The Effects of the Traditional Data Augmentation Techniques on Long Bone Fracture Detection

Abstract

Image collection and preparation phases are highly costly for machine learning algorithms. They require the majority of labeled data. Hence, the image pre-processing method, data augmentation, is commonly used. Since there are so many proposed methods for the augmentation task, this comparison study is presented to be a supporting guide for the researchers. In addition, the lack of studies with animal-based data sets makes this study more valuable. The study is investigated on a comprehensive medical image data set consists of X-ray images of many different dogs. The main goal is to determine the fracture of the long bones in dogs. Many traditional augmentation methods are employed on the data set including flipping, rotating, changing brightness and contrast of the images. Transfer learning is applied on both raw and augmented data sets as a feature extractor and Support Vector Machine (SVM) is utilized as a classifier. For the classification task, the experimental study shows that changing the contrast is the outstanding method for accuracy manner, while the rotation method has the best sensitivity value.