Development and evaluation of a deep learning model for computer-aided diagnosis of neonatal pneumothorax on chest radiographs.

Abstract

Background: Neonatal pneumothorax can progress rapidly, yet detection on supine chest radiographs remains challenging.

Objective: To develop and evaluate an artificial intelligence model for detecting pneumothorax on supine neonatal chest radiographs.

Materials and methods: This retrospective single-center study included neonates admitted to the neonatal intensive care unit between January 2011 and December 2024. The dataset comprised 648 radiographs from 288 neonates with pneumothorax and 5,511 radiographs from 3,377 controls. A ResNet-18-based model was developed using publicly available normal adult chest radiographs for pre-training and knowledge distillation. Data were split at the patient level into training, validation, and test sets. Eight-fold cross-validation determined a fixed probability threshold applied for the test set. Gradient-weighted class activation mapping (Grad-CAM) was used for lung-level localization, and logistic regression identified factors associated with misclassification.

Results: In the test dataset, the model achieved an area under the curve of 0.975 (95% confidence interval (CI), 0.961-0.986; sensitivity, 87.6%; specificity, 95.3%; accuracy, 94.5%; positive predictive value, 68.5%; negative predictive value, 98.5%). Grad-CAM localization corresponded to the clinically determined side in 84 of 92 right-lung instances (91.3%) and 48 of 71 left-lung instances (67.6%) (P<0.001). Transient tachypnea of the newborn and gestational age were independently associated with misclassification (OR, 1.88; 95% CI, 1.14-3.12; P=0.03 and OR, 1.10; 95% CI, 1.01-1.19; P=0.03, respectively).

Conclusion: The model showed promising diagnostic performance for detecting pneumothorax on supine neonatal chest radiographs. Underlying pulmonary abnormalities may influence model performance, and external validation is warranted before clinical implementation.