Automated multidimensional deep learning platform for referable diabetic retinopathy detection: a multicentre,...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss.
• Early detection and treatment are crucial for preventing blindness.
• Current screening methods can be resource-intensive.

Purpose of the Study:

• To develop and validate a real-world, guideline-based deep learning (DL) system for detecting referable diabetic retinopathy (DR).
• To assess the performance of the DL system against human experts in a multicentre setting.

Main Methods:

• A DL model was developed using retrospective, cross-sectional retinal image data from three hospitals and a mobile screening program.
• The model incorporated five classifiers: image quality, retinopathy, maculopathy gradability, maculopathy, and photocoagulation, based on UK DR screening guidelines.
• Performance was evaluated using metrics such as accuracy, F1 score, sensitivity, specificity, AUROC, and AUPRC, and compared with DR experts.

Main Results:

• The DL system demonstrated high accuracy across multiple classifiers in the external validation set (e.g., accuracy 0.915-0.980, AUROC 0.9639-0.9944).
• Referable DR detection achieved high performance (e.g., accuracy 0.918-0.967, AUROC 0.9848-0.9931).
• The DL system's performance was comparable to that of three DR experts (Cohen's κ: 0.86-0.93 vs. 0.89-0.96).

Conclusions:

• The developed DL system accurately performs multidimensional classifications for referable DR detection based on UK guidelines.
• The system's high accuracy and expert-level performance make it suitable for large-scale, real-world diabetic retinopathy screening programs.