Bridging the Gap: A Slice-to-Volume Registration Technique for Converting 2D Single-Shot LGE images into 3D Volumetric Data (RF_FR_348)

Late gadolinium enhancement (LGE) is a well-established cardiac MRI technique for assessing myocardial fibrosis and diagnosing cardiomyopathies. While 3D free-breathing LGE provides high-resolution isotropic imaging, it suffers from long scan times, respiratory gating inefficiencies, and motion corruption [1, 2]. In contrast, 2D single-shot LGE is less affected by respiratory motion but has lower signal-to-noise ratio (SNR) and slice misregistration, leaving gaps in coverage. These issues worsen in patients with intracardiac devices, where high bandwidth lowers SNR, leading to longer acquisition times and increased risk of misregistration and coverage gaps. We previously proposed using a 2D stack of overlapping slices with a Karhunen-Loeve Transform-wavelet (KW) filter to enhance SNR and provide whole-heart coverage with shorter scan times compared to motion-corrected (MOCO) slice averaging [3]. Building on this, we present a novel slice-to-volume registration method that integrates 2D and 3D LGE strengths.

Methods:

This retrospective study included 74 patients with intracardiac devices scanned at 1.5T (Siemens MAGNETOM Sola), generating 1,258 myocardial segments based on the 17-segment AHA model. Single-shot, free-breathing 2D LGE images were acquired using a wideband inversion pulse. Short-axis images covering the whole heart, with 6 mm slice thickness and 4 mm overlap, resulted in a stack of 57 ± 6.5 slices. A KW filter was applied to enhance SNR.

Filtered 2D stacks were registered into a 3D dataset using a novel slice-to-volume approach with sliding window technique (Fig. 1). This involved registering the center slice to adjacent slices, averaging displacement fields for robust pixel transport estimation [4], and aligning the center slice, generating high-resolution 3D volumes without a pre-existing target (Figs. 2 and 3).
A cardiologist, blinded to patient data, evaluated LGE presence in standard 2D short and long-axis LGE views; interactive multi-planar reformatting (MPR) was used to review the 3D LGE data. LGE presence was confirmed if visible in corresponding orthogonal short and long-axis segments, with results categorized as "positive" (LGE), "negative" (no LGE), or "indeterminate" (seen in one view only).

Results: The cardiologist identified 861 negative, 198 positive, and 199 indeterminate segments in the 2D images, while 938 negative, 320 positive, and 0 indeterminate segments were identified in the 3D datasets. Using the 2D positive and negative segments as the gold standard, the 3D protocol achieved 81.3% sensitivity and 92.2% specificity. The slight reduction in sensitivity may be because the 3D data were acquired later after contrast injection.

Conclusion: The 3D volumetric data, generated through a novel registration technique from free-breathing overlapping 2D LGE slices, enabled the interactive generation of long-axis views and eliminated indeterminate segments, reducing the need for additional long-axis acquisitions.