Automated retinal disease classification using deep learning and AlexNet with statistical models analysis.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Diabetic Retinopathy, Cataract, and Glaucoma are leading causes of vision loss.
• Early detection is crucial for preventing irreversible visual impairment.

Purpose of the Study:

• To develop and evaluate a deep learning framework for automated classification of retinal images.
• To classify images into Normal, Diabetic Retinopathy, Cataract, and Glaucoma categories.

Main Methods:

• Utilized a dataset from IDRiD and HRF retinal imaging databases.
• Evaluated four Convolutional Neural Network (CNN) architectures: EfficientNet-B0, EfficientNet-B7, a custom model, and AlexNet.
• Employed performance metrics including accuracy, sensitivity, specificity, PPV, NPV, F1-score, and R2 via five-fold cross-validation.

Main Results:

• AlexNet achieved the highest performance with 93.65% accuracy, 94.39% sensitivity, and 98.05% specificity.
• EfficientNet-B7 also showed strong performance (92.82% accuracy), highlighting transfer learning benefits.
• AlexNet demonstrated robustness with a mean R2 of 0.8891 and provided interpretable results using SHAP, highlighting key retinal features.

Conclusions:

• Interpretable deep learning models can accurately and consistently classify retinal diseases.
• The proposed framework supports AI-assisted ophthalmic screening for early detection and intervention.
• AlexNet offers a computationally efficient and explainable solution for retinal image analysis.