DenseNet based skin lesion classification and melanoma detection

Abstract

In the language of medical science, the most harmful variant of skin cancer that may develop in human cells is distinguished as melanoma. The principal reasons behind developing melanoma in human skin are still unknown. However, scientists assume that the risk of developing melanoma increases due to exposure to ultraviolet radiation emitting from the sun. The increased rate of melanoma cancer is now a threat to the medical sector to cope with the increasing number of patients. Many scientists have already researched and tried to develop different projects to identify melanoma efficiently. Skin lesions are the best approach to find the symptoms of melanoma and predict the possibility of cancer growing in the skin. In this research paper, the main objective is to classify different types of lesions and find melanoma from skin lesion images using DenseNet-121 which is a densely connected CNN-based algorithm. We evaluated on 5066 imbalanced test images from ISIC 2019 Challenge dataset for initial classification of lesion images. We also organized the dataset into a balanced dataset by over sampling and downsampling where 600 test images were used for validation. The evaluation of imbalanced and balanced datasets results in respectively 80% and 84% accuracy for lesion images classification. Moreover, we normalized the dataset into two different classes which consists of melanoma and non-melanoma lesion images to perform binary classification. In this stage, we executed our model on 2000 test images and got an accuracy of 89% for classifying melanoma accurately.