Deep Learning Algorithms for the Diagnosis and Staging of Retinopathy of Prematurity: A Systematic Review and...

The objective of this systematic review and meta-analysis is to evaluate artificial intelligence (AI) algorithms that diagnose and stage retinopathy of prematurity (ROP). Diagnostic accuracy studies reporting metrics of algorithms using fundus images for ROP identification and staging were included and evaluated. Meta-analysis was conducted for ROP identification studies. Fourteen studies were included in the systematic review and ten studies in the meta-analysis. In total, the studies included in the meta-analysis dealt with 37,057 images. AI showed a high performance in detecting ROP with pooled sensitivity equal to 95% (95% CI 91.5%-97.1%), while the respective value concerning specificity was 97.4% (95% CI 95.3%-98.6%) for ROP identification. The heterogeneity index I² showed a moderate level of heterogeneity, equal to 0.75. Most of the studies lacked bias control for participants' selection to allow clinical setting applicability. Our research demonstrates that AI algorithms have the potential to become a useful tool in accurately diagnosing and grading ROP. ROP screening can be enhanced by the application of AI but requires special attention regarding bias and real-world implementation.