E2389. A New Deep Learning-Based Model for Detecting Pneumothorax on Infantograms in Patients Admitted to the Neonatal Care Units
Authors
Jae Yeon Hwang;
Pusan National University Yangsan Hospital, College of Medicine
Young Hun Choi;
Seoul National University Hospital, Seoul National University College of Medicine
Objective:
To develop and evaluate the diagnostic performance of a deep learning-based model for the automated detection of pneumothorax on infantograms
Materials and Methods:
In this retrospective and multicenter study, a hybrid neural network consisting of a convolutional neural network (EfficientNet) and a hierarchical vision transformer using shifted windows (Swin Transformer) was fine-tuned to detect pneumothorax. A total of 1233 infantograms of patients admitted in neonatal care units from two institutions were used to construct a dataset that was divided into training (n = 1120) and internal test sets (n = 113). An external test set consisted of 245 infantograms from a third institution. Two board-certified pediatric radiologists with 12 years of experience in interpreting pediatric radiography made the final diagnosis, and five radiologists (two board-certified pediatric radiologists and three radiology residents) reviewed infantograms to label the presence of pneumothorax. Receiver operating characteristics (ROC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were calculated to evaluate the model and compare the diagnostic performances.
Results:
The area under the ROC curve (AUC) of the deep learning-based model and radiologists were 0.995 and 0.997 – 0.979, respectively. The sensitivity, specificity, PPV, NPV, accuracy of the deep learning-based model were 97.6%, 97.3%, 97.1%, 97.8%, and 97.5%, respectively. There was no intergroup difference between the AUC of the deep learning-based model, the board-certified pediatric radiologists, and the third-year radiology residents. The deep learning-based model showed higher AUC than the two second-year radiology residents.
Conclusion:
The proposed deep learning-based model showed an excellent diagnostic performance in detecting pneumothorax on infantograms of patients admitted to neonatal intensive care units, which was comparable to the diagnostic performance of experienced pediatric radiologists
