Detection of prognostically significant coronary artery disease in stress echocardiography using artificial intelligence

Thesis

Coronary artery disease (CAD) is a major cause of morbidity and mortality worldwide, affecting 2.3 million people in the UK. Whilst CAD can be diagnosed through a variety of imaging modalities, stress echocardiography is the most commonly used technique, used in over 50% of hospitals in the UK. However, the current clinical standard for stress echo is very subjective and is only 81 % sensitive and 82% specific. Therefore, there is interest in the development of robust imaging biomarkers which can reliably and reproducibly provide a predictive index for prognostically significant CAD.

The aim of this thesis is to develop and test methods that improve clinical diagnosis of coronary artery disease in stress echocardiograms using machine learning for 1) image contouring 2) producing features from the contours and 3) predicting coronary artery disease.

A framework for auto contouring of stress echocardiograms was developed, initially using two and four chamber cardiac views for both contrast enhanced and noncontrast echocardiograms, which was later expanded to short-axis (SAX) views. Images were obtained from the prospective multisite EVAREST trial imaging database. The contouring technology was based on a U-net convolutional neural network (CNN) and was trained using contours of the endocardial border created manually, comprised of 5692 and 2182 frames for the contrast and non-contrast application, respectively. The autocontouring achieved a dice coefficient score of 93% for contrast images and 92% for non-contrast images.

Following autocontouring, features were generated from the 2-chamber (AP2) and 4- chamber (AP4) contour output. The feature generation methodology was validated by performing a retrospective study in 146 patients to compare the automated outputs of ejection fraction (EF) and global longitudinal strain (GLS) against equivalent measures generated using standard TomTec analysis software. EF demonstrated a root mean square error (RMSE) of 7.36. GLS demonstrated a RMSE of 2.76.

Following validations of contouring and feature calculation these features and additional novel features were inputted into a machine learning classifier to produce a binary high/low risk prediction for CAD. The training dataset consisted of 633 patients. The final trained model was an ensemble. The area under a Receiver Operator Characteristic Curve (AUROC) of 0.93 was obtained on the training data using a nested cross validation. The model was then validated in an entirely independent dataset and a similar AUROC of 0.93 was obtained. These results demonstrate the utility of artificial intelligence for automated echocardiography analysis and, for the first time, an ability to predict clinical outcome directly from echo images.

Authors: Upton, R

• DOI: 10.5287/ora-d5eddv5kw
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English