Using Artificial Intelligence to Diagnose Skin Cancers
Posted: Monday, March 25, 2019
Although cancerous skin lesions can be identified from photos using artificial intelligence (AI), there is still a lot of research to be done before the technology can be implemented in dermatology practices. A scientific session led in part by Roger S. Ho, MD, MPH, FAAD, of NYU Langone Health, at the 2019 American Academy of Dermatology (AAD) Annual Meeting discussed AI and its use in the detection of skin cancers (Abstract D004).
“AI systems for skin cancer detection are still in very early stages. Nothing is 100% clear-cut yet,” said Dr. Ho in an AAD press release.
Several problems with the use of AI were explored during this discussion session. One problem is that the algorithm places scores on the lesions it sees. A dermatologist would then need to interpret the score. These systems also need to be trained to interpret any picture. The machine needs to be shown photos of benign and malignant lesions in many different settings.
“Just because the computer can read these validated data sets with near-100% accuracy doesn’t mean they can read any image,” emphasized Dr. Ho. “Everyone has a different phone, lighting, background.”
Additionally, the training systems so far have used lighter-skinned patients alone, potentially altering the ability of the system to recognize malignancies on darker-skinned patients. To become more accurate, photos shown to the machine must be from all different parts of the body as well as from a variety of patient populations.
AI is not intended to replace board-certified physicians, but the hope is that the technology will support dermatologists in diagnosis. Presenters emphasized that board-certified dermatologists can consider additional factors in their evaluation of a lesion, such as other spots on the body, the individual’s skin type, risk factors, and sun protection habits. Potentially, the use of AI in conjunction with physicians may help to reduce disparities in care and work.
Disclosure: The session participants’ full disclosures can be found at aad.org.
