A deep learning based model for diabetic retinopathy grading.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a primary cause of blindness.
• Manual DR image analysis is time-consuming and error-prone.
• Current automated methods lack adaptability due to reliance on handcrafted features.

Purpose of the Study:

• To develop an automated system for early detection and grading of diabetic retinopathy severity.
• To create an efficient deep neural network for classifying DR stages.
• To improve upon existing automated DR detection methods.

Main Methods:

• Developed RSG-Net (Retinopathy Severity Grading), a deep neural network.
• Utilized the Messidor-1 dataset for training and testing.
• Applied preprocessing techniques: Histogram Equalization and denoising.
• Employed data augmentation (flipping, rotation, zooming, color/contrast/brightness adjustment) to address class imbalance.
• Incorporated convolutional layers, batch normalization, max pooling, dropout, and fully connected layers.

Main Results:

• RSG-Net achieved 99.36% accuracy, 99.79% specificity, and 99.41% sensitivity for 4-grade DR classification.
• RSG-Net achieved 99.37% accuracy, 100% sensitivity, and 98.62% specificity for 2-grade DR classification.
• The proposed model outperformed existing state-of-the-art methodologies.

Conclusions:

• RSG-Net demonstrates high efficacy in automated diabetic retinopathy detection and grading.
• The deep learning approach offers a time-efficient and accurate alternative to manual analysis.
• The system shows potential for widespread clinical application in early DR screening.