Generative Adversarial Networks for Melanoma Detection

Melanoma is famously known as the most dangerous skin cancer as it is capable of transferring from tissue to tissue within an individual. Therefore, if this skin cancer is identified and treated in the early stages, then the patient's survivability increases manifolds. However, visual inspection conducted by many medical professionals (the method used conventionally) leaves a lot of scope for errors, and the whole process can be time-consuming. In this study, the application of Generative Adversarial Networks GAN for such a critical operation is studied and evaluated. The GAN module provides a sophisticated data augmentation process for the Convolutional Neural Network (CNN) classifier. CNN classifiers eventually learn and differentiate melanoma images from the rest. In this implementation, an accuracy of 93% is observed. Additionally, the scope for several other kinds of classifiers is also studied, such as Support Vector Machine (SVM), Random Forest Classifier, K Nearest Neighbours (KNN), and Bayesian Classifier, etc. HAM10000 dataset provided around 11000 dermoscopic images of different kinds of skin lesions. This dataset is used almost exclusively for studies conducted on melanoma or skin cancer. However, a very high-class imbalance exists in this dataset. This study aims to develop a reliable and fast-working model that can identify melanoma at the point of care for patients. If such a model is practically used then the quality of life and survivability of melanoma patients can be expected to increase.