부산대학교 도서관

Over a span of two years, the Novel Coronavirus has evolved into a highly contagious and perilous disease. Originating in Wuhan city, it swiftly extended throughout the entire country, with the potential for person-to-person transmission through the air. Presently, it has become a global pandemic, impacting regions worldwide. The primary means of confirming a COVID-19 infection is through tests known as Reverse Transcription Polymerase Chain Reaction (RT-PCR). However, these tests have some limitations, including time-consuming processes and relatively high costs. Consequently, there is an urgent need for alternative, rapid, and cost-effective diagnostic methods. Inspired by recent research that correlates COVID-19 detection with findings from CT scans, this paper proposes a methodology that leverages established Deep Learning models, such as ResNet, and employs Data Augmentation techniques on comprehensive datasets. The goal is to categorize these scans as either true positive or true negative for the presence of COVID-19. The significance of this approach extends to both patients and healthcare professionals, particularly in regions that may lack access to laboratory testing kits. This approach becomes even more critical in countries with limited resources. By utilizing CT scan data for diagnosis, it offers a viable alternative to traditional testing methods. To develop a robust and generalized model, various datasets from diverse sources have been amalgamated. This diverse dataset ensures the model's versatility and consistent performance across different scenarios. Statistical data associated with this approach reveals that the proposed model achieves an impressive sensitivity of approximately 90%, indicating its ability to correctly identify COVID-19 cases. Additionally, the model exhibits a specificity of around 85%, highlighting its capability to accurately detect true negatives. This results in an overall accuracy rate of 87%, making it a promising tool for COVID-19 diagnosis. The F1 score, a measure of the model's precision and recall, stands at 0.88, emphasizing the methodology's reliability and effectiveness.