Determining the best convolutional neural network for identifying tuberculosis and pneumonia in chest x-rays

Tuberculosis (TB) and pneumonia are commonly misdiagnosed respiratory conditions associated with high rates of mortality. Chest X-rays (CXRs) are an inexpensive method to identify respiratory conditions. Thus, a model used to distinguish between CXRs depicting lungs classified as normal, pneumonia, and tuberculosis would lead to accurate diagnoses of these respiratory conditions. This need is fulfilled by the recent explosion in deep learning, and new models with robust performance are constantly developed. However, these models have varying strengths and weaknesses which allow them to excel at certain tasks and struggle with others. Therefore, testing these models is essential to find the most suited model. In this study, we trained and applied six convolutional neural networks, the InceptionV3, ResNet50, ResNet152, InceptionResNetV2, DenseNet121, and AlexNet, to the diagnosis of TB and pneumonia. We hypothesized that InceptionResNetV2 would perform best for this task due to its combination of inception blocks that reduce the dimensionality of the CXR images and residual blocks that allow for deeper models by eliminating vanishing gradient. After training on a combination of five datasets from the Guangzhou Women and Children’s Medical Center, Shenzhen, Montgomery County, Belarus, and ChestX-ray8, it was found that various models excelled in predicting different diseases shown in the datasets. The results displayed that there was no clear superior model but instead significant superiority within certain diseases.