A CNN Framework for Classification of Melanoma and Benign Lesions on Dermatoscopic Skin Images

Melanoma is the most common type of skin cancer due to a genetic predisposition. In recent years, it has been determined that the number of different types of skin cancer has increased worldwide and caused a large number of deaths. Some skin cancers, such as melanoma and its derivatives, can be prevented, but early and accurate diagnosis is very important for treatment. Image processing techniques in medical applications are frequently used in the diagnosis, follow-up, and treatment processes of skin cancer. However, manual control of medical images is laborious and time-consuming and is vulnerable to expert errors in the interpretation of images. Developing a safe and autonomous classification system for medical applications is a fundamental need. In this study, a CNN-based deep learning framework has been developed in which the HAM10000 dataset, a dermatoscopic clinical skin image collection, has been classified for skin cancer detection. Classification preprocessing using contrast limited adaptive histogram equalization is demonstrated by the accuracy results that improve the recognition of subtle features of class labels. A 45-layer model is proposed for classification. With this developed model, an accuracy rate of 99.69% has been achieved. The results show that the proposed model achieves high accuracies and F-measures with low false-negative compared to known classifiers. This CNN model showed the best two-level performance classifying melanoma and benign cases as nevi and non-nevi. It has emphasized that skin cancer can be detected early with the proposed model and can contribute to the execution of the treatment process.