Analysis of Data Augmentation on Skin Lesion Classification by Using Deep Learning Models

Abstract

Skin lesions may occur without any trauma, as well as due to prolonged exposure to the sun, ultraviolet rays, solarium use, advancing age, family history of skin disease, and physical and genetic factors. Early and accurate diagnosis is important, these lesions may be insignificant or may be the precursors of many dermatological diseases. For this reason, several studies have been carried out for the detection and diagnosis of skin lesions, especially with the developments in the field of deep learning, and many data sets have been created and used. The skin lesions in these data sets were classified by dermatology specialists in terms of their sub-types. However, due to its nature, some classes have relatively few examples compared to others. This can lead to inconsistent results in computational models. This study performs an analysis to observe the effects of data augmentation methods on the performances of various deep learning models on skin lesion classification. The models were evaluated based on performances on the original and balanced data sets. While building of deep learning architectures, well-known pre-trained models were used to take advantage of the transfer learning approach along with the original models. For the first time, one of the recent model, EfficientNet V2, was applied for skin lesion classification problem. We observed that there is a positive contribution of data augmentation methods to the deep learning models on skin lesion classification task.