Hybrid deep learning models for diabetic retinopathy stage classification using fundus images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a primary cause of vision loss in diabetic patients.
• Early detection and accurate staging of DR are crucial for preventing severe visual impairment.
• Automated classification systems can aid clinicians in timely DR diagnosis and management.

Purpose of the Study:

• To develop and evaluate hybrid deep learning and machine learning models for automated diabetic retinopathy stage classification.
• To assess the performance of different hybrid architectures using retinal fundus images.
• To enhance model interpretability using visualization techniques.

Main Methods:

• Utilized deep learning (MobileNetV2, VGG16) for feature extraction from retinal fundus images.
• Combined deep learning features with traditional classifiers (Support Vector Machine, Random Forest).
• Applied Grad-CAM and Score-CAM for visualizing image regions influencing predictions.

Main Results:

• Hybrid models achieved high performance on the APTOS 2019 dataset.
• MobileNetV2+SVM and MobileNetV2+RF both attained 85% accuracy.
• The models demonstrated reliable and interpretable predictions for DR staging.

Conclusions:

• Lightweight Convolutional Neural Networks (CNNs) coupled with traditional classifiers offer effective DR stage prediction.
• The proposed hybrid approach provides accurate, transparent, and efficient tools for clinical support.
• Future research will focus on improving generalizability and clinical integration.