If we want to train algorithms to detect skin cancer we need more images (especially of darker skin)

If we want to train algorithms to detect skin cancer we need more images (especially of darker skin)

According to new research presented at the NCRI Festival and published in Lancet Digital Health , the images and attachments available to train artificial intelligence (AI) to detect skin cancer are insufficient and include very few images of darker skin .

The more examples the better

AI is increasingly used in medicine as it can make the diagnosis of diseases such as skin cancer faster and more effective. However, AI must be "trained" by looking at data and images from a large number of patients in whom the diagnosis has already been established, so an AI program is highly dependent on the information it is trained on.

That is why David Wen of the University of Oxford, and his colleagues carried out the first review of all freely available data sets on skin lesions worldwide. They found 21 sets that included more than 100,000 photographs .

Diagnosing skin cancer typically requires a photo of the worrisome lesion, as well as a photo taken with a special hand-held magnifying glass, called a dermatoscope, but only two of the 21 data sets included images taken with both methods. The data sets were also missing other important information: how the images were chosen for inclusion and evidence of ethical approval or patient consent.

Fourteen of the 21 data sets provided information on which country they came from, and of these, nine contained images of European countries. Only a small percentage of images were accompanied by information on the skin color or ethnic origin of the patients . Among the images in which skin color was indicated (2,436 images), only ten were brown skin and only one was brown or black skin. Among the images in which ethnic origin was indicated (1,585 images), none were of people of African, Afro-Caribbean or South Asian origin:

This has implications for the programs developed from these images, due to the uncertainty about how they may work in different groups of people, especially those who are not well represented in the data sets, such as those with darker skin. This can potentially lead to the exclusion or even harm of these groups due to AI technologies. Although skin cancer is rarer in people with darker skin, there is evidence that those who develop it may have a worse disease or are more likely to die from the disease.