Artificial Intelligence-Assisted Lesion-Based Urgent Referral Triage of Ultra-Widefield Retinal Images

Ultra-widefield retinal imaging is increasingly used for retinal disease screening and referral triage. Many vision-threatening retinal abnormalities require timely identification and referral to retinal specialists.

The AI system evaluated in this study is designed as a lesion-based triage tool rather than a disease-diagnosis system. The model identifies predefined urgent referral retinal findings and generates referral recommendations based on lesion-level evidence.

Urgent referral findings include:

* Retinal detachment
* Untreated retinal tear or retinal hole
* Vitreous hemorrhage
* Pre-retinal hemorrhage
* Subretinal hemorrhage
* Retinal neovascularization
* Optic disc neovascularization
* Tractional fibrovascular membrane Treated retinal tears associated with laser barricade scars are classified as non-urgent referral findings.

A total of 600 UWF retinal images acquired using Zeiss and Optos imaging systems will be included.

Participating ophthalmologists will independently evaluate images in randomized AI-assisted and unassisted settings.

The primary objective is to determine whether AI assistance improves lesion-based urgent referral triage accuracy.

• AI-Assisted UWF Lesion-Based Triage System (Diagnostic test): Readers interpret UWF retinal images with lesion-level AI findings and urgent referral recommendations.
• Unassisted Interpretation (Diagnostic test): Readers interpret UWF retinal images without AI assistance.

Primary outcomes

• Correct Lesion-Based Urgent Referral Triage Rate [Through study completion, up to 2 months]: Proportion of reader referral decisions consistent with expert-adjudicated lesion-based urgent referral classifications.

Secondary outcomes

• Sensitivity for Urgent Referral Findings [Through study completion, up to 2 months]: Sensitivity for correctly classifying non-urgent referral images according to expert-adjudicated lesion-based triage labels.
• Specificity for Urgent Referral Findings [Through study completion, up to 2 months]: Specificity for correctly classifying non-urgent referral images according to expert-adjudicated lesion-based triage labels.
• False-Negative Rate for Urgent Referral Findings [Through study completion, up to 2 months]: Proportion of urgent referral images incorrectly classified as non-urgent referral by readers.
• False-Positive Rate for Urgent Referral Findings [Through study completion, up to 2 months]: Proportion of non-urgent referral images incorrectly classified as urgent referral by readers.
• Reader Confidence Score [Immediately after image interpretation.]: Reader-reported confidence level for referral decisions measured using a 5-point Likert scale, ranging from 1 (very uncertain) to 5 (very confident).
• Change in Correct Urgent Referral Decisions After AI Assistance [Through study completion, up to 2 months]: Number and proportion of cases in which AI assistance changed an incorrect referral decision to a correct referral decision.

Sex: All, ages: 18 Years to any, accepts healthy volunteers

Inclusion Criteria:

* Licensed ophthalmologists
* Willing to participate as readers
* Completion of study training

Exclusion Criteria:

* Retinal specialists involved in establishing gold-standard labels
* Prior access to gold-standard labels
* Incomplete study participation