Abstract 16599: Deep Learning Prediction of Cardiac Chamber Enlargement on Chest Radiographs

Background: Chest radiographs (CXRs) are the most commonly performed radiographic exam. The heart is routinely assessed on frontal CXRs using the cardiothoracic ratio (CTR) to identify cardiomegaly, but the predictive performance of CTR is limited by variability in patient position, lung volumes, and radiographic technique. Thus, CXR may miss a substantial proportion of patients with cardiac disease.

Hypothesis: A deep learning (DL) algorithm can accurately identify cardiac chamber enlargement on CXRs and exceed the predictive performance of cardiothoracic ratio.

Methods: Frontal CXRs (n=6,228 cases from unique patients, single-site, dated 2014-2022), obtained on the same day as transthoracic echocardiography (reference standard for cardiac chamber enlargement), were randomly grouped for DL model training (n=4,969), validation (n=627), and testing (n=634). A pre-trained EfficientNet-B6 convolutional neural network was modified to predict the presence of one or more enlarged cardiac chambers. The DL model's performance was benchmarked against manual CTR measurements by two radiologists on a random subset of 150 anteroposterior CXRs from the test set.

Results: The DL model demonstrated an AUROC (Area under the Receiver Operating Characteristic curve) of 0.78 in the validation set and an AUROC of 0.79 in the unseen test set. The model achieved 73% accuracy, 77% sensitivity, and 67% specificity on the validation set and showed similar results on the test set with 72% accuracy, 77% sensitivity, and 64% specificity. The model outperformed manual CTR measurements for identification of chamber enlargement on the random subset of test set CXRs (AUROC of 0.85 vs. 0.75; p < 0.006, Figure 1).

Conclusion: The DL model significantly outperformed traditional CTR assessment for cardiac chamber enlargement on CXR. DL models may improve radiologist identification of cardiac disease on CXR and prompt formal cardiac evaluation.