Loading ...

TOPLINE:

A smartphone-based AI app (CaptureTumor) analyzed images of the eyes captured by users and detected malignancies on the eye surface nearly as accurately as a specialist, identified many previously undiagnosed cancers, and was associated with a simpler and faster referral pathway to expert care.

METHODOLOGY:

• Researchers conducted a nonrandomized clinical trial across China to develop and validate a smartphone app that allowed individuals to self-screen for malignancies of the eye surface.
• They included 614 participants aged 4-87 years (median age, 46 years; 49% female) between December 2022 and June 2023 via television, social media, and internet hospitals. The final analysis included 805 images from 535 participants.
• A deep learning model was trained on 12 years of slit-lamp images taken by specialists and fine-tuned with smartphone photos that passed the app’s framing, focus, and exposure checks.
• The app was deployed as a WeChat mini-program that provided photo guidance in real time, ran automatic image quality checks, processed images in the cloud, had clinicians review uploads within 24 hours, and referred those at high risk to specialist centers.
• Ophthalmologists labeled every image using histopathology when available; otherwise, clinical diagnosis or telemedicine was utilized. The primary outcome was area under the receiver operating characteristic curve (AUC) for distinguishing malignant lesions from benign ones.

TAKEAWAY:

• The app achieved an AUC of 0.945 for distinguishing malignant lesions from benign ones on prospectively collected specialist slit-lamp images. In real-world screening, the app achieved an AUC of 0.977 with 89.3% sensitivity and 95.9% specificity.
• The app achieved an AUC of 0.905 for distinguishing malignant lesions from benign ones when people used the app with in‑app photo guidance.
• The app prompted 58 referrals. A total of 20 malignancies were confirmed by histopathology, 19 of which were newly diagnosed, and none required removal of the eye or surrounding orbital tissue.
• Patients required a mean of 3.69 referrals before definitive treatment before using the app, which reduced to 1.02 after referral through the app (P < .001). The app projected a fivefold increase in cases detected per center, although the projection required validation.

IN PRACTICE:

“This ‘closed-loop’ model, which integrates public education, AI-guided triage, and specialist referral, is a noteworthy achievement in its own right and offers a compelling proof of concept for decentralized rare disease screening,” according to the authors of an invited commentary on the journal article. “However, this tool’s success moving forward depends on three key questions: how the model generalizes to a globally diverse patient population, who within any community engages with the technology, and how reliably the headline performance figures transfer to autonomous deployment at scale.”

SOURCE:

The study was led by Ruixin Wang, MD, PhD; Shaowei Bi, MM; Duoru Lin, MD, PhD; and Mingyuan Li, MEng, Sun Yat-sen University, Guangzhou, China. It was published online on June 4 in JAMA Ophthalmology.

LIMITATIONS:

The study included only a small cohort of non-Chinese individuals, limiting generalizability. The smartphone-based screening might have left older adults out, and age-friendly app designs might be needed. The study reported short-term screening results only.

DISCLOSURES:

The study received support from grants provided by the National Natural Science Foundation of China, Guangdong Natural Science Fund for Distinguished Young Scholars, and Guangzhou Medical Highland Construction Special Subsidy Fund. The authors did not report any conflicts of interest.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.