Screening Referable Diabetic Retinopathy Using a Semi-automated Deep Learning Algorithm Assisted Approach.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss.
• Early detection of vision-threatening DR is crucial for timely intervention.
• Automated methods for DR screening are being developed to improve efficiency.

Purpose of the Study:

• To evaluate the accuracy and efficacy of a semi-automated deep learning algorithm (DLA)-assisted approach for detecting vision-threatening diabetic retinopathy (DR).

Main Methods:

• A two-step approach combining DLA screening and human grading for referable DR in fundus photographs.
• Initial grading by DLA, followed by human re-grading of high-risk and a sample of negative images.
• Validation using baseline and long-term follow-up datasets from the Lingtou Cohort Study.

Main Results:

• The DLA achieved high performance metrics (AUC 0.953, sensitivity 0.970, specificity 0.879) on baseline images.
• Validation on follow-up datasets showed strong performance (AUC 0.914, sensitivity 0.852, specificity 0.853).
• The semi-automated approach offered substantial time (75.6%) and economic cost (90.1%) savings compared to fully manual grading.

Conclusions:

• The DLA demonstrates high accuracy, sensitivity, and specificity for grading referable DR.
• The semi-automated DLA-assisted approach effectively identifies suspect DR cases while optimizing safety, time, and cost.
• This method shows promise for efficient and accurate DR screening in clinical practice.