Introduction

Intravoxel incoherent motion (IVIM) is a diffusion imaging technique utilizing a range of diffusion weightings (b-values) to estimate fast and slow water movement within a tissue compartment¹. IVIM is of interest for non-invasive study of the human placenta, as placental perfusion difficulties are thought to contribute to diseases of pregnancy such as inter-uterine growth restriction². However, when using IVIM to assess the human placenta in utero, volumes representing individual b-values may become misaligned during imaging due maternal respiration, uterine contraction, and fetal movements. Registration algorithms (e.g. Elastix³) exist to correct these misalignments by re-orienting and scaling individual volumes to a reference image. Volumes with higher b-values are more challenging to register because of their low SNR and differing tissue representation (e.g. amniotic fluid signal) relative to the reference volume. Ben-Amitay et al⁴ suggested a method of high b-value registration in the brain by fitting an initial diffusion curve without the high b-values and applying this model of signal decay to the reference volume, resulting in a “template” reference volume similar in features to the high b-value volume to be registered. The purpose of this study was to test the efficacy of a similar approach in registering high b-value diffusion-weighted volumes of the human placenta.

Methods

With approval of our local research ethics board, 3 pregnant volunteers (31.6yo, 29-36; gestational age 33.9 wks; 31.1-36.4) were scanned free-breathing at 1.5T (GE Optima MR450W, Milwaukee, WI) using a 32-coil body array. The IVIM dataset (TR/TE: 7000/71.4 ms, axial, FOV 50cm, 128x128 matrix, 5mm slice, A-P direction, 11 b-values 0-750 s/mm², 3-4 acquisitions each) was motion-corrected by aligning each b>0 volume to the b=0 volume using a rigid-affine-nonlinear (B-spline) processing stream with Mattes mutual information cost function within Elastix^3,5. A voxel-wise biexponential fit was performed with 10 b-values 0-300 s/mm², each with 3-4 data points, using Matlab (2015a, The Mathworks, Natick, MA) (Fig. 1). The biexponential fit at each voxel location was then extrapolated from the b=0 volume to estimate the signal decay at higher b-values (b=500 and 750 s/mm²), producing a “pseudo” volume for each in the reference space of the b=0 (Fig. 1). The acquired high b-value volumes were then registered to the “pseudo” volumes. Regions-of-interest (ROIs) representing the placental boundary were drawn on 3 slices per volume representing the ¼, ½, and ¾ planes in a stack of axial images for the b=0 reference volume, the b=50,300,500,750 volumes post-registration to original b=0, and the b=500,750 volumes post-registration respective to their “pseudo” volumes. ROI drawing and calculation of ROI overlap between registered and reference volumes was performed with FSL⁶. The Dice coefficient⁷ was used to assess the overlap of b=0-registered and “pseudo”-registered volumes relative to the b=0 reference volume. A one-way ANOVA and Tukey’s test for multiple comparisons were performed between the Dice coefficients of each b-value and registration strategy using Prism (v7.03, GraphPad, San Diego, CA).

Results

The “pseudo” volumes show more visual similarity to the high b-dependant tissue features than the b=0 image originally used as a registration reference (Fig. 2). Final registrations to the “pseudo” volumes show improvement in alignment to the reference space compared to registering directly to the b=0 reference image (Fig. 3). The mean Dice coefficients when using the b=0 reference image were 0.83+/-0.07, 0.81+/-0.06, 0.80+/-0.06, and 0.71+/-0.05 for b=50, 300, 500, and 750, respectively (Fig. 4). The mean Dice coefficient when using the “pseudo” reference image was 0.82+/-0.06 for b=500 and 0.78+/-0.09 for b=750. The ANOVA was significant (p<0.005), and Tukey’s test revealed that b=750, when registering to the original b=0, achieved a significantly worse Dice coefficient than b=50 or b=300.

Discussion

Alignment of the human placenta across multiple MRI volumes during a free-breathing acquisition continues to be a challenge for available registration methods. The proposed technique of using a “pseudo”-reference image significantly improved the registration accuracy in high-b-value, low-SNR diffusion volumes to the level of accuracy obtained with higher SNR volumes. Calculation of the “pseudo”-reference relies upon the accuracy of registration and subsequent biexponential fit of b-values 0-300 s/mm², and thus could be corrupted in cases of extreme motion artifact within these volumes.

Conclusion

A registration reference volume was created with signal intensities and tissue features appropriate to high diffusion-weighted volumes of the uterus and placenta, and demonstrated an improvement in registration of these images compared to use of the full signal b=0 reference volume. This technique shows promise in alleviating some of the challenges involved in quantification of the human placenta in utero using intra-voxel incoherent motion.

Acknowledgements

Grant support from the Children’s Health Research Institute, National Institutes of Health, U01 HD087181-01 and Canadian Institutes of Health Research, MOP-209113.