Iterative Motion Correction
Srikant Kamesh Iyer, Research Scientist, Canon Medical Research USA, Inc.
Hassan Haji-Valizadeh, Research Scientist, Canon Medical Research USA, Inc.
Sampada Bhave, Research Scientist, Canon Medical Research USA, Inc.
Saurav Sajib, Research Scientist, Canon Medical Research USA, Inc.
Jennifer Wagner, Senior Clinical Research and Applications Specialist, Canon Medical Research USA, Inc.
Samir Sharma, Manager, MR Methods, Canon Medical Research USA, Inc.
Magnetic resonance imaging (MRI) is a relatively slow imaging modality. This can make motion unavoidable during patient scans. Motion can be sporadic, such as involuntary coughing and sneezing, or it can be continuous, such as breathing and cardiac motion. Such motion leads to image quality (IQ) degradation in the form of blurring, ghosting, and replication. In a study of 192 clinical examinations, Andre JB. et al.1 found that 16.4% of scans were repeated due to poor diagnostic quality caused by moderate to severe motion. An example of IQ degradation due to motion is shown in Figure 1. The presence of motion artifacts can hinder clinical diagnosis. This can lead to rescans or patient callback, which further decrease patient comfort. Hence, there is a strong need to develop motion correction methods for MRI.
Hassan Haji-Valizadeh, Research Scientist, Canon Medical Research USA, Inc.
Sampada Bhave, Research Scientist, Canon Medical Research USA, Inc.
Saurav Sajib, Research Scientist, Canon Medical Research USA, Inc.
Jennifer Wagner, Senior Clinical Research and Applications Specialist, Canon Medical Research USA, Inc.
Samir Sharma, Manager, MR Methods, Canon Medical Research USA, Inc.
Magnetic resonance imaging (MRI) is a relatively slow imaging modality. This can make motion unavoidable during patient scans. Motion can be sporadic, such as involuntary coughing and sneezing, or it can be continuous, such as breathing and cardiac motion. Such motion leads to image quality (IQ) degradation in the form of blurring, ghosting, and replication. In a study of 192 clinical examinations, Andre JB. et al.1 found that 16.4% of scans were repeated due to poor diagnostic quality caused by moderate to severe motion. An example of IQ degradation due to motion is shown in Figure 1. The presence of motion artifacts can hinder clinical diagnosis. This can lead to rescans or patient callback, which further decrease patient comfort. Hence, there is a strong need to develop motion correction methods for MRI.
Motion correction methods are designed to maintain diagnostic IQ in the presence of motion, which will in turn reduce rescans and callbacks.
Motion correction techniques can be broadly classified into (a) prospective and (b) retrospective methods. Prospective motion correction (PMC) techniques adjust acquisition position in real-time when motion is detected during scanning. PMC methods typically rely on navigators or external devices, such as cameras and fiducial markers, to estimate motion. PMC has high accuracy and is efficient because little to no data are discarded, however, it has limitations.
Implementing PMC is challenging because it requires real-time calculation of motion parameters and real-time acquisition adjustments. Further, navigators are constrained by sequence design, and camera-based PMC can impede workflow by relying on external hardware and calibrations. Moreover, PMC methods only correct for rigid-body motion.
Retrospective motion correction (RMC) performs motion correction during image reconstruction. RMC needs neither external hardware to detect motion nor real-time feedback to update the acquisition process. Moreover, RMC can correct both rigid and non-rigid motion. Like PMC, data navigators can be combined with RMC for motion correction. RMC methods typically require little to no modification of workflow. These advantages make RMC a suitable choice for routine clinical use.
Motion correction techniques can be broadly classified into (a) prospective and (b) retrospective methods. Prospective motion correction (PMC) techniques adjust acquisition position in real-time when motion is detected during scanning. PMC methods typically rely on navigators or external devices, such as cameras and fiducial markers, to estimate motion. PMC has high accuracy and is efficient because little to no data are discarded, however, it has limitations.
Implementing PMC is challenging because it requires real-time calculation of motion parameters and real-time acquisition adjustments. Further, navigators are constrained by sequence design, and camera-based PMC can impede workflow by relying on external hardware and calibrations. Moreover, PMC methods only correct for rigid-body motion.
Retrospective motion correction (RMC) performs motion correction during image reconstruction. RMC needs neither external hardware to detect motion nor real-time feedback to update the acquisition process. Moreover, RMC can correct both rigid and non-rigid motion. Like PMC, data navigators can be combined with RMC for motion correction. RMC methods typically require little to no modification of workflow. These advantages make RMC a suitable choice for routine clinical use.
Figure 1: Example of axial T2W images without (A) and with (B) motion artifacts. In the case of (B), patient motion led to artifacts and non-diagnostic image quality.
Canon’s JET is an RMC implementation for numerous clinical applications. JET uses radial acquisitions with overlapping blades. Data are acquired in parallel lines rotating around the center of the k-space. The overlapping data allows for the correction of motion. Despite its benefits, the radial acquisition produces artifacts such as streaking, image blurring, and a change of image contrast. Furthermore, radial sampling increases scan time compared to Cartesian sampling and can be applied to only a limited selection of contrast weightings. Therefore, alternative approaches for RMC are sought.
