Deep learning model accurately detects neonatal pneumothorax on supine chest X-rays

AI model detects neonatal pneumothorax on supine chest radiographs with AUC 0.975, sensitivity 87.6%, specificity 95.3%. PPV 68.5%, NPV 98.5%. Grad-CAM matched clinical side in 91.3% of right-lung vs 67.6% left-lung cases. External validation needed.

• Retrospective single-center study using 648 pneumothorax and 5,511 control neonatal chest radiographs. ResNet-18 model achieved AUC 0.975 (95% CI 0.961-0.986), sensitivity 87.6%, specificity 95.3%.
• Grad-CAM localization matched clinical side in 91.3% right vs 67.6% left (P<0.001).
• Limitations: no external validation; transient tachypnea and gestational age independently associated with misclassification (OR 1.88, 1.10).

Automated summary

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