A Comprehensive Approach for Learning-based Fully-Automated Inter-slice Motion Correction for Short-Axis Cine Cardiac MR Image Stacks

Giacomo Tarroni, Ozan Oktay, Matthew Sinclair, Wenjia Bai, Andreas Schuh, Hideaki Suzuki, Antonio de Marvao, Declan P. O'Regan, Stuart A. Cook, Daniel Rueckert

Introduction

In clinical cardiac MRI, short-axis (SA) images are considered the gold standard for assessing heart function. However, these images are typically acquired as a stack of 2D slices over multiple breath-holds. Inconsistent breath-hold positions lead to inter-slice motion artifacts, often appearing as a "staircase" effect in 3D reconstructions. This paper presents a fully automated, learning-based framework to correct these artifacts using Hybrid Decision Forests and Probabilistic Segmentation Maps (PSMs).

The Challenge of Inter-slice Motion

The primary challenge in cardiac MRI is the misalignment of SA slices, which complicates volumetric calculations and 3D visualization. Manual correction is labor-intensive and prone to error, while existing automated methods often struggle when high-quality reference images are unavailable or themselves misaligned. This research addresses the need for a robust system that can operate effectively even in the absence of standard reference images.

Proposed Methodology

The authors utilize Hybrid Decision Forests to generate Probabilistic Segmentation Maps (PSMs) rather than binary masks. PSMs represent the likelihood of a pixel belonging to the left ventricular (LV) cavity, providing a smoother signal for registration algorithms. Binary masks often have jagged edges that can confuse alignment; probabilistic maps provide a gradient that guides slices into place more effectively.

Overview of the motion correction pipeline using Long-Axis (LA) images and Probabilistic Segmentation Maps (PSMs).

The system operates in two primary modes:

  1. With Long-Axis (LA) Images: The model predicts the LV cavity and identifies landmarks like the apex and mitral valve to guide registration.
  2. Without LA Images (3D PSM): The system uses a 3D model trained on motion-free data to predict a "correct" geometry for registration.

A key innovation is the constraint mechanism: the system only attempts to move slices that have a high probability of containing the heart and fall between the detected apex and mitral valve. This prevents the algorithm from creating spurious alignments in regions where the heart is not present.

Results and Analysis

The study utilized 350 clinical scans for training and 500 UK Biobank scans for generalization testing. Evaluation was performed on a subset of 24 subjects with significant motion artifacts. The proposed method (MC LA PSMs) significantly outperformed traditional intensity-based registration and state-of-the-art "hard" segmentation methods.

Key performance metrics include:

  • Accuracy: Mean Absolute Distance (MAD) was reduced from 3.1 mm (unaligned) to 1.9 mm.
  • Reliability: Successfully improved the alignment of 92% of the motion-corrupted slices.
  • Efficiency: The process takes approximately 25–36 seconds on a standard 6-core CPU, making it suitable for clinical workflows.
Visual comparison of motion correction results across Experiment A and B against the A3D reference.

As shown in the visual results, the PSM-based approach effectively smooths the "staircase" artifacts, bringing the SA stack closer to the anatomical 3D (A3D) reference. The method maintained high performance even when tested on completely different datasets (Experiment B), proving its robustness against variations in scanners and protocols.

Conclusion

By prioritizing probabilistic maps over hard segmentations and using anatomical landmarks as "guardrails," this method offers a robust and clinically viable solution for cardiac MRI motion correction. While the study focused on healthy subjects, the results demonstrate a significant step toward fully automated, reliable cardiac image processing. Future work may involve extending this framework to pathological cases and integrating modern deep learning architectures.

Summary updated

Abstract

In the clinical routine, short axis (SA) cine cardiac MR (CMR) image stacks are acquired during multiple subsequent breath-holds. If the patient cannot consistently hold the breath at the same position, the acquired image stack will be affected by inter-slice respiratory motion and will not correctly represent the cardiac volume, introducing potential errors in the following analyses and visualisations. We propose an approach to automatically correct inter-slice respiratory motion in SA CMR image stacks. Our approach makes use of probabilistic segmentation maps (PSMs) of the left ventricular (LV) cavity generated with decision forests. PSMs are generated for each slice of the SA stack and rigidly registered in-plane to a target PSM. If long axis (LA) images are available, PSMs are generated for them and combined to create the target PSM; if not, the target PSM is produced from the same stack using a 3D model trained from motion-free stacks. The proposed approach was tested on a dataset of SA stacks acquired from 24 healthy subjects (for which anatomical 3D cardiac images were also available as reference) and compared to two techniques which use LA intensity images and LA segmentations as targets, respectively. The results show the accuracy and robustness of the proposed approach in motion compensation.
BibTeX
@inproceedings{tarroni2018comprehensive,
  author = {Giacomo Tarroni and Ozan Oktay and Matthew Sinclair and Wenjia Bai and Andreas Schuh and Hideaki Suzuki and Antonio de Marvao and Declan P. O'Regan and Stuart A. Cook and Daniel Rueckert},
  title = {{A Comprehensive Approach for Learning-based Fully-Automated Inter-slice Motion Correction for Short-Axis Cine Cardiac MR Image Stacks}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2018},
  doi = {10.1007/978-3-030-00928-1\_31},
}
Figure