Automated detection of severe diabetic retinopathy using deep learning method

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss.
• Early and accurate diagnosis of severe DR is crucial for timely intervention.
• Color fundus photography is a common imaging modality for DR screening.

Purpose of the Study:

• To develop and validate an AI system for intelligent diagnosis of severe DR.
• To incorporate lesion recognition into the DR diagnosis model.
• To evaluate the performance of the AI system using a large dataset of fundus images.

Main Methods:

• Utilized a large Kaggle dataset of 53,576 fundus images for testing and 28,101 for training.
• Employed the Inception V3 architecture for image classification.
• Compared performance using 299x299 and 896x896 pixel input image resolutions.
• Evaluated models using ROC curve, AUC, sensitivity, specificity, and harmonic mean.

Main Results:

• The 896x896 input model demonstrated superior performance over the 299x299 model.
• Achieved high performance metrics for severe DR detection with the 896x896 model: 0.925 sensitivity, 0.907 specificity, 0.916 harmonic mean, and 0.968 AUC.
• Identified that moderate non-proliferative diabetic retinopathy (NPDR) cases were prone to misclassification.
• Hard exudates, cotton wool spots, preretinal hemorrhage, and vitreous hemorrhage influenced classification accuracy; IRMA was the most challenging lesion to recognize.

Conclusions:

• Developed an effective AI-based system for severe DR diagnosis from color fundus photography.
• The AI system shows promise for enhancing the accessibility and efficiency of severe DR screening.
• Higher resolution images (896x896) improved diagnostic performance compared to lower resolution (299x299).