Related Experiment Videos
Feature-reconstructed diabetic retinopathy classification using variational autoencoder with disentanglement factor
Ashu Priya1, Niranjana Banerjee1, Manas Ranjan Prusty2
1School of Computer Science and Engineering, Vellore Institute of Technology, Chennai, India.
Frontiers in Medical Technology
|August 4, 2026
Summary
A new Variational Autoencoder (VAE) framework accurately classifies Diabetic Retinopathy (DR) using retinal images. This automated method aids early detection and grading of DR, a leading cause of blindness.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Diabetic Retinopathy (DR) is a primary cause of vision loss in working-age adults.
- Early detection and classification of DR are critical challenges in healthcare.
- Existing deep learning models often lack efficient feature extraction mechanisms.
Purpose of the Study:
- To propose a novel framework for classifying Diabetic Retinopathy (DR).
- To develop a lightweight, feature-focused learning mechanism for DR classification.
- To enable both binary and multilevel classification of DR severity.
Main Methods:
- A Variational Autoencoder (VAE) with a disentanglement factor (Beta) was employed.
- The VAE encoder compresses retinal image features into a latent vector.
- Logistic Regression was used for classification based on the VAE output.
Main Results:
- Achieved 98.64% accuracy for binary classification on the APTOS 2019 dataset.
- Achieved 97.83% accuracy for binary classification on the DDR dataset.
- Recorded 97.80% and 97.23% accuracy for multilevel classification on APTOS 2019 and DDR datasets, respectively.
Conclusions:
- The proposed VAE-Logistic Regression framework demonstrates high accuracy in DR classification.
- This method offers a potential solution for automated DR screening and severity grading.
- The approach effectively extracts and utilizes fine features for improved diagnostic accuracy.