Canon has developed a new approach for robust RMC using iterative motion correction (IMC). IMC aims to improve IQ by incrementally updating the image. Canon’s IMC focuses on maintaining diagnostic IQ in the presence of sporadic motion and can correct both rigid and non-rigid motion. An attractive feature of IMC is its use of Cartesian trajectories for data acquisition, which mitigates some of the challenges associated with radial acquisitions.
Canon released the initial version of IMC in their software release Version 8.0. The initial release of IMC supported motion correction for brain FLAIR and cervical spine T2W imaging. Canon’s latest software release (V9.0) extends IMC to several new contrasts such as T2W and T1W brain imaging as well as T1W and STIR C-spine imaging.
The latest IMC release includes many technical improvements that are designed to expand its clinical applicability and reduce scan and reconstruction times. To improve robustness to motion, the shuffle encoding pattern used in the initial IMC release was retained. An example of the motion robust shuffle encoding pattern is shown in Figure 2. Conventional sampling approaches have uniform spacing of phase encoding lines for each shot, which causes coherent artifacts in the presence of motion (Figure 2A). As seen in Figure 2B, the shuffle encoding pattern has nonuniform spacing of phase encoding lines for each shot, and hence produces incoherent artifacts in the presence of motion. This is similar to the use of nonuniform sampling pattern in compressed sensing (CS). The latest IMC release reduces scan time by acquiring navigator data more efficiently and reduces reconstruction time by decreasing the reconstruction algorithm’s complexity. Lastly, the latest IMC release complements the improved model-based approach with a new machine learning (ML)-based motion correction.
Canon has developed a new approach for robust RMC using iterative motion correction (IMC). IMC aims to improve IQ by incrementally updating the image. Canon’s IMC focuses on maintaining diagnostic IQ in the presence of sporadic motion and can correct both rigid and non-rigid motion. An attractive feature of IMC is its use of Cartesian trajectories for data acquisition, which mitigates some of the challenges associated with radial acquisitions.
Canon released the initial version of IMC in their software release Version 8.0. The initial release of IMC supported motion correction for brain FLAIR and cervical spine T2W imaging. Canon’s latest software release (V9.0) extends IMC to several new contrasts such as T2W and T1W brain imaging as well as T1W and STIR C-spine imaging.
The latest IMC release includes many technical improvements that are designed to expand its clinical applicability and reduce scan and reconstruction times. To improve robustness to motion, the shuffle encoding pattern used in the initial IMC release was retained. An example of the motion robust shuffle encoding pattern is shown in Figure 2. Conventional sampling approaches have uniform spacing of phase encoding lines for each shot, which causes coherent artifacts in the presence of motion (Figure 2A). As seen in Figure 2B, the shuffle encoding pattern has nonuniform spacing of phase encoding lines for each shot, and hence produces incoherent artifacts in the presence of motion. This is similar to the use of nonuniform sampling pattern in compressed sensing (CS). The latest IMC release reduces scan time by acquiring navigator data more efficiently and reduces reconstruction time by decreasing the reconstruction algorithm’s complexity. Lastly, the latest IMC release complements the improved model-based approach with a new machine learning (ML)-based motion correction.
Figure 2: Standard sampling (A) and shuffle encoded sampling (B) are shown. Shuffle encoded sampling (B) exhibits lower motion artifacts compared to standard sampling (A) for in-vivo scanning (red arrows).
Several model-based motion correction techniques have been developed 2-4. These methods integrate a physics model in the image reconstruction pipeline to reduce motion artifacts. Modelling motion parameters preserves the acquired data, which maintains signal to noise ratio (SNR). Despite their advantages, model-based methods can be limited by their long reconstruction times and the inability to correct for all motion types.
Recently, the strength and wide applicability of ML methods has been realized in several MRI applications 5-7. ML offers a data driven approach for rapid motion correction. It does not rely on external hardware. Further, the modeling of motion, which can be computationally complex, is done only during the training stage. During runtime (inference), the ML network rapidly processes the image using the model that was learned during training. These advantages have driven the development of many promising ML-based solutions for motion artifact suppression 8-11. However, ML may show reduced performance for motion not seen during training and may struggle to produce diagnostic IQ for cases with high motion.
Canon’s new approach to motion correction combines model-based and ML-based approaches. This synergistic combination overcomes some of the limitations of each individual component. Additionally, this combined approach allows for greater robustness to motion than model-based or ML-based approaches applied individually. The new IMC release uses a complex-valued residual U-Net (Res U-Net) for ML (Figure 3). The ML network was trained on pairs of with/without motion datasets in a supervised learning framework.
Recently, the strength and wide applicability of ML methods has been realized in several MRI applications 5-7. ML offers a data driven approach for rapid motion correction. It does not rely on external hardware. Further, the modeling of motion, which can be computationally complex, is done only during the training stage. During runtime (inference), the ML network rapidly processes the image using the model that was learned during training. These advantages have driven the development of many promising ML-based solutions for motion artifact suppression 8-11. However, ML may show reduced performance for motion not seen during training and may struggle to produce diagnostic IQ for cases with high motion.
