Comparison of pre-trained deep learning model classification performance of COVID-19 and normal chest X-ray images

INTRODUCTION: The aim of this study was to compare the accuracy and performance of 12 pre-trained deep learning models for classifying covid-19 and normal chest X-ray images from Kaggle. MATERIALS: a desktop computer with an Intel CPU i9-10900 2.80GHz and NVIDIA GPU GeForce RTX2070 SUPER, Anaconda3 software with 12 pre-trained models including VGG16, VGG19, DenseNet121, DenseNet169, DenseNet201, RestNet50V2, RestNet101V2, RestNet152V2, InceptionRestnetV2, InceptionV3, XceptionV1 and MobileNetV2, covid-19 and normal chest X-ray from Kaggle website. METHODS: the images were divided into three sets of train, test, and validation sets using a ratio of 70:20:10, respectively. The performance was recorded for each pre-train model with hyperparameters of epoch, batch size, and learning rate as 16, 16 and 0.0001 respectively. The prediction results of each model were recorded and compared. RESULTS: from the results of all 12 pre-trained deep learning model, five models that have highest validation accuracy were DenseNet169, DenseNet201, InceptionV3, DenseNet121 and InceptionRestNetV2, respectively. CONCLUSION: The top-5 highest accuracy models for classifying the COVID-19 were DenseNet169, DenseNet201, InceptionV3, DenseNet121 and InceptionRestnetV2 with accuracies of 95.4%, 95.07%, 94.73%, 94.51% and 93.61% respectively.