A recent study published in JAMA Ophthalmology, “Smartphone-Based Proactive Self-Screening for Ocular Surface Malignancies: A Nonrandomized Clinical Trial,” explored the impact of smartphone technology to detect signs of ocular cancers.
About the research
Ocular surface malignancies are rare but can threaten both vision and survival. The study authors said these lesions are often mistaken for benign growths, which can delay diagnosis and treatment. In advanced cases, this can lead to more extensive surgery.
In response to those challenges, researchers developed CaptureTumor, a smartphone-based artificial intelligence screening system designed to help members of the public photograph suspected lesions at home. Patients receive an immediate risk assessment and are referred for specialist evaluation when needed.
Methodology
This was a nonrandomized clinical trial conducted across China from December 2022 to June 2023. Researchers first trained and validated a deep learning model using 12 years of multicenter slit lamp images. They then adapted the system for smartphone photography after finding that image quality issues, including poor focus, suboptimal exposure and inadequate lesion framing, reduced performance on phone-captured images.
The final tool was deployed as a WeChat Mini-Program that did not require installation. Public outreach through television, social media, internet hospitals, community posters and seminars directed people to the app. After electronic consent and a brief questionnaire, participants used the app’s guided photography features to capture lesion images. The app then generated immediate benign-or-malignant and multiclass risk assessments.
All uploaded images were reviewed within 24 hours by ophthalmologists at Zhongshan Ophthalmic Center. Patients with concerning findings were contacted and referred for clinical evaluation. Histopathology was used to confirm diagnoses in excised lesions.
Results
Multimedia outreach reached 256,053 people. Of those, 13,243 accessed the screening application and 614 completed at-home self-screening. After image quality review, 535 participants with 805 images were included in the final analysis. The median age was 46, and 49% of participants were female.
In prospective slit lamp testing, the model achieved an AUC of 0.945 for binary classification of malignant vs. benign lesions. When first applied to retrospective smartphone images, performance dropped to an AUC of 0.787. After the smartphone imaging protocol and app-based quality guidance were introduced, prospective smartphone testing reached an AUC of 0.905. The authors said this was comparable to the slit lamp-based model.
In the real-world screening cohort, the app achieved an AUC of 0.977, with sensitivity of 89.3% and specificity of 95.9% for malignancy detection. The system initially flagged 47 participants as high risk. Expert review identified 11 additional suggestive cases, leading to 58 referrals.
Among those referred, 20 malignancies were confirmed by pathology after surgery: 14 basal cell carcinomas and six malignant melanomas. Nineteen of the 20 malignancies were newly diagnosed. The study reported that no patient required enucleation or orbital exenteration.
The researchers also reported improved referral efficiency. Compared with a mean of eight malignancies diagnosed annually per center in conventional slit lamp-based screening in 2021 and 2022, the mobile-based approach was projected to increase annual detection to 40 per center, a fivefold increase. App-referred patients also needed fewer prior referrals before definitive management than historical surgical patients at the study center.
What this means for eyecare providers
For eyecare providers, the study suggests that smartphone-based AI screening may help identify suspicious ocular surface lesions earlier and channel patients to specialty care more directly. The system was designed to support, not replace, clinical evaluation, with ophthalmologists reviewing uploaded images and pathology confirming surgical cases.
The findings may be most relevant for providers who see patients with limited access to subspecialty care or who practice outside major metropolitan areas, since 60% of app users in the final analysis were from non-tier one cities. The study also suggests that image quality guidance is critical. Smartphone performance improved only after the app incorporated instructions on framing, focus and lighting.
At the same time, the authors noted several limitations, including the need for larger multinational studies, questions about how well the model will perform across broader populations and the need for longer-term data on outcomes, cost-effectiveness and scalability.
For more news from AI in Eye Care, click here
Author
