Performance Analysis of VGG-19 Deep Learning Model for COVID-19 Detection

Various clinical studies and researchers have established that chest CT scans provide an accurate clinical diagnosis on the detection of COVID-19. The traditional gold standard RT-PCR testing methodology might give false positive and false negative results than the desired rates. AI has proven to be the driving force in developing various COVID-19 management tools. Provided with the situation of lack of datasets, we applied a transfer learning approach to detect COVID-19 from chest CT images. The previous work observed that the VGG-19 has better performance with medical image data compared to other deep learning models such as VGG-16, InceptionV3, DenseNet121, which showed overfitting in the initial epochs. This study determined the best performing parameters for the VGG-19 transfer learning model to classify COVID-19 cases and healthy cases. We experimented with the model against three parameters: activation function, loss function, and training batch size. After the analysis, we found that the VGG-19 model with SoftMax activation function, Categorical cross-entropy loss function, and training batch size as 32 has the highest accuracy of 93%.