Unsupervised Motion-Compensated Decomposition for Cardiac MRI Reconstruction via Neural Representation

TL;DR

MoCo-INR tackles unsupervised reconstruction of highly undersampled dynamic CMR by integrating implicit neural representations with a motion-compensated framework. It models cardiac motion as continuous functions for a DVF $u_t(p)$ and a canonical image $x_{cano}(\tilde{p})$, implemented by a tailored INR architecture with coarse-to-fine hash encoding and a CNN-based decoder to stabilize optimization. A differentiable forward model and a combined loss $\mathcal{L} = \mathcal{L}_{DC} + \mathcal{L}_{DVF}$ promote data consistency and plausible motion fields, enabling fast convergence and high-fidelity reconstructions, even at ultra-high accelerations (e.g., $20\times$ Cartesian, $69\times$ non-Cartesian). Experiments on retrospective and prospective CMR data show superior performance and robust DVF/canonical-image estimates, with ablations confirming the necessity of the encoder/decoder design and regularization. The work advances real-time, unsupervised CMR reconstruction with potential clinical impact for free-breathing imaging and rapid motion-resolved analysis.

Abstract

Cardiac magnetic resonance (CMR) imaging is widely used to characterize cardiac morphology and function. To accelerate CMR imaging, various methods have been proposed to recover high-quality spatiotemporal CMR images from highly undersampled k-t space data. However, current CMR reconstruction techniques either fail to achieve satisfactory image quality or are restricted by the scarcity of ground truth data, leading to limited applicability in clinical scenarios. In this work, we proposed MoCo-INR, a new unsupervised method that integrates implicit neural representations (INR) with the conventional motion-compensated (MoCo) framework. Using explicit motion modeling and the continuous prior of INRs, MoCo-INR can produce accurate cardiac motion decomposition and high-quality CMR reconstruction. Furthermore, we introduce a new INR network architecture tailored to the CMR problem, which significantly stabilizes model optimization. Experiments on retrospective (simulated) datasets demonstrate the superiority of MoCo-INR over state-of-the-art methods, achieving fast convergence and fine-detailed reconstructions at ultra-high acceleration factors (e.g., 20x in VISTA sampling). Additionally, evaluations on prospective (real-acquired) free-breathing CMR scans highlight the clinical practicality of MoCo-INR for real-time imaging. Several ablation studies further confirm the effectiveness of the critical components of MoCo-INR.

Figures (10)

• Figure 1: Overview of the proposed MoCo-INR framework. Given any spatial coordinate $\boldsymbol{p}=(x,y)$ in the physical space and temporal coordinate $t$, our deformation network predicts the corresponding time-varying displacement vector field (DVF). Adding this displacement to the spatial coordinate in physical space yields the associated location in the canonical space. Then, the canonical network maps these warped coordinates to the dynamic image $\boldsymbol{x}_t$. Finally, the two networks are jointly optimized by minimizing the data-consistency loss (Eq. \ref{['eq:Loss']}) and DVF regularization loss (Eq. \ref{['eq:Loss']}).
• Figure 2: Illustration of the proposed coarse-to-fine hash encoding strategy. Given any input coordinate $\boldsymbol{p}$, the low-frequency feature (i.e., $\boldsymbol{\gamma}_1$) is learned first and then frozen. As the optimization proceeds, higher-frequency features (i.e., $\boldsymbol{\gamma}_2$ and $\boldsymbol{\gamma}_3$) are progressively optimized.
• Figure 3: Qualitative results of retrospective reconstructions obtained with the compared methods. The figure displays the reconstructed image, its profile line over time (the $y\text{-}t$ plane), and the corresponding error map. The selected $y$-axis is indicated by a white dashed line, and zoom-in boxes highlight regions of interest at the end-diastole (ED) and end-systole (ES) phases. The upper part shows results for SAX slice acquired using a VISTA sampling pattern with an acceleration factor of AF=20. The bottom part shows results for LAX slice acquired using a golden-angle radial sampling pattern with 3 spokes.
• Figure 4: Qualitative results of prospective reconstruction under free-breathing scans.
• Figure 5: Visualization of the estimated DVFs and canonical image of MoCo-INR at the diastolic and systolic phases.