Investigating Image Augmentation for Classification of Chest X-Ray Images

Convolutional neural network (CNN) plays a significant role with different tasks in computer vision task, specifically with medical imaging. Most of computer vision tasks achieved superior performance by utilizing CNN-based models in a supervised manner. However, Deep Learning (DL) models learn better with larger dataset which is sometimes hard in most of fields especially when it comes to medical imaging with rare diseases. To overcome data shortage, data augmentation is a solution to increase the amount of data used for training a DL model. This work investigates three commonly used augmentation techniques: horizontal flip, rotation and shearing. Experiments were conducted with three different scenarios using different training data. First scenario, the model is trained with original data without augmentation. Second scenario called mixed method where the training data are half augmented and half original. Third scenario called all method where whole training data augmented with one of the selected augmentation technique. As result, six training-sets were prepared. In this study Covid-19 X-ray images was used as case study. ResNet50 pre-trained architecture was used for a classification task on chest X-ray images to classify them into “Covid-19” or “normal”. The results show that using mixed method is better than all method. Moreover, the horizontal flip technique shows the highest score in the model performance among other techniques.