Deep Learning-Based Classification for Neonatal Respiratory Diseases on Chest Radiographs in Neonatal intensive Care Units

초록

Background and Aims: Chest radiographs serve as the primary imaging modality for assessing respiratory distress and play a pivotal role in determining clinical management decisions for critically ill newborn infants in NICUs. We introduce a fully automated deep-learning system designed to classify respiratory diseases in chest radiographic images to enhance the efficiency and accuracy of computer-aided diagnostic systems. Methods: Portable supine chest X-ray images for 6 classes—healthy normal (no abnormal finding), RDS, TTN, air leak syndrome (ALS), atelectasis, and BPD—were prospectively collected from the NICUs of 10 university hospitals in Korea. Ground truth of manual classification for the diagnosis of respiratory diseases was generated by 20 neonatologists and subsequently validated by neonatologists from different hospitals. The dataset was divided into Train, Validation, and Test sets, with 34,598, 4,370, and 4,370 images, respectively. A ResNet50-based deep-learning model was used for automated classification, with gestational age and birth weight serving as additional input channels alongside image data. Results: The automatic classification algorithm showed high concordance with human-annotated classification for normal and 5 respiratory diseases (overall testing accuracy = 86.89% and F1-score = 86.89%). The F1-scores, which balance precision and recall for the positive class, were 87.38% for “normal,” and 92.19%, 90.65%, 90.30%, 86.56%, and 70.84% for “BPD,” “ALS,” “RDS,” “atelectasis,” and “TTN,” respectively. Conclusions: Our deep-learning–based automated algorithm accurately identifies normal conditions and various respiratory diseases on chest radiographs. Artificial intelligence analysis for chest radiographic images can assist neonatologists in making more precise diagnoses and informing clinical decisions in the NICU.