Tackling the class imbalanced dermoscopic image classification using data augmentation and GAN

Dermoscopy is a noninvasive way to examine and diagnose skin lesions, e.g. nevus and melanoma, and is a critical step for skin cancer detection. Accurate classification of dermoscopic images can detect skin cancer at an early stage and bring social and economic impact to patients and communities. Using deep learning methods to classify dermoscopic images has shown superior performance, but existing research often overlooks the class imbalance in the data. In addition, although a handful of public datasets are available for skin cancer research, these datasets are generally not large enough for deep learning algorithms to produce accurate results. In this paper, we propose to use data augmentation and generative adversarial networks (GAN) to tackle class-imbalanced dermoscopic image classification. Our main objectives are to determine (1) how state-of-the-art fine-tuned deep learning models perform on class-imbalanced dermoscopic images, (2) whether data augmentation and GAN can help alleviate class imbalances to improve classification accuracy, and (3) which method is more effective in addressing the class imbalance. By using public datasets and a carefully designed framework to generate augmented images and synthetic images, our research provides clear answers to these questions. Code and data used in the study are available at: https://github.com/mjan2021/Dermoscopic-image-classification.git