Abstract 13246: Utility and Validity of Deep Learning Based Fully Automated Echocardiogram Interpretation in Patients With Congenital and Structural Heart Disease: Importance of Congenital Specific Algorithms

Abstract

Background: Automated echocardiographic image interpretation has the potential to transform clinical practice by improving efficiency and supporting serial assessment of structural abnormalities or cardiac function. We hypothesized that neural networks (NN) developed for image segmentation in general/non-congenital cohorts may underperform in the setting of abnormal cardiac anatomy and that deep learning (DL) algorithms trained specifically in this cohort may have superior precision in this setting. Methods: Consecutive patients with various forms of congenital or structural heart disease (CHD/SHD) were used to validate an existing convolutional NN trained on 14,035 echocardiograms for automated viewpoint classification (Zhang et al. Circulation 2018). In addition, a new specific deep convolutional network for viewpoint classification was trained and tested in patients with CHD/SHD. Results: Overall, 9,793 imaging files (8,363 GE Medical, 1,421 Philips Medical) from 262 patients with CHD/SHD (mean age 48 years, 59.5% male) and 62 normal controls (mean age 45 years, 50,0% male) were included. Congenital diagnoses included among others, Tetralogy of Fallot (30), Ebstein anomaly (18), transposition of the great arteries (TGA) (48), congenitally corrected TGA (34). Assessing correct view classification based on 284,250 individual frames revealed that the non-congenital model had an overall accuracy of 54.2% for correct view classification in patients with CHD/SHD (66.7% in normal controls), compared to 84% in the original derivation sample. In contrast our newly trained convolutional networks for echocardiographic view detection based on over 139,910 frames and tested on 35,614 frames from CHD/SHD patients achieved an accuracy comparable to the general cardiology model, exceeding 90% for standard view classification. Conclusions: The current study is the first to validate CHD/SHD specific echocardiographic DL networks and demonstrate its marked superiority compared to generic models which appear to have acceptable accuracy in general cardiology patients, only.