Synopsis

Motion is common in MR image acquisitions. It causes artefacts in image quality that requires repeated scans, increasing the burden for patients and providers. There have been advances on hardware and software fronts in the quest to avoid such issues. The latter include data consistency methods which require no additional installation but can fail to detect high frequency outliers in k-space. We propose a Deep CNN approach to detect motion-corrupted phase encode lines, coupled with a low-rank reconstruction. This approach improves outlier detection in comparison to low-rank only methods and accelerates reconstruction time.

Introduction:

Patient motion affects MR images globally and can degrade their diagnostic value. A recent estimate indicates that motion artefacts cost providers $2Billion per year for all scans¹. Motion-control approaches employ hardware (i.e. respiratory belts, camera tracking), signal navigators and data consistency^2-7. The latter has the convenience of being implementable without any additional installations. Data consistency based approaches typically consist of joint iterative phases involving a) outlier detection and b) data replacement exploit redundancies from low rank representation of the data^8-10.

However, relative to the significant redundancy in MR data sets, robust outlier reconstructions can only tolerate a small degree of data corruption (say, 25%) which compares with typical scan time reductions of 3-4x (66-75% data loss). If the location of the corrupted data could be known beforehand, then reconstruction of a 25% corrupted dataset would be essentially trivial.

In this study, we present a deep learning based framework to first detect outliers in motion-artefacted k-space data then use a second step to reconstruct images using low-rank based reconstruction (i.e. SAKE²) . The goal is to improve the performance of outlier detection, reconstruction speed and overall quality of the reconstructed image.

Methods:

Although head motion often corrupts the acquired phase encode lines during MR scans, we usually do not have information about the lines that are affected. Thus it is inevitable that artificially synthesized data is required for CNN training. 500 complex-valued 2D brain images deemed to be free of motion artefact were acquired as part of clinical brain scans over a six-month time-span. Data were contaminated with motion artefacts corresponding to translational motion in 35 to 45 percent of the phase encode lines. Artefacts were dispersed randomly to simulate the effect of desynchronized k-space sampling¹¹.The corruption process was repeated 4 times for each dataset which increased the training size to 2000 complex images. Phase and normalized magnitude values of k-space from 8 coil sets of the data were extracted and fed as the inputs to the deep CNN.

Figure.1 shows our proposed platform to classify the phase encode lines into corrupted and clean data. The outliers were removed after the detection and a final image was reconstructed by using an auto calibration low rank based method (SAKE²).

The performance of the network in detecting outliers was tested using 50 new corrupted datasets and compared with an outlier-robust, trimmed variant of SAKE (TAKE¹⁰) that uses low-rank property to detect outliers. The performance of the CNN method in correcting the motion corruption was compared statistically with TAKE.

Results:

Figure.2 shows the confusion matrix for our proposed CNN approach and TAKE. As can be seen, the proposed deep learning based solution for outlier detection is fifteen percent more accurate than TAKE.

Figure.3 shows two examples of final corrected images from the Deep CNN + SAKE method and TAKE. As noted in Figure.3, TAKE can suffer from high frequency residual artefacts owing to failure to reject outliers at the high frequencies in k-space. This is substantially reduced by Deep CNN + SAKE approach.

Figure.4 (a) summarizes the SSIM and nRMSE values for both methods as a box plot for all test datasets. Paired-sample t-test comparisons between the SSIM and nRMSE values are reported in the Figure.4 (b). Overall there was a significant difference (p<0.05) between the SSIM and nRMSE scores for Deep CNN + SAKE relative to TAKE motion corrected images. Typical reconstruction time for TAKE was 15min/image compared to 1min/image for the Deep CNN+SAKE method.

Discussion:

Low rank outlier detection (such as TAKE) is based on a joint process of rejecting outliers (corrupted phase encoding lines) and filling the matrix by exploiting low rank properties. Finding and rejecting the outliers is the rate-determining process in TAKE algorithm, explaining its need for thousands of iterations in its motion-correction calculations. Deep CNN avoids this time consuming process.

Another potential advantage of deep learning outlier detection is that it can detect features that are not considered in low rank models, e.g. the change in noise variance in receiver coils under different loads5. Any changes arising from motion can be trained into a CNN given suitable training data of real motion artefacts.

By combining Deep CNN outlier detection with a separate reconstruction method, such as SAKE, we were able to reduce reconstruction times and improve SSIM and nRMSE.

Acknowledgements

References

[1] J. B. Andre et al., “Toward Quantifying the Prevalence, Severity, and Cost Associated With Patient Motion During Clinical MR Examinations,” J. Am. Coll. Radiol., vol. 12, no. 7, pp. 689–695, 2015.

[2] P. J. Shin et al., “Calibrationless parallel imaging reconstruction based on structured low‐rank matrix completion,” Magn. Reson. Med., vol. 72, no. 4, pp. 959–970, 2014.

[3] M. Zaitsev, J. Maclaren, and M. Herbst, “Motion artifacts in MRI: a complex problem with many partial solutions,” J. Magn. Reson. Imaging, vol. 42, no. 4, pp. 887–901, 2015.

[4] M. Bydder, D. J. Larkman, and J. V Hajnal, “Detection and elimination of motion artifacts by regeneration of k‐space,” Magn. Reson. Med. An Off. J. Int. Soc. Magn. Reson. Med., vol. 47, no. 4, pp. 677–686, 2002.

[5] A. Andreychenko et al., “Thermal noise variance of a receive radiofrequency coil as a respiratory motion sensor,” Magn. Reson. Med., vol. 77, no. 1, pp. 221–228, 2017.

[6] A. A. Samsonov, J. Velikina, Y. Jung, E. G. Kholmovski, C. R. Johnson, and W. F. Block, “POCS‐enhanced correction of motion artifacts in parallel MRI,” Magn. Reson. Med., vol. 63, no. 4, pp. 1104–1110, 2010.

[7] T. Hilbert, T. Kober, J.-P. Thiran, R. Meuli, and G. Krueger, “Phase-encode ghosting detection using multi-channel coil arrays,” in ISMRM 2016, ISMRM 24rd Annual Meeting & Exhibition, SMRT 25th Annual Meeting, 2016.

[8] A. E. Campbell‐Washburn et al., “Using the robust principal component analysis algorithm to remove RF spike artifacts from MR images,” Magn. Reson. Med., vol. 75, no. 6, pp. 2517–2525, 2016.

[9] K. H. Jin, J. Um, D. Lee, J. Lee, S. Park, and J. C. Ye, “MRI artifact correction using sparse+ low‐rank decomposition of annihilating filter‐based hankel matrix,” Magn. Reson. Med., vol. 78, no. 1, pp. 327–340, 2017.

[10] M. Bydder, S. Rapacchi, O. Girard, M. Guye, and J.-P. Ranjeva, “Trimmed autocalibrating k-space estimation based on structured matrix completion,” Magn. Reson. Imaging, vol. 43, pp. 88–94, 2017.

[11] S. Weick et al., “Desynchronization of Cartesian k‐space sampling and periodic motion for improved retrospectively self‐gated 3D lung MRI using quasi‐random numbers,” Magn. Reson. Med., vol. 77, no. 2, pp. 787–793, 2017.