Beqiri, Arian and Parker, Andrew and Mumith, Angela and Hassanali, Neelam and Upton, Ross and Woodward, Gary and Dockerill, Cameron and Woodward, William and Alsharqi, Maryam and McCourt, Annabelle and others.
A video presentation of the poster.
Background
Automated echocardiographic image analysis removes inter-operator variability which represents a major paradigm shift in current clinical practice and will facilitate the adoption of echocardiographic quantification and impact serial assessment of cardiac function by nonexperts in primary care and direct consumer settings. By using AI methods we built a fully automated and scalable analysis pipeline, robust enough for clinical application, to quantify echocardiographic measurements including global longitudinal strain (GLS) and ejection fraction (EF).
Methods
Machine learning methods were developed using echocardiographic images to build a fully automated, AI platform for image processing that can segment the left ventricle (LV), select cycles and frames and quantify novel and standard metrics (i.e. GLS and EF) reproducibly. The automation framework was developed for 2 and 4 chamber views and short-axis (SAX) cardiac views using data obtained from a prospective multisite trial. The model to differentiate between contrast and non-contrast data was constructed using 86,941 images. The automated contouring model was developed on 5,692 and 2,182 frames for the contrast and non-contrast images, respectively. The training dataset for the SAX views consisted of 2,197 images. Cross-validation (CV) and hold out test data was used to evaluate the AI platforms. Segmentation reproducibility between 5 operators was assessed and compared to the variability observed using
Results
A nominal AUROC of 0.998 was produced for the contrast/non-contrast detection model and the Dice Coefficient for auto-contouring accuracy was 0.932 and 0.924 for contrast and non-contrast images, respectively. In all cases of processing of the same images between operators, no variation was observed (0 RMSE, Bias and Limits of Agreement for EF, GLS and Volumes). Further, this inter-operator performance was far below that reported in the literature. A greater degree of inter-operator variation was observed for the semi-automated application under the same conditions (same images processed between operators).
Conclusion
Our pipeline demonstrates feasibility of a fully automated AI-based methodology to quantify echocardiography images in a robust platform that is scalable for analysis of echocardiograms within healthcare systems. The application is the first of its kind capable of processing contrast enhanced echocardiograms that completely eliminates inter-operator variability on image analysis.
