Intensity-based 3D motion correction for cardiac MR images

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

TL;DR

This paper introduces a subject-specific optimization algorithm that corrects 3D slice misalignments in cardiac MRI (CMR) by maximizing intensity agreement at slice intersections. Unlike traditional methods, it requires no anatomical priors or pre-trained segmentation models, offering a robust, GPU-accelerated solution that aligns all imaging planes simultaneously in under 30 seconds.

What Problem Does This Paper Address?

Cardiac MRI is typically acquired as a series of 2D "cine" slices. Because the heart is a moving target, patients are asked to hold their breath during acquisition to ensure all slices are captured in the same anatomical position. However, humans are imperfect: variations in breath-hold depth and patient movement between scans lead to 3D misalignments.

When these 2D slices are stacked to create a 3D volume, the resulting "jagged" anatomy makes accurate volumetric measurements (like stroke volume or ejection fraction) difficult. Existing solutions often rely on complex shape models or require high-resolution 3D reference volumes that aren't always available in standard clinical workflows. The authors sought a "contour-free" method that works using only the raw image data available in a standard scan.

![Fig.

  1. The top row shows how short-axis (SA) slices (red lines) are often misaligned relative to the long-axis (LA) view. The bottom row shows the "jagged" 3D reconstruction before and after the proposed motion correction.](/figures/art_2004ab2674434435a21cc62138e6275b)

Methods

The core idea is elegant: if two imaging planes (e.g., a Long Axis and a Short Axis view) are perfectly aligned, the pixel intensities along their line of intersection must be identical.

The researchers formulated this as a joint optimization problem:

  1. Rigid Transformation: They assume each slice has undergone a rigid 3D transformation (rotation and translation) due to motion.
  2. Intersection Analysis: For every pair of intersecting slices, the algorithm calculates the intersection line in 3D space.
  3. Intensity Matching: It samples intensities along these lines across all time frames of the cardiac cycle.
  4. Loss Function: The algorithm minimizes the sum of squared (L2) intensity differences between these sampled lines.

By using the Adam optimizer and a GPU-accelerated framework (PyTorch), the system can simultaneously adjust the position and orientation of all slices to find the global "best fit" where all intersections match up.

![Fig.

  1. A visualization of the intensity matching. The red and green lines show intersections before and after optimization. The vertical bars show the sampled intensities; notice how the intensities "align" much better in the green (optimized) version.](/figures/art_81534c77ee32461daf07f5ed71220b02)

Datasets

The method was evaluated using 10 motion-free CMR datasets from the UK Biobank.

  • Each dataset included 9–13 Short Axis (SA) slices and three Long Axis (LA) views (2-chamber, 3-chamber, and 4-chamber).
  • To test the algorithm, the researchers synthetically "broke" these clean datasets by applying random rotations (up to ±22.5°) and translations (up to ±22.5 mm) to each slice.
  • The goal was to see if the algorithm could "undo" these synthetic deformations and return the slices to their original, gold-standard positions.

Key Insights and Ideas

  • Anatomy-Agnostic: The most significant insight is that you don't need to know what a heart looks like to align it. By focusing purely on the physics of intersecting planes, the model avoids the bias and training requirements of segmentation-based methods.
  • Simultaneous Alignment: Rather than fixing one slice and moving another (which can lead to error propagation), the algorithm treats the entire stack as a single system to be optimized.
  • Rotation vs. Translation: The authors found that rotation is actually easier to correct than pure translation. This is because rotations maintain some intensity overlap near the axis of rotation, providing a "gradient" for the optimizer to follow. Pure, large translations can move intersections so far that the optimizer "loses the trail" of similar intensities.

Key Results

The algorithm demonstrated impressive robustness and speed:

  • Convergence: In almost all test cases, the algorithm converged to the motion-free parameters in approximately 30 seconds (1000 iterations).
  • Accuracy: Even with massive initial misalignments (22.5°/22.5mm), the final mean absolute error for rotation remained extremely low.
  • Robustness: The system successfully recovered the original alignment across a wide range of perturbation magnitudes, as shown in the error distribution plots.

![Fig.

  1. Distribution of errors after optimization. Even with large initial perturbations, the average errors (dark green/red bars) remain very close to zero.](/figures/art_3923978889174c1983d935c2c7213e85)

Limitations and Future Work

While the results are promising, the authors noted a few areas for improvement:

  • Local Minima: In scenarios with only large translations and no rotation, the gradient descent can get stuck. The authors suggest exploring stochastic optimization techniques to help the algorithm "jump out" of these local traps.
  • Intensity Consistency: The current L2 loss assumes intensities are identical across slices. However, different MR acquisitions can have slight variations in brightness. Future iterations might use Mutual Information as a loss function, which is more robust to these intensity shifts.
  • Non-Rigid Motion: The current model only handles rigid motion (rotation/translation). It does not account for scaling or shearing, though these are less common in inter-slice breath-hold errors.

Summary updated

Abstract

Cardiac magnetic resonance (CMR) image acquisition requires subjects to hold their breath while 2D cine images are acquired. This process assumes that the heart remains in the same position across all slices. However, differences in breathhold positions or patient motion introduce 3D slice misalignments. In this work, we propose an algorithm that simultaneously aligns all SA and LA slices by maximizing the pair-wise intensity agreement between their intersections. Unlike previous works, our approach is formulated as a subject-specific optimization problem and requires no prior knowledge of the underlying anatomy. We quantitatively demonstrate that the proposed method is robust against a large range of rotations and translations by synthetically misaligning 10 motion-free datasets and aligning them back using the proposed method.
BibTeX
@inproceedings{stoltanso2024intensity,
  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