A Lightweight Deep Learning Model for COVID-19 Detection

Abstract

COVID-19 is a contagious disease that has caused more than 230,000 deaths worldwide at the end of April 2020. Within a span of just a few months, it has infected more than 4 million peoples across the globe due to its high transmittance rate. Thus, many governments have tried their best to increase the diagnostic capability of their hospitals so that the disease can be identified as early as possible. However, in most cases, the results only come back after a day or two, which directly increases the possibility of disease spreadness because of the delayed diagnosis. Therefore, a fast screening method using existing tools such as x-ray and computerized tomography scans can help alleviate the burden of mass diagnosis tests. A chest x-ray is one of the best modalities in diagnosing a pneumonia symptom, which is the primary symptom for COVID-19. Hence, this paper proposes a lightweight deep learning model to screen the possibility of COVID-19 accurately. A lightweight model is important, as such it allows the model to be deployed on various platforms that include mobile phones, tablets, and normal computers without worrying about the memory storage capacity. The proposed model is based on 14 layers of convolutional neural network with a modified spatial pyramid pooling module. The multiscale ability of the proposed network allows it to identify the COVID-19 disease for various severity levels. According to the performance results, the proposed SPP-COVID-Net achieves the best mean accuracy of 0.946 with the lowest standard deviation among the training folds accuracy. It comprises of just 862,331 total number of parameters, which uses less than 4 MegaBytes memory storage. The model is suitable to be implemented for fast screening purposes so that better-targeted diagnoses can be performed to optimize the test time and cost.