Accelerated Dictionary-Based Motion Correction for Free-Breathing Cardiac Multi-Parametric Mapping

Description: Background

Free-breathing cardiac multi-parametric mapping is valuable for noninvasive myocardial tissue characterization yet requires robust motion correction (MoCo). The DM+LR MoCo method [1], integrating dictionary-based image synthesis with signal low-rankness constraints, has shown promising accuracy. However, its clinical adoption faces barriers from computational bottlenecks in Extended Phase Graph (EPG)-based dictionary generation [2].



Methods

We propose modified DM+LR (mDM+LR) that replaces the EPG simulations with a learnable signal simulator for rapid scan-specific dictionary generation (Figure 1). The signal simulator processes tissue parameters and RR history to predict simulated signals through a Multi-Layer Perceptron (MLP) [3]. The MLP contains three 200-width linear layers with leaky ReLU activation, followed by a linear output layer. A simulated signal dataset of 12.5 million samples was generated using EPG, covering diverse cardiac rhythms and tissue parameters. The signal simulator was trained by minimizing the mean squared errors between the predicted and ground-truth signals. mDM+LR employs the trained signal simulator to construct dictionaries, of which each entry contains a unit-length signal with specific tissue parameters and RR intervals. mDM+LR then alternates between dictionary matching and image registration until convergence. By dictionary matching, the method produces motion-free synthetic images with similar contrasts to those acquired. In the registration step, mDM+LR aligns the acquired images to the synthetic images using the sum-of-squares errors, with the signal low-rankness constraints posed.

The institutional review board approved the study. We performed ECG-gated cardiac Multimapping [4] for joint T1 and T2 quantification in 19 subjects on a 3T scanner (uMR 790, United Imaging Healthcare, Shanghai, China) after obtaining written informed consent. For each subject, we scanned 3 short-axis slices under free-breathing and breath-holding conditions. Baselines included non-MoCo and pTVreg [5].



Results

mDM+LR achieved an average runtime of 25 s per sample, with dictionary generation under 1 s (vs. ~300 s using EPG). Although slower than pTVreg (6 s), mDM+LR is reasonably efficient for clinical utilization. Figure 2 shows some representative parametric maps. mDM+LR better suppressed the motion artifacts than non-MoCo and pTVreg, and demonstrated high agreement with the breath-holding measurements. Figure 3 presents correlation evaluation results against the breath-holding reference. mDM+LR obtained stronger T1/T2 correlation and lower T1/T2 biases than non-MoCo and pTVreg.



Conclusions

mDM+LR outperforms the baselines in motion correction accuracy and reaches feasible efficiency for clinical deployment. Its open-source implementation supports community adoption and development.

Access Instructions: The source code and example data are available at: https://github.com/SJTU-CMRLab/Model-Based-MoCo-for-Cardiac-Multi-Parametric-Mapping.

Figure1. Schematic of the proposed mDM+LR. The method consists of two modules: (1) An MLP-based signal simulator that maps tissue parameters (e.g., T1 and T2) and RR history to simulated signals to enable fast dictionary construction. (2) A MoCo module that alternately performs dictionary-based image synthesis and image registration until convergence. This alternating process outputs motion-corrected weighted images and parametric maps.
Figure 2. Comparison of T1 and T2 maps pre- and post-correction. The arrows indicate motion artifacts and blurs. mDM+LR better suppressed the motion-induced errors compared to non-MoCo and pTVreg, agreeing well with the breath-holding reference.
Figure 3. Correlation evaluation of myocardial (A) T1 and (B) T2 values against the breath-holding reference through linear regression and Bland-Altman analyses. The Bland-Altman plots depict the mean differences (black lines) and 95% limits of agreement (red lines). The fitted linear models, Pearson correlation coefficients (r), p-values, and Intraclass Correlation Coefficients (ICC) were labeled. mDM+LR achieved stronger T1/T2 correlation relative to the breath-holding reference than non-MoCo and pTVreg.