Objective:

To comprehensively assess the diagnostic accuracy of fundus image-based deep learning models in the grading of diabetic retinopathy.

Approach:

• Study Selection: Systematic search of PubMed, Embase, Web of Science, and Cochrane Library for studies published up to October 28, 2025, focusing on diagnostic accuracy of DL algorithms for DR grading.
• Data Extraction: Literature screening and data extraction performed independently by two researchers, with bias risk assessed using the QUADAS−2 tool.

Key Findings:

• 41 studies included, covering various DL architectures and datasets.
• Pooled sensitivities for five-class classification: 95.19% (95% CI: 93.00%–97.00%) for no DR, 72.06% (95% CI: 62.06%–81.09%) for mild NPDR, 84.33% (95% CI: 78.90%–89.10%) for moderate NPDR, 75.84% (95% CI: 68.42%–82.57%) for severe NPDR, and 78.82% (95% CI: 71.76%–85.13%) for PDR.
• In four-class classification, sensitivities improved: 96.85% (95% CI: 90.18%–99.93%) for stage 0, 92.94% (95% CI: 79.50%–99.72%) for stage 1, 92.75% (95% CI: 79.31%–99.61%) for stage 2, and 88.19% (95% CI: 68.99%–98.93%) for stage 3.

Interpretation:

Deep learning shows high sensitivity for DR grading, especially in identifying no DR and vision-threatening DR, but differentiating adjacent non-proliferative stages remains challenging.

Limitations:

• Observed heterogeneity in study results, indicating variability in diagnostic performance across studies.

Conclusion:

Future research should focus on enhancing clinical utility and generalizability of DL models for DR grading.

Sources:

• Systematic Review Registration