Deep learning model for chest radiograph assessment of pulmonary vascularity in congenital heart disease outperforms structured clinician evaluation

A deep learning model analyzing chest radiographs in congenital heart disease predicted pulmonary-to-systemic flow ratio (Qp:Qs) with AUC 98% for Qp:Qs>1.5 (98% sensitivity, 90% specificity) and Qp:Qs<0.9 (95% sensitivity, 92% specificity). Model concordance for increased or dec…

• Intraclass correlation between DLM-predicted and Fick-derived Qp:Qs was 0.782.
• For interpreting increased pulmonary vascularity, DLM concordance was 88% vs 75% for clinicians; for decreased, 98% vs 66% (P=0.03).
• Clinician agreement was highest for the number of end-on vessels; other structured criteria had only modest interobserver agreement.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.