Description

The world has been fighting against COVID-19 for more than two years. Given the urgency of early and accurate detection of COVID-19 cases, deep learning is a promising technique for identifying COVID-19 cases and preventing the infection from spreading. This project aims to use the DenseNet121 architecture, a dense convolutional neural network, to diagnose COVID-19 patients from chest X-ray images. The project will compare the performances of various optimization algorithms, such as SGD, Adam, and their variants. Through tuning different sets of hyperparameters for each optimizer, applying early stopping based on alidation error during training, and using Binary Cross Entropy as the loss function, the best non-adaptive and adaptive optimization algorithms are selected according to validation and test errors. They are SGD with Nesterov momentum and Adamax. A trade-off between convergence speed and classification error for the two best performing optimizers is observed. Compared to Adamax, SGD with Nesterov momentum has superior performance in terms of validation and test accuracies, but has a slower convergence speed. Moreover, our results show that SGD with Nesterov momentum has the highest test accuracy of 94%, arriving at a different conclusion compared to other papers performing similar optimizer evaluations on COVID-19 chest X-ray images.