AbstractCongenital heart disease (CHD) is the most common birth defect. Fetal survey ultrasound is recommended worldwide, including five views of the heart that together could detect 90% of complex CHD. In practice, however, sensitivity is as low as 30%. We hypothesized poor detection results from challenges in acquiring and interpreting diagnostic-quality cardiac views, and that deep learning could improve complex CHD detection. Using 107,823 images from 1,326 retrospective echocardiograms and surveys from 18-24 week fetuses, we trained an ensemble of neural networks to (i) identify recommended cardiac views and (ii) distinguish between normal hearts and complex CHD. Finally, (iii) we used segmentation models to calculate standard fetal cardiothoracic measurements. In a test set of 4,108 fetal surveys (0.9% CHD, >4.4 million images, about 400 times the size of the training dataset) the model achieved an AUC of 0.99, 95% sensitivity (95%CI, 84-99), 96% specificity (95%CI, 95-97), and 100% NPV in distinguishing normal from abnormal hearts. Sensitivity was comparable to clinicians’ task-for-task and remained robust on external and lower-quality images. The model’s decisions were based on clinically relevant features. Cardiac measurements correlated with reported measures for normal and abnormal hearts. Applied to guidelines-recommended imaging, ensemble learning models could significantly improve detection of fetal CHD and expand telehealth options for prenatal care at a time when the COVID-19 pandemic has further limited patient access to trained providers. This is the first use of deep learning to ∼double standard clinical performance on a critical and global diagnostic challenge.
Prediction of Pulmonary to Systemic Flow Ratio in Patients With Congenital Heart Disease Using Deep Learning–Based Analysis of Chest Radiographs
