Related Experiment Video
Updated: Dec 7, 2025
07:23
Retinal Vascular Reactivity as Assessed by Optical Coherence Tomography Angiography
Published on: March 26, 2020
8.6K
DcardNet: Diabetic Retinopathy Classification at Multiple Levels Based on Structural and Angiographic Optical
IEEE Transactions on Bio-Medical Engineering
|September 28, 2020
Summary
This study introduces a novel deep learning framework using Optical Coherence Tomography (OCT) and OCT Angiography (OCTA) for automated diabetic retinopathy (DR) classification. The DcardNet model achieved high accuracy in detecting referable DR, aiding early diagnosis and vision preservation.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Diabetic retinopathy (DR) poses a significant threat to vision, necessitating early detection and diagnosis.
- Optical coherence tomography (OCT) and OCT angiography (OCTA) offer advanced imaging capabilities for DR assessment.
- Existing automated DR classification frameworks lack completeness, especially when integrating both OCT and OCTA data.
Purpose of the Study:
- To propose a comprehensive, automated diabetic retinopathy classification framework utilizing both en face OCT and OCTA data.
- To develop and evaluate a novel convolutional neural network (CNN) for accurate DR grading.
Main Methods:
- A densely and continuously connected neural network with adaptive rate dropout (DcardNet) was designed for DR classification.
- Adaptive label smoothing was employed to mitigate overfitting during model training.
- The framework generated three classification levels based on the International Clinical Diabetic Retinopathy scale: referable/non-referable, NPDR/PDR, and detailed severity grading.
Main Results:
- The DcardNet model achieved classification accuracies of 95.7% for referable DR detection, 85.0% for distinguishing between non-proliferative and proliferative DR, and 71.0% for detailed DR severity grading.
- Performance was validated using 10-fold cross-validation on 10% of the dataset.
Conclusions:
- The proposed automated classification framework demonstrates high reliability, sensitivity, and specificity for DR detection.
- This AI-driven approach can serve as a key technology to facilitate timely referral to ophthalmologists, ultimately reducing vision loss associated with DR.
