SKIN CANCER DETECTION USING KNOWLEDGE DISTILLATION

Abstract

Skin cancer is one of the most lethal forms of cancer, this is because it is among the cancers that are known to be invasive and can easily spread if early treatment is not sought. Photodermatitis worsens the erythema that occurs due to exposure to light and also enhances the rate at which skin cells divide. Therefore, the need to design a technique that will automate the diagnosis of skin lesion is essential for facilitation of diagnosis thus saving on time and lives. This research suggests the development of a skin cancer classification system that is automatic in nature and is aimed to differentiate seven types of skin cancer with melanocytic nevus and angioma. The implementation uses what is known as Knowledge Distillation, a method that aspires to enhance the current form of detection by transferring knowledge from a big and complicated model (technically termed as ‘teacher’) to a new model that is much simpler and effective, or what is famously referred to as ‘student’. The dataset comprises images of nine different skin conditions: Actinic keratosis, basal cell carcinoma, benign keratosis, dermatofibroma, melanoma of the skin, new, seborrheic keratosis, squamous cell carcinoma, and vascular lesion. The proposed method works; the feature extraction and classification of Knowledge Distillation ensure that it delivers reasonably good results. These findings show that the implementation of CNN-based automated systems for image analysis in the diagnosis of skin cancer enables physicians to diagnose the cancer in its preliminary stages from medical images, which leads to higher rates of patient survival and higher effectiveness of usable treatment techniques.