Unsupervised GAN for Melanoma

Abstract

Computer vision technology is being used more often as time progresses in the medical imaging world. The neural network-based classification methods play a better role in Melanoma pattern identification to recognize an exact pattern in images. However, images need to be preprocessed such that artifacts, highlight shadows, and noises may be removed. The GAN (Generative Adversarial Network) is being applied to Melanoma’s deadliest form of skin cancer images. The GAN helps preprocess images so that finding correct patterns and focusing on identifying and distinguishing malignant and benign melanoma lesions becomes easy. For that reason, our contribution is to build the unsupervised GANs and make it easy and seamless early identification in order that patients may be diagnosed in the early stage of Melanoma. First, the publicly accessible ISIC (International Skin Imaging Collections) dataset will be preprocessed using unsupervised GAN. After that, the preprocessing images will be classified using deep neural network-based algorithms - CNN, RNN, and XG-Boost. The performance of these algorithms will be measured using six performance metrics - loss, accuracy, recall, precision, ROC, and F1 score. Finally, these performance metrics will be compared with those without preprocessed images. The outcome of this work will be helpful for dermatologists, geneticists, and health care scientists. Also, people can be diagnosed with Melanoma earlier and receive treatment before cancer becomes lethal.