Abstract 17157: Improving Cardiac MRI Ventricular Contouring In Tetralogy Of Fallot Using Machine Learning

IntroductionCardiac magnetic resonance imaging (CMR) evaluates ventricular volumes in tetralogy of Fallot (TOF) and is key in deciding timing of pulmonary valve replacement. However, ventricular contouring is time-consuming and has inherent intraobserver and interobserver variability. Recently, machine learning neural networks (MLNN) have been developed to automate contouring, but are mostly trained on hearts with normal morphology. These networks work best when trained on data similar to that being contoured. We evaluated whether adding TOF data to the training set would improve ventricular contouring for TOF.HypothesisWe hypothesized that adding TOF datasets to the training set would improve TOF contouring compared to a MLNN trained only on structurally normal hearts.MethodsThe initial MLNN was trained on short axis cine CMR scans of morphologically normal hearts from the UK Biobank, and patients with dilated and hypertrophic cardiomyopathies and myocardial infarction. Left ventricular (LV) epicardial and endocardial borders, and right ventricular (RV) endocardial borders, at end systole (ES) and diastole (ED), were manually contoured by experts for both training and TOF CMRs. Agreement between gold standard manual and MLNN contours was evaluated using Dice similarity coefficient (DSC); DSC=1 is perfect overlap and DSC=0 is no overlap. Wilcoxon rank sum tests were used with p