Arm Injury Classification on a Small Custom Dataset Using CNNs and Augmentation

Abstract

Some situations do not have a wide public access to medical expertise, poor health care systems, or a shortage of physicians. Therefore, the development of computer-aided diagnosis (CAD) systems that automatically estimate the extents of patients' injuries is required to reduce the burden of diagnosis on physicians and to provide early diagnosis and early treatment of patients. This study presented convolutional neural networks (CNNs) and image data augmentation for classifying external arm injuries. The arm injury classification is a three-class problem: healthy, wound, and bruises. With the limited number of data available, image data augmentation of a perspective transformation was used to improve an overtraining problem of CNNs. The experimental results showed that the CNN with augmentation had a higher average accuracy.