A study of hyperparameters effect on CNN performance for chest X-ray based COVID-19 detection

Intellectual Systems and Technologies

COVID-19 disease has been spreading around the world for the last four years. Different generations of corona viruses appeared: Alpha-, Beta-, Gamma-, and Delta variants. Thus, COVID-19 changed human lifestyle and affected economic development of many countries. According to clinical studies, most of the positive cases of COVID-19 patients suffer from lung infection. For this, a lot of efforts were aimed at developing fast and accurate detection methods. Thanks to the Deep Learning techniques that facilitate the process of identifying COVID-19 based on the chest images of the patients. X-ray and CT scan images are commonly used to evaluate corona virus lung infection. X-ray images are adopted by many researchers since they place less financial burden on the patient. In this work, we used chest X-ray images to develop eight CNN-based detection models. Three sets of images, i.e., COVID-19, pneumonia and normal cases were used for the training and testing. The performance of each model was optimized based on different hyperparameters to come up with the best results in terms of high detection accuracy, recall, precision and f1 score. These hyperparameters include Number of CNN layers, filters, dense layers, and number of nodes per dense layer. Our findings show that increasing both the CNN layers and number of filters result in high precision and f1 score of the positive samples, while increasing the number of dense layers leads to low precision recall and f1 score.