An Efficient Approach for Automatic Melanoma Detection Based on Data Balance and Deep Neural Network

Abstract

One of the most serious types of skin cancer is Melanoma, which can be fatal if it is not detected in its early stages. Patients need to visit a dermatologist to diagnose infected skin and determine if it is Melanoma or not. The traditional method for a dermatologist is more complicated and requires extensive experience to look at the skin with a dermatoscope and then provide a biopsy report for diagnosis. Instead of traditional methods, artificial intelligence, especially deep learning, provides powerful results in experience-based problems without the need for experts in the specific field of the problem. For this reason, deep neural network architectures can be useful for dermatologists and patients in the early stages of identifying melanoma skin cancer. This paper offers a proposed approach for automatically classifying Melanoma using convolution neural network (CNN) architectures VGG19 and GoogleNet. From data balance for input images, which makes a huge difference in results to preprocessing images and testing VGG19, GoogleNet in the feature extraction process and final binary classification with class 1 means Melanoma and class 0 means nonmelanoma. A dataset was used from the international skin imaging collaboration datastores (ISIC 2019) with 7146 total used images. Proposed approach results show that GoogleNet accuracy is 80.07 % and 81.28% in the training and testing dataset, and VGG19 accuracy is 85.57 % and 78.21 % in the training and testing dataset.