Table 1.
Dataset used in the experiment.
Fig 1.
Network architecture and workflow.
Our network architecture (left) follows the structure of CNN with an Xception backbone to extract features and a fully connected stage. Our training and validation workflow (right) exploits three curated and preprocessed databases. The model was trained on the adult dataset, and then we carried out independent validation on the adult and pediatric datasets. Further analysis was performed in the pediatric dataset for model explainability.
Fig 2.
Heatmaps and corresponding pneumonia probabilities.
Images were obtained when training the network with and without thorax segmentation.
Fig 3.
ROC plot for the adult dataset.
ROC curve depicting the performance of our model for pneumonia discrimination for the Adult Dataset. An AUC of 0.95 (95%CI: 0.94–0.95) was obtained, showing the good performance of the presented network.
Fig 4.
ROC plots for the pediatric dataset.
ROC curve for pneumonia discrimination was obtained using all images from the Pediatric Dataset. Although the images in this dataset were distinct from the training dataset, our model achieved an AUC of 0.82 (95%CI: 0.8–0.83), highlighting the model’s promising performance in accurately distinguishing pneumonia cases within pediatric populations.
Fig 5.
Examples of pneumonia predictions on images from the pediatric dataset.
Images corresponding to the Pediatric dataset were provided by the proposed neural network. Upper images reflect areas where the network provided accurate predictions based on the network probabilities. Lower images reflect cases where the network was less confident about the decision. (Left) We observe two clear chest X-ray images. The upper image exhibits lung fields without abnormalities, while the lower one, although lacking signs of pneumonia, displays some bilateral perihilar infiltrates, thus, potentially explaining why our proposed networks give a mid-range probability value despite being a healthy case. (Middle) These two chest X-rays indicate bacterial pneumonia, typically more severe than viral pneumonia. The upper image displays pronounced bilateral infiltrates, particularly notable in the right hemithorax, whereas the lower image exhibits subtle right paramilitary infiltrates, less pronounced than the upper ones, explaining the lower confidence of the network as reflected by the predicted probability. (Right) These two cases illustrate viral pneumonia, which typically spreads more diffusely across both lung fields. The upper image reveals a bilateral diffuse infiltrate, with a more pronounced presence in the right hemithorax, while the lower image displays a similar pattern but with lesser intensity. The images in the lower row, despite their difference in disease status, reveal noticeable radiographic similarities with subtle paramilitary infiltrates without distinct consolidations, explaining why they share similar prediction probabilities.
Table 2.
Validation results for pneumonia discrimination of the neural network (trained with adult data) on both the adult and the pediatric datasets.