Intensity-Based 3D Motion Correction for Cardiac MR Images

Nil Stolt Ansó, Vasiliki Sideri-Lampretsa, Maik Dannecker, Daniel Rueckert

Abstract

Cardiac Magnetic Resonance (CMR) imaging is the gold standard for non-invasive cardiac assessment, yet its accuracy is frequently compromised by respiratory motion. Because 3D volumes are typically reconstructed from multiple 2D slices acquired during separate breath-holds, inconsistencies in diaphragm position lead to slice misalignment. This paper presents a fast, intensity-based optimization algorithm designed to correct these 3D alignment errors without the need for pre-trained models or manual labeling. By maximizing pixel-level agreement at the intersections of different imaging planes, the method achieves robust alignment in approximately 30 seconds.

Introduction

In clinical practice, Cardiac Magnetic Resonance (CMR) imaging provides unparalleled detail of the heart's structure and function. However, the acquisition process is inherently sensitive to patient motion. To capture the full volume of the heart, clinicians typically acquire a series of 2D "slices." Because these slices are gathered over multiple breath-holds, the heart's position often shifts between acquisitions.

The resulting "stack" of images is frequently misaligned—a phenomenon often compared to a nudged deck of cards. These misalignments introduce significant errors in volumetric calculations and hinder the creation of accurate 3D heart reconstructions. Traditional correction methods often rely on time-consuming 3D reference scans or complex deep learning models that require extensive manual annotation. This research proposes a "plug-and-play" alternative: a subject-specific optimization approach that utilizes raw image data to resolve alignment errors on the fly.

Methodology

The proposed method treats motion correction as a mathematical optimization problem, focusing on the inherent geometric relationships between different imaging views.

1. Rigid Transformation Model

The algorithm assumes that each slice acts as a rigid body. The goal is to find the optimal rotation ($\theta$) and translation ($t$) parameters in 3D space for every slice simultaneously.

2. Intersection Logic and Loss Function

The core of the approach lies in the intersection of the Short Axis (SA) and Long Axis (LA) planes. In a perfectly aligned 3D volume, the pixels along the line where an SA slice and an LA slice intersect should represent the same anatomical location and thus have nearly identical intensity values.

The researchers developed a loss function based on the $L_2$ (squared) difference between intensities along these intersection lines. This calculation is performed across all time frames of the cardiac cycle (cine MRI), ensuring that the alignment is consistent throughout the heartbeat.

3. Simultaneous Optimization

Rather than aligning slices one by one, the algorithm optimizes the positions of all slices at once. By leveraging the Adam optimizer and GPU acceleration via PyTorch, the system can process the complex registration of an entire cardiac scan in roughly 30 seconds on consumer-grade hardware (e.g., an NVIDIA RTX 2070).

Experimental Setup

To validate the algorithm, the researchers utilized 10 motion-free CMR datasets from the UK Biobank. Each dataset consisted of 9–13 Short Axis slices and three Long Axis views (2-chamber, 3-chamber, and 4-chamber), with 50 time frames per scan.

Because these datasets were originally "clean," the researchers applied synthetic, random rotations and translations to create "broken" scans. This allowed for a precise quantitative evaluation of the algorithm’s ability to recover the original, correct alignment.

Key Insights

Anatomy-Agnostic Alignment

A significant finding of this work is that anatomical knowledge—such as heart contours or landmarks—is not strictly necessary for alignment. By focusing purely on the intensity agreement at plane intersections, the algorithm naturally recovers the correct anatomy.

The "Translation Trap" and Rotational Synergy

The study revealed that correcting pure translation is often more challenging than correcting rotation. If a slice is shifted significantly, the intersection line may move into entirely different tissue (e.g., from the blood pool to the lung). This creates a "flat" gradient where the optimizer lacks the information needed to find the correct direction. Interestingly, the researchers found that the presence of rotation can actually aid the optimizer in resolving translation errors by preserving more overlapping intensity information during the search.

Results

The optimization method demonstrated high robustness and accuracy:

  • Recovery Range: The algorithm successfully recovered alignments even when slices were synthetically displaced by up to 22.5 mm and rotated by 22.5 degrees.
  • Quantitative Performance: Error distribution analysis showed a significant drop in absolute error for both rotation and translation across all tested ranges following optimization.
  • Visual Fidelity: Post-correction 3D reconstructions of the left ventricle showed a transition from "jagged," misaligned edges to smooth, anatomically plausible surfaces.

Limitations and Future Work

While the results are promising, several areas for improvement remain:

  • Local Minima: The optimizer can occasionally become trapped in sub-optimal solutions if the initial misalignment is extreme. Future iterations may incorporate stochastic optimization techniques to improve global convergence.
  • Intensity Consistency: The current $L_2$ loss assumes identical brightness across views. However, MRI artifacts can cause intensity variations. Transitioning to a Mutual Information metric may provide better results by focusing on statistical patterns rather than raw intensity.
  • Deformable Motion: The current model is restricted to rigid transformations. It does not yet account for "through-plane" motion or the stretching and deformation of the heart tissue between breath-holds.

Conclusion

This intensity-based optimization offers a fast, effective, and label-free solution for 3D motion correction in cardiac MRI. By transforming a registration problem into a GPU-accelerated optimization task, the method provides a practical tool for improving the accuracy of cardiac volume measurements and 3D modeling in clinical settings.

Summary updated

BibTeX
@inproceedings{anso2024intensity,
  author = {Nil Stolt Ansó and Vasiliki Sideri-Lampretsa and Maik Dannecker and Daniel Rueckert},
  title = {{Intensity-Based 3D Motion Correction for Cardiac MR Images}},
  booktitle = {ISBI},
  year = {2024},
  doi = {10.1109/isbi56570.2024.10635803},
}
Figure