Deep Learning for Two-Step Classification of Malignant Pigmented Skin Lesions

Abstract

Skin cancer is one of the most common types of cancer. Its early detection drastically improves outcomes and saves human lives. Well known skin cancer types are melanoma, basal cell carcinoma and squamous cell carcinoma. Melanoma is melanocytic malignant while basal cell carcinoma and squamous cell carcinoma are non-melanocytic malignant. Even though the diagnosis of these cancer types is done by a skin biopsy, automatic detection of skin cancer using computerized methods may lead to a faster and a more accurate diagnosis. The majority of automated skin cancer detection methods proposed by researchers so far concentrated only on melanocytic malignant type melanoma. Non-melanocytic malignant skin lesions could not be investigated in detail due to the lack of available datasets with different lesion classes. In this paper, an automatic detection of malignant pigmented skin lesions is investigated. For this, the two-step skin lesion diagnostic procedure of the dermatologists is followed. Using a deep learning model, the skin lesion is first classified as melanocytic or non-melanocytic and then malignant types are detected using other deep learning models. The performance evaluations show that melanocytic and non-melanocytic skin lesions are detected with the highest accuracy. They also show that melanocytic malignant skin lesions can be classified with a higher accuracy than non-melanocytic malignant skin lesions.