Skin Cancer Detection Using Deep Learning

Abstract

Melanoma becomes the most serious types of skin cancer, and the fury and speed with which it spreads has resulted in a high fatality rate among those who are affected when the malignancy is not treated. The surgical excision of melanoma can be used to cure it if it is found early enough before it spreads to other parts (metastases). However, advanced malignant melanoma is difficult to cure and has a high fatality rate. So, It is necessary to differentiate between malignant melanoma or benign early on to improve survival rate. Thus, the introduction of a system to automatically classify melanoma cancer as malignant or benign is needed. Previous approaches still have problems in categorizing melanoma cancer. In this paper, the research topic aims to address the SIIM-ISIC melanoma classification problem presented by the Kaggle competition. This paper presents a melanoma classification method derived from MobileNetV2 networks; this is the first time MobileNetV2 networks have been used for the classification of melanoma skin lesions. Thus, at the first stage of training, pre-processing of samples and the use of heavy augmentation to address the problem of a highly unbalanced dataset are used for diversification. On the second level, MobileNetV2 that is a well-known deep learning network is applied for melanoma classification. Then, the effectiveness of the proposed MObileNetV2 model is assessed on the ISIC Challenge 2020 dataset. From the 3500 test images, the presented deep learning reaches an accuracy of 94 % on the ISIC-2020 dataset. Moreover, the latest technique is compared with the previous models and the accuracy of the latter technique is augmented as well as the computing cost is reduced. Keywords: Melanoma, Skin cancer, Deep learning ISIC-2020 dataset, MobileNetV2.