Abstract
<title>Abstract</title> <p>Background/Aims Pneumothorax is a potentially life-threatening condition in neonates that requires rapid detection and accurate severity assessment to guide timely intervention. In neonatal intensive care units (NICUs), manual interpretation of supine chest radiographs is challenging, as it is time-consuming, labor-intensive, and reliant on clinician expertise, often resulting in inter-observer variability and diagnostic delays. Although deep learning–based algorithms have shown high performance in detecting pneumothorax, most focus on binary classification rather than clinically relevant quantification. This study aimed to develop and validate a deep learning–based system for the quantitative severity assessment of pneumothorax in neonatal chest radiographs, including cases with extrathoracic extension, using pixel-level semantic segmentation. The system is designed to enhance diagnostic accuracy and facilitate timely clinical decision-making. Methods In this multicenter retrospective study, 3,443 neonatal chest radiographs with pneumothorax collected from 11 university hospitals were manually annotated and verified by three pediatric experts. Two independent UNet + + models with ResNet-34 encoders were trained to segment pneumothorax and thoracic cavity regions. Pneumothorax severity was quantified using the pneumothorax-to-thorax area ratio derived from pixel-level masks. Images were divided into training, validation, and test sets, and segmentation performance was evaluated using the Dice similarity coefficient (DSC). Pneumothorax severity was stratified into four ratio-based categories, and intrathoracic and extrathoracic pneumothoraces were analyzed separately. Results In the test dataset (n = 345), the model achieved a median DSC of 0.9088 for pneumothorax segmentation, with performance improving as pneumothorax size increased. Median DSCs were 0.8366, 0.9191, 0.9452, and 0.9742 for pneumothorax-to-thorax ratios of < 10%, 10–20%, 20–30%, and ≥ 30%, respectively. Extrathoracic pneumothorax was identified in 50 cases, with robust segmentation performance (median DSC 0.9531) maintained across severity groups. The algorithm consistently provided objective, reproducible severity estimates across a wide spectrum of pneumothorax burden. Conclusion This deep learning–based system enables accurate detection, localization, and quantitative severity assessment of pneumothorax on neonatal chest radiographs, including extrathoracic extension. This quantitative approach offers reproducible severity assessment that may facilitate earlier recognition of clinically significant pneumothorax, enhance severity stratification, and support timely, informed management.</p>