Context

Skin cancer is the most common human malignancy and its incidence has been increasing in the last decade. Within the general category of skin cancers, melanoma constitutes the main cause of death. According to the latest statistics, cutaneous melanoma is currently the sixth most common type of cancer in Europe, with more than 144,000 new cases diagnosed in 2018.

Fortunately, melanoma may be cured if treated at an early stage. Mortality increases with increasing growth into the skin. More than 90% of melanoma patients are still alive after 5 years, if treated early. If distant spread of cancer cells has occurred (metastatic melanoma), the proportion of patients alive after 5 years may be 23% or lower. For these reasons, rapid diagnosis is essential to ensure treatment is undertaken before local and metastatic spreading occur.

The treatment of melanoma is based on surgical removal (excision) of the primary skin lesion. From a technical viewpoint, the excision of skin lesions suggestive of melanoma is fairly trivial in the majority of cases. While excision of melanoma is quite simple, early-stage melanoma detection is not easy even for expert dermatologists because it often resembles a common skin lesion (“mole”). In high-risk patients with many atypical moles (dysplastic naevi), a high number of moles need to be excised for one melanoma to be detected. Thus, the risk of missing a melanoma remains significant even with a large number of excisions of benign lesions in every patient.

The proliferation of hand-held dermoscopes has remarkably improved the diagnostic accuracy for melanoma when used by dermatologists with specialized training. Concurrently, Artificial Intelligence (AI) systems for identification of melanomas have seen tremendous growth in the last 3 years, driven by the availability of massive new datasets, with deep learning (DL) systems achieving expert-level classification accuracy. However, most studies comparing human to AI performance present a key constraint: they attempt to differentiate skin lesions by using just the images at hand, without any clinical context, and in this situation, DL has shown to outperform the average dermatologist. When experienced dermatologists have access to this clinical data, their performance notably improves. But even with that evidence, most DL systems are still relying on the imagery only, ignoring the complementary clinical data that is available.