Advancement in deep learning for skin detection: A comprehensive review

Skin cancer is a crucial health issue worldwide, with a high mortality rate for those affected. The detection of skin cancer is now being advanced with technology, specifically through computer vision, enabling machines to see and comprehend visual data like humans by involving image processing and analysis. The main goal of this paper is to analyze the best model trained across 6 benchmark datasets: ISIC, PH2, HAM10000, DermIs, DermNet, and Academic Torrents. This article presents a comprehensive review of recent methods and state-of-the-art (SOTA) approaches for skin cancer classification, including both melanoma and non-melanoma types. The evaluation is based on an analysis of 24 peer-reviewed studies that specifically focus on classification techniques using various deep learning models such as Basic CNN, Inception, ResNet, Inception-ResNet, DenseNet, AlexNet, VGG, and other advance model to determine their effectiveness in detecting skin cancer from image data. Deep CNN (SkinNet) consistently achieved high accuracy across multiple benchmark datasets, reaching up to 0.988. However, more advanced approaches demonstrated even higher performance, with transformer-based models and ensemble methods achieving 0.990 and 0.991 respectively, hybrid models and optimization techniques outperforming with an accuracy of up to 0.998. Deep learning has been proven to be capable of producing significant performance in skin cancer detection based on the data used, the architecture created, and the hyperparameters selected.