Automated Prescription of Aortic Flow Views in CMR Using a Deep Learning Model

Background: Cardiac magnetic resonance (CMR) imaging demands that radiographers accurately prescribe double-oblique imaging views while supervising patients throughout a prolonged procedure that includes multiple breath-holds. This process requires extensive expertise and training, which restricts broader access to CMR. Introducing an automated system for plane prescription could streamline the workflow, shorten scan times, and enhance image consistency [1,2].

In earlier work, we presented a prototype capable of automatically prescribing axial, long-axis, short-axis and aortic valve cardiac views. This study aims to assess a Deep Learning model that builds on that foundation by enabling automated prescription of aortic flow views.

Methods: A prototype convolutional neural network was developed to identify the centerline of the aortic outflow in left ventricular outflow tract (LVOT) images. The model was trained on 1,000 bSSFP Cine LVOT images, each annotated by CMR experts. Geometric and intensity augmentation was applied, resulting in a 12x expansion of the training set. The network architecture was based on a U-Net with five resolution levels and ~493k trainable parameters. Model optimization was performed using the Adam optimizer with Dice loss.

A cubic spline was fitted to the model output to approximate the centerline of the ascending aorta and used to prescribe orthogonal flow views at a user-defined distance from the valve. The accuracy of the automated centerline positions and view angles was evaluated against ground truth landmarks annotated by an experienced cardiologist. Validation was performed on datasets from 21 healthy volunteers and 11 clinical cases (71±10 years; 74±14 kg), acquired using 1.5T and 3.0T GE HealthCare MRI systems.

Results: The inference completed successfully on all series, without significant differences between healthy volunteer and clinical data. Figure 1 shows typical inference results, in comparison with the corresponding ground truth. Table 1 shows the average distance and view angle at the four reference points defined by the cardiologist (aortic valve, sinotubular junction, and 2x ascending aorta). On average, the centerline remained within 1mm of the ground truth, and the view angle below 10 degrees. The performance slowly degraded beyond 40mm of the valve, as the training set did not include landmarks beyond that region. Preliminary measurements indicate that manual prescriptions by an experienced CMR technologist can vary by 8±5 degrees for a given patient.

Conclusion: The evaluation results indicate that the proposed model effectively replicates human performance in prescribing flow views for the sinus and ascending aorta. The model has been integrated into a prototype for deployment on clinical scanners. Ongoing studies are assessing its impact on workflow efficiency and measurement repeatability. Future work will focus on enriching the training dataset and evaluating the effect of angulation and location differences on quantitative aortic flow.

Table 1: Mean centerline distance and angle difference between manually and automatically prescribed aortic flow views at four locations.

Location	Centerline distance [mm]	View angle [deg]
Aortic valve	0.10 ± 0.02	3.80 ± 3.37
Sinotubular junction	0.84 ± 0.65	4.59 ± 3.38
Ascending aorta 1	0.59 ± 0.58	4.17 ± 2.88
Ascending aorta 2	0.76 ± 0.78	5.34 ± 4.24

Figure 1: Example LVOT views showing the ground truth flow measurement landmarks prescribed by the cardiologist (solid red lines) and corresponding automated prescriptions (dashed blue lines). Notice how the AI performance degrades with the distance to the valve.
Figure 2: Box-whisker plots of the centerline distance and the flow measurement view angle discrepancy with respect to the four ground truth landmarks defined by the cardiologist.