Generating Synthetic Medical Images by Using GAN to Improve CNN Performance in Skin Cancer Classification

One of the main reasons of slow progress in using deep learning methods for cancer detection is the lack of data, especially the annotated data which is usually used for supervised learning algorithms. This paper presents a Convolutional Neural Network (CNN) to detect skin cancer. The primary database which is used to train the designed CNN algorithm has 97 members (50 benign and 47 malignant), which are collected from the International Skin Imaging Collaboration (ISIC). In order to compensate the lack of data for training the proposed CNN algorithm, a Generative Adversarial Network (GAN) is designed to produce synthetic skin cancer images. The classification performance of the designed trained CNN without the obtained synthetic images is near 53%, but by adding the synthetic images to the primary database the performance of the model is increased to 71%.