HOG-CNN: Integrating Histogram of Oriented Gradients with Convolutional Neural Networks for Retinal Image...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Manual interpretation of retinal fundus images for diagnosing diseases like diabetic retinopathy (DR), glaucoma, and age-related macular degeneration (AMD) is time-consuming and resource-intensive.
• Automated diagnostic tools are needed to improve efficiency and accessibility in clinical settings.

Purpose of the Study:

• To develop an automated and interpretable clinical decision support framework for retinal disease detection using a hybrid feature extraction model.
• To integrate handcrafted Histogram of Oriented Gradients (HOG) features with deep learning features from EfficientNetB3 for enhanced image analysis.

Main Methods:

• A hybrid feature extraction model, HOG-CNN, was developed by combining HOG features with deep features from a pretrained EfficientNetB3 convolutional neural network.
• The model was evaluated on three public benchmark datasets: APTOS 2019 (DR), ORIGA (Glaucoma), and IC-AMD (AMD).

Main Results:

• HOG-CNN achieved high performance across datasets, including 98.5% accuracy and 99.2 AUC for binary DR classification, and 92.8% accuracy, 94.8% precision, and 94.5 AUC for AMD diagnosis.
• The model demonstrated competitive performance for Glaucoma detection (83.9% accuracy, 87.2 AUC) and outperformed several state-of-the-art methods on the IC-AMD dataset.
• Appendix studies confirmed the complementary strengths of combining HOG and CNN features.

Conclusions:

• The HOG-CNN framework provides a robust and scalable solution for automated retinal disease screening.
• Its lightweight and interpretable design makes it suitable for deployment in resource-constrained clinical environments, enhancing early detection and diagnosis.