LIVERPOOL — Artificial intelligence (AI) identified squamous cell carcinoma (SCC) and basal cell carcinoma (BCC) with very high sensitivity and specificity using the Deep Ensemble for the Recognition of Malignancy (DERM) image analysing algorithm (Skin Analytics), showed data from the DERM-003 phase 3 study.
Specificity results also indicated that the AI-Digital Health Technology (DHT) could aid the management of benign lesions, while DERM's ability to correctly identify melanomas was high, and pre-malignant and benign lesions was good, said Stephanie Austin, MBBS, dermatologist from University Hospitals Dorset NHS Trust in Poole, as she reported study results at this year's British Association of Dermatologists 103rd annual meeting.
"With threshold settings applied [for optimisation to primary or secondary care use] the sensitivities and specificities are at a level that are clinically useful," said Dr Austin, adding that "DERM generalises well across different phone cameras, but specific threshold settings to optimise sensitivity may be needed for each camera."
Shortage of Dermatologists
The registrar added that AI tech might benefit dermatologists' workflow. "Overall, AI is getting to where it wants to be, but this particular platform is the best I've seen in dermatology," she said, "It frees me up to do an extra paediatric clinic or skin surgery, for example, that we don't have time to do otherwise."
Dr Austin explained why AI image analysing algorithms were in great need in today's NHS: "There's a huge shortage of dermatologists, yet referrals for suspected skin cancers have doubled to over 700 per 100,000 of the general population in recent years compared with a decade ago. This, together with an ageing population who love the sun is driving demand for dermatologist time," she remarked.
AI-development firm, Skin Analytics, have created an algorithm and device that enables the photography and analysis of skin lesions such that it recognises the most common malignant, premalignant, and benign skin lesions. The DERM algorithm provides the user with a probability estimate that a named lesion type will be confirmed by a dermatologist.
The DERM-003 study aimed to establish the effectiveness of DERM AI technology to identify melanoma, SCC, and BCC based on dermoscopic images of skin lesions, with a primary endpoint of the area under the receiver operating characteristic curve (AUROC), which effectively represents the extent to which the AI algorithm identifies the lesion correctly as matched by clinical and histopathological reading.
A total of 572 patients were recruited across four UK NHS sites. Mean age was 68 years, half were female, and 94% White with Fitzpatrick skin type II (57%). Patients with at least one suspicious skin lesion that was suitable for photography were eligible. Each lesion was photographed with three smartphone cameras (iPhone 6S, iPhone 11, and Samsung 10) with a dermoscopic lens attached. Each image was assessed by the AI-DHT and compared with histopathology-confirmed diagnosis if a biopsy was available, or clinical diagnosis otherwise.
In total, 592 lesions had images from all three cameras available; only one lesion had no images available. Altogether, 395 lesions had a histopathology result, enabling comparisons between 394 iPhone 6S images, 394 Samsung 10 images, and 389 iPhone 11 images.
The majority of lesions were on the head and scalp, followed by back, arms and legs; mean lesion size was 8.6 mm and final diagnosis was BCC (176), SCC (44), benign lesions (297). Clinical assessment of skin cancer resulted in unlikely (216), likely (203), highly likely (109), and equivocal (59).
The AUROCs for images taken by the iPhone 6S were 0.88 (95% CI 0.83 to 0.94) for SCC and 0.89 (95% CI 0.86 to 0.92) for BCC. For Samsung 10, the AUROCs were 0.85 (95% CI 0.79 to 0.99) and 0.87 (95% CI 0.83 to 0.90); and for iPhone 11, they were 0.89 (95% CI 0.84 to 0.93) and 0.89 (95% CI 0.86 to 0.93) for SCC and BCC, respectively. By comparison, an experienced dermatologist would achieve an AUROC of around 0.85, according to Dr Austin.
"The AI algorithm works comparably across all three smart phones despite the fact they all create slightly different images, for example lighting might vary," said Dr Austin. "DERM works by comparing our input clinical images to its own huge database of lesion images, eliminating some images, while matching to others."
Using images taken on iPhone 6S of biopsied-only lesions, the sensitivity rates in detecting SCC and BCC were 98% (88, 100) and 94% (90, 97), respectively; the specificity rates were 38% (33, 44) and 28% (21, 35), respectively.
All 12 lesions diagnosed as Bowen's disease were classified correctly by the AI-DHT. Of the 61 lesions diagnosed as actinic keratosis, 52 were correctly classified with AUROCs between 79 and 89; and of the 22 lesions diagnosed as dysplastic, 18 were correctly classified by the AI-DHT with AUROCs between 77 and 94. All 16 lesions diagnosed as melanoma were classified as such by the AI-DHT with AUROCs between 92 and 97.
AI Results Comparable to Clinical Practice
The AI-DHT has the potential to support the diagnosis of non-melanoma skin cancer and premalignant lesions. "These results are comparable to what we would see in clinical practice, so the results suggest that we could use this AI alongside our clinical practice to help us manage our workload," said Dr Austin.
"I see so many referrals but very few are malignant. If we use this technology, clinicians could reassure patients, and give them an information leaflet before even coming to hospital. Some [dermatological AI technology] will be used by GPs and some by secondary care."
David De Berker, MBBS, dermatologist at the University Hospitals Bristol and Weston NHS Foundation Trust co-moderated the session. He explained that dermatology was in an education phase with AI currently. "Right now, AI is all being tested in secondary care because the images need to be standardised for the AI systems to learn what is being viewed accurately. Some patient-captured images might be blurred and at funny angles which are not good for educating the AI."
"Also, in the UK, dermatology is going through the process of developing confidence in using the technology to do things we would previously have done ourselves," Dr De Berker pointed out.
"Hospital dermatologists hold the ambition that AI will be used in primary care, so it effectively becomes a filter in primary care, meaning that only patients most likely to need our help are referred. Currently about 50% of patients sent on the 2-week wait pathways don't have anything worrying at all. They don't need the anxiety and, respectfully, we don't need to spend our time on them."
Dr De Berker has no relevant personal financial disclosures. His university runs a project with Skin Analytics. Dr Austin received travel expenses to attend BAD 2023.