A Motion-Robust Dual-Echo 3D LGE Reconstruction Framework

Background: CMR-based late gadolinium enhancement (LGE) provides high-resolution myocardial scar visualization and fibrosis quantification, enabling risk stratification and management across cardiomyopathies.^1,2 While conventional 2D LGE is hindered by multiple breath-holds, limited coverage, and poor through-plane resolution, free-breathing 3D LGE delivers isotropic coverage, enhancing scar detection and patient comfort.^3,4 State-of-the-art 3D implementations rely on image navigation (iNAV) to either align images across heartbeats⁵ or perform respiratory bining⁶ to improve data efficiency from a ~5 to 10 minutes scan. Nonetheless, low signal-to-noise ratio and residual motion blurring remain a challenge in 3D LGE. Moreover, these motion-correction or compensation techniques cannot accommodate patients with frequent arrhythmia. By extending our recently proposed highly accelerated real-time cine ML-DIP framework,⁷ we propose a motion-robust reconstruction framework for 3D LGE, called multi-dynamic low-rank multi-encoding deep image prior with sensitivity maps estimation (MLE-DIPS).

Methods: We extended ML-DIP to 3D LGE by jointly estimating sensitivity maps and enabling dual-echo reconstruction from a shared spatial basis to exploit inter-echo redundancy (Fig. 1). We validated the resulting MLE-DIPS in five ventilated Yucatan minipigs and five human subjects scanned on a 3T scanner (MAGNETOM Vida, Siemens Healthineers) using a ~5-minute single-shot free-breathing dual-echo iNAV sequence.⁵ In minipigs, rupture of the mitral valve chordae tendineae induced mitral regurgitation with left atrial enlargement. Human patients were randomly selected from those clinically referred for CMR-LGE. Imaging parameters for both studies included spatial resolution of 1.0×1.0×1.25 to 1.25×1.25×1.30 mm³, temporal footprint of 180–240 ms, and per-frame acceleration rates of 768–1672. The coil data were compressed to six virtual channels using ROVir⁸ to enhance the signal from the cardiac region. Reconstructions were compared against an iNAV-based inline reconstruction⁵ and LR-DIP⁹ extended to 3D (Fig. 2, Table 1). Fat-water separation was performed post-reconstruction using IDEAL-CE.¹⁰ For all subjects, a CMR expert scored each method (3D movies and static slices) on a five-point Likert scale (1: non-diagnostic, 5: excellent).

Results: In both animal and human studies (Fig. 2), iNAV reconstructions appeared noisy and LR-DIP overly smooth, whereas MLE-DIPS effectively suppressed motion artifacts while preserving detail. Quantitative scores (Table 1) showed a consistent advantage for MLE-DIPS over iNAV.

Conclusion: We introduce MLE-DIPS, a self-supervised deep-learning reconstruction framework for motion-robust 3D LGE, demonstrating feasibility in animal and human studies using a ~5-minute free-breathing acquisition. Future work will integrate Dixon modeling in MLE-DIPS for improved image synthesis, robust fat-water separation, and more confident scar-fat differentiation, with validation in a larger cohort.

Table 1: CMR Expert Scoring of Reconstructions in Terms of Overall Image Quality in Animal and Human Studies.

	iNAV	LR-DIP	MLE-DIPS
Pigs (5)	4.2	3.4	4.4
Humans (5)	3.8	2.8	4
Both (10)	4	3.1	4.2

The first row reports mean image-quality scores across five pigs, and the second row across five human subjects, each reviewed in a blinded manner. The third row shows combined averages over all ten subjects. Italic values indicate the best score.Fig. 1: An overview of the proposed MLE-DIPS framework. (A) Motion estimation: A CNN decoder maps a static code vector z1 to a 3D deformation basis consisting of building blocks, which are aggregated using time-varying weights generated by a motion MLP driven by trainable dynamic code vectors (or self-gating) to form three deformations corresponding to three spatial directions, represented by Dx, Dy, and Dz. (B) Dual-echo images and sensitivity maps synthesis: With input z2, a Y-Net generates spatial basis as well as smooth sensitivity maps. In addition, a spatial MLP generated echo- and time-specific weights to synthesize in-phase and out-of-phase images from the shared spatial basis. (C) Forward model and data consistency: The echo images from (B) are spatially warped (“o”) with estimated motion fields from (A), multiplied (“x”) with generated sensitivity maps from (B), applied Fourier and sampling operations, and finally made consistent with acquired data. This joint formulation avoids data binning/rejection and can potentially remain robust in arrhythmia.
Fig. 2: Qualitative comparison in animal and human studies. (A) Two representative slices from a Yucatan minipig. (B) A representative slice from a human subject: the first row shows a selected slice, while the second row includes a slice matching 2D SAX reference. Each two-column subgroup corresponds to a competing 3D reconstruction method (iNAV, LR-DIP, and MLE-DIPS), showing both water and fat images. iNAV images look sharp but suffer from noise amplification, LR-DIP blurs details, and MLE-DIPS effectively suppresses noise while preserving sharp myocardial detail, as highlighted by red arrows.