PIN92 PEDIATRIC BACTERIAL PNEUMONIA CLASSIFICATION THROUGH CHEST X-RAYS USING TRANSFER LEARNING

Abstract

Listen

Translate article

Chest X-ray (CXR)-confirmed pneumonia in children is more specific for bacterial pneumonia than symptom-based diagnosis, and it may be more responsive for assessing the impact of vaccines that target respiratory bacterial pathogens. However, the variability of CXR interpretation by radiologists makes it costly and difficult to use CXR-confirmed pneumonia as an endpoint in large-scale epidemiology research. We aim to develop an automated system to interpret pediatric CXR. We applied a deep neural network (DNN) to classify suspected bacterial pneumonia from pediatric CXRs. A DNN model was first trained using the largest publicly available CXR dataset, ChestX-ray14 (N=112,120). A transfer learning technique was applied to train this pre-trained model on two smaller pediatric datasets, Guangzhou Women and Children’s Hospital CXR (GZ-CXR, N=5,856, 47% bacterial pneumonia) and World Health Organization pediatric training CXR (WHO-CXR, N=410, 39% primary endpoint consolidation (PEP)), separately. Data augmentation techniques (i.e., horizonal flipping, minor rotation, translation and shear) were applied to boost model robustness. We assessed eight DNN models using the above framework, and we evaluated results using the average area under the curve (AUC) from 10-fold cross-validation. While the classification accuracy across the eight DNN models was similar in GZ-CXR, there was more variation in WHO-CXR due to the small sample size. DenseNet-121 had the best trade-off between speed and accuracy. It achieved AUC 0.91 and 0.92 for classifying bacterial pneumonia on GZ-CXR and for PEP on WHO-CXR, respectively. DenseNet-121 was also computationally efficient, with average training time about twenty and four minutes in GZ-CXR and WHO-CXR, respectively. It was two-to-four times faster than NASNetLarge and InceptionResNetV2, with similar accuracy. We developed a CXR-based pneumonia classification framework which achieved high classification accuracy. With substantial reduction in human time required, computer-aided reading of CXR-confirmed pneumonia may facilitate future disease burden and vaccine impact studies of pneumonia in children.