Which resolution is needed for a 3D MRA based automated CMR planning

Background: Cardiac MR (CMR) planning is complex and typically performed by experienced technologists. An AI-based method using a 3D MR angiography (3D MRA) localizer has shown comparable performance to expert planning [1], offering a potential solution to standardize and streamline CMR exams. This study aims to evaluate and compare the planning accuracy of three 3D MRA localizer protocols with different resolutions.

Methods: Volunteers were scanned with three isotropic 3D MRA protocols based on T2prep GRE Dixon research sequence [2] on 1.5T MRI systems (MAGNETOM Sola, Siemens Healthineers, Forchheim, Germany) and 18-channel coil. BioMatrix Beat Sensor was used for cardiac synchronization [2]. Variants acquired: single breathhold, resolution 2.0 mm³; and two free breathing with navigator gating, resolutions 1.2 mm³ and 1.8 mm³. AI-based cardiac views and FOV planning were generated and scanned for each variant. Acquisition, image recon, and AI processing times displaying cardiac views in MPR were recorded and compared using repeated-measures ANOVA. AI planning was performed for left ventricular (4-2 and 3 chamber, short-axis, and LVOT), right ventricular (RV 2 and 3 chamber, RVOT), and valve planes (mitral, tricuspid, pulmonary, aortic, and 1 cm above aortic valve). Two blinded readers independently graded planning quality using 3-point Likert scale (0 = incorrect, 1 = acceptable, 2 = excellent). Cases scored 0 were re-evaluated using the original AI MPR views from the 3D MRA (Fig1). Planning quality differences for the most experienced reader were assessed using Friedman’s test. Interobserver agreement evaluated using weighted kappa statistics.

Results: Twelve volunteers (8 female, age 23-64, height 173±10cm weight 82±23kg) were scanned. The 1.2 mm³ 3D MRA produced the highest quality (Figure 2) but longest acquisition times. Planning quality was comparable across all. Mean Likert scores were 1.72 (2.0 mm³), 1.80 (1.8 mm³), and 1.73 (1.2 mm³),(Table 1).

From 468 planned views, 14 were initially rated score = 0, on re-evaluation using MPR AI output and expert evaluation on whether a better view could have been established, rating was changed from incorrect to excellent in 13 cases. Differences between AI generated MPR planning and cine output could be explained by different breathing pattern. Interobserver agreement ranged from slight to substantial, highest for LV 3-chamber, lowest for pulmonary valve.

Conclusion: AI-based planning of CMR views using 3D MRA localizers showed consistent performance across resolutions. While 1.2 mm³ MRA had better image quality, it didn’t improve planning accuracy. The 1.8 mm³ protocol provided the best planning scores with shorter scan time compared to the 1.2 mm³ 3D MRA (Table 1), delivering still diagnostic MRA images, supporting its use as an efficient and clinically viable option. These findings support the integration of AI-assisted planning into routine CMR workflows, particularly in time-constrained or low-expertise settings.

Table 1: Acquisition, reconstruction and AI detection times as well as mean Likert scores (0: Incorrect, 1: Acceptable, 2: Excellent) for each detected cardiac view.

	2.0 mm3	1.8 mm3	1.2 mm3
Acquisition time [min:sec]	0:14 ± 0:03	1:49 ± 0:32	6:18 ± 1:34
Reconstruction time [min:sec]	0:44 ± 0:01	0:42 ± 0:01	1:40 ± 0:03
AI detection time [min:sec]	0:17 ± 0:02	0:19 ± 0:03	0:26 ± 0:04
	Mean Likert scores			p	Weighted kappa	Percent agreement
	2.0 mm3	1.8 mm3	1.2 mm3
LV 4-chamber	1.58	1.92	1.17	0.02	0.48	83
LV 2-chamber	1.67	2.0	2.0	0.02	0.21	92
LV 3-chamber	1.67	1.67	1.67	0.87	0.65	92
LVOT	1.75	2.0	2.0	0.05	0.47	97
Short-axis	2.0	2.0	2.0	-	-	100
RV 2-chamber	1.83	2.0	1.75	0.26	0.34	89
RV 3-chamber	1.83	2.0	1.67	0.37	0.52	94
RVOT	1.92	2.0	2.0	0.37	0.07	94
Tricuspid valve	1.42	1.33	1.50	0.74	0.1	92
Mitral valve	2.0	1.58	1.75	0.09	0.1	100
Aortic valve	1.83	1.58	1.67	0.37	0.24	100
Above aortic valve	1.42	1.92	1.75	0.07	-	97
Pulmonary flow	1.42	1.42	1.50	0.90	0.02	97
Mean of all views	1.72	1.80	1.73

Figure 1: Side-by-side comparison of AI-calculated cardiac views as MPRs of the 3D MRA and a cine or flow scan of the same cardiac view. For scoring, the cine/flow images were rated first. If a view was rated as “Incorrect”, the corresponding MPR was then checked to differentiate breath-hold mismatches from AI misdetection.
Figure 2: Examples of the 3D MRA image quality of two volunteers for all resolution variants. Volunteer 12 did not hold their breath until the end of the acquisition for the 2 mm^3 breath-hold variant, which resulted in lower image quality.