Abstract 10614: An Automated View Identification Model for Pediatric Echocardiography Using Artificial Intelligence

• Published in: Circulation (New York, N.Y.) N.Y.), 2021-11, Vol.144 (Suppl_1), p.A10614-A10614
• Main Authors: Gearhart, Addison, Goto, Shinichi, Powell, Andrew J, Deo, Rahul C
• Format: Article
• Language: English
• ISSN: 0009-7322; 1524-4539
• DOI: 10.1161/circ.144.suppl_1.10614

IntroductionA fully automated artificial intelligence driven pipeline to perform cardiac structural and functional measurements and disease detection requires accurate identification of individual echocardiographic views. However, there are no view classification algorithms for pediatric echocardiograms where there is a broader range in body size, anatomy, view types, and sweep length than in adult studies. HypothesisTo develop and test a machine learning model to automatically perform view classification on individual pediatric echocardiogram images. MethodsUsing a derivation dataset of 12,067 echocardiogram cine and still images from patients 0 to 19 years of age, a convolutional neural network (CNN) model was trained and evaluated for the automated identification of 28 pre-selected standard pediatric echocardiogram views that included anatomic sweeps, color Doppler, and Doppler tracings. The model was validated using an additional 6,197 images and trained for 150 epochs. Classification accuracy was evaluated using a test dataset composed of 9,684 images obtained from 100 different patients equally distributed across the age range. The model was subsequently deployed on studies of 524 children with leukemia to identify 6 pre-selected views pertinent to left ventricular function. ResultsThe model identified the 28 preselected views with 90% accuracy. Accuracy was comparable across age groups (89% for 0-4 yr, 91% for 4-9 yr, 90% for 9-14 yr, and 91% for 14-19 yr). Accuracy was 91% for the sweeps with color Doppler, 83% for sweeps without color Doppler, and 91% for Doppler tracings. Among the leukemia cohort, the model identified the 6 preselected views on a per study basis with a positive predictive value of 99% and an accuracy of 81-97%. A second pass of the model with a threshold prioritizing sensitivity achieved 100% detection of the missing views. ConclusionsWe constructed a CNN model for view classification of pediatric echocardiograms that was accurate across the age spectrum and for cine images, anatomic sweeps, color Doppler images, and Doppler traces. This model lays the groundwork for automated quantitative analysis and diagnostic support to promote more efficient, accurate, and scalable analysis of pediatric echocardiograms.