JMIR Publications

ABSTRACT

Background:

Cataracts, with high prevalence and blindness-inducing potential, necessitate effective approaches for early diagnosis, underscoring the clinical significance of this investigation.

Objective:

To evaluate the performance of deep learning (DL) in cataract diagnosis and assess its potential as an effective tool for automated diagnosis.

Methods:

A systematic search was conducted in Web of Science, Embase, IEEE Xplore, PubMed, and Cochrane Library until April 1, 2025 for studies on image-based DL for cataract diagnosis or clinical subtype classification. The included studies were assessed for the risk of bias (RoB) using Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2). Bivariate mixed-effects models were utilized for data analyses, and publication bias was assessed by Deek's funnel plots.

Results:

Sixty-three studies were finally included. The RoB was high or unclear in six studies in patient selection, and high or unclear in five studies in index test due to no predefined thresholds. No studies had high or unclear RoB in reference standard. Moreover, high or unclear applicability concerns were found in only one study in patient selection. Image-based DL achieved a sensitivity (SE) of 95% (95% confidence interval [CI] 0.94–0.97) and a specificity (SP) of 97% (0.96–0.98) for cataract detection, with an area under the ROC curve (AUC) of 0.99 (0.98–1.00). For cataract classification, the SE and SP of image-based DL were 93% (0.91–0.94) and 97% (0.96–0.98), respectively, with an AUC of 0.98 (0.97–0.99).

Conclusions:

DL algorithms demonstrate superior accuracy to traditional machine learning, perform comparably to human experts in cataract detection, and exhibit substantial potential as tools for automated diagnosis. However, these algorithms can only serve as adjuncts to cataract diagnosis currently due to the moderate quality and high heterogeneity of existing evidence. Clinical Trial: Not applicable