State-of-the-art skin disease classification: a review of deep learning models

Skin disease classification and detection have gained much research attention over the years, considering that skin disease, a prevalent medical concern due to the vulnerability of our body's outermost layers, can become life-threatening. Hence, timely detection of skin diseases is vital, as it can prevent them from progressing and becoming life-threatening. Though the research community has covered quite a number of skin diseases, little is known about how accurately deep-learning models have performed in this domain. We present a systematic review of articles covering the state-of-the-art application of deep learning models in skin disease classification. We explored articles published between 2019 and 2023 to uncover the trends, performance of deep learning models, and limitations to inform future work in this domain. In view of this, we collected 6934 articles from ScienceDirect, IEEE, PubMed, Scopus, and other databases. Results from our review of 63 skin diseases collected from these articles show that deep learning models, on average, have attained 86.20% accuracy predictions. In addition, deep learning models have shown significant sensitivity and specificity values over 90%. Nevertheless, we found some limitations with studies employing deep learning models, including non-generalizability of models developed and bias towards one skin disease compared to the other and other related limitations. Overall, we present recommendations for improving on these limitations in future work, including an improved design, implementation, and testing of skin disease applications in a real-world setting.