A New Classification Model COVID-SERA-NeXt for COVID-19 CXR Images

Abstract

The Chest X-ray (CXR) images of COVID-19 patients are different from those of normal people, which has been an effective base for making correct diagnosis. It is an important way to help medicine doctors to make the fast and accurate diagnosis for patients by using computer aided automatic classification technique based on the patient chest X-ray images. The new COVID-SERA-NeXt model was proposed in this paper for classifying COVID-19 CXR images by introducing the cross-stacked channel attention module and residual attention module, as well as the proposed dimensional reduction module, into the ResNeXt model. The performance of the proposed COVID-SERA-NeXt model was tested on the open accessed COVIDx dataset by extensive experiments. The experimental results show that the proposed COVID-SERA-NeXt model is superior to its base model ResNeXt. It achieves the accuracy and Macro_Recall of 96.11% and 95.46%, respectively. Further experiments demonstrate that the proposed COVID-SERA-NeXt model achieves better performance to classify COVID-19 CXR images when it is pre-trained using ChestX-ray8 dataset.

Keywords

• covid-19
• computer aided diagnosis
• attention mechanism
• deep convolutional neural networks
• cxr images
• classification