Grading of diabetic retinopathy using a pre-segmenting deep learning classification model: Validation of an automated...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss.
• Accurate and efficient DR grading is crucial for timely intervention.
• Deep learning (DL) algorithms offer potential for automated medical image analysis.

Purpose of the Study:

• To validate the performance of an autonomous deep-learning (DL) algorithm for diabetic retinopathy (DR) grading.
• To compare the DL algorithm's performance against a human grader and a gold standard.
• To assess the clinical feasibility of autonomous DR classification.

Main Methods:

• 500 retinal images were graded by an expert ophthalmologist (gold standard) using the International Clinical Diabetic Retinopathy Disease Severity Scale (DR levels 0-4).
• Agreement was measured using weighted kappa for a human grader (with/without DL assistance) and the autonomous DL algorithm.
• Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated, with discrepancies analyzed.

Main Results:

• The autonomous DL algorithm achieved a weighted kappa of 0.72, sensitivity of 78%, and specificity of 81% compared to the gold standard.
• Extrapolated to a 23.8% real-world DR prevalence, the PPV was 57% and NPV was 92%.
• Discrepancies included artefact detection, missed microaneurysms, and segmentation/classification inconsistencies.

Conclusions:

• The autonomous DL algorithm's performance was comparable to a human grader in some aspects for high-risk populations.
• The high NPV (92%) in a real-world population suggests clinical feasibility for autonomous identification of non-DR patients.
• Further refinement may improve accuracy for DR classification.