Tuning of Data Augmentation Hyperparameters to Covid-19 Detection in X-Ray Images with Deep Learning

Abstract

The Covid-19 pandemic has been declared in 2020 by the World Health Organization. One of the most relevant aspects of this respiratory disease is the fact that the infection caused by the new coronavirus has a high rate of spread. Thus, rapid and accurate diagnosis can contribute to reducing the transmission rate. In this aspect, in the literature, Deep Learning techniques are studied for application in the detection of this disease through X-ray images of the patient’s lung. However, one of the challenges in this area is the training of Convolutional Neural Network models with a database with few samples. One possibility is the generation of artificial images through Data Augmentation techniques. Thus, the objective of this work is to propose a careful methodology for the tuning of Data Augmentation hyperparameters for the classification of lung X-ray images in Covid-19 detection with Deep Learning. The proposed method consists of analyzing the accuracy of 36 Data Augmentation transformations applied to generate new images for training with balanced and unbalanced database. After the selection of hyperparameters, the classifier system achieved accuracies up to 100% on the testing stage, both for combinations and individual transformations with balanced database. Therefore, it is recommended to use a balanced database with the use of zoom, rotation, brightness in combination or individually, for Covid-19 versus Normal and Covid-19 versus Pneumonia classification.