Melanoma Detection using Advanced Deep Neural Network

Melanoma is a type of skin cancer that starts in the cells (melanocytes) that govern the color of your skin. Melanoma is the most lethal one among all other skin diseases and the only reason for 77% deaths due to skin cancer. The best way to reduce these deaths is to detect cancer at its early stages so it can be treated and cured with minor treatment or surgeries. To speed up and improve the process of early detection, we propose an automatic classification method for melanoma cancer using an advanced deep neural network. Deep learning models require a large dataset to work efficiently, but due to limited time and the heavy workload of doctors, there is a lack of annotated skin can-cer images. Therefore, the proposed model introduces adversarial training for achieving better accuracy even with a small amount of data. The model removes unnecessary details and noise from the image and amplifies the depth and gradient in the dimensions and the shade of the image. This proposed adversarial method uses the gradients of the loss with respect to the input image to create a new adversarial example image that maximizes the loss for an input image. The synthetically generated images are used in the classification system for training and testing purposes. A comparative analysis of training with an adversarial approach and without an adversarial approach on different pre-trained models, namely VGG16, VGG19, Densenet121, and Resnet101, is also introduced in this work. ResNet101 with adversarial training has shown a state-of-the-art accuracy performance of 84.77% for melanoma classification. Therefore, the proposed approach can be considered an efficient method for classifying benign and malignant melanoma.