A pre-processing pipeline for bias field correction and rigid and elastic motion compensation in Placental-Fetal BOLD MRI
Esra Abaci Turk1,2, Jie Luo1,2, Borjan Gagoski 1, Carolina Bibbo3, Julian N. Robinson3, P. Ellen Grant1, Elfar Adalsteinsson2,4,5, Polina Golland4,6, and Norberto Malpica2,7

1Fetal-Neonatal Neuroimaging and Developmental Science Center, Boston Children's Hospital, Harvard Medical School, Boston, MA, United States, 2Madrid-MIT M+Vision Consortium in RLE, Massachusetts Institute of Technology, Cambridge, MA, United States, 3Maternal and Fetal Medicine, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, United States, 4Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, Cambridge, MA, United States, 5Harvard-MIT Health Sciences and Technology, Massachusetts Institute of Technology, Cambridge, MA, United States, 6Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology, Cambridge, MA, United States, 7Medical Image Analysis and Biometry Laboratory, Universidad Rey Juan Carlos, Madrid, Spain


Abnormal oxygen transport through the placenta is thought to be a major etiologic factor in intrauterine growth restriction (IUGR). Blood-oxygen-level-dependent (BOLD) magnetic resonance imaging (MRI) with oxygen exposure is a non-invasive technique that can estimate oxygenation changes in specific organs. However, applying this technique to placentae and fetal organs can be challenging due to non-uniform signal and unpredictable fetal and maternal motion. This study presents a preprocessing pipeline to mitigate signal non-uniformities and motion, to enable automated regional analysis of placenta and fetal body for BOLD MRI studies.


Blood-oxygen-level-dependent (BOLD) time series demonstrate signal changes in placentae and fetal organs during maternal hyperoxia.1,2 Accurate quantification of signal changes is difficult due to complex fetal and maternal movements. Recent studies proposing different motion correction pipelines highlighted the importance of this problem.3,4 Here, we propose a robust preprocessing pipeline compensating rigid and elastic motion within a volume and between volumes for better regional analysis of placentae and fetal organs on BOLD MRI during maternal hyperoxia.


In this IRB approved study, six twin pregnancies with gestational age range 26 to 34 weeks were recruited. BOLD imaging of the placentae and fetal bodies was performed on a 3T Skyra scanner (Siemens Healthcare, Erlangen, Germany) with a maternal oxygenation protocol, designed as three consecutive 10-minute episodes: initial normoxic episode (room air, 21% O2), hyperoxic episode (15 l/min), and a final normoxic episode. Data was acquired with single shot GRE-EPI: TR=5.8-8s, TE=32-38ms, FA=90°, isotropic voxels of 3×3×3mm3, and, total acquisition time 30 minutes. Preprocessing pipeline: 1) Signal non-uniformity correction: 3D-N45 bias correction implementation in ANTS registration suite was used with a mask covering whole uterus. To improve robustness, a single bias field was estimated from the average of selected volumes collected in first 10 minutes and was applied to each frame. 2) Motion correction within each volume: Fetal and placental volumes were acquired with an interleaved slice acquisition resulting in severe motion artifacts between even and odd slices. To compensate these artifacts, first even and odd slice components were separated into two sub-volumes with doubled slice thickness. Then, a group-wise B-spline transformation approach was applied as proposed by Guyader et al.6 to align even and odd slice volumes at each time point. 3) Motion correction between volumes: Mean Square Error (MSE) difference of these motion corrected volumes were computed and the volume with the least MSE difference was chosen as a reference volume. To correct maternal bulk motion, an initial rigid body transformation was created as a mapping from the reference volume to the remaining volumes. Next, a B-spline transformation with mutual information was used for motion correction of the placentae and fetal livers while using the previous rigid transformation as an initialization component. B-spline transformation parameters were adjusted in a subset of volumes to prevent undesired regional volume increase or decrease (e.g. keeping voxel-wise change <50%). In order to avoid unrealistic deformations of the fetal head, a rigid body transformation was applied to the fetal head. Note that Elastix7 toolbox was used for motion correction steps in the pipeline. 4) BOLD Signal Change Visualization: Regions of interest (i.e. placenta, fetal liver and brain) were manually delineated in ITK-SNAP using the reference frame under the supervision of an experienced radiologist. Averaged signal intensities were calculated in these regions of each registered volume. 5) Outlier rejection: A two-step outlier rejection approach was followed. As a first step, after motion correction within a volume, volumes including voxels with negative determinants of the Jacobian of the B-spline transformation were rejected. A second outlier rejection step was performed after the motion correction between volumes. The difference between the intensities of each voxel within a ROI at time t and the next time point t+1 were compared to the mean signal intensity change in the entire ROI at time t. When the difference was larger than this mean value, the voxel at time t+1 was marked as an outlier and not used in the mean signal calculation. However, its value was replaced with the value at time t for the evaluation at the next time point. Volumes with > 5% outlier voxels in the ROI were rejected.


