Performance Analysis of Deep Convolutional Features using Support Vector Machines for COVID-19 Diagnosis on X-ray Images

Since the first case of COVID-19 appeared in Wuhan city, China, in December 2019, the disease has affected more than millions of people worldwide. Therefore, early detection of COVID-19 is important to prevent transmission to more people. One method widely used to detect COVID-19 through X-ray images is Convolutional Neural Networks (CNN). However, CNN needs large amounts of image data to build models with high accuracy, while the medical image has limited amounts of data. To overcome this problem, transfer learning technique where CNN is used as a feature extraction method is usually be chosen as an alternative. However, most studies use the extraction results of the final layers such as fully connected layer or the last convolutional layer. In this study, all layers will be used by turns to analyze how the extraction results affect the performance of classification method. The CNN models used are pre-trained models VGG16 and VGG19, while the classification method used is Support Vector Machines (SVM). Based on the results of the study, the extraction results by the initial layer gave a better performance on SVM compared to the layers that are deeper in the selected CNN architecture. Several layers in CNN model did not analyze due to limited source capability in doing computation. Therefore, as the future work, the rest layers of CNN in this study can be analyzed as well as the other CNN models and the classification method.