EDLDR: An Ensemble Deep Learning Technique for Detection and Classification of Diabetic Retinopathy

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss, necessitating early and accurate detection.
• Retinal fundus images are crucial for identifying DR-related abnormalities.
• Current detection methods require improvement in accuracy and efficiency.

Purpose of the Study:

• To propose an automated ensemble deep learning model for the detection and classification of diabetic retinopathy.
• To enhance diagnostic performance by ensembling two advanced deep learning architectures.
• To evaluate the model's effectiveness on diverse datasets and classification schemes.

Main Methods:

• An ensemble model combining modified DenseNet101 and ResNeXt deep learning architectures was developed.
• Data preprocessing included CLAHE for histogram equalization, and GAN-based augmentation addressed class imbalance.
• Ensembling involved normalization over classes and maximum a posteriori for final class label computation.

Main Results:

• The proposed model achieved high accuracy, precision, and recall on the APTOS19 and DIARETDB1 datasets.
• For two-class classification, accuracy reached 96.98% with precision and recall of 0.97.
• For five-class classification, accuracy was 86.08%, with precision of 0.76 and recall of 0.82.

Conclusions:

• The automated ensemble deep learning model demonstrates superior performance in detecting and classifying diabetic retinopathy.
• This approach offers a promising tool for early DR detection, potentially preventing vision loss.
• The model's effectiveness is validated across different classification granularities and datasets.