MBSS-T1: Model-Based Subject-Specific Self-Supervised Motion Correction for Robust Cardiac T1 Mapping

TL;DR

MBSS-T1 tackles motion artifacts in cardiac T1 mapping by embedding a physics-informed signal-recovery model within a subject-specific, self-supervised registration framework guided by myocardial segmentation. The architecture jointly estimates deformation fields and sequence-specific T1 parameters, enforcing realistic signal decay and anatomical plausibility via a three-term loss and Dice-based segmentation constraints. Evaluated on public STONE and in-house MOLLI datasets, MBSS-T1 consistently outperforms baseline deep-learning registrations in model fit ($R^2$), segmentation accuracy (Dice), and boundary fidelity (HD), while achieving higher radiologist-rated clinical scores. This approach enables robust free-breathing cardiac T1 mapping without large annotated datasets and shows strong potential for broader clinical deployment across protocols.

Abstract

Cardiac T1 mapping is a valuable quantitative MRI technique for diagnosing diffuse myocardial diseases. Traditional methods, relying on breath-hold sequences and cardiac triggering based on an ECG signal, face challenges with patient compliance, limiting their effectiveness. Image registration can enable motion-robust cardiac T1 mapping, but inherent intensity differences between time points pose a challenge. We present MBSS-T1, a subject-specific self-supervised model for motion correction in cardiac T1 mapping. Physical constraints, implemented through a loss function comparing synthesized and motion-corrected images, enforce signal decay behavior, while anatomical constraints, applied via a Dice loss, ensure realistic deformations. The unique combination of these constraints results in motion-robust cardiac T1 mapping along the longitudinal relaxation axis. In a 5-fold experiment on a public dataset of 210 patients (STONE sequence) and an internal dataset of 19 patients (MOLLI sequence), MBSS-T1 outperformed baseline deep-learning registration methods. It achieved superior model fitting quality ($R^2$: 0.975 vs. 0.941, 0.946 for STONE; 0.987 vs. 0.982, 0.965 for MOLLI free-breathing; 0.994 vs. 0.993, 0.991 for MOLLI breath-hold), anatomical alignment (Dice: 0.89 vs. 0.84, 0.88 for STONE; 0.963 vs. 0.919, 0.851 for MOLLI free-breathing; 0.954 vs. 0.924, 0.871 for MOLLI breath-hold), and visual quality (4.33 vs. 3.38, 3.66 for STONE; 4.1 vs. 3.5, 3.28 for MOLLI free-breathing; 3.79 vs. 3.15, 2.84 for MOLLI breath-hold). MBSS-T1 enables motion-robust T1 mapping for broader patient populations, overcoming challenges such as suboptimal compliance, and facilitates free-breathing cardiac T1 mapping without requiring large annotated datasets. Our code is available at https://github.com/TechnionComputationalMRILab/MBSS-T1.

Figures (9)

• Figure 1: Schematic representation of cardiac T1 mapping for a single voxel. (a) T1-weighted myocardial images acquired at $N-1$ sequential time points. (b) Fitting an inversion recovery curve $f\left(\Theta,t\right)$ of the longitudinal magnetization across time points $t$ to estimate the corresponding parameters $\Theta$. In this STONE sequence with the two-parameter model, $\Theta = \{M_0, T1\}$ and $f\left(\Theta,t\right) = M_0 \cdot \left(1 - 2 \cdot e^{-t/T1} \right)$, where $M_0$ represents the equilibrium magnetization, which is the tissue magnetization before any preparation, and $T1$ reflects the actual T1 value. (c) Computed T1 map visualizing the estimated T1 values across the myocardium.
• Figure 2: Overview of the MBSS-T1 model architecture: (a) The motion component is responsible for deformable image registration, predicting deformation fields that align the acquired images across different time points. (b) The parametric mapping component estimates the signal recovery model parameters and generates synthetic aligned T1-weighted (T1W) images. (c) The segmentation and the confidence component uses a pre-trained segmentation network to extract the myocardium segmentations, calculates the deformed segmentations, and retains the ones with high confidence. The network aims to minimize the distance between the motion-corrected T1W and synthetic images while maximizing the Dice score between the deformed segmentations.
• Figure 3: The distribution of the selected reference time points across different acquisition methods (BH-MOLLI, FB-MOLLI, and FB-STONE) is presented in the figure below. The x-axis represents the time points within the respective sequences, while the y-axis indicates the percentage of cases in which each time point was chosen. The percentage for MOLLI (BH-MOLLI, FB-MOLLI) is calculated based on a sequence length of 8 time points, whereas for STONE (FB-STONE), the sequence length is 11 time points.
• Figure 4: Comparison of deformation grids and registered images using different motion correction methods across multiple frames. The top row illustrates images without motion correction, followed by deformation grids and registered images produced by SynthMorph. The third row presents results from the MBSS-T1$_\textrm{STONE}$ method, including the deformation grids and corresponding registered images.
• Figure 5: Representative T1 maps computed with the different approaches. Our approach (MBSS-T1$_\textrm{STONE}$) demonstrates a clearer delineation between the blood and the muscle with a reduced partial volume effect, resulting in a more homogeneous mapping of the myocardium. The maps are presented in the colormap recommended by fuderer2025color.