Artificial Intelligence for skin image analysis using Total Body Photography

In the last decade the application of artificial intelligence (AI) algorithms in dermatology to classify skin lesions, particularly melanoma, has advanced rapidly. Large international computer skin image analysis challenges have successfully drawn attention to the potential for AI to aid the detection of skin cancers.

A landmark study, reported by Esteva et al1 in 2017, compared the accuracy of their algorithm against 16 expert Dermatologists. The algorithm outperformed the average dermatologist score for detecting both melanomas and keratinocyte carcinomas. Since then, there have been several other reports of algorithms that have outperformed dermatologists in reader studies.

While reported algorithm accuracy and performance are promising, they do not represent a real-world clinical setting, where examination and diagnosis involve context beyond the pixels of a single lesion image. Furthermore, the image datasets generally used for machine learning (ML) can lack generalisability to the day-to-day skin lesions examined by dermatologists, as image archives often capture skin lesions that were deemed ‘interesting enough’ to warrant a picture. In a clinical setting, a dermatologist would consider several factors relating to the patient’s clinical and family history, photodamage of the skin, naevi characteristics of the whole patient, age, and potentially genetic information when available.

The growing use of Total Body Photography (TBP) systems represents both challenges and opportunities to standardise both the image acquisition and labelling of clinical metadata for ML. The opportunity to train algorithms to not just consider the image pixels of a single lesion, but also phenotype information of the whole patient, along with medical records and genetic information will produce the next generation of AI algorithms for dermatology.

With funding from the European Union 2020 Horizons programme, the iToBoS consortium are developing an Intelligent Total Body Scanner with integrated Computer Aided Diagnostic (CAD) tools. To achieve this objective, a multi-site clinical trial for data acquisition is taking place in both Barcelona, Spain, and Brisbane, Australia. This clinical trial will collect TBP image data, patient health information, and genetic risk scores to create an image training dataset with labelled metadata, to develop algorithms that consider the whole patient to better reflect a dermatologist’s diagnosis.

In recent years, it has been a common research practice to organize international competitions or challenges in which the algorithms of different researchers can be benchmarked on publicly released datasets. Over the period of the iToBoS project, the wider consortium will organise two competitive challenges where world-leading groups can participate in solving new problems on: 1) lesion detection and boundary segmentation in regional body images, and 2) on lesion classification. These challenges will facilitate advancements in the development of AI and CAD tools and contribute to the knowledge dissemination in the field.



  1. Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, Thrun S. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017 Feb 2;542(7639):115-118. doi: 10.1038/nature21056. Epub 2017 Jan 25. Erratum in: Nature. 2017 Jun 28;546(7660):686. PMID: 28117445; PMCID: PMC8382232.