Evaluation of a Deep Learning model for the automated prescription of aortic valve planes in CMR (RF_FR_345)

Cardiac magnetic resonance (CMR) exams require radiographers to carefully set up double oblique cardiac planes while closely monitoring patients during a lengthy process involving many breath-hold manoeuvres. This task needs a high level of training and skill, limiting the availability of CMR. An automated tool for prescribing these planes could make cardiac imaging more accessible, reduce exam time, and improve image quality [1,2].

We have previously presented a prototype enabling the automated prescription of axial, long- and short-axis cardiac views. The goal of this study was to evaluate a Deep Learning model that extends this to the prescription of left ventricle outflow tract and aortic valve views.

Methods:

A prototype convolutional neural network was trained to identify the aortic valve insertion points on 3-chamber (3CH) and left ventricle outflow tract (LVOT) localizer images (figure 1). The training database consisted of 3288 balanced steady-state free precession (bSSFP) images, manually annotated by CMR experts at GE Healthcare and the University of California San Diego. Data augmentation was applied to achieve a 12-fold increase of the training set. The model, a U-net with 4 resolution levels and 492k trainable parameters, was optimized using the Adam algorithm with a Dice metric.

The landmarks identified by the model were used to prescribe an LVOT localizer from 3CH images, and aortic valve (AoV) views from LVOT images, following the recommendations in [1]. The orientation of the resulting views was compared with corresponding views manually defined by an experienced technologist. Evaluation of the algorithm consisted of a separate test set obtained from 21 clinical patients (14M/7F, 69±11 years, 72±16 kg) acquired on a GE HealthCare 1.5T Signa Explorer.

Results:

The inference completed successfully on 41 of 42 series, except for one outlier on a 3CH view in a patient with severe aortic stenosis. Excluding the outlier image, the average dihedral angle between automated prescriptions and the corresponding manually prescribed views was 13.36 ± 9.92 degrees for LVOT and 10.22 ± 8.95 degrees for AoV views. A summary of the comparison can be found in table 1. Preliminary measurements indicate that manual prescriptions can vary up to 15 degrees. Figure 2 shows a typical inference result, in comparison with the AoV and LVOT views prescribed by an experienced operator. In four cases, the differences were found to be due to the manual prescription: In one case due to operator error, and in three cases due to intentionally not prescribing the LVOT view orthogonal to the valve. Removing these cases had a minor impact on the results: 12.44 ± 10.66 degrees for LVOT and 8.90 ± 4.45 degrees for AoV views.

Conclusion:

The evaluation results demonstrate that the proposed model can replicate the performance of experienced CMR operators for the prescription of LVOT and AoV views. The training database should be expanded with additional pathological cases to ensure robust performance in a clinical setting.

These findings indicate that the automated prescription prototype can effectively aid non-expert radiographers in conducting diagnostic examinations, thereby enhancing access to cardiovascular magnetic resonance imaging. The proposed model has been integrated in a prototype that can be deployed on clinical scanners and ongoing work is under way to evaluate its workflow impact.