Pre-trained Deep Learning Models for Chest X-Rays' Classification: Views and Age-Groups

It is important for a generic diagnostic deep learning model in chest radiology to be independent of manual inputs such as patients' age-group and the view of Chest X-rays. Moreover, an accurate radiological diagnosis should take into consideration the patient age-group, as well as the Chest X-ray view. In this paper, we aim to find the optimal classification deep learning model to classify Chest X-rays by age-group and by view. We trained seven pre-trained deep learning models on customized Chest X-ray datasets, and we found that deep learning models based on residual blocks are the optimal models for the required classification. MobileNetV2 and ResNet50 were the best two models for classifying Chest X-rays into Adults or Pediatrics classes, and into AnteriorPosterior (AP) or Posterior-Anterior (PA) views. On the other hand, Xception was the least favored deep learning model for the two tasks. The drawn conclusions can help add an optimal preliminary classification head to a pulmonary diseases detection model, in order to optimize its performance.