Computer programme beats dermatologist at cancer diagnosis
A computer programme, designed to teach itself, can spot skin cancers as successfully as a panel of trained skin specialists, engineers in the US showed this week.
Skin cancers are one of the world's commonest malignancies with millions of cases diagnosed every year. It's also the case that many of these diagnoses are made "late" because people don't realise that they have a problem until it's too late to seek curative treatment.
The bottleneck is access to the skilled eye of a dermatologist, who can spot problem lesions quickly. So can a computer programme trained to educate itself help to surmount the problem? Stanford engineering PhD student Andre Esteva and his colleagues, writing in this week's edition of Nature, believe so.
They have built machine-learning system based on an architecture developed by Google which can examine images of skin lesions and pick out danger signs that indicate disease.
The machine was trained by feeding in more than 100,000 clinical images for which the diagnosis was known. The computer algorithm automatically adjusted itself as it learned to pick out salient features that were reproducibly present in malignant versus benign skin lesions.
Having trained the programme the Stanford team then pitted the wits of their "AI-dermatologist" against 21 medical doctors using a set of previously unseen medical images.
The diagnoses for these images had been verified by carrying biopsies beforehand. The images that they used consistent of benign and malignant pigmented skin lesions: for instance moles and melanomas. They also looked at seborrhoeic keratoses, which are common in adults. They form benign, dark blobs on the skin and the team challenged their system to tell these apart from skin carcinomas (cancers).
In all cases the machine was as successful, or better, than the panel of human specialists.
According to Esteva, the basis approach could be re-purposed to look at a raft of other conditions too, and also translated to a simple app that could operate on a smart-phone. A user could snap an image of an iffy-looking mole and analyse it themselves, helping them to decide whether to follow it up with a trip to the medical centre for a second (human) opinion or not.
That said, an interaction with a computer is not a replacement for the human touch, cautions Oregon-based dermatologist Sancy Leachman, who points out that a computer programme will only answer the question put to it.
"The mole on your arm might be okay, but what about the one on your back that's cancerous that you've missed? A 'human' doctor would look and probably spot that..."