Smartphone-Enabled Imaging Part of Strategy in Screening Rare Ocular Malignancies

Addressing gaps in early detection of rare ocular malignancies could be done through the use of smartphone-enabled imaging, targeted media outreach, and diagnostics driven by artificial intelligence (AI) according to a new study published in JAMA Ophthalmology.¹ Equitable care delivery and early detection could be better addressed through these mobile health initiatives, which can improve prognosis in these rare diseases.

Effective screening for rare ocular diseases is plagued by systemic challenges, including limited availability of specialists and misdiagnoses. Although rare, eye cancers can be debilitating if not treated, causing vision loss or changes in the eye’s appearance if not treated.² Innovative means of diagnosing ocular surface tumors are required to address these gaps. This study aimed to test the CaptureTumor (CaT) AI diagnostic system that works in the smartphone to help diagnose rare ocular conditions.

A 3-Stage Approach to AI-Driven Screening

The CaT deep-learning model was trained on slit lamp images from clinical settings as part of the first stage of developing and validating the system. The second stage of the study involved adapting the model through a human-computer interface and optimizing image preprocessing. The third and final stage involved implementing the model to assess its diagnostic performance.

To train the model, data were retrospectively collected, with data from January 1, 2010, to December 31, 2021, included. All data were collected from the Zhongshan Ophthalmic Center (ZOC) and 6 other partner centers. The first stage included images from slit lamps that were collected by clinicians, the second stage included pictures from smartphones that were captured by trained operators, and the third stage included self-captured images by the public. Internal training of the CaT application was done by training the application on the dataset from ZOC.

Smartphone Screening Matches Specialist-Level Accuracy

There were 535 participants who completed at-home screening through the mobile app and were included in the final analysis. The participants had a median (IQR) age of 46 (11) years, and 49% were women.

A binary classification model that identified between malignant and benign tumors achieved an area under the curve (AUC) of 0.959, whereas a multiclass classification between the top 5 most prevalent lesion types had a mean AUC of 0.926, both through internal testing. The AUC for binary classification was reduced to 0.945, and multiclass classification was reduced to 0.874 after the evaluation of 535 slit-lamp images from 239 patients, but both maintained efficacies.

Performance of the model worsened when the model was applied to retrospective smartphone images, with binary classification AUC declining to 0.787 and the multiclass AUC declining to 0.676. The researchers determined that focus blur, insufficient lesion-to-image ratio, and suboptimal exposure were all factors in performance. Implementing changes to address these factors, the model improved to an AUC of 0.905 for binary classification and 0.866 in multiclass classification.

The researchers included educational digital workflows to help participants use the app easily, including guided instructions for self-examinations, access to a program for image capture and uploading, and immediate risk stratification with follow-up guidance. After submission of 805 images from 535 participants, the researchers found that the model achieved an AUC of 0.977 for binary classification, with a sensitivity of 89.3% and specificity of 95.9%. An AUC of 0.810 was achieved for multiclass classification. Notably, 19 of the 20 confirmed malignancies (95%) were newly diagnosed through the app-based screening, and no patient required enucleation or orbital exenteration.

Looking Ahead: Validation and Long-Term Outcomes

There were some limitations to this study. The study included a Chinese cohort, which necessitates further studies in other countries. Older adults may find mobile health more difficult than accessible, which requires further investigation into designs that help older patients. Long-term outcomes of those screened were not assessed in this study.

“This mobile health model offers a potentially scalable, accessible, and affordable strategy for early detection of rare, vision- and life-threatening diseases,” the authors concluded. “Further validation and long-term assessment will be essential to determine its sustained potential to influence patient outcomes and health systems worldwide.”

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. Eye cancer. Cleveland Clinic. Updated December 8, 2022. Accessed June 5, 2026. https://my.clevelandclinic.org/health/diseases/17292-eye-cancer