Deep-Learning-Based Multi-Class Classification for Neonatal Respiratory Diseases on Chest Radiographs in Neonatal Intensive Care Units.

Introduction: Accurate and timely interpretation of chest radiographs is essential for assessing respiratory distress and guiding clinical management to improve outcomes of critically ill newborns. This study aimed to introduce a deep-learning-based automated algorithm designed to classify various neonatal respiratory diseases and healthy lungs using a large dataset of high-quality, multi-class labeled chest X-ray images from neonatal intensive care units.

Methods: Portable supine chest X-ray images for six common conditions (healthy lung, respiratory distress syndrome [RDS], transient tachypnea of the newborn [TTN], air leak syndrome [ALS], atelectasis, and bronchopulmonary dysplasia [BPD]) and demographic variables (gestational age and birth weight) were retrospectively collected from 10 university hospitals in Korea. Ground truth for manual classification of these conditions was generated by 20 neonatologists and validated by others from different hospitals. The dataset, consisting 34,598 for training, 4,370 for validation, and 4,370 for testing, was used to train a modified ResNet50-based deep-learning model for automatic classification.

Results: The automatic classification algorithm showed high concordance with human-annotated classifications, achieving an overall testing accuracy of 83.96% and an F1 score of 83.68%. The F1 score for each condition was 87.38% for "healthy lung" and 92.19% for "BPD," 90.65% for "ALS," 90.30% for "RDS," 86.56% for "atelectasis," and 70.84% for "TTN."

Conclusion: We introduced a deep-learning-based automated algorithm to classify neonatal respiratory diseases using a large dataset of high-quality, multi-class labeled chest X-ray images, incorporating non-imaging data, which could support neonatologists in making timely and accurate decisions for critically ill newborns.