Deep learning for evaluating left atrium stress echocardiography: a proof of principle study

Abstract Introduction Left atrial (LA) size is a dynamic variable that changes during stress echocardiography (SE) and provides valuable information within beyond coronary artery disease. However, its measurement remains subjective, time-consuming based on manual tracings. Recent advances in deep learning might save analysis time deflate variability by removing the subjectivity of LA assessment. Purpose In this proof principle study we aim to validate potential an automated machine system for evaluation SE. Methods From image data bank Stress Echo 2030 study, selected 20 consecutive patients who underwent SE (using variety methods) 2 recruiting centers. Imaging was acquired DICOM format anonymized. The studies were reviewed expert cardiologist trained apical four chamber (A4Ch) view images rest phases marked end-systolic (ES) end-diastolic frames be used further evaluation. Endocardial borders traced ES. tracings repeated three different evaluators (expert [C1], from external center [C2] [ML] model – convolutional neural network unrelated set images). area (LAA), volume (LAV) using area-length formula proportional LAV between (ΔLAV = [rest − stress] / rest) calculated. Each evaluator blinded each other's measurements. Results total, 40 A4Ch (20 at stress) which all sufficient quality performing measurements system. Pearson correlation coefficients (R) LAA 0.95 (C1-ML), 0.96 (C2-ML) 0.88 0.79 stress. C1-C2 pair had R 0.98 0.86 also showed good raters with values 0.90 0.94 (C2-ML), (C1-C2) 0.87 0.84 Root mean squared errors (RMSEs) 13.48 ml 7.44 10.26 14.03 8.53 13.69 There high level correlations ΔLAV 0.93, C1-ML, C2-ML pairs respectively. Mean (95% CI) −20.1 (−51.59 +11.39), −14.95 (−43.77 +13.87) −14.37 (−34.46 +5.72) ML, C1 C2. Comparison one-way ANOVA did not show significant differences operators (p=0.94). Conclusions Automated ML produces are comparable human can reduce need tracing. tendency have lower levels agreement compared rest, showing additional standardization operators. Funding Acknowledgement Type funding sources: None.