Generative Adversarial Network Image synthesis method for skin lesion Generation and classification

Abstract

Skin cancer is the most commonly diagnosed cancer in today's growing population. One of the common limitations in the treatment of cancer is in the early detection of this disease. Mostly, skin cancer is detected in its later stages, when it has already compromised most of the skin area. Early detection of skin cancer is of utmost importance in increasing the chances for successful treatment, thus reducing mortality and morbidity. Currently, most dermatologists use a special microscope to examine the pattern and the affected area. This method is time-consuming and is prone to human errors, so there is a need for detecting skin cancer automatically. In this study, we investigate the automated classification of skin cancer using the Deep Convolution Generative Adversarial Network(DCGAN).In this work, Deep Convolutional GAN is used to generate realistic synthetic dermoscopic images, in a way that could enhance the classification performance in a large dataset and to evaluate whether the classification accuracy is enhanced or not, by generating a substantial amount of new skin lesion images. The DCGAN is trained using images generated by the Generator and then tweaked using the actual images and allow the Discriminator to make a distinction between fake and real images. The DCGAN might need slightly more fine-tuning to ripe a better return. Hyperparameter optimization can be utilized for selecting the best-performed hyperparameter combinations and several network hyperparameters, namely number of iterations, batch size, and Learning rate can be tweaked, for example in this work we decreased the learning rate from the default 0.001 to 0.0002 and the momentum for Adam optimization algorithm from 0.9 to 0.5, in trying to reduce the instability issues related to GAN models. Moreover, at each iteration in the course of the training process, the weights of the discriminative and generative network are updated to balance the loss between them. This pretraining and fine-tuning process is substantial for the model performance.