Skin cancer: an AI that reads a photo like a dermatologist

Why is skin cancer hard to spot?

Skin cancer is the most common cancer in humans. It is first diagnosed by eye: an initial clinical check, sometimes a look through a special magnifier called dermoscopy, then a biopsy (a small piece of skin is removed) and finally an examination under the microscope.

The difficulty is that skin lesions often look very alike from one disease to another. A harmless spot and a melanoma, the most dangerous form of skin cancer, can look close to an untrained eye. Sorting these images automatically was seen as a very hard task for a computer.

This is why screening depends heavily on the doctor's experience and on access to a specialist. In many places, seeing a dermatologist takes time, yet spotting a dangerous lesion early is exactly what matters.

How did the researchers train the AI?

The team used a convolutional neural network: a program that learns to spot patterns in images, a bit like learning to recognise faces by seeing many of them. It was given only the pixels of the photos and the name of the matching disease, with no hand-written rules.

To train it, they gathered 129,450 clinical images covering 2,032 different diseases, a database far larger than any used before. They then tested the program against 21 board-certified dermatologists, using images whose diagnosis had been confirmed by biopsy.

What did the AI manage to do?

Two situations were tested. The first was telling keratinocyte carcinomas (the most common skin cancers) apart from seborrhoeic keratoses, which are harmless. The second was telling malignant melanomas apart from harmless moles, which is about the deadliest skin cancer.

In both cases, the neural network performed on par with all the experts tested. The authors conclude that this artificial intelligence can classify skin cancer with a level of skill comparable to dermatologists.

The authors also point out that the program learned from pixels and disease labels only, with no hand-made rules about what a melanoma looks like. In other words, it worked out the useful visual clues by itself from a very large number of examples.

What could this change for patients?

The authors see this result as a way to extend access to dermatologists beyond the clinic. They point out that smartphones are widespread and that, fitted with this kind of program, they could offer low-cost, broad access to important diagnostic care.

This does not mean a photo sent from home replaces a consultation. For now, it is a serious demonstration that a computer can learn to recognise these lesions, and a reason to keep researching.

What are the limits of this study?

The test used selected images with a diagnosis already confirmed by biopsy, and it asked two precise questions. That is not the same as real life, where photos vary in quality and the doctor has plenty of other information, such as the patient's history.

The study does not show that patients do better thanks to the tool. That would need trials with real patients followed over time. Diagnosis and treatment decisions remain for doctors to make.