Deformable Slice-to-Volume Registration for Motion Correction of Fetal Body and Placenta MRI

Alena Uus, Tong Zhang, Laurence H. Jackson, Thomas A. Roberts, Mary A. Rutherford, Joseph V. Hajnal, Maria DeprezIEEE TMI 2020 · OpenReview

TL;DR

This paper introduces Deformable Slice-to-Volume Registration (DSVR), a novel method for correcting motion artifacts in fetal body and placenta MRI scans. DSVR addresses the limitations of existing rigid registration techniques by incorporating deformable transformations and a hierarchical approach, leading to significantly improved reconstruction quality for these complex, non-rigidly moving organs.

What Problem Does This Paper Address?

Imaging fetuses in utero using MRI presents a significant challenge due to constant motion. While techniques like Slice-to-Volume Registration (SVR) have been successful in correcting rigid motion, particularly for the fetal brain, they fall short when dealing with the fetal body and placenta. These organs undergo local, non-rigid deformations (bending, stretching) caused by fetal movement and maternal breathing. Rigid registration methods fail to accurately align these structures, leading to blurred features, loss of detail, and overall degradation of the reconstructed 3D volumes. This limits the diagnostic potential of fetal MRI for assessing the body and placenta.

Methods

The core innovation of this paper is the Deformable Slice-to-Volume Registration (DSVR) method. It extends the established rigid SVR framework by introducing several key components:

  • Hierarchical Deformable Registration: Instead of a single rigid transformation, DSVR employs a multi-stage approach. It starts with global rigid and Free-Form Deformation (FFD) registration to align entire stacks to a template. Then, it iteratively refines this using deformable registration between the reconstructed volume and individual slices. This hierarchical process allows for gradual correction of both in-plane and out-of-plane deformations.
  • B-spline FFD for Deformable Transformations: The paper utilizes B-spline Free-Form Deformations (FFD) for capturing non-rigid shape changes. The resolution of these deformations is progressively refined throughout the DSVR iterations by decreasing the B-spline control point spacing.
  • Structure-Based Outlier Rejection: To mitigate the impact of registration errors, DSVR incorporates a two-tiered outlier rejection strategy:
  • Global Structure-Based Outlier Rejection: This step uses normalized cross-correlation (NCC) to identify and reject slices with poor global alignment to the reconstructed volume.
  • Local Structure-Based Outlier Rejection: This step employs the Structural Similarity Index Measure (SSIM) to detect and reject regions within slices that exhibit low structural similarity to the corresponding regions in the reconstructed volume. This is crucial for handling regional misregistrations.
  • Integration with Super-Resolution (SR): DSVR is integrated into an iterative SR reconstruction pipeline, allowing for the generation of high-resolution, motion-free volumes.

Datasets

The study utilized two main types of data:

  • Simulated Experiments: A high-quality reference volume was created from a minimal motion dataset. Motion-corrupted stacks were then generated by applying simulated deformable and rigid transformations. This allowed for controlled evaluation of the method's performance against known ground truth.
  • Real Fetal MRI Data:
    • 20 fetal MRI datasets from 28-31 weeks gestational age (GA) were used for quantitative evaluation. These datasets were categorized into minor and severe motion groups based on visual assessment.
  • 100 fetal body cases from 20-34 weeks GA were used for qualitative evaluation. These were graded by clinicians for image quality.
  • 10 iFIND fetal MRI cases were used for placenta reconstruction evaluation.

Key Insights and Ideas

  • Deformable motion is the key challenge: The paper strongly emphasizes that rigid registration is insufficient for fetal body and placenta MRI due to their inherent non-rigid deformations. This highlights a critical gap in existing motion correction techniques for these specific anatomical regions.
  • Hierarchical refinement is crucial: The multi-stage approach of DSVR, starting with global alignment and progressively refining with local deformable registration, is essential for handling the complex deformations. This suggests that a single-step deformable registration might struggle with the large range of motion.
  • Robust outlier rejection is vital: The introduction of both global and local structure-based outlier rejection significantly improves the robustness of the reconstruction, preventing erroneous registrations from degrading the final volume. This is particularly important in cases of severe motion.
  • DSVR generalizes SVR: The work effectively extends the successful SVR paradigm from rigid motion (fetal brain) to deformable motion (fetal body and placenta), opening up new possibilities for detailed fetal imaging.

