Multiclass Skin Disease Classification using Generative Adversarial Networks and Neural Network Models

Identification of skin diseases and treating it accordingly is one of the recent research areas that is being worked upon and every improvement in it is beneficial for the mankind. This work focuses on multiclass classification of skin diseases using neural network models. There are nine classes in the dataset. The number of samples are very less for each class. Generative Adversarial Network (GAN) is an algorithm which generates fake images and, in this work, GAN has been used to increase the number of samples. The GAN algorithm generates synthetic images in such a way that it has its own generator and discriminator. The generator generates synthetic images and the discriminator identifies whether the images are fake or not after comparing the fake images with real ones. The objective is to increase the sample size of the dataset using GAN. The comparison is performed between Augmented dataset and original dataset. Results show that dataset with GAN images resulted in significantly better accuracy on the test dataset than the non-GAN images dataset. The obtained accuracy was 96% compared to the 46% obtained without GAN which is highly encouraging and the proposed work is more robust.