Smartphone artificial intelligence tool could support earlier detection of ocular surface malignancies

A mobile phone-based artificial intelligence (AI) tool may help identify people with suspicious ocular surface lesions before they reach specialist care, according to findings from a nonrandomized clinical trial published in JAMA Ophthalmology.¹

Ocular surface malignancies are uncommon but clinically important causes of ocular morbidity and, in some cases, mortality. Malignant conjunctival tumors include ocular surface squamous neoplasia (OSSN), conjunctival melanoma and lymphoma, while pigmented ocular surface lesions may range from benign nevi to primary acquired melanosis and melanoma.^2,3 OSSN is the most common non-pigmented malignancy of the ocular surface and can resemble benign conjunctival or corneal disease, making timely recognition important.⁴ Conjunctival melanoma is rarer but potentially aggressive, with a risk of local recurrence, metastasis and disease-related death.^3,5

What did the study investigate?

The study assessed CaptureTumor, an AI-enabled mobile application developed to support self-screening for pigmented lesions on the ocular surface. The researchers first built the model using slitlamp images collected from multiple centers over a 12-year period, then adapted the system so that it could work with photographs taken on smartphones.

The app was designed to guide users through image capture and then classify lesions according to their likelihood of malignancy. Cases considered more concerning by the system were directed for ophthalmologist review and potential referral. Images uploaded through the app were assessed by ophthalmologists within 24 hours, and patients referred for clinical assessment underwent slitlamp examination. Histopathology was used to confirm the diagnosis in excised lesions.

What do the data show?

Public outreach across television, social media and internet hospital platforms reached 256,053 people. Of these, 614 completed at-home screening using the app. Participants had a median age of 46 years (interquartile range [IQR], 11 years), with an overall age range of 4–87 years.

In prospective smartphone testing, CaptureTumor achieved an area under the receiver operating characteristic curve (AUC) of 0.905 (95% confidence interval [CI], 0.837–0.973) for distinguishing malignant from benign lesions. The slitlamp-based version of the model achieved an AUC of 0.945 (95% CI, 0.918–0.972).

In the population-level screening trial, the app achieved an AUC of 0.977 (95% CI, 0.964–0.990), sensitivity of 89.3% (95% CI, 86.7–91.9) and specificity of 95.9% (95% CI, 94.2–97.6) for malignancy detection. Overall binary accuracy was 95.5% (95% CI, 93.8–97.3).

Among the 614 people screened, 20 were confirmed to have malignant disease after pathological assessment. Most of these cases were newly detected, with 19 of 20 patients (95%) receiving a new diagnosis through the screening pathway. No patient in this group underwent enucleation.

The study also suggested that the app could reduce the number of steps before specialist assessment. The authors reported that most patients referred through the app required one clinical visit before reaching specialist care, compared with a mean of 3.8 previous referrals among historical surgical patients at the study center.

Why is this relevant for ophthalmologists?

For ophthalmologists, the findings are relevant because ocular surface malignancies can be difficult for patients and non-specialists to recognize, particularly when lesions appear subtle or are assumed to be benign. A tool that prompts earlier review of suspicious lesions could help shorten the path to specialist assessment.

The study is also relevant to the wider use of AI in ophthalmology. CaptureTumor was tested not only as an image-classification model, but as part of a screening pathway that included public engagement, image capture, AI assessment, ophthalmologist review and referral.

However, the findings should be interpreted with some caution. The study was conducted in China, and further studies will be needed to determine whether the model performs similarly across other populations, imaging conditions and healthcare systems. The authors also noted that smartphone-based screening may be less accessible to some older adults, and that longer-term outcomes, cost-effectiveness and scalability require further evaluation.

Clinical takeaway

This study suggests that smartphone-based AI screening may help direct people with suspicious ocular surface lesions toward timely ophthalmic assessment. The approach is not a substitute for clinical examination or histopathological confirmation, but it could have value as a triage tool, particularly in settings where access to ocular oncology expertise is limited.

References

1. Wang R, Bi S, Lin D, et al. Smartphone-based proactive self-screening for ocular surface malignancies: a nonrandomized clinical trial. JAMA Ophthalmol. Published online June 4, 2026. doi:10.1001/jamaophthalmol.2026.1609
2. Shields CL, Alset AE, Boal NS, et al. Conjunctival tumors: review of clinical features, risks, biomarkers, and outcomes. Asia Pac J Ophthalmol (Phila). 2017;6:109–20.
3. Oellers P, Karp CL. Management of pigmented conjunctival lesions. Ocul Surf. 2012;10:251–63.
4. Cicinelli MV, Marchese A, Bandello F, Modorati G. Clinical management of ocular surface squamous neoplasia. Ophthalmol Ther. 2018;7:179–94
5. Virgili G, Parravano M, Gatta G, et al. Incidence and survival of patients with conjunctival melanoma in Europe. JAMA Ophthalmol. 2020;138:601–8.