Using AI to improve skin cancer detection in primary care: the vision and barriers

Richard Armitage is a GP and Honorary Assistant Professor at the University of Nottingham’s Academic Unit of Population and Lifespan Sciences. He is on X: @drricharmitage

The cancer site with the highest rate of Urgent Suspected Cancer (USC) referral in England in 2022/23 was skin (634,970 total referrals, or 1,345.2 referrals per 100,000 people). In the same year, 6.0% of all USC referrals (all cancer sites) resulted in a diagnosis of cancer, including 6.2% of skin USC referrals. Accordingly, 595,759 benign lesions were referred along the skin USC pathway in 2022/23.¹

The potential for AI in primary care – specifically deep learning-based lesion classifiers – to increase the early detection of skin cancer, improve the accuracy of its diagnosis, and reduce the rate of benign lesion USC referral is currently being explored. To date, most AI technologies to detect skin cancer constitute patient-facing smartphone apps used to triage images of concerning skin lesions in secondary care,^2,3 and the NHS is considering the deployment of these tools to triage skin USC referrals and discharge clearly benign lesions⁴ (evidence showing the value of benign lesion assessment is pending).⁵ However, primary care has been identified as the stage in the skin cancer pathway with the greatest potential for the use of AI to increase early detection.⁶

It has been suggested that AI-driven binary benign/malignant lesion image classifier technologies could be used by GPs as decision support tools to assess and triage concerning skin lesions and to inform USC referral decision-making in primary care.⁶ Such patient-facing AI technologies, which could be contained within a simple smartphone app, would have to be developed to analyse macroscopic images (in addition to dermatoscopic images, which are not ubiquitously available across primary care), have a very high sensitivity to avoid delaying diagnosis, and have a very high specificity to prevent substantial increases in skin USC referral rates.

The placement of such AI tools in primary care could bring about improved rates of early diagnosis and fewer benign lesion USC referrals. However, various barriers must first be overcome for this vision to be realised.

Image quality

Images submitted by patients for remote GP consultations are often of low quality (such as being out of focus, poorly lit, or the lesion itself comprising a small proportion of the image), which will impact the AI’s performance. Additionally, image classifiers that analyse both macroscopic and dermatoscopic images are thought to be most promising.⁶ While GPs could generate macroscopic images during in-person consultations, not all GPs have the relevant skills to produce dermatoscopic images. Accordingly, patient-facing AI technologies would have to be developed to analyse macroscopic images alone,⁶ or additional equipment would have to be purchased by, and further training conducted in, primary care to generate the necessary data for an AI tool to analyse.

Non-visual data

Image classifiers cannot consider other relevant factors including changes in lesion characteristics (such as size, shape and colour), sensation changes, oozing, UV exposure, and family history. This might impact the accuracy of these technologies, thereby causing false positives (which induce patient distress and unnecessary USC referrals) and false negatives (which delay diagnoses). Close GP oversight – the general practice expression of the increasingly popular human-in-the-loop approach to clinical AI development⁸ – of these technologies would therefore be required, which might ultimately increase rather than decrease GP workload.

Data privacy and security

Images taken in primary care must be stored securely (for example, they must not be taken using an AI-enabled app on the GP’s personal smartphone), meaning additional equipment and digital storage may need to be purchased. Furthermore, images must not be used for secondary purposes (such as AI training) without the patient’s informed consent, which must be discussed with the patient and documented accordingly, thereby prolonging the consultation when such images are taken.

Patient demand

The availability and apparent usability of such AI technologies in primary care might substantially increase patient demand to “have their moles checked” in the absence of concerning symptoms, including as secondary issues during consultations for unrelated problems. This would raise GP workload and the rate of false positives, the latter of which would both increase USC referrals and the consequent patient distress.

GP trust and acceptability

GPs must be assured of these technologies’ accuracy (which requires continuous performance monitoring and evaluation), their congruence with clinical workflows, and the clinical accountability of their use (is the technology developer, the NHS, or the individual GP accountable for AI-assisted clinical decision-making?).⁸ Sufficient explainability of the technology’s outputs is also required to foster GP trust. Multiclass lesion image classifiers, which identify various kinds of skin lesion rather than simple binary benign/malignant outputs, would generate greater GP trust through increased explainability. A human-in-the-loop approach, rather than complete delegation of decision-making to the technology, would also be needed to promote GP trust.

Patient consent and maintaining workforce skills

Patients must be fully aware of and consent to the use of these technologies in clinical decisions regarding their care, and alternative processes that do not use AI must be available for those who do not provide this consent. This means that GPs must continue to be sufficiently trained to recognise suspicious skin lesions and to not defer entirely to AI technologies.

For these barriers to be addressed, prospective clinical trials in primary care, and qualitative studies of patient and GP sentiment regarding the use of these technologies in this setting, are required before they can be deployed, accepted, and scaled in primary care.

References

1. NHS Digital. Urgent suspected cancer referrals. https://digital.nhs.uk/ndrs/data/data-outputs/cancer-data-hub/urgent-suspected-cancer-referrals [accessed 15 December 2024]
2. D Wen, A Coombe, M Charalambides, et al. DS03: A scoping review of commercially available artificial intelligence-based technologies for diagnosis or risk assessment of skin cancer. British Journal of Dermatology July 2022; 187(S1): 159. DOI: 10.1111/bjd.21446
3. OT Jones, RN Matin, M van der Schaar, et al. Artificial intelligence and machine learning algorithms for early detection of skin cancer in community and primary care settings: a systematic review. The Lancet Digital Health 2022; 4: e466-e476
4. National Institute for Health and Care Excellence. DAP76: Artificial Intelligence technologies for assessing skin lesions: early value assessment. July 2024. https://www.nice.org.uk/guidance/gid-hte10047/documents/committee-papers-2 [accessed 15 December 2024]
5. National Institute for Health and Care Excellence. Artificial Intelligence technologies for assessing skin lesions selected for referral on the urgent suspected cancer pathway to detect benign lesions and reduce secondary care specialist appointments: early value assessment. https://www.nice.org.uk/guidance/indevelopment/gid-hte10047 [accessed 14 December 2024]
6. OT Jones, RN Matin, and FM Walter. Using artificial intelligence technologies to improve skin cancer detection in primary care. The Lancet Digital Health 10 December 2024. DOI: 10.1016/S2589-7500(24)00216-4
7. S Bakken. AI in health: keeping the human in the loop. Journal of the American Medical Informatics Association 2023; 30(7): 1225-1226
8. Papanikitas A, Rees S. Automation, machine learning, and AI: considerations for commissioners, providers, and recipients of health care. Br J Gen Pract. 2024 Dec 26;75(750):28-29. doi: 10.3399/bjgp25X740325. PMID: 39725529; PMCID: PMC11684434. https://bjgp.org/content/75/750/28.long