Multitask learning for the prediction of the pulmonary artery development in children patients with congenital heart disease on chest radiographs: an external validation study

Pulmonary artery development is an important indicator in the assessment and management of children with congenital heart disease. This study aimed to develop and validate a deep learning (DL) model that can predict the development of pulmonary artery from chest radiographs. Children patients who had undergone cardiac CT examinations in our hospital and other local medical center from December 2007 to March 2023 were retrospectively included. Our study of chest radiographs involved an internal dataset (1326 cases) and an external dataset (126 cases). A multitask DL model was developed to predict the McGoon ratios and classify disease severity. Model performance was compared with that of radiologists. Gradient-weight class activation mapping (Grad-CAM) was performed to visualize how the model predicts the pulmonary artery development. CT scans and chest radiographs were performed in 1452 patients. The correlation and the Bland-Altman mean error in the internal and external groups of the McGoon ratio measured by CT manually and the McGoon ratio predicted by DL algorithm were 0.797 ( P < 0.001) and 0.468 ( P < 0.001), -0.05 ± 0.51 and − 0.06 ± 0.37 respectively. The AUCs for identifying a McGoon ratio less than 2.0 were 0.926 and 0.719 in the internal and external cohorts, respectively. DL can provide a feasible approach for estimating pulmonary artery development from chest radiographs. The model is intended as a screening and triage aid for preliminary quantitative assessment, not a replacement for established cardiovascular imaging. Further validation is required before clinical application.