An Enhanced Convolution Neural Network Model based on Depthwise-separable Convolutions for Pneumonia Classification

This paper is an attempt to customize a lightweight model to classify pneumonia images by integrating depthwise-separable convolutions with typical CNN model, and focus on the performance of DSCNN in comparison with typical CNN model based on X-ray images. The experimental result shows that in our four-layer structure, DSCNN reduce around 50,000 parameters compared to CNN. But DSCNN had a relative low recall on COVID-19(89.23%). However, with proper means of optimization such as focal loss and data augmentation, there was a slight increase in test accuracy of DSCNN(from 95.25% to 96.14%), and a significant increase in recall on COVID-19(from 89.23% to 94.61%). And this model also performed well on the rest two labels.