A new algorithm that classifies skin cancers from photos is described in a paper published online in Nature this week. The study highlights the potential of artificial intelligence to support, simplify and extend the reach of skin disease diagnostics.
Skin cancer, the most common human malignancy, tends to be diagnosed visually, and then confirmed with follow-up biopsies and histological tests. Researchers have tried to develop automated classification systems before, but this has been difficult because skin lesions vary greatly in appearance. Andrew Esteva and colleagues have overcome this problem by developing a deep-learning algorithm that they trained using a dataset of over 129,000 images representing more than 2,000 different skin diseases. They assessed the ability of the system to recognise both the most common and most deadly types of skin cancer - malignant carcinomas and melanomas, respectively - and found that the system performed on par with a group of 21 specialist clinicians.
The authors caution that their system has yet to be validated in a real-world clinical setting, but its potential to affect primary care is extensive. The system could be extended to other areas including ophthalmology, radiology and pathology, and if installed on mobile phones it could offer low-cost, universal access to vital diagnostic care, they conclude.
Please note that there will be an accompanying Nature Video about this research. A preview of this will be available on the Nature press site on Friday.
Genetics: Correcting for genetic associations between alcohol and diseaseNature Communications
Biomedical engineering: Tiny device goes with the (blood) flowNature Communications