Artificial intelligence and external photographs in ophthalmology: a systematic review

OR 'pterygium'.The PRISMA guidelines were followed without language restrictions.The quality of included studies was assessed using the Quality Assessment of Diagnostic Accuracy Studies 2.Inclusion criteria were (1) clinical trials, prospective and retrospective cohort studies, comparative studies, validation, or evaluation studies, and epidemiologic studies;(2) studies involving facial images or external photographs in AI for detecting ophthalmic conditions; and (3) original articles.Exclusion criteria were (1) studies for which the full text was unavailable; (2) studies reporting the use of AI for non-ophthalmic conditions; (3) studies involving images taken from optical coherence tomography (OCT), fundus photography, or oculus keratography; and (4) review articles without original data.Two reviewers independently searched the literature and screened the abstracts of identified studies.Disagreements were resolved by discussion with four senior reviewers.The two reviewers then independently assessed the full text of eligible studies, and data were extracted and crossvalidated.In cases of unresolved disagreement, a thorough discussion involving all reviewers was conducted until a consensus was reached.The study selection process based on the PRISMA flowchart 14 is shown in the Figure .A customized form was used to record year of publication, study population, types of ophthalmic conditions, number and focal region of external photographs, number of test sets, types of deep neural network, and key findings.The risk of bias was assessed by two reviewers using the Quality Assessment of Diagnostic Accuracy Studies 2. ResultsIn total, 48 studies were included in the systematic review (Table 1).Among the 48 studies, 14 exhibited low risk for all four risk of bias parameters and eight displayed low risk for all three applicability concerns parameters (Table 2).