Deep Learning Reads Chest X-rays to Gauge Pulmonary Artery Growth in Children with Heart Defects

Pulmonary artery development is a key indicator in assessing and managing children with congenital heart disease, and a new retrospective study reports that a multitask deep learning model can estimate it directly from chest radiographs. The work, published in BMC Medical Imaging as an external validation study, developed the model on a large internal dataset and then tested it on cases from a separate medical center.

The researchers included children who had undergone cardiac CT examinations at their hospital and another local medical center between December 2007 and March 2023. The imaging study involved an internal dataset of 1326 cases and an external dataset of 126 cases, for a total of 1452 patients with both CT scans and chest radiographs available.

The multitask deep learning architecture was designed to do two things at once: predict the McGoon ratio, a quantitative measure of pulmonary artery development normally derived from CT, and classify disease severity. To understand what the model was looking at, the team applied gradient-weighted class activation mapping, or Grad-CAM, to visualize how the network arrived at its predictions of pulmonary artery development.

Performance was benchmarked against radiologists and against manually measured CT values. In the internal cohort, the correlation between the McGoon ratio predicted by the algorithm and the ratio measured manually on CT was 0.797, while in the external group it was 0.468, both statistically significant. The Bland-Altman mean errors were -0.05 plus or minus 0.51 internally and -0.06 plus or minus 0.37 externally, indicating small average bias but appreciable variability, particularly in the external set.

For the clinically relevant task of identifying patients whose McGoon ratio falls below 2.0, the model achieved an area under the curve of 0.926 in the internal cohort.

Importantly, the study’s authors position the model as a screening and triage aid for preliminary quantitative assessment, not a replacement for established cardiovascular imaging such as cardiac CT. They also note that further validation is required before any clinical application, and the external results reported here underscore that need.

The study is retrospective and single-region in design, drawing on cases from one hospital and one other local medical center, so performance in broader and more diverse populations remains to be established. As an externally validated but pre-clinical tool, the model represents a step toward quantitative, AI-assisted reading of pediatric chest radiographs rather than a deployed clinical system.

Subject of Research: A multitask deep learning model that predicts pulmonary artery development in children with congenital heart disease from chest radiographs

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

Article References: Yao, X., He, X., Nie, C., OuYang, R., Guo, C., Hu, L., Liu, J., Wang, Q., Wang, Q., & Zhong, Y. (2026). Multitask learning for the prediction of the pulmonary artery development in children patients with congenital heart disease on chest radiographs: an external validation study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02986-w

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02986-w

Keywords: deep learning, chest radiographs, pulmonary artery, congenital heart disease, McGoon ratio, external validation, pediatric imaging, screening, multitask learning, cardiac CT, Multitask, learning