Diabetic Retinopathy Fundus Image Classification and Lesions Localization System Using Deep Learning

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a diabetes complication causing irreversible retinal damage and a leading cause of blindness.
• Early detection of DR is crucial for preventing vision loss, as current treatments only slow progression.
• Automated, high-efficiency computer-based systems are vital for early DR diagnosis.

Purpose of the Study:

• To develop and evaluate a fully automatic deep learning system for diagnosing diabetic retinopathy.
• To classify DR images into five stages and localize affected lesions on the retina.
• To improve diagnostic accuracy, reduce misdiagnosis, and enhance efficiency compared to manual methods.

Main Methods:

• A deep learning system comprising two models: CNN512 for DR stage classification and YOLOv3 for lesion detection and localization.
• CNN512 processed whole images, achieving 88.6% accuracy on the DDR dataset.
• YOLOv3 achieved a 0.216 mAP for lesion localization on the DDR dataset.

Main Results:

• The fused system achieved 89% accuracy, 89% sensitivity, and 97.3% specificity in classifying DR images and localizing lesions.
• The system demonstrated superior performance compared to state-of-the-art results on public datasets (DDR and APTOS Kaggle 2019).
• The automated system effectively avoids misdiagnosis, saving time, effort, and cost.

Conclusions:

• The proposed fully automatic deep learning system offers a highly accurate and efficient solution for diabetic retinopathy diagnosis.
• The system's ability to classify stages and localize lesions surpasses current manual and automated techniques.
• This technology holds significant potential for early intervention and preventing blindness caused by diabetic retinopathy.