Deep-Learning-Based Groupwise Registration for Motion Correction of Cardiac $$T_1$$ Mapping

Yi Zhang, Yidong Zhao, Lü Huang, Liming Xia, Qian TaoLecture Notes in Computer Science 2024 · DOI · OpenAlex

Abstract

Quantitative $T_1$ mapping by MRI is an increasingly important tool for clinical assessment of cardiovascular diseases. The cardiac $T_1$ map is derived by fitting a known signal model to a series of baseline images, while the quality of this map can be deteriorated by involuntary respiratory and cardiac motion. To correct motion, a template image is often needed to register all baseline images, but the choice of template is nontrivial, leading to inconsistent performance sensitive to image contrast. In this work, we propose a novel deep-learning-based groupwise registration framework, which omits the need for a template, and registers all baseline images simultaneously. We design two groupwise losses for this registration framework: the first is a linear principal component analysis (PCA) loss that enforces alignment of baseline images irrespective of the intensity variation, and the second is an auxiliary relaxometry loss that enforces adherence of intensity profile to the signal model. We extensively evaluated our method, termed "PCA-Relax", and other baseline methods on an in-house cardiac MRI dataset including both pre- and post-contrast $T_1$ sequences. All methods were evaluated under three distinct training-and-evaluation strategies, namely, standard, one-shot, and test-time-adaptation. The proposed PCA-Relax showed further improved performance of registration and mapping over well-established baselines. The proposed groupwise framework is generic and can be adapted to applications involving multiple images.

Introduction

Quantitative $T_1$ mapping is a vital MRI technique for diagnosing heart disease, as it allows clinicians to measure tissue properties directly. To create these maps, clinicians acquire a series of "baseline" images over time. Because the scanner settings change for each image to highlight different tissue properties, the contrast and brightness vary significantly across the series.

A major challenge in this process is patient motion—both respiratory and cardiac—during the scan. For an accurate $T_1$ map, every pixel across the entire series must represent the exact same anatomical location. Traditional "pairwise registration" methods rely on selecting a single image as a template to which all others are aligned. However, due to the high contrast variability, selecting an optimal template is difficult, and poor choices often lead to registration failure and inaccurate quantitative maps.

Methodology

To address these limitations, this research introduces a groupwise registration framework. Unlike pairwise methods, this approach treats all images equally, aligning them into a common "mean shape" space simultaneously without requiring a fixed template.

The "PCA-Relax" method utilizes two novel loss functions to drive the registration process:

  1. PCA Loss: Based on the intuition that perfectly aligned images should exhibit a low-rank intensity pattern at any given pixel, this loss minimizes the spread of eigenvalues in a PCA decomposition of the image series. This encourages the model to maximize shared structural information across the images.
  2. Auxiliary Relaxometry Loss: This physics-informed loss utilizes a pre-trained module to estimate $T_1$ parameters. It ensures that the registered pixels adhere to the known physical signal model for $T_1$ relaxation, acting as a constraint to maintain physical consistency.
Overview of the PCA-Relax framework, illustrating the registration module, PCA loss calculation, and the auxiliary relaxometry mapping for quantitative maps.

The framework is built upon a modified VoxelMorph (U-Net) backbone and incorporates a cyclic consistency loss to prevent the transformation from collapsing into degenerate solutions.

Experimental Setup

The framework was evaluated using an in-house cardiac MRI dataset of 50 subjects scanned on a 3.0T Philips system. The dataset included both pre-contrast and post-contrast MOLLI (Modified Look-Locker Inversion recovery) sequences. The model was primarily trained on pre-contrast data (36 subjects) to evaluate its generalization capabilities to the distinct contrast profiles of post-contrast images.

Three deployment strategies were tested: Standard (pre-trained), One-shot (trained per patient), and Test-Time Adaptation (TTA) (pre-trained then fine-tuned). TTA was identified as a highly effective strategy, balancing accuracy with a practical computation time of approximately 20 seconds per sequence.

Results and Discussion

Performance was quantified using the Fitting Standard Deviation (SD) in the myocardium, where lower values indicate better alignment with the physical model. PCA-Relax consistently outperformed baseline methods, including standard pairwise VoxelMorph (VM-P) and groupwise VoxelMorph (VM-G).

Quantitative comparison of SD values across different registration methods and evaluation strategies for (a) Pre-Gd and (b) Post-Gd sequences.

In the pre-contrast TTA scenario, PCA-Relax achieved an SD of 32.88 ms, a significant improvement over the raw data's 45.89 ms and the pairwise baseline's 44.84 ms. Notably, the PCA-based approach demonstrated strong robustness, generalizing well to post-contrast images despite being trained on pre-contrast data. While the relaxometry module showed some sensitivity to domain shifts in the standard post-contrast setting, the overall framework provided a substantial performance boost.

Conclusion

The proposed PCA-Relax framework offers a robust, template-free solution for motion correction in cardiac $T_1$ mapping. By integrating global structural alignment via PCA with physics-informed relaxometry constraints, the method achieves superior registration accuracy and more reliable quantitative maps. The framework's efficiency and generic design suggest potential for broader application in other multi-image MRI tasks, such as $T_2$ mapping or diffusion imaging.

Summary updated

BibTeX
@inproceedings{zhang2024deep,
  author = {Yi Zhang and Yidong Zhao and Lü Huang and Liming Xia and Qian Tao},
  title = {{Deep-Learning-Based Groupwise Registration for Motion Correction of Cardiac $$T\_1$$ Mapping}},
  booktitle = {Lecture Notes in Computer Science},
  year = {2024},
  doi = {10.1007/978-3-031-72069-7\_55},
}

Loading related content...

Figure