Explainable AI for Skin Lesion Detection in a Clinical Environment

Project Details

Description

A quarter of the population in England and Wales consults their GP annually for dermatological conditions, overwhelming dermatologists with referrals. Yet, only about 6% of these referrals lead to a skin cancer diagnosis, highlighting inefficiencies in the referral process.

Our team developed a novel AI algorithm using anonymised NHS data curated by the Dermatology Department at the University Hospital of Wales, Cardiff. This algorithm identifies clinically significant features of skin lesions—such as shape, colour, asymmetry, border irregularity, and dermoscopic structures—and combines them with individual risk factors in a probabilistic model to predict skin cancer likelihood.

Ongoing work focuses on enhancing the interpretability of this algorithm and refining the probabilistic model to ensure clinical reliability. Designed to support, not replace, human expertise, this tool could reduce unnecessary referrals, improve service capacity, and alleviate NHS backlogs following extensive clinical validation, offering significant benefits for primary care dermatological diagnostics.
StatusActive
Effective start/end date1/06/2231/12/28