Deep Learning-Based Classification for Grading of Respiratory Distress Syndrome on Neonatal Chest Radiographs

초록

Introduction: The aim of this study was to develop and validate a deep learning-based multi-class classification model for automated grading of respiratory distress syndrome (RDS) severity on neonatal chest radiographs. Methods: A total of 23,210 radiographs, including normal and RDS cases, manually annotated by trained neonatologists, were divided into training, validation, and external test sets using patient-level splitting. Lung regions were segmented using UNet++, and RDS severity was classified into five ordered grades using a ResNet-50-based model. Results: The model achieved a quadratic weighted kappa of 0.696, with 85.7% of predictions within one grade and five-class accuracy of 0.575. AUROCs were 0.966 for detecting RDS and 0.866 for clinically significant RDS (grade >= 3). Gradient-weighted Class Activation Mapping demonstrated attention to lung regions with reduced aeration and granular opacities in severe RDS. Conclusions: This model may provide objective and interpretable radiographic assessment of RDS severity and support more consistent radiographic evaluation in neonatal intensive care units.

키워드