Forum for Nordic Dermato-Venereology No 3, 2019 Telemedicine | Page 14

Systematic Review Artificial Intelligence in Dermatology – A Systematic Review C arsten S auer M ikkelsen 1 , K ristian B akke A rvesen 2 , P eter B jerring 3 and L uit P enninga 4 Research Lab, Department of Dermatology, University of Aal- borg, and Private Dermatology Practice, Brønderslev, Denmark, 2 Resident in dermato-venereology, Department of Dermato-Venereology, Aarhus University Hospital, Denmark, 3 Professor, dr.med, Department of Dermatology, Aalborg University Hospital, Denmark. 4 Specialist in Surgery and Surgical Gastroenterology, PhD, Ilulissat Hospital, Avannaa Health Region, Greenland. E-mail: [email protected]. 1 A rtificial intelligence (AI) is the science of training a ma- chine or computer to perform human tasks. The term AI was first used in 1955, and AI has recently become increasingly popular (1). Advanced algorithms, sophisticated computers with large power and storage options, and increased data volumes, have contributed to the increase in interest in AI. AI applications can be used in many aspects of society, and are expanding into areas that were previously considered tasks for human experts. AI applications are being developed and used within health- care, and the question arises as to whether AI might gradually change medical practice. A key future problem within global healthcare is the immense predicted shortage of healthcare workers. The World Health Organization (WHO) has estimated a shortfall of almost 13 million healthcare worker worldwide by 2035 (2). Within the field of dermatology a lack of special- ists is already evident in the UK, with only 650 dermatologists for a population of over 66 million (3, 4). Dermatologists often work under pressure, with long waiting times and insufficient time to spend with patients. New technology, and especially AI, might be useful in this field, as diagnosis in dermatology to a great extent depends on the visual recognition of patho- logical structures. We performed a systematic review to identify current appli- cations of AI within dermatology. Current and future appli- cations, benefits and harms are reported here, together with the safety of different applications. M ethods The Cochrane Library, Cochrane Central Register of Controlled Trials (CENTRAL), MEDLINE, Embase, Science Citation Index Expanded were searched until May 2019 using the search terms “artificial intelligence” or “computer-assisted”, “com- puter-aided” and/or “dermatology” or “dermoscopy”. Relevant articles were selected, including randomized controlled trials and review articles. The reference sections of relevant articles were also searched for relevant publications. 98 R esults Relevant publications on applications of AI for diagnosis in dermatology were identified. The selected articles were primarily review articles, Cochrane Reviews, and larger pro- spective studies (5–23). Furthermore, some expert and user opinions were selected (1). The vast majority of publications deal with the diagnosis of malignant melanoma. This is un- derstandable as malignant melanoma contributes to 80% of skin-cancer-related deaths (17). The annual incidence of ma- lignant melanoma has increased dramatically in the last few decades, and the risk of melanoma appears to be increasing in people under the age of 40 years, especially among women (17). Furthermore, early and correct diagnosis of malignant melanoma is important, as it enables treatment by surgical resection, which is more likely to result in cure, whereas more advanced stages of the disease have a worse prognosis (17). A rtificial intelligence - based applications The available AI-based applications identified by the search can be divided into 2 groups: (i) applications with the potential to alert people, through the use of a mobile phone or smart device, when they may need to see a doctor (6, 7); and (ii) ap- plications that help dermatologists to increase the accuracy of diagnosing malignant melanoma (5, 9–16). Most applications in the first group and some in the second group use AI based on fractal analysis and machine learning algorithms (1). These algorithms allow comparison of a stored photograph against numerous photographs of melanoma and benign lesions, or allow comparison of the stored photograph against numerous benign and melanoma lesion characteristics learned from analysing a very high number of photographs, in order to assess the likelihood of melanoma. Fractal analysis is based on a natural phenomenon that exhib- its a repeating pattern at every scale. It can provide a quan- titative measure of irregularity where regularity is expected (18). With regard to melanoma, this includes irregularities in the physical characteristics of a lesion, such as those used in Forum for Nord Derm Ven 2019, Vol. 24, No. 3