AI Driven Edge Device for Screening Skin Lesion and Its Severity in Peripheral Communities

It is vital to treat any skin disorder as early as possible. Neglecting the rudimentary skin disease may lead to acute skin cancer. Skin diseases are mostly neglected in peripheral regions because of a lack of awareness and accessibility to dermatologists. Therefore, it is important to diagnose skin diseases promptly and take countermeasures to treat them effectively. The paper presents a novel method to detect skin disease and its severity using a mobile application. A dataset consisting of four classes- Eczema, Measles, Leprosy, and Healthy Normal Skin is chosen in this proposed work. The images are passed through two modified CNN architectures. The first layer is a modified Mobile Net V2 architecture that aids in predicting the type of skin disease, which is referred to as SkinLesion Net. The next layer that predicts the severity of the disease is named SeverityNet. Observant comparison on four different CNN architectures - VGGl6,Inception V3, Xception and the proposed SkinLesion Net, is performed using this image dataset. Skin Lesion Net outperforms the other networks with 94.32% accuracy, 93.02% F1-Score, 93.53% Precision and 92.76% Recall. The model is lightweight, about 14 MB in size, which is appropriate for deployment into an Android application.