A new artificial intelligence model may outperform existing methods for identifying skin cancer, a new study shows.
Researchers from Anglia Ruskin University, Check4Cancer, University of Essex and Addenbrooke’s Hospital worked on the AI model which was trained on data from 53,601 skin lesions from 25,105 patients. They used machine learning and combination theory to distil 22 clinical features down to the seven most important that predict if a skin lesion might be suspicious or not. These features include: whether the lesion has recently changed size, color or shape; whether the lesion was pink or inflamed; and hair color at age 15.
Researchers then applied proportional weighting to these seven features to create the new C4C Risk Score which has an accuracy of 69%. In the study it significantly outperformed existing methods such as 7PCL (62%) and Williams score (60%).
Some of the new risk factors they discovered, such as lesion age, pinkness, and hair color, were important for all types of skin cancer but were not included in the older methods, which only focused on melanoma, a specific type of skin cancer.
“This study shows the importance of using clinical data in skin lesion classification, which should help to improve the detection of skin cancer,” says study author Professor Gordon Wishart, Visiting Professor of Cancer Surgery at Anglia Ruskin University and Chief Medical Officer at Check4Cancer, in a news release: “This study shows the importance of using clinical data in skin lesion classification, which should help to improve the detection of skin cancer.”
“Our new AI model, which combines the C4C risk score together with skin lesion images, could lead to a reduction in the need for patient referrals for biopsies, shorter waiting times for skin cancer diagnosis and treatment, and improved outcomes for patients,” adds Consultant Plastic Surgeon Per Hall, who recently retired from Addenbrooke’s Hospital. “The added value that this paper brings is the ability to help identify patients whose skin lesions are suspicious enough to justify onward referral for face-to-face analysis. “It is hoped that regulatory approval for the AI model can be given in 2025.
The study appears in Scientific Reports.
