Diagnostic accuracy and clinical performance of deep learning models for grading diabetic retinopathy: a systematic review and meta-analysis - Report - MDSpire
Advertisement
Diagnostic accuracy and clinical performance of deep learning models for grading diabetic retinopathy: a systematic review and meta-analysis
Clinical Report: Evaluating the Diagnostic Precision of Deep Learning for DR
Overview
This systematic review and meta-analysis assess the diagnostic accuracy of deep learning models for grading diabetic retinopathy (DR).
Background
Diabetic retinopathy is a major cause of preventable blindness globally, necessitating accurate grading for effective management. Traditional grading methods face challenges such as inter-observer variability and reliance on specialist expertise.
Data Highlights
Classification Stage
Pooled Sensitivity
95% CI
No DR (Stage 0)
95.19%
93.00%–97.00%
Mild NPDR (Stage 1)
72.06%
62.06%–81.09%
Moderate NPDR (Stage 2)
84.33%
78.90%–89.10%
Severe NPDR (Stage 3)
75.84%
68.42%–82.57%
PDR (Stage 4)
78.82%
71.76%–85.13%
Key Findings
The pooled sensitivities of DL models varied significantly across DR severity levels.
In a simplified four-class classification, sensitivities improved across all grades compared to the five-class classification.
High sensitivity was noted for detecting no DR and vision-threatening DR.
Challenges remain in differentiating between adjacent non-proliferative stages.
There is a need for methodological standardization and rigorous external validation of DL models.
Clinical Implications
The high sensitivity of deep learning models for diabetic retinopathy grading suggests their utility in clinical screening settings.
Conclusion
Deep learning approaches demonstrate potential for diabetic retinopathy grading.
