Hawkes, William and Beqiri, Arian and Mumith, Angela and Parker, Andrew and Upton, Ross and McCourt, Annabelle and Woodward, William and Dockerill, Cameron and Heitner, Stephen and Yadava, Mrinal and others.
A video presentation of the poster.
Background
Strain imaging during stress echocardiography (SE) has proven to be a sensitive measure of left ventricular contractile function in suspected coronary artery disease (CAD). Using speckle tracking for strain imaging is limited at the higher heart rates and contrast enhanced imaging typically required for SE. Thus, we sought to train an Artificial Intelligence (AI) platform with fully automated contouring and strain analysis capabilities for use during SE.
Methods
Bespoke machine learning methods were developed using SE images acquired at the different stages of a stress examination within a prospective multisite study. A fully automated AI platform for left ventricle (LV) segmentation, and cycle and frame selection was developed to allow quantification of standard metrics (i.e. GLS and ejection fraction). The automation framework was developed for two and four chamber cardiac views, both with and without contrast enhancement. The automated contouring model was developed on 5,692 and 2,182 frames for the contrast and non-contrast images, respectively. The ability of the auto-strain model to evaluate GLS during SE was then studied by comparison of mean GLS in patients with and without CAD, based on clinical and angiographic outcome.
Results
Calculation of GLS in dataset 1 (578 patients, 85 with clinically proven CAD) demonstrated significantly reduced (p<0.01) GLS at peak stress (-14.7±5.4 vs -19.7±5.4) and smaller increases in GLS from rest to peak stress (ΔGLS = GLSpeak – GLSrest, -0.9±5.5 vs -2.8±5.8) in CAD positive patients. In dataset 2 (154 patients, 45 with CAD positive outcomes), containing only contrast enhanced images, CAD patients also exhibited significantly reduced peak GLS (-12.1±3.3 vs -17.8±4.11) and ΔGLS (2.0±5.0 vs -1.7±5.0). ROC curve evaluation for diagnostic performance demonstrated AUROC values of 0.76 (sensitivity=0.78, specificity=0.62) and 0.85 (sensitivity=1.0, specificity=0.49) for peakGLS using -18.3 as the cut-off.
Conclusion
We have demonstrated that AI processing of images acquired during routine SE is possible and allows automatic contouring and quantification of GLS. This ability to conduct a fully automated SE strain analysis presents an exciting opportunity for the reduction variability in clinical interpretation and decision making, in cases suspected CAD.
