Detection of COVID-19 pneumonia on computed tomography images using a Lightweight Deep Learning Model / Serena Low Woan Ching

Serena Low , Woan Ching (2023) Detection of COVID-19 pneumonia on computed tomography images using a Lightweight Deep Learning Model / Serena Low Woan Ching. PhD thesis, Universiti Malaya.

	PDF (The Candidate's Agreement) Restricted to Repository staff only Download (306Kb)
	PDF (Thesis PhD) Download (1840Kb)

Abstract

SARS-CoV-2, also known as COVID-19, is a novel contagious respiratory disease discovered in 2019 that caused a worldwide pandemic that claimed many lives. The virus epidemic was initially discovered in December 2019 in Wuhan, China. It immediately escalated into an international crisis, causing widespread illness, death, and significant socio-economic disruptions. The rapid and accurate identification of COVID-19 is crucial to the ongoing global epidemic. One of the various medical devices is computed tomography (CT) imaging, which has shown promise in its application to detect distinctive patterns associated with lung tissue deterioration. The common practice is relying on radiologists to diagnose the CT images, which is time-consuming. Various advanced CNN architectures can detect and classify CT images. However, most require high computational costs and are not designed for commercial use. The study aims to automatically detect and classify ‘COVID-19 pneumonia’, 'normal', and ‘pneumonia’ lung CT images using transfer learning of the pre-existing CNN models and the proposed model. The study proposed a DL model inspired by ResNeXt and Inception to create a synergistic effect. The research conducted binary classification to compare the results of the existing models and multi-classification to compare the existing models and the proposed model. The existing models are DenseNet 201, GoogLeNet, ResNet 50, ResNet 101, ResNet 152, and ResNeXt 101. The dataset was collected from medRxiv, bioRxiv, NEJM, JAMA, Lancet, and the China National Centre of Bioinformatics. A comprehensive dataset was subdivided into training, validation, and testing. The images were then pre-trained using existing CNN architectures. Pre-trained models that had been fine-tuned extracted the features from the CT images. The research study applied transfer learning and deterministic concepts to the existing and proposed models to evaluate and compare their results. The proposed model is also designed to capture nuanced features indicative of COVID-19 infection, inspired by ResNeXt 101 and Inception, to achieve a lightweight and efficient CNN model. Data augmentation was applied to the dataset to introduce variations of unseen data to the proposed model. The study indicated the results for binary classification, ResNeXt 101, which obtained the best results among the existing CNN architectures. It acquired the highest sensitivity, specificity, precision, negative predicted value (NPV), accuracy, and F1-score of 0.9571, 1.0000, 1.0000, 0.9639, 0.9800, and 0.9781. The experimental results for multi-class classification showcase the efficacy of the proposed model, achieving an impressive accuracy, precision, recall, and F1-score of 0.9980. The proposed model has 7,724,523 parameters, 11 times less than ResNeXt 101 while having similar accuracy. The research depicted the possibilities of the proposed model in aiding medical diagnosis, especially in COVID-19 pneumonia detection using CT images. In conclusion, the proposed model and other developed models from the thesis offer promising tools for healthcare professionals to identify and detect COVID-19 pneumonia CT images early. It encourages the application of CNN architectures as a diagnostic aid for radiologists, particularly during the pandemic.

Actions (For repository staff only : Login required)