Reduction of Feature Extraction for COVID-19 CXR using Depthwise Separable Convolution Network

Abstract

A Convolutional Neural Network (CNN) classifier is generally utilized to classify an image tensor according to the mapped labels. The simplification of the classifier causes CNN to be often used to classify images, especially in the biomedical field. Thus, CNN is widely used to classify computer tomography (CT) and chest X-ray (CXR) images against the mapped labels. Several transfer learning models were implemented to classify CXR images for preliminary detection of COVID-19 infection, e.g., ResNet, Inception, Xception, etc. However, a transfer learning model has a maximum and minimum input resolution. Thus, the computational cost tends to be huge and unable to be optimized. Therefore, A custom CNN model can be a solution to reduce computational costs by configuring the feature extraction layers. This study proposed an efficient reduction of feature extraction for COVID-19 CXR namely Depthwise Separable Convolution Network. Furthermore, numerous strategies were adopted to lower the computational cost while retaining accuracy, including customizing the Batch Normalization (BN) layer and replacing the convolution layer with a separable convolution layer. The proposed model successfully reduced the feature extraction represented by the decreases in trainable parameters from 28.640 trainable parameters to 4.640 trainable parameters. The depthwise separable convolution effectively retains the performance accuracy 72.96%, loss 12.43%, recall 74.67%, precision 77.67%, and F1-score 75.33%. The CXR augmentation is also successfully increase the performance accuracy 74.55%, loss 11.37%, recall 77.67%, precision 79.56%, and F1-score 78.33%.