COVID-19 detection using CXR images: simple CNN vs. transfer learning

Abstract

We propose a novel convolutional neural networks (CNN) model for efficient COVID-19 detection using CXR images without recourse to pre-training. We study to what extent the proposed model compares with pre-trained-based transfer learning approach; this latter being based on the exploitation of the capabilities of a model, initially trained on a large dataset, but different from smaller available datasets commonly under use. The proposed model compares well with two state-of-the-art pre-trained models: ResNet50V2 and Inception V3;even outperforming both, in terms of validation accuracy with 93.80% as compared with 92.00% and 92.60%, for the other two. Although the cost of avoiding pre-training incurs time-delay in the testing phase, it is, in the worst-case scenario, less than 2.5% w.r.t. its competitors - all confusion matrix-based criteria combined.