Key Results

  • Quantitative Improvement: DSVR consistently outperformed rigid SVR and Patch-to-Volume Registration (PVR) in simulated experiments and real fetal MRI datasets.
  • In simulated experiments, DSVR achieved significantly lower Normalized Root Mean Square Error (NRMSE) and higher Peak Signal-to-Noise Ratio (PSNR) and Normalized Cross-Correlation (NCC) compared to SVR and PVR.
  • Target Registration Error (TRE) for DSVR was significantly lower than for SVR, indicating more accurate recovery of non-rigid transformations.
  • In leave-one-out analysis on fetal body reconstruction, DSVR showed superior NCC and PSNR values compared to SVR and PVR, especially in severe motion cases.
  • DSVR also demonstrated better performance in placenta reconstruction, achieving higher NCC values.
  • Qualitative Improvement:
    • DSVR reconstructions exhibited well-defined local anatomy, including the spine, heart, and abdominal organs, with significantly less blurring and loss of texture compared to SVR and PVR.
  • Qualitative evaluation of 100 fetal body cases showed that DSVR provided sufficient quality for interpretation in the majority of cases (94%), with average grades of 3.09±0.78.

Limitations and Future Work

  • Large Amplitude Motion: The current DSVR implementation, relying on gradient descent-based optimization, struggles with very large amplitude motions, such as significant rotations and bending. In such cases, it relies heavily on outlier rejection.
  • Future Work: The authors suggest exploring Convolutional Neural Network (CNN)-based methods, which have shown promise for rigid SVR, to address this limitation and handle larger motions.
  • Volume Preservation: The paper acknowledges that the question of volume preservation in DSVR reconstructions remains open.
  • Future Work: Future research could focus on developing model-based constraints for deformation fields and automating template selection and masking to improve volume preservation.
  • Maternal Motion Decoupling: For improved correction of large amplitude motion, particularly related to maternal movement, further decoupling of maternal motion is suggested.

Summary updated

Abstract

In in-utero MRI, motion correction for fetal body and placenta poses a particular challenge due to the presence of local non-rigid transformations of organs caused by bending and stretching. The existing slice-to-volume registration (SVR) reconstruction methods are widely employed for motion correction of fetal brain that undergoes only rigid transformation. However, for reconstruction of fetal body and placenta, rigid registration cannot resolve the issue of misregistrations due to deformable motion, resulting in degradation of features in the reconstructed volume. We propose a Deformable SVR (DSVR), a novel approach for non-rigid motion correction of fetal MRI based on a hierarchical deformable SVR scheme to allow high resolution reconstruction of the fetal body and placenta. Additionally, a robust scheme for structure-based rejection of outliers minimises the impact of registration errors. The improved performance of DSVR in comparison to SVR and patch-to-volume registration (PVR) methods is quantitatively demonstrated in simulated experiments and 20 fetal MRI datasets from 28–31 weeks gestational age (GA) range with varying degree of motion corruption. In addition, we present qualitative evaluation of 100 fetal body cases from 20–34 weeks GA range.
BibTeX
@article{uus2020deformable,
  author = {Alena Uus and Tong Zhang and Laurence H. Jackson and Thomas A. Roberts and Mary A. Rutherford and Joseph V. Hajnal and Maria Deprez},
  title = {{Deformable Slice-to-Volume Registration for Motion Correction of Fetal Body and Placenta MRI}},
  journal = {IEEE Trans. Medical Imaging},
  year = {2020},
  url = {https://openreview.net/forum?id=l4kTZvW1Fg},
}
Figure