Angus Bezzina
Angus Bezzina is a writer from Sydney, Australia.
Harmless mole or potential melanoma? A new computer algorithm can make that diagnosis from photos as accurately accredited dermatologists.
The system, developed by Andre Esteva from Stanford University in the US and his colleagues and published in Nature, has the potential to support, simplify and extend the reach of skin disease diagnostics beyond the confines of hospitals and medical clinics.
This is particularly important for more lethal forms of skin cancer such as melanoma, in which early detection is critical to survival rate. The estimated five-year survival rate for melanoma patients drops from more than 99% when detected in its earliest stages to about 14% if detected in its latest stages.
But while less threatening forms of skin cancer such as basal cell carcinoma and squamous cell carcinoma are relatively easy to visually identify, it is still difficult to pick out melanomas this way.
Pablo Fernandez-Peñas, a dermatologist at Westmead Hospital in Sydney, Australia, explains that this is because a number of variables go into melanoma diagnosis.
{%recommended 788%}
“For [the diagnosis of] melanoma we have to go through the ‘A, B, C, D, E’ criteria, which is very important.
“This includes whether the lesion is asymmetrical, has irregular borders, displays multiple colours, whether the diameter of the lesion is greater than six millimetres and if there has been any evolution”, such as changing colour, shape or size.
Of course, Fernandez-Peñas explains, even if skin cancer is suspected, a biopsy and histological tests must still be undertaken to confirm a diagnosis.
To help identify the more serious cases of skin cancer that require further testing from those that do not, scientists have tried to create automated classification systems.
But due to the immense variation in skin lesion appearance, developing a system that can determine whether one is cancerous or not has proved elusive.
To overcome this issue, Esteva and his team developed a convolutional neural network – a type of powerful machine-learning algorithm associated with artificial intelligence – that they trained by presenting it with a set of 129,450 images of 2,032 different skin diseases.
When they tested the algorithm by showing it biopsy-proved images of the most common and most lethal forms of skin cancer – malignant carcinomas and melanomas respectively – the researchers found its performance was on par with that of 21 certified dermatologists in the US.
While the algorithm still needs to be validated in a real-world clinical setting, Esteva and his team suggest the neural network could be turned into an app for mobile devices and provide low-cost and accessible primary diagnostic care.
Similar systems could conceivably be adopted for other fields such as ophthalmology, radiology and pathology.
For more, check out the video below.
