Related Experiment Video
Updated: Dec 11, 2025
07:18
Evaluation of Capillary and Other Vessel Contribution to Macular Perfusion Density Measured with Optical Coherence Tomography Angiography
Published on: February 18, 2022
2.0K
Ensemble Deep Learning for Diabetic Retinopathy Detection Using Optical Coherence Tomography Angiography.
Morgan Heisler1, Sonja Karst2, Julian Lo1
1School of Engineering Science, Simon Fraser University, Burnaby, British Columbia, Canada.
Translational Vision Science & Technology
|August 21, 2020
Summary
Ensemble learning with deep learning significantly improves diabetic retinopathy (DR) classification accuracy on optical coherence tomography angiography (OCTA) images. These advanced methods offer promising clinical solutions for DR diagnosis.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Diabetic retinopathy (DR) is a leading cause of vision loss.
- Accurate classification of DR is crucial for timely intervention.
- Optical coherence tomography angiography (OCTA) provides detailed microvascular information.
Purpose of the Study:
- To evaluate ensemble learning techniques combined with deep learning for DR classification.
- To assess the performance of these methods on OCTA images and co-registered structural data.
- To compare deep learning ensemble models with traditional feature engineering approaches.
Main Methods:
- Utilized 463 OCTA volumes from 380 eyes.
- Developed component neural networks using VGG19, ResNet50, and DenseNet architectures.
- Ensembled networks via majority soft voting and stacking.
- Compared results against manually engineered features and single data-type networks.
- Employed Class Activation Maps (CAMs) and Grad-CAM for visualization.
Main Results:
- VGG19-based networks outperformed deeper architectures.
- Ensemble networks achieved accuracies of 0.92 (voting) and 0.90 (stacking).
- Ensemble methods surpassed single data-type and hand-crafted feature classifiers.
- Grad-CAM effectively highlighted disease-affected areas.
Conclusions:
- Ensemble learning enhances Convolutional Neural Networks (CNNs) for referable DR classification on OCTA datasets.
- Proposed methods show potential for clinical DR diagnosis solutions.
- OCTA image-based deep learning offers superior diagnostic accuracy compared to manual features.
