Abstract:
Pneumonia causes death every year worldwide, especially under five-year-old children 15 percent of death are accounted for it. Chest x-rays are primarily used for the diagnosis of pneumonia disease. However, there is a low number of trained radiologists. It is a challenging task to examine chest x-rays when there is a high number of patients, particularly in the sub-Saharan region. There is a need to improve the diagnosis accuracy and reduce radiologists' caseload. In this work, an efficient and generalizable model for the detection of pneumonia trained on chest x-ray images was developed. It could support radiologists in their decision-making process. A modern approach based on convolutional neural network (CNN) and DenseNet201 pre-trained models were used to solve this problem. Both methods are supervised learning approaches in which the network
predicts the result based on the quality of the dataset used. Data augmentation techniques were used to augment the training dataset in a more balanced way. DenseNet201 model outperformed the CNN model. Finally, the model is evaluated with F1-score as it helps to bring balance in case there is a class imbalance. The final DenseNet201 model achieved an F1-score of 95.59 on the unseen data from the Guangzhou Women and Children’s Medical Center pneumonia dataset and
was able to generalize better to different data with an F1-score equal to 94.29 on patients with age less than 30 in the National Institute of Health (NIH) chest x-ray image dataset. The developed model is generalizable to new data, especially at young ages. Therefore, it can be used for a pneumonia diagnosis and can reduce the caseload for the radiologists.
