Challenges and Opportunities in Malignant Image Reconstruction Using GAN: A Review

Melanoma, the most common form of skin cancer, occurs when abnormal cells grow in the skin. It could be detected in the early stage to cure and save patients' lives. However, early detection is possible from dermoscopic images to train, test, and validate through deep learning models [1]–[6]. However, our current dataset is unbalanced - fewer malignant images in comparison to benign images. Therefore, this paper reviewed different Generative Adversarial Networks (GAN) algorithms to generate synthetic malignant images. The variety of GAN algorithms, including Deep Convolutional GAN (DCGAN), Conditional GAN (cGAN), Progressive GAN (PGAN), Wasserstein GAN (WGAN), GFP (Generative Facial Prior) GAN, Radio GAN, and Cycle GAN, etc. These algorithms provide many opportunities and have many challenges in the reconstruction of natural and correct patterns of images. The major challenges include different sizes of images, preprocessing of raw images, correct number of epochs, appropriate training of generator and encoder, discriminator methods of GANs, etc. These GAN methods generated images are useful in achieving better performance of image classification.