Deep learning for automated left ventricular outflow tract diameter measurements in 2D echocardiography

Background Left ventricular outflow tract diameter (LVOTd) is routinely measured to calculate stroke volume and estimate aortic valve area by the continuity equation. Despite LVOTd being regularly measured clinically, significant inter- and intraobserver variability is evident. This variability is highly impactful on both aortic stenosis evaluation and cardiac output calculation due to the squaring of the LVOT radius. Purpose We aimed to investigate if LVOTd measurements from clinical echocardiographic examinations could be used in a deep learning (DL) model to automatically perform LVOTd measurements with equivalent accuracy and improved consistency compared to current practice. Methods Data was collected from clinical echocardiographic examinations performed on 656 consecutive patients admitted to the cardiac catheterization laboratory at a university hospital in January – December 2018. Parasternal views with cardiologist annotated LVOTd coordinates were assessed for 1314 echocardiographic still images. The quality of the still image and annotated LVOT ground truth were individually graded as high, medium and low by experienced cardiologists to establish a rigorous training basis. Spatial geometry data was preserved for each still image in order to distinguish between different degrees of image zoom. Data was randomly split into training, validation and testing sets (68%, 17%, 15%). A fully convolutional network based on the Resnet50 architecture was used with a custom loss function with heatmap regression. Image augmentations were added to extend the dataset. Results When including echocardiographic images of any quality (n=1314) in the model training and inference, the median absolute difference between cardiologist LVOTd and DL LVOTd was 0.97 mm (95% Confidence interval (CI) 0.79–1.14). Using only high and medium quality still images and ground truth (n=869) in the training and inference, median absolute difference decreased to 0.81 mm (95% CI 0.60–0.96). Adding image augmentations to this dataset further improved the model, resulting in a median LVOTd absolute difference of 0.66 mm (95% CI 0.51–0.78). The LVOTd error in inference increased with decreasing image quality, as shown in Figure 1, with two predictions (0.9%) failing completely. Conclusion Deep learning models are capable of measuring LVOTd with comparable accuracy to cardiologists when trained on clinical data. Data quality affects both training and inference. Even with a slightly lower accuracy when used on lower quality echocardiographic images, DL-assisted LVOT measurement has a clear potential to increase repeatability and consistency of LVOTd measurements. Funding Acknowledgement Type of funding sources: Public Institution(s). Main funding source(s): Research Council of Norway (Norges forskningsråd) Figure 1