, one corrupted shot and two normal shots were selected using the k-space entropy metric as in the previous work6, and these three shots are interleavedly combined to simulate the multi-shot data set. The k-space central portion with 21 ky lines was cropped as navigator data. On this data set, four different reconstruction methods were implemented and compared: 1) conventional CK-GRAPPA without corrupted data rejection; 2) GRAPPA (each shot with R=3) with corrupted data rejection; 3) proposed method 1 which uses data rejection before the FK-GRAPPA interpolation; 4) proposed method 2 which uses data rejection after the CK-GRAPPA interpolation. The RMSE was calculated in brain region using the average from 40 good shots as reference. For all k-space reconstruction methods, the kernel size was 3×5 (kx×ky) and the λ=1e-6 for calibration regularization.Results and Discussion
Fig. 3 shows the reconstruction from two typical slices and Table. 1 shows the RMSE of different methods. For CK-GRAPPA without corrupted data rejection, the error maps show large deviation (the brainstem, the cerebellum and the corpus callosum in Fig. 3, yellow arrowheads) from the reference. The error area corresponds with the corrupted shot at the right column. For the GRAPPA with data rejection, the erroneous deviation is reduced, but the RMSE is still large in some case. For the proposed two methods, the RMSEs are both smaller than both CK-GRAPPA without corrupted data rejection and GRAPPA with data rejection (P<0.01, paired t-test). Between the two methods, the proposed method 2 (data rejection after the k-space based reconstruction) shows smaller (P=0.032) RMSE than the proposed method 1 (data rejection before the k-space based reconstruction).
In this simulation, the data from single-shot acquisition were used to generate multi-shot data, in order to keep the gold standard reference and simulate the real pulsatile motion. The results show that the corrupted data identification and rejection is an indispensable procedure. Using corrupted data identification and rejection after the CK-GRAPPA reconstruction provides the result with minimum error compared with the reference. Future work should be focused on testing the algorithm on different b values, different organs (cervical spine, liver, heart, etc.) and different multi-shot acquisition settings.
Conclusion
Implementing corrupted data identification and rejection in k-space based multi-shot diffusion reconstruction can reduce signal errors. Using corrupted data identification and rejection after the CK-GRAPPA reconstruction is potentially a robust choice for diffusion imaging reconstruction.Acknowledgements
No acknowledgement found.References
1. Liu W, Zhao X, Ma Y, Tang X, Gao JH. DWI using navigated interleaved multishot EPI with realigned GRAPPA reconstruction. Magn. Reson. Med. 2016;75:280–286 doi: 10.1002/mrm.25586.
2. Ma X, Zhang Z, Dai E, Guo H. Improved multi-shot diffusion imaging using GRAPPA with a compact kernel. Neuroimage 2016;138:88–99 doi: 10.1016/j.neuroimage.2016.05.079.
3. Zhang Z, Ma X, Dai E, Zhang H, Guo H. Self-calibrated K-space Phase Correction Method for Multi-shot Diffusion Imaging. In: Proc. Intl. Soc. Mag. Reson. Med. 24 (2016); 2016. p. 3020.
4. Dong Z, Wang F, Reese TG, et al. Tilted-CAIPI for highly accelerated distortion-free EPI with point spread function (PSF) encoding. Magn. Reson. Med. 2018:20–22 doi: 10.1002/mrm.27413.
5. Chang LC, Jones DK, Pierpaoli C. RESTORE: Robust estimation of tensors by outlier rejection. Magn. Reson. Med. 2005;53:1088–1095 doi: 10.1002/mrm.20426.
6. Zhang Z, Guo H, Hu Z, Liu Y, Wang Y, Yuan C. Comparison of Different Methods for Motion-induced Data Corruption Detection Using k-space Information in Diffusion Imaging. In: Proc. Intl. Soc. Mag. Reson. Med. 26 (2018); 2018. p. 5325.
Figures
