Comparison of simple augmentation transformations for a convolutional neural network classifying medical images

Simple image augmentation techniques, such as reflection, rotation, or translation, might work differently for medical images than they do for regular photographs due to the fundamental properties of medical imaging techniques and the bilateral symmetry of the human body. Here, we compare the predictions of a convolutional neural network (CNN) trained for binary classification by using either no augmentation or one of seven usual types augmentation. We have 11 different medical data sets, mostly related to lung infections or cancer, with X-rays, ultrasound (US) images, and images from positron emission tomography (PET) and magnetic resonance imaging (MRI). According to our results, the augmentation types do not produce statistically significant differences for US and PET data sets, but, for X-rays and MRI images, the best augmentation technique is adding Gaussian blur to images.