Comparative Study of Multiple CNN Models for Classification of 23 Skin Diseases

Abstract

Cutaneous disorders are one of the most common burdens world-wide, that affects 30% to 70% of individuals. Despite its prevalence, skin disease diagnosis is highly difficult due to several influencing visual clues, such as the complexities of skin texture, the location of the lesion, and presence of hair. Over 1500 identified skin disorders, ranging from infectious disorders and benign tumors to severe inflammatory diseases and malignant tumors, that often have a major effect on the quality of life. In this paper, several deep CNN architectures are proposed, exploring the potential of Deep Learning trained on “DermNet” dataset for the diagnosis of 23 type of skin diseases. These architectures are compared in order to choose the most performed one. Our approach shows that DenseNet was the most performed one for the skin disease classification using DermNet Dataset with a Top-1 accuracy of 68.97% and Top-5 accuracy of 89.05%.