Advancing Cardiac MRI Accessibility: ML-DIP-based 3D Cine Imaging at 0.55T

Background: The accessibility and utility of cardiac magnetic resonance imaging (CMR) continue to expand with improvements to real time sequences and low field, wide bore systems. Two-dimensional (2D) real time imaging serves as a valuable alternative for patients who struggle with breath-holding or have poor gating. However, these sequences still have limitations, such as through-plane motion and increased slice thickness. The recent development of a multi-dynamic low-rank deep image prior (ML-DIP) framework enables highly undersampled three-dimensional (3D) imaging from a free-breathing, ungated acquisition (1). Phantom and human data have previously been assessed at 1.5T (2). In this study, we applied the ML-DIP approach to lower field (0.55T) cine imaging.

Methods: Five healthy volunteers underwent imaging on a 0.55T system (MAGNETOM Free.Max, Siemens Healthcare, Forchheim, Germany). The standard Cartesian balanced steady-state free precession (bSSFP) 2D breath-held, ECG-gated cine sequence was acquired in 2, 3, and 4 chamber views as well as a short axis stack covering the left ventricle. A 3D ML-DIP bSSFP acquisition, as described by Chen et al. (2), was acquired covering the entire heart in approximately five minutes. To optimize the acquisition for a low magnetic field setting, the ML-DIP loss function was changed from mean-squared loss to Huber loss to suppress artifacts. Table 1 summarizes the scan parameters. Matched long- and short-axis slices were reconstructed from the 3D volume and functional assessments of both datasets were performed using suiteHeart (NeoSoft, LLC, Pewaukee, WI).

Results: The volunteers had a mean age of 44 years (range: 20-67) and a mean body mass index (BMI) of 27.1 kg/m² (range: 21.8-32.5). Both 2D and 3D scans were successfully completed for all subjects, and all images were rated as diagnostic by an experienced reader. Figure 1 presents a comparison of resulting 2D and 3D images acquired for one subject. Table 2 summarizes the functional assessment from both datasets. The 3D ML-DIP reconstructions showed excellent overall agreement with the standard 2D cine data, with the 3D acquired data underestimating stroke volume by only an average of 3%.

Conclusion: As lower field, wide-bore systems become increasingly available for CMR in patients who cannot be scanned using standard bore systems, the development of rapid, reliable acquisition methods is critical. Free-breathing, non-gated 3D ML-DIP cine imaging enables volumetric acquisition covering the entire heart in approximately five minutes. Reformatting this data into 2D slices allows for detailed visual and quantitative evaluation of cardiac function, yielding results comparable to standard 2D breath-held, ECG-gated techniques. Additionally, this technique results in five minutes of continuous real time imaging, visualizing several hundred sequential cardiac cycles, rather than a single composite cycle as generated by other 3D cine sequences.

Table 1: Protocol parameters

Sequence	2D bSSFP	3D ML-DIP
Flip Angle (˚)	100	75-110
TE (ms)	1.82	1.72
TR (ms)	4.3	3.88
Asymmetric Echo Setting	Strong	Weak
Bandwidth (Hz/pixel)	930	807
Segments	5-7	11
Temporal Resolution (ms)	25.8-45.1	42.7-46
Number of Slices	11-14	80-120
Slice Thickness (mm)	8	2-2.2
Acquisition Matrix	154 x 194	108 x 144 x 80-120
Field of View (mm)	280-336 x 290-420	217-262 x 290-350 x 176-264
Acquired Resolution	2.0-2.2 x 2.0-2.2	2.0-2.4 x 2.0-2.4 x 2.0-2.2
Acceleration Factor	3.7-5.1	758-864
Average Total Scan Time	Long and Short axis Views 3.5-4 minutes	Single 3D Volume  5.7 minutes

Table 2: Functional assessment comparison

		Ejection Fraction	Stroke Volume (mL)	End Diastolic Volume (indexed, mL)	End Systolic Volume (indexed, mL)	Heart Rate	Cardiac Output (L/min)	Left Ventricle Mass (g)
Volunteer 1	2D bSSFP	60	76.8	74.8	30.1	64	4.9	82
	3D ML-DIP	58	73.5	73.4	30.7	63	4.6	81
Volunteer 2	2D bSSFP	62	104.0	75.8	29.1	63	6.1	97
	3D ML-DIP	59	101.0	76.5	31.1	63	6.4	96
Volunteer 3	2D bSSFP	63	66.8	61.4	22.6	91	6.1	61
	3D ML-DIP	62	66.4	62.2	23.6	90	6.0	63
Volunteer 4	2D bSSFP	57	80.8	75.6	32.7	63	5.1	91
	3D ML-DIP	55	77.3	74.6	33.5	63	4.9	90
Volunteer 5	2D bSSFP	60	80.4	74.0	29.4	88	7.1	76
	3D ML-DIP	60	76.5	71.2	28.7	88	6.7	80

Figure 1