3D Real-time Cine in Arrhythmia Patients

Background: Existing binning-based 3D methods often fail in subjects with frequent arrhythmias. Three-dimensional real-time (RT) imaging can accommodate irregular rhythms and enable beat-specific functional assessment. However, due to very high acceleration rates (R > 1,000), 3D RT cine has not been feasible. We recently proposed a scan-specific framework, multi-dynamic deep image prior (ML-DIP), for 3D RT cine [1]. Here, we validate ML-DIP in patients with frequent premature ventricular contractions (PVCs).

Methods: Five patients with a history of frequent PVCs were scanned on a 1.5 T scanner (MAGNETOM Sola, Siemens). A 5-minute 3D acquisition was appended to the end of their clinical scan. Three patients had frequent irregular PVCs, one had bigeminy, and one had no PVCs during the scan. Data were collected using Cartesian sampling with 11 readouts per 3D frame [2]. Key parameters: TE/TR 1.3/3.1 ms, temporal resolution 32–34 ms, acceleration R = 1,047, matrix 190×144×80, voxel size 1.4–1.8×1.5–1.8×2.0 mm³, flip 34–40°, ~9,000 frames, bSSFP readout. The same 3D data were reconstructed with compressed sensing after cardiorespiratory binning (5D-Cine) [3, 4]. For comparison, 2D RT bSSFP (3 seconds/slice) was also acquired.

Image quality was assessed by two Level-3 CMR cardiologists using one short-axis and one long-axis cine from each method. Cines were presented as randomized movies, and each reader assigned a score from 1 (worst) to 5 (best). The interpolated short-axis stacks from 5D-Cine and ML-DIP were analyzed in suiteHEART (NeoSoft, Pewaukee, Wisconsin) along with the 2D RT stack. Left-ventricular (LV) and right-ventricular (RV) end-diastolic volume (EDV), end-systolic volume (ESV), stroke volume (SV), and ejection fraction (EF) were computed from 2D RT cine, 5D-Cine, and ML-DIP using the short-axis stacks. Quantification was performed only for subjects whose average image quality score across both readers and both views was at least 3.

Results: In five PVC patients, ML-DIP achieved an average image quality score of 4.1 versus 2.4 for 5D-Cine and 4.5 for 2D RT cine, the latter benefitting from in-flow enhancement. ML-DIP produced diagnostic-quality images (score >3) in all five patients, whereas 5D-Cine was interpretable in only two (Table 1). LV and RV volumes from ML-DIP closely agreed with 2D cine, with mean absolute differences (MAD) of 8.7 mL (EDV), 6.0 mL (ESV), 7.3 mL (SV), and 2.7% (EF) (Figure 1). Representative images from a patient with irregular PVCs and another with bigeminy are shown in Figure 2, where ML-DIP clearly depicts PVCs corroborated by the ECG signal.

Conclusion: ML-DIP enabled free-breathing 3D RT cine in PVC patients, preserving beat-to-beat variability and reconstructing irregular rhythms that conventional binning-based 3D methods failed to capture.

Table 1: Image quality scoring. For an individual dataset, each number represents an average over two cardiac views and two expert readers.

Subject	5D-Cine	2D RT	ML-DIP
#1	1.00	4.00	4.00
#2	3.75	4.75	4.00
#3	2.25	4.50	3.75
#4	2.00	4.25	4.25
#5	3.00	5.00	4.50
Avg.	2.40	4.50	4.10

Figure 1: Left and right ventricular quantification from 5D-Cine and ML-DIP, with 2D RT used as a reference. In the case of 5D-Cine, only two (out of five) PVC patients are included because the image quality from three other patients was not adequate for analysis.
Figure 2: Representative short-axis images from 5D-cine, 2D RT, and ML-DIP. Space-time (x-t) profiles along the dashed green lines are shown to the right of the images. The ML-DIP x–t profiles span 200 frames (6.6 s), while the x-t profiles span 3 seconds for 2D RT cine and one cardiac cycle for 5D-Cine. The 3D reconstructions from 5D-Cine and ML-DIP were interpolated along the 2D plane defined by the 2D RT acquisition. Significant motion artifact is observed in 5D-Cine. The arrows indicate PVC beats. The synchronous ECG is plotted in yellow.