Radiological assessment of pulmonary vascularity in congenital heart disease: standardized clinician assessment versus deep learning–based prediction

Background Assessment of pulmonary vascularity on chest radiographs (CXRs) in congenital heart disease (CHD) is limited by subjectivity, and existing criteria lack sufficient validation. Artificial intelligence-based deep learning model (DLM) analysis may improve accuracy. We developed and validated structured criteria and a DLM to evaluate pulmonary vascularity and compared it with conventional clinical interpretation. Methods A total of 400 deidentified CXRs from CHD patients with Fick-derived pulmonary-to-systemic flow ratio (Qp:Qs) were evaluated by three experts and three trainees in pediatric cardiology and radiology using nine prespecified criteria. A separate set of 1,088 CXRs was used to train the DLM. The clinician-interpreted CXRs were subsequently assessed by the DLM to predict the Qp:Qs. Results Intraclass correlation between Fick- and DLM-derived Qp:Qs was 0.782 ( P < 0.001), with a root mean squared error of 0.395 ( P < 0.001). The DLM showed a strong agreement with Fick-derived classification of pulmonary vascularity (κ = 0.770, P < 0.001; predicting Qp:Qs > 1.5: 90% specificity, 98% sensitivity, area under the receiver operating characteristic curve [AUROC] 98%; predicting Qp:Qs < 0.9: 92% specificity, 95% sensitivity, AUROC 98%). For interpreting pulmonary vascularity as increased or decreased, clinician concordance was 75% and 66%, respectively, whereas DLM concordance was 88% and 98%, respectively ( P = 0.03). Among the structured criteria, the number of end-on vessels demonstrated strong interobserver agreement, whereas agreement for individual peripheral vessels, the size of end-on vessels, branch pulmonary arteries with their respective bronchi, and opacity of lung fields was modest. Conclusion Deep learning–based analysis of CXRs enables accurate evaluation of pulmonary vascularity and outperforms structured assessment by clinicians.