Comparison of Traditional Transformations for Data Augmentation in Deep Learning of Medical Thermography

Abstract

Convolutional neural networks (CNN) models demonstrate high performance in image-based applications such as image classification, segmentation, noise reduction and object recognition. However, balanced and sufficient data are required to effectively train a CNN model but this is not always possible. Since conditions of both hospitals and patients are not always appropriate for collecting data and patients with the same disease are not always available, the problem of collecting balanced and sufficient data are often occurred in medical fields. In this study, comparison of traditional data augmentation methods such as rotating, mirroring, zooming, shearing, histogram equalization, color changing, sharpening, blurring, brightness enhancement and contrast changing were performed by using neonatal thermal images. These images belonged to 19 unhealthy and 19 healthy neonates were obtained from Selcuk University, Faculty of Medicine, Neonatal Intensive Care Unit. A combination of three different augmentation methods were implemented to original images in each one of the 10 different comparisons accomplished and CNN was used to classify the these comparisons in the study. When contrast changing, sharpening and blurring methods were used, increased the sensitivity rate by 23.49%, the specificity rate by 29.09% and the accuracy rate by 26.29%. The obtained results show that simple and low-cost traditional data enhancement methods can improve the performance of classification.