A smartphone-based artificial intelligence (AI) platform could offer a new route to earlier detection of rare but potentially sight- and life-threatening ocular surface malignancies, according to a nationwide clinical trial conducted in China.
Published in JAMA Ophthalmology, the researchers developed "CaptureTumor" (CaT), an AI-powered screening system that combines smartphone photography, automated image analysis, public health outreach, and specialist referral pathways. The goal was to address a longstanding challenge in ocular oncology: the late diagnosis of ocular surface tumors, which are frequently mistaken for benign lesions and often present only after significant disease progression.
In their study, the investigators trained the deep-learning system using more than a decade of slit-lamp photographs collected from multiple ophthalmic centers before adapting the algorithm for smartphone-based image capture. The final system was deployed as a WeChat Mini Program, enabling users to photograph suspicious pigmented ocular lesions at home and receive an immediate risk assessment.
To support image quality, the app incorporated real-time guidance on lighting, focus, and framing. Users were then classified into benign or malignant categories, with high-risk cases referred for specialist evaluation.
The public-facing screening campaign reached more than 256,000 people through television, social media, internet hospitals, and community outreach. Of these, 13,243 individuals accessed the application and 614 completed the self-screening process. After quality control filtering, 535 participants were included in the final analysis.
The smartphone-based system achieved an area under the receiver operating characteristic curve (AUC) of 0.905 during prospective smartphone testing, approaching the performance of the original slit-lamp-based model (AUC 0.945). In the real-world screening cohort, performance improved further, with an AUC of 0.977, sensitivity of 89.3 percent, and specificity of 95.9%.
During the real-world screening, the platform identified 20 histopathologically confirmed malignancies, including 14 basal cell carcinomas and six malignant melanomas. Nineteen of these cases represented previously undiagnosed disease. Notably, none of the patients required enucleation or orbital exenteration at presentation, suggesting that earlier detection may have enabled intervention at a more manageable stage.
The study authors also report substantial gains in referral efficiency. Historically, patients treated surgically at the investigators' center had undergone an average of nearly four referrals before reaching definitive care. By contrast, patients identified through the app typically required only a single visit to access specialist services.
The study highlights the growing role of consumer technologies in ophthalmic screening. While AI tools have shown promise in hospital-based image analysis, relatively few have successfully transitioned into large-scale public health applications. By integrating awareness campaigns, AI-enabled self-screening, and structured referral pathways, the authors describe a "closed-loop mobile health ecosystem" that may help address disparities in access to specialist eye care.
