# Diagnostic accuracy of deep learning using ultra-widefield fundus imaging for retinal detachment: a systematic review and meta-analysis

## TL;DR

This study reviews and analyzes how well deep learning models using ultra-widefield retinal images can detect retinal detachment, finding high accuracy but noting limitations in consistency and generalizability.

## Contribution

The paper provides a systematic review and meta-analysis of deep learning's diagnostic accuracy for retinal detachment using ultra-widefield imaging.

## Key findings

- Pooled sensitivity and specificity of deep learning models for retinal detachment detection were 0.95 and 0.99, respectively.
- Subgroup analysis showed higher sensitivity in external evaluations compared to internal ones.
- High heterogeneity and inconsistent reporting limit the generalizability of the findings.

## Abstract

Retinal detachment (RD) requires prompt detection to prevent vision loss. Ultra-widefield (UWF) imaging captures the peripheral retina, and deep learning (DL) may enable automated RD detection. We aimed to systematically review and meta-analyze the diagnostic accuracy of DL applied to UWF images for detecting RD.

We systematically searched PubMed, Web of Science, and reference lists (last search 22 May 2025) for diagnostic-accuracy studies evaluating DL models for retinal detachment on UWF images with extractable 2 × 2 data. Two reviewers independently selected studies, extracted data, and assessed risk of bias and concerns regarding applicability using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2). Sensitivity, specificity, and area under the curve (AUC) were pooled using random-effects models, with subgroup analyses by dataset origin (internal vs. external) and retinal detachment spectrum.

We included 11 studies (2017–2024) using UWF imaging and DL, with test set sizes ranging from 89 to 6,222 images. The pooled sensitivity and specificity were 0.95 (95% CI, 0.94–0.96) and 0.99 (95% CI, 0.99–0.99); the AUC of the summary receiver operating characteristic (SROC) = 0.9962. Heterogeneity was high (I² = 92% for sensitivity; 90% for specificity). In subgroup analyses, external evaluations showed higher sensitivity than internal ones (0.97 vs. 0.92), with similarly high specificity (both ≈ 0.99). Heterogeneity remained substantial within subgroups. QUADAS-2 indicated a low risk of bias in most domains, with unclear index test risk common due to non-prespecified thresholds.

DL applied to UWF imaging shows high diagnostic accuracy for RD, with pooled sensitivity and specificity of 0.95 and 0.99, respectively, and an AUC of 0.9962. However, the evidence is limited by substantial heterogeneity, inconsistent index-test reporting, and variation in case spectrum and sample size, which may constrain generalizability. Overall, these findings suggest that DL combined with UWF imaging is likely to serve as a valuable adjunctive tool for RD detection and triage, particularly in settings where rapid, wide-field assessment is needed.

UMIN-CTR UMIN000057903; PROSPERO CRD420251058209.

The online version contains supplementary material available at 10.1186/s12886-025-04605-8.