1780
Application of Probabilistic Modeling to Motion Correction of Neonatal Brain Resting-State BOLD Data1Department of Biomedical Informatics, University of Pittsburgh, Pittsburgh, PA, United States, 2Department of Radiology, Children's Hospital of Pittsburgh of UPMC, Pittsburgh, PA, United States
Synopsis
Functional connectivity studies commonly use resting-state BOLD MR images to study the neurodevelopment of healthy and at-risk neonates. BOLD images are highly sensitive to motion; post-acquisition motion correction techniques can be applied to BOLD data to compensate for motion. We compare the corrective performance of two motion correction techniques on a cohort of 17 healthy neonates: the traditional correction to the first volume technique and a novel, HMM-based motion correction technique. We evaluate the corrected images in terms of the Power et al. thresholds and show the HMM-based technique can be used to recover neonatal BOLD data corrupted by motion.
Introduction
Over the last decade, there has been an increase in the number of functional connectivity studies in neonates for healthy and clinical at-risk populations (infants with prematurity, congenital heart disease, intrauterine growth retardation, etc). Patients in these populations often undergo resting-state blood-oxygen-level-dependent (BOLD) magnetic resonance imaging (MRI) scans, which are highly sensitive to motion: any movement causes artifacts that can lead to spurious results. Some motion artifacts in BOLD images can be corrected using post-acquisition motion correction techniques, though established techniques are not able to correct all artifacts. The use of probabilistic modeling concepts for post-acquisition motion correction shows promise in recovering previously unusable images. As we are focused on motion in neonatal images, we define a “recovered image” as an image that meets the Power et al. displacement and RMS intensity change thresholds after correction1. Here, we compare the traditional “correct to the first volume” technique to a novel hidden Markov model (HMM) based motion correction technique by applying both techniques independently to 17 healthy neonatal BOLD images.Methods
A cohort of 17 healthy, full-term neonates were recruited for an ongoing, IRB approved study of chronic heart disease. The subjects were scanned on either a 3T Skyra (Siemans) or a 3T Achieva (Philips) using the following scanning parameters: FOV = 240 mm, TE/TR = 32/2020 ms, and an isotropic voxel size of 4.0 mm. The acquired sequences consisted of 150 volumes of size 64x64x32 voxels.
The first motion correction method was the correction to the first volume method 2,3. When this method is applied to an image, every volume in the image sequence is registered to the first volume. This method minimizes the difference between the first volume and all other volumes, but not the difference between any other pair of volumes. The second motion correction method was based on the idea of an HMM 4. The BOLD image itself is viewed as an HMM with a collection of deformations as the hidden states. The deformations are the unknown transformation between the first volume in the sequence and each volume (hidden state), and can be estimated as a byproduct of the volume registration process. The Markov property of the deformations is used during the volume registration process: each volume is prealigned to the template volume using the previously estimated transformation. This process minimizes the differences between all volumes in the image. All volume registrations for both methods used affine registrations and were performed using ANTS 5.
Results
Motion in the cohort was characterized using the correlation ratio similarity metric between every pair of volumes in each BOLD image as calculated by FSL 6-8. The correlation ratio matrices for the subjects with the highest average correlation ratio and the median average correlation ratio can be seen in Figures 1 and 2. The mean and standard deviation of the correlation ratio for every original and corrected image in the cohort can be seen in Figure 3. These figures show that, for images with average correlation ratios above 0.035, both motion correction techniques were able to reduce the amount of motion in the images. The corrected images were compared in terms of the displacement and RMS intensity changes between pairs of neighboring volumes as calculated using FSLMotionOutliers 7,8. The Kolmogorov-Smirnov test showed that the distributions of displacement and RMS intensity changes for the different correction methods (Figures 4 and 5) are statistically different for p < 0.001, with the HMM-based correction having a lower average, median, and standard deviation for both the displacement and RMS intensity changes.Discussion
Overall, the HMM-based method showed better corrective ability than the first volume method on this dataset. However, both motion correction techniques address only the impact of motion on patient position in the images, not on the changes of the magnetic gradients (spin history) within the patients. Additional work is needed to account for this component of motion. Furthermore, both techniques take a significant amount of time to run: parallelizing the HMM-based method could produce a technique with the same corrective ability that runs in a fraction of the time. We believe the HMM-based technique has the ability to correct unpredictable motion, such as the motion in fetal images, better than traditional techniques can. We plan to apply the HMM-based technique to a fetal resting-state BOLD MRI dataset in future work.Conclusion
We show for the first time that an HMM-based method has the ability to recover resting-state BOLD data corrupted by motion artifacts in healthy neonates.Acknowledgements
No acknowledgement found.References
1. Power J, Barnes K, Synder A, et al. Spurious by systematic correlations in functional connectivity MRI networks arise from subject motion. NeuroImage. 2012;59(3):2142-2154.
2. Friston K, Williams S, Howard R, et al. Movement-related effects in fMRI time-series. Magnetic Resonance in Medicine. 1996;35(3):346-355.
3. Liao R, Krolik J, McKeown M. An information-theoretic criterion for intrasubject alignment of fMRI time series: Motion corrected independent component analysis. IEEE Transactions on Medical Imaging. 2005;24(1):29–44.
4. Liao R, Turk E, Zhang M, et al. Temporal registration in in-utero volumetric MRI time series. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics). 2016;9902:54–62.
5. Avants B, Tustison N, Stauffer M, et al. The Insight ToolKit image registration framework. Frontiers in Neuroinformatics. 2014.
6. Woolrich M, Behrens T, Beckmann C, et al. Multilevel linear modelling for FMRI group analysis using Bayesian inference. NeuroImage. 2004;21:1732-1747.
7. Smith S, Jenkinson M, Woolrich M, et al. Advances in functional and structural MR image analysis and implementation as FSL. NeuroImage. 2004;23:208-219.
8. Roche A, Malandain G, Pennec X, et al. The correlation ratio as a new similarity measure for multimodal image registration. Lecture Notes in Computer Science: Medical Image Computing and Computer-Assisted Intervention (MICCAI'98). 1998;1496.
Figures
