[Paper Review] PVR: Patch-to-Volume Reconstruction for Large Area Motion Correction of Fetal MRI

Amir Alansary, Bernhard Kainz+9 more|arXiv (Cornell University)|Nov 22, 2016

Fetal and Pediatric Neurological Disorders21 references4 citations

TL;DR

This paper proposes Patch-to-Volume Reconstruction (PVR), a novel motion correction method for fetal MRI that reconstructs high-resolution 3D volumes from motion-corrupted, overlapping 2D slices using patch-wise optimization, super-resolution, and automatic outlier rejection. PVR significantly improves reconstruction accuracy by ~30% over standard slice-to-volume registration (SVR) in synthetic experiments and enables robust, automatic reconstruction of the whole uterus, fetus, and placenta in real clinical data.

ABSTRACT

In this paper we present a novel method for the correction of motion artifacts that are present in fetal Magnetic Resonance Imaging (MRI) scans of the whole uterus. Contrary to current slice-to-volume registration (SVR) methods, requiring an inflexible anatomical enclosure of a single investigated organ, the proposed patch-to-volume reconstruction (PVR) approach is able to reconstruct a large field of view of non-rigidly deforming structures. It relaxes rigid motion assumptions by introducing a specific amount of redundant information that is exploited with parallelized patch-wise optimization, super-resolution, and automatic outlier rejection. We further describe and provide an efficient parallel implementation of PVR allowing its execution within reasonable time on commercially available graphics processing units (GPU), enabling its use in the clinical practice. We evaluate PVR's computational overhead compared to standard methods and observe improved reconstruction accuracy in presence of affine motion artifacts of approximately 30% compared to conventional SVR in synthetic experiments. Furthermore, we have evaluated our method qualitatively and quantitatively on real fetal MRI data subject to maternal breathing and sudden fetal movements. We evaluate peak-signal-to-noise ratio (PSNR), structural similarity index (SSIM), and cross correlation (CC) with respect to the originally acquired data and provide a method for visual inspection of reconstruction uncertainty. With these experiments we demonstrate successful application of PVR motion compensation to the whole uterus, the human fetus, and the human placenta.

Motivation & Objective

• To address the challenge of non-rigid, unpredictable motion in fetal MRI that degrades image quality and hinders diagnosis.
• To overcome the limitations of traditional slice-to-volume registration (SVR), which assumes rigid motion and struggles with large, non-rigidly deforming structures like the uterus and placenta.
• To develop an automatic, segmentation-free method that reconstructs high-resolution 3D volumes from multiple motion-corrupted 2D slice stacks without manual intervention.
• To enable real-time, clinically feasible motion correction through efficient GPU-accelerated parallel implementation.
• To evaluate reconstruction quality beyond single metrics by introducing visual uncertainty inspection and structural similarity (DSSIM) heat maps.

Proposed method

• Patches are extracted from overlapping 2D slices and mapped into a 3D volume using motion-corrected, patch-wise optimization to reconstruct a high-resolution (HR) volume from low-resolution (LR) inputs.
• The method employs a super-resolution model where each LR image is modeled as a downsampled, blurred, and motion-corrupted version of the HR volume: $ x_i = W_i y + n_i $, with $ W_i = D B T_i $, combining sub-sampling, blurring, and transformation.
• A parallelized, GPU-accelerated optimization framework enables efficient computation, with patch selection based on fixed-size squares or multi-scale superpixels to improve efficiency and accuracy.
• Automatic outlier rejection is performed using an EM-based algorithm to identify and exclude inconsistent or corrupted data across overlapping patches.
• The method supports both fixed-size and multi-size patch configurations, with multi-size superpixels showing improved performance on non-rigid regions like the placenta and uterus.
• Reconstruction uncertainty is visualized using DSSIM heat maps to enable qualitative assessment of reliability across the 3D volume.

Experimental results

Research questions

• RQ1 Can a patch-based reconstruction method achieve superior motion correction over standard SVR in the presence of large, non-rigid fetal and maternal motion?
• RQ2 How does the choice of patch type (square vs. superpixel) and size (fixed vs. multi-scale) affect reconstruction accuracy and computational efficiency?
• RQ3 To what extent can PVR reconstruct complex, non-rigidly deforming structures such as the placenta and whole uterus without manual segmentation or region-of-interest definition?
• RQ4 How does PVR perform in clinical scenarios involving multiple fetuses or extreme limb movements?
• RQ5 Can a visual uncertainty metric like DSSIM heat mapping effectively reveal reconstruction reliability in motion-corrupted fetal MRI?

Key findings

• PVR improves reconstruction accuracy by approximately 30% compared to conventional SVR in synthetic experiments with affine motion artifacts.
• PVR achieves superior performance on non-rigidly deforming regions such as the placenta and uterus, where superpixel-based patches outperform square patches.
• The method successfully reconstructs the whole uterus, fetal brain, placenta, and multiple fetuses in twin pregnancies without requiring manual segmentation or region identification.
• Multi-scale superpixel patches are significantly more computationally efficient than overlapping square patches while maintaining comparable reconstruction accuracy.
• For extreme limb movements (>2 cm between slices), PVR fails to reconstruct fine structures due to lack of structural consensus, highlighting a limitation of intensity-based optimization.
• Visual inspection using DSSIM heat maps reveals spatial variations in reconstruction quality, enabling clinicians to identify unreliable regions in the 3D volume.