Improved COVID-19 Diagnosis Using a Hybrid Transfer Learning Model with Fuzzy Edge Detection on CT Scan Images

Abstract

CT imaging provides detailed and comprehensive visualization of lung abnormalities associated with the disease, aiding in accurate and timely diagnosis. Medical specialists often recommend the use of CT scans for COVID-19 diagnosis, particularly in the second week of illness when there is a high suspicion of infection, due to the unique lung patterns observed in COVID-19 patients. Consequently, there is a need for a quick recognition system based on deep learning techniques to detect COVID-19 infection and prevent its spread. In this study, a novel fusion method incorporating dropout, data augmentation, transfer learning, and edge detection using the fuzzy technique was employed, resulting in improved model accuracy and reduced overfitting. In addition, CT images of 300 patients from the Fariabi Hospital in Kermanshah Province were collected and classified into two groups based on lung involvement. Critical frames, essential for decision-making, were separated and used for disease diagnosis by the designed models under specialist supervision. The findings demonstrate that the proposed hybrid transfer learning model utilizing the fuzzy technique achieves significant performance improvement, with 97% accuracy on the training set and 88% accuracy on the test set.