Abstract
Purpose :
Infectious keratitis (IK) stands as the fifth-leading cause of global blindness. Timely and precise diagnosis, coupled with appropriate treatment, is pivotal for achieving favorable outcomes in IK cases. Within the realm of artificial intelligence (AI), deep learning (DL) models have demonstrated diagnostic accuracy comparable to, if not better than clinicians, in various posterior segment diseases such as glaucoma and diabetic retinopathy. This study aimed to evaluate the performance of various DL models in diagnosing IK.
Methods :
We searched EMBASE and MEDLINE (up to April 23, 2023) with two independent reviewers. Eligible studies evaluated the utilization of DL models for diagnosing suspected IK of various etiologies (viral, bacterial, fungal, and protozoal) through corneal imaging tests. Our primary outcome was the sensitivity and specificity of DL models in distinguishing IK from other non-IK corneal pathologies. We also compared the accuracy in differentiating causes of IK. We extracted sufficient information to build 2x2 contingency tables for calculating the sensitivity and specificity. We conducted a bivariate meta-analysis for the overall accuracy and a bivariate meta-regression to assess the accuracy of two subgroups [imaging types: slit lamp photography versus in vivo confocal microscopy (IVCM)] following Cochrane guidance on diagnostic test accuracy meta-analyses. We used Stata for all analyses.
Results :
We screened 745 unique records and included 30 studies, comprising a collective sample of 91742 images and 37764 patients. We included 7 studies and 11 studies in meta-analysis for the primary and secondary outcomes, respectively. Overall, the pooled sensitivity and specificity for differentiating IK from non-IK [97.7% (95% confidence interval (CI): 93.8%-99.2%) and 97.2% (90.7%-99.2%)] was better than for differentiating causes of IK [82.6% (95% CI: 73.5%-89.1%) and 80.0% (67.0%-88.8%)] (Figure). For both outcomes, the sensitivity and specificity appeared better for IVCM imaging than slit lamp imaging (Table).
Conclusions :
This meta-analysis demonstrates promising diagnostic performances of DL models in distinguishing IK from non-IK corneal pathologies, and to a lesser extent, in differentiating the underlying causes of IK. However, further studies are required to fully evaluate their generalizability and potential for clinical deployment.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.