CNN Approaches for Early Detection and Classification of Skin Diseases: A Systematic Review

Abstract

The introduction of Convolutional Neural Networks (CNNs) has altered the way skin disease detection and classification are carried out through the analysis of dermatological images. Surveying the state of the art for CNN driven systems in dermatology is a necessity for method variety, data set dependency, and/or use in practice. Here we survey the current state of the art with a focus on these four aspects. Transfer learning, and hybrid CNN ViTs outperform networks trained from scratch. HAM10000 and ISIC benchmark networks see significant improvement. Mobile and teledermatology see the use of lightweight networks like MobileNet. In hospital settings, deeper networks see more use. Issues remain with the accessibility of data sets, especially concerning skin of color. CNNs lack interpretability and have high computational costs. Ethics and bias in AI pose additional concerns. The need for diverse data sets and explainability frameworks is crucial. Federated learning, multimodal systems, and mobile optimized CNNs provide new and exciting opportunities. This survey highlights the successful areas of cross-disciplinary work. There are so many possibilities that have yet to see clinical implementation.