Computers can classify skin cancers as successfully as human experts, according to the latest research attempting to apply AI to health (The Guardian, 2017).

Researchers at Stanford University said the new system, which is based on image recognition, could be developed for smartphones, increasing access to screening and providing a low-cost way to check whether skin lesions are cause for concern. Andre Esteva, an electrical engineering PhD student from Stanford University and co-author of the research, said “We hope that this is a first step towards early detection.”

According to the World Health Organisation, skin cancer accounts for one in every three cancers diagnosed worldwide, with global incidence on the rise. In the UK alone, 131,772 cases of non-melanoma skin cancer were recorded in 2014. In the same year there were 15,419 new cases of the deadliest skin cancer, melanoma, making it the fifth most common cancer, according to Cancer Research UK.

As the disease is often initially spotted by a visual examination, Andre Esteva teamed up with colleagues in fields ranging from dermatology to artificial intelligence to create a computer system that would aid screening. Their approach, described in Nature, is based on deep learning. When fed with a large set of ready-sorted data these algorithms pick out and “learn” patterns and relationships. Once trained, the algorithms can then be used to categorise new, unsorted data.

To create the system, the team harnessed a deep learning algorithm built by Google that had already been presented with 1.28 million images of objects such as cats, dogs and cups. Andre Esteva and his colleagues then fed the system more than 127,000 clinical images of skin lesions, each already labelled, encompassing many different skin diseases. Once trained, the team then tested the system’s ability to classify skin cancer by presenting it with just under 2,000 previously unseen images of skin lesions, whose nature had previously been determined by biopsy, and further compared the results for nearly 400 of the images against the judgement of at least 21 dermatologists.

The results reveal that the system is on a par with, if not better than, the experts in telling apart carcinomas from common benign skin growths and melanomas from moles. For melanomas, the average dermatologist classified around 95% of malignant lesions and 76% of harmless moles correctly. By comparison, the algorithm is capable of correctly classifying 96% of malignant lesions, and correspondingly 90% of benign lesions. Andre Esteva said “The aim is absolutely not to replace doctors nor to replace diagnosis. What we are replicating [is] sort of the first two initial screenings that a dermatologist might perform.”

While Andre Esteva and colleagues admit the system needs further testing in clinical settings they believe the approach has great promise, suggesting it could be applied to a host of other medical fields.

Boguslaw Obara, a computer scientist at Durham University and expert in image processing, said that the size and complexity of the dataset used to train the system was impressive. The work, he added, shows it is likely that algorithms will start cropping up more and more in everyday life.

Dr Anjali Mahto, consultant dermatologist and spokesperson for the British Skin Foundation also welcomed the research. She said “This is an exciting new technology that has the potential to increase access to dermatology at a time where there is a national shortage in this speciality and the rates of skin cancer continue to rise.” But she warned the system will need to be carefully assessed for its benefits before it can be rolled out. She added the approach is also unlikely to replace the role of dermatologists, pointing out that during a full-body examination experts often discover skin cancer at different sites to those that initially concerned the patient. She said “There is therefore a possibility that if you rely on people to self-report what they are worried about, other skin cancers – particularly in hard to see sites, e.g. the back – may be missed.”