Figure 1 and 2 demonstrate the results of each step in the pipeline for a single data set. Intensity-time plots in Figure 2 illustrate that not only rejecting volumes with outlier voxels but also updating ROIs after rejecting outlier voxels within the ROI decreased the motion related signal variance between time frames. In order to quantify the alignment accuracy, ROIs for placentae, livers and brains were manually delineated on several time points (excluding outlier volumes) and Dice similarity coefficients were computed with respect to the initial masks and the updated masks after the outlier voxel rejection. For 6 datasets, dice coefficients and the percentage of the outlier volumes were given in Table 1. Updating ROIs by rejecting outlier voxels significantly improved the similarity (p<0.001).


It is shown that the proposed preprocessing pipeline can decrease the signal non-uniformity and motion related signal variance between time frames in placental and fetal BOLD MRI.


We would like to thank Dr. Clare Tempany, Dr. Arvind Palanisamy, Javier Pascau, PhD and Ata Turk, PhD. This project is supported by NIH U01 HD087211, NIH R01 EB017337 and the Comunidad de Madrid.


1) Sørensen, A. et al. Changes in Human Fetal Oxygenation during Maternal Hyperoxia as Estimated by BOLD MRI. Prenatal Diagnosis 2013, 33, 141–145.

2) Luo, J. et al. Human placental and fetal response to maternal hyperoxygenation in IUGR pregnancy as measured by BOLD MRI. In ISMRM, Toronto, Canada; 2015.

3) You, Wonsang, et al. Robust motion correction and outlier rejection of in vivo functional MR images of the fetal brain and placenta during maternal hyperoxia. SPIE Medical Imaging, International Society for Optics and Photonics, 2015.

4) Abaci Turk, E. et al. Automated ROI Extraction of Placental and Fetal Regions for 30 minutes of EPI BOLD Acquisition with Different Maternal Oxygenation Episodes. In ISMRM; Toronto, Canada, 2015

5)Tustison NJ et al. N4ITK: improved N3 bias correction. IEEE Trans Med Imaging. 2010;29(6):1310-20.

6) Guyader, Jean Marie, et al. Influence of image registration on apparent diffusion coefficient images computed from free breathing diffusion MR images of the abdomen. Journal of Magnetic Resonance Imaging (2014).

7) Klein, Stefan, et al. elastix: a toolbox for intensity based medical image registration. IEEE Transactions on Medical Imaging, vol. 29, no. 1, pp. 196 - 205, January 2010.


Table 1: Quantitative evaluation of alignment accuracy and percent rejected volumes for placentae, livers and brains.

Figure 1: A. Effect of signal non-uniformity correction, B. Sagittal (left), coronal (middle) and axial (right) views before and after motion correction within a volume, C. Slice view of an averaged volume before and after motion correction and intensity profile along the red line as a function of time.

Figure 2: Intensity vs. time plots for the placenta, fetal livers and brains of a single subject (S1): 2D images in upper left of liver plots demonstrate the updated ROIs (red) and rejected voxels (green) for the liver. Dashed horizontal black lines demark beginning and end of oxygenation exposure.

Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)