Effects of Data Augmentation Techniques on Classification Performance in Knee MRIs

Abstract

- The role of the amount of data used in increasing the effectiveness of deep learning models is very important. Due to the insufficient publicly available data in some sub-fields of health, data augmentation is vital. This study proposes an approach to improve the experimental work process in deep learning-based injury detection using data augmentation techniques on knee Magnetic Resonance (MR) images. The study is also one of the few in the body of literature to examine the impact of data augmentation in hard tissues. The effect of data augmentation on classification performance is tested using various transfer learning models, and the highest success rates in this study are determined for three classes. Forecasting achievements: the accuracy is 89.98% in the abnormal, 80.35% in the anterior cruciate ligament, and 76.66% in the meniscus classes. As a result of the experiments, it has been seen that the AutoAugment architecture works faster and generally gives more successful results than other augmentation methods.