Development of a Deep Learning Model to Classify X-ray of Normal, COVID-19, and Pneumonia-affected Patients

Law, Boon Kai (2021) Development of a Deep Learning Model to Classify X-ray of Normal, COVID-19, and Pneumonia-affected Patients. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Law Boon Kai.pdf Restricted to Registered users only Download (4MB)

Abstract

The chest radiography (X-ray) is one of the commonly used radiological examinations that help doctors to have a better view of a patient’s internal organs such as the heart, lungs, trachea and others. Through the exploring of x-ray images, doctors can spot abnormalities in the internal organs. In additions, the x-ray service enables pneumonia to be confirmed. Nowadays, the morbidity and mortality of pneumonia-affected patients are under serious concern due to coronavirus (COVID-19), a global pandemic. The effective way to combat all of the pneumonia diseases is through early recognition that can lead to the fast administration of empirical antibiotic therapy. However, the delay in recognition and misdiagnosis due to the shortage of expert and human inspection will lead to fatal consequences. The traditional way of inspection, such as manual inspection which requires radiologists to assess the chest x-rays according to previous training, knowledge as well as experience, is susceptible to errors. This is because human inspection alone is undependable, and there might be imprecision of eyesight during an assessment. Besides, the cost to hire a human expert for inspection is expensive. To overcome this issue, a deep learning CNN model to classify X-ray of COVID-19, normal and pneumonia-affected patients is developed in this paper. The highest training accuracy achieved in this experiment is 97.38% by using the VGG-16 model with a learning rate of 10-2. The highest weighted average accuracy achieved in the experiment is 94% by using the VGG-19 model with a learning rate of 10-3 and VGG-16 model with a learning rate of 10-2. The accuracy for the SqueezeNet model with a learning rate of 10-3 is 93% in both cases of training with and without image augmentation. For those users that have no restriction on the hardware requirements, memory and strictly to the accuracy of the model, then the VGG-19 model with a learning rate of 10-3 and the VGG-16 model with a learning rate of 10-2 is suitable for them whereas the SqueezeNet model with a learning rate of 10-3 is a balanced option for those users that are facing hardware and memory restriction.