Sequential Transfer Learning Models with Additional Layers for Pneumonia Diagnosis

Pneumonia is a disease caused by numerous microorganisms affecting the alveoli sac of the lungs, filling them with pleural fluid. It affects many children and adults and can be fatal to them, hence early identification is essential. X-ray is a non-invasive and cheapest method available for the diagnosis of pneumonia. For this study, the impact of custom sequential models with additional layers for pneumonia diagnosis was investigated. The research used a dataset of 5856 X-ray images from the Kermany dataset to train deep learning models for computer-aided diagnosis of pneumonia. Transfer learning was employed with pre-trained classifiers, and the results were compared between a basic approach of using only a classification layer on extracted features and a second approach that involved adding additional layers after the pre-trained classifiers. The performance of the deep learning models was evaluated using accuracy, precision, specificity, recall, F1 score, and ROC curve. The results showed that the second approach, which included additional layers, led to an increase in the AUC of all classifiers. The Vgg16 model performed particularly well, displaying an F1 score of 89.61%.