Utilizing deep learning architectures for early detection of lung diseases in chest X-ray images

Abstract

"COVID-19 and viral Pneumonia are types of lung disease that are highly infectious, and have a tendency of causing an outbreak, that then puts stress on the healthcare sector. When the COVID-19 pandemic first began there were still some working-class people who could not work remotely and so, had to expose themselves to the virus. The COVID-19 pandemic has presented the research community with a complicated problem because of the significant human and financial costs. Consequently, a suitable system is required, that will try to predict if a patient is infected or not by analyzing their X-Ray report. Before the final result of their lung disease is provided to the doctors, patients who are at risk and, having other lung related issues might benefit from having this component of patient evaluation as the results from the tests can be sent through the internet to reduce the danger of infection while traveling to get tested. This thesis has trained 5 models (ResNet18, VGG16, Alexnet, Inception-v3 and Densenet169) to detect COVID-19 and viral Pneumonia in chest X-ray images from the COVID-QU-Ex dataset which yielded 90-95% prediction accuracy. Afterwards, we created an ensemble model using model-level fusion, comprising of pre-trained CNNs ResNet18, VGG16, and Inception-v3 which were the top-performing models. Afterwards, the performance of the model is evaluated by using multiple metrics such as precision, recall, and F1-score was generated. The ensemble yields a prediction accuracy of 95.002% with a recall of 94.543% and main focus of the ensemble model is to utilize it as a supplementary tool by medical professionals for early detection, while ensuring its credibility through explanation generation. This explanation generation is created using AI tools, the code generates a prediction accuracy score, a prediction mask, and a saliency map in an effort to better explain the predictions made by our ensemble model. The saliency map, which highlights the areas in an image that have the greatest impact on a machine learning model’s prediction, is created by further processing the predicted mask. The severity and location of lung infections inside the X-ray image are shown on this map, which also illustrates how confident the model is in its predictions."