Clinical Feasibility of two Cardiac Deep Learning Cine MRI Sequences: Single-Breath-Hold and Free-Breathing Motion-Corrected Approaches

Background: Cardiac cine MRI faces the challenges of prolonged examination times and repeated breathhold (BH). This study evaluated the clinical feasibility of the deep learning (DL)-accelerated cine sequences, which shorten the scanning time and achieve comparable image quality (IQ) and function.

Methods: This prospective study included patients who underwent 3T cardiac MRI from August 2024 to March 2025. The examination included three cine sequences [2D segmented cine, 2D single-BH DL cine, and 2D free-breathing motion-corrected DL cine (FB-MOCO DL cine)]. The actual total acquisition time were recorded. The overall IQ, blood pool myocardial signal ratio (BMC), edge sharpness, three-dimensional volumetric mesh contour quality, biventricular cardiac function parameters, and left ventricular (LV) strain parameters were evaluated. The Friedman test was used to compare the above parameters among the three cine sequences. Correlation analysis and Bland-Altman analysis were used to evaluate the correlation and consistency between the two cine sequences.

Results: Eighty-six patients were evaluated (52.98±14.34 years, 79% male). Compared with segmented cine [239.70 (224.55, 260.15)]s, the total acquisition time of single-BH DL cine [63.55 (60.98, 66.00)]s and FB-MOCO DL cine [90.65 (79.43, 103.80)]s decreased by 73% and 62% respectively. There were no statistically significant differences in overall IQ, biventricular functional parameters among the three cine sequences. The three-dimensional volumetric mesh contour scores of the single-BH DL cine and FB-MOCO DL cine were higher than those of the segmented cine (p < 0.001 and 0.04), but the edge sharpness and BMC were lower than those of the segmented cine. The LV strain of the two fast cine sequences was lower than that of the segmented cine.

Conclusion: Compared with traditional segmented cine, DL-accelerated cine enables ventricular imaging in a shorter acquisition time, with preserved IQ and quantitative cardiac function results. Scan times are just 63 seconds (single BH) or 91 seconds (free-breathing).

Table 1. Imaging parameters for the three cine sequences

Parameters	segmented cine	single-BH DL cine	FB-MOCO DL cine
ECG mode	Retrospective	Retrospective	Retrospective
TR/TE (msec)	3.32/1.57	2.93/1.36	2.93/1.36
Flip angle (º)	55	55	55
FOV (mm2)	320×360	360×360	360×360
Spatial resolution (mm3)	1.61×1.61×8	1.88×1.88×8	1.88×1.88×8
Image matrix	199×224	192×192	192×192
Temporal resolution (seconds)	53	44	44
Number of slices	9	9	9
Slice thickness (mm)	8	8	8
Bandwidth (Hz/pixel)	1000	1000	1000
Reconstructed cardiac phases	25	25	25
Number of BHs	9	1	0
Data acquisition	9 heartbeats per slice	1 heartbeat per slice	1 heartbeat per slice
VPS	16	15	15
Acceleration factor	2	12.8	14.7

Table 2. Comparison of LV and RV function parameters between the three cine sequences

Parameters	Segmented cine	Single-BH DL cine	FB-MOCO DL cine	χ² (degrees of freedom)	P
LV EDV	143.65 (123.94, 170.99)	141.68 (124.06, 163.50)	144.90 (124.07, 172.34)	5.46 (2)	0.07
LV ESV	64.88 (55.92, 86.03)	65.75 (57.95, 85.48)	67.28 (57.60, 85.69)	0.13 (2)	0.94
LV SV	73.81 ± 17.30	73.47 ± 15.57	73.97 ± 16.68	0.22 (2)	0.81
LV EF	53.39 (43.50, 57.94)	52.40 (44.39, 57.36)	52.93 (44.05, 57.83)	2.26 (2)	0.32
LV CO	4.83 (4.29, 5.48)	4.75(3.99, 5.49)	4.80 (4.10, 5.55)	5.93 (2)	0.05
LV MASS	104.42 (88.84, 126.31)	100.11 (84.95, 121.56)	99.91 (86.84, 118.21)	5.28 (2)	0.07
RV EDV	121.63 ± 28.60	120.03 ± 29.02	118.56 ± 30.90	0.40 (2)	0.67
RV ESV	61.43 ± 18.81	59.65 ± 18.72	59.13 ±19.52	0.35 (2)	0.71
RV SV	60.20 ± 18.55	59.02 ± 17.68	58.19 ± 19.71	0.15 (2)	0.86
RV EF	49.89 (44.27, 56.42)	49.96 (45.09, 56.95)	50.24 (44.84, 56.70)	5.40 (2)	0.07
LV GCS	-18.49 (-20.77, -15.05)	-15.78±4.57	-15.95 (-18.30, -12.93)	49.07 (2)	<0.001
LV GRS	53.61±19.64	42.37 (32.52, 60.12)	44.16 (34.19, 59.78)	9.85 (2)	0.007
LV GLS	-20.08 (-22.15, -16.21)	-17.00 ± 4.54	-16.84 ± 4.67	43.16 (2)	< 0.001

Figure 1. Overall IQ and three-dimensional volumetric mesh contour scores. Median (interquartile range) are indicated in the white circles, while percentages of patients scored from 1-5 are shown in the pie charts. DL, deep learning; BH, breath-hold; FB, free breathing; MOCO, motion correction; IQ, image quality.
Figure 2. Sampling patterns and deep learning reconstruction of cine imaging. GRU, Gated Recurrent Unit; CNN, Convolutional Neural Network; Res-CRNN, Residual Convolutional Recurrent Neural Network.
Figure 3. Bland-Altman plots of LVEDV, LVESV, LVSV and LVEF showing agreement between segmented-, single-BH DL cine and FB-MOCO DL cine. Middle black solid line represent the average difference, and black solid lines on both sides indicate 95% confidence intervals.