Deep learning for effective triaging of skin disease in the NHS

Resource within primary and secondary NHS care (dermatology and plastic surgery) cannot meet the growing demand from increasing numbers of referrals for lesions suspicion of skin cancer, estimated at 100,000 p.a. for NHS Scotland.

Although skin cancer prevalence has increased by 5% per annum for the last quarter century, most referrals requesting a specialist opinion are benign lesions.

Image-based AI assisting human experts has huge potential to support effective diagnostic triage by GPs, dramatically reducing referrals to secondary care whilst delivering safe care nearer to home and rapidly alleviating patient anxiety.

To develop an AI system operating robustly and reliably within NHS referral structures, it is essential to train the system using images representative of those produced in primary care or by patients.

Our pilot project developed software able to distinguish benign skin lesions from common skin cancers using preselected and pre-processed image datasets, with clear and limited categories, but these are not truly representative of the distribution of NHS clinical data. We propose to develop a tool of high diagnostic accuracy within a real-world NHS setting.

Our primary aim, preventing unnecessary referral of benign skin lesions, requires a robust system with near-zero false negative rate.

This 12-month project will prepare clinical image datasets and use them to develop and validate a prototype designed for integration into existing clinical workflow.

This will incorporate deep learning ensembles optimised for high sensitivity with mechanisms to reject low confidence images and provide visual explanations.

We will disseminate technology development and pave the way for subsequent development and integration within the whole NHS, allowing more patients immediate reassurance, and reducing attendance at GPs where one fifth of all appointments are for skin disease.

When achieved, this is predicted to reduce dermatology and plastic surgery workloads by 20%, saving £10M per annum for NHS Scotland.