Canon’s new approach to motion correction combines model-based and ML-based approaches. This synergistic combination overcomes some of the limitations of each individual component. Additionally, this combined approach allows for greater robustness to motion than model-based or ML-based approaches applied individually. The new IMC release uses a complex-valued residual U-Net (Res U-Net) for ML (Figure 3). The ML network was trained on pairs of with/without motion datasets in a supervised learning framework.
Figure 3: Architecture of the complex residual U-Net used for ML. The residual U-Net combines 3D residual block processing with the U-Net architecture to enable motion artifact suppression.
An example of the improvement in IQ using the combination of model-based and ML-based approaches is shown in Figure 4. A coronal T2W image acquired in the presence of motion is seen in Figure 4A. Using only the ML network for motion correction does not produce a diagnostic quality image (Figure 4B) and motion artifacts and blurring still remain. Similarly, in Figure 4C, it is seen that some residual motion artifacts remain when only model-based motion correction is used. The best IQ is achieved when a combination of model-based correction and ML is used (Figure 4D).
Figure 4: Coronal T2W image acquired in the presence of motion. Reconstructions are shown without IMC (A), with ML processing only (B), with IMC without ML (C), and with IMC (D). ML-only processing did not produce good image quality. Main features are blurred out (red arrowheads). The model-based method has remaining residual motion artifacts (red arrows). Combining model-based and ML processing produced the best IQ. (E)-(H) shows the zoomed-in image of a small region (indicated by the dashed red box in (A)) for (A)-(D) respectively.
IMC performance for several different contrasts is shown in Figures 5, 6, and 7. In Figure 5, an example of an axial T2 brain image is shown where the input image is corrupted due to motion. The use of IMC helped remove the motion artifacts and improved IQ. The different structures and features in the brain are better visualized in the IMC output. In Figure 6, the IQ benefits of IMC are shown in an axial T1W brain case demonstrating that the ringing artifacts caused by motion have been effectively removed with IMC. IMC can also be used for cervical spine imaging, an example of a sagittal STIR C-spine image is shown in Figure 7. The cord has significant motion artifacts when reconstructed without IMC. IMC recovered the IQ from the motion corrupted input image. These results show the effectiveness of Canon’s latest IMC release for brain and cervical spine imaging applications. //
Figure 5: Axial T2W brain images processed without and with IMC. Without IMC (A), the image has artifacts because of patient motion (red arrows). Using IMC (B), IQ is restored, and the image is of diagnostic quality.
Figure 6: Axial T1W brain images processed without and with IMC. Without IMC (A), the image has artifacts because of patient motion (red arrows). IMC (B) significantly resolves the motion artifacts, leading to a diagnostic quality image.
Figure 7: Sagittal STIR C-spine images without and with IMC. Without IMC (A), the image has artifacts due to motion during the scan (red arrows). IMC (B) significantly resolves the motion artifacts, leading to a diagnostic quality image.
References
1 Andre JB, et al. Towards quantifying the prevalence, severity, and cost associated with patient motion during clinical MR examinations. J Am Coll Radiol 2015; 12: 689-695.
2 Batchelor PG, et al. Matrix description of general motion correction applied to multishot images. Magn Reson Med. 2005; 54: 1273–80.
3 Atkinson D, et al. Automatic correction of motion artifacts in magnetic resonance images using an entropy focus criterion. IEEE Trans Med Imag. 1997; 16: 903–10.
4 Manke D, et al. Novel prospective respiratory motion correction approach for free-breathing coronary MR angiography using a patient-adapted affine motion model. Magn Reson Med. 2003; 50: 122–131.
5 Montalt-Tordera J, et al. Machine learning in Magnetic Resonance Imaging: Image reconstruction. Physica Medica. 2003; 83: 79-87.
6 Chen Y, et al. AI-Based Reconstruction for Fast MRI—A Systematic Review and Meta-Analysis. Proc of the IEEE. 2022; 110(2): 224 – 245.
7 Zeng C, et al. Review of Deep Learning Approaches for the Segmentation of Multiple Sclerosis Lesions on Brain MRI. Front Neuroinform 2020.
8 Küstner T, et al. Retrospective correction of motion-affected MR images using deep learning frameworks. Magn Reson Med. 2019; 82(4): 1527-1540.
9 Pawar K, et al. Suppressing motion artefacts in MRI using an Inception-ResNet network with motion simulation augmentation. NMR in Biomed. 2019; 35(4).
10 Tamada D, et al. Motion Artifact Reduction Using a Convolutional Neural Network for Dynamic Contrast Enhanced MR Imaging of the Liver. Magn Reson Med Sci. 2020; 19(1): 64-76.
11 Qi H, et al. End-to-end deep learning nonrigid motion-corrected reconstruction for highly accelerated free-breathing coronary MRA. Magn Reson Med. 2021: 86(4); 1983-1996.
