A hybrid approach for diabetic retinopathy stages classification using spatial and textural features

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Diabetic Retinopathy Diagnosis

Background:

• Diabetic Retinopathy (DR) is a complication of diabetes damaging retinal blood vessels.
• Current DR detection methods are computationally intensive and lack accuracy.
• Automated, efficient DR monitoring is crucial for diabetic patients.

Purpose of the Study:

• To develop and evaluate computationally efficient machine learning models for multi-class DR classification.
• To compare the performance of various models including Random Forest, Logistic Regression, and deep learning architectures.
• To assess the utility of spatial, textural, and hybrid features for DR stage identification.

Main Methods:

• Extracted spatial features using Convolutional Neural Networks (CNN) and textural features using Grey Level Co-occurrence Matrix (GLCM).
• Developed hybrid features by combining spatial and textural information.
• Trained and evaluated Random Forest (RF), Logistic Regression (LR), Decision Tree (DT), Gaussian Naive Bayes (GNB), CNN, EfficientNet, PyramidCNN, and Pyramid Vision Transformer (PVT) models.

Main Results:

• The RF model with hybrid features achieved the highest accuracy (98.00%), outperforming existing methods by 1%.
• EfficientNet demonstrated competitive performance with 97.00% accuracy.
• Machine learning models showed computational efficiency in training and inference times.

Conclusions:

• The proposed RF model with hybrid features offers a highly accurate and efficient approach for DR classification.
• Machine learning models are suitable for deployment in low-resource clinical settings for automated DR monitoring.
• This study highlights the potential of AI in improving diabetic eye care and disease management.