Deep Learning-Based Diabetic Retinopathy Severity Grading System Employing Quadrant Ensemble Model.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) causes retinal blood vessel deterioration, leading to vision loss.
• Early detection of DR is crucial for effective treatment and preventing complications.
• Automated DR diagnosis systems support ophthalmic experts in timely identification.

Purpose of the Study:

• To propose a quadrant ensemble automated DR grading approach using deep learning.
• To enhance the accuracy and generalization ability of DR detection models.
• To provide an effective tool for early and accurate diagnosis of diabetic retinopathy.

Main Methods:

• Implementation of the InceptionResnet-V2 deep neural network framework.
• Incorporation of image preprocessing techniques: histogram equalization, optical disc localization, and quadrant cropping.
• Utilization of data augmentation to improve network performance and validation on diverse datasets (MESSIDOR, IDRiD).

Main Results:

• Achieved a superior accuracy of 93.33% on the MESSIDOR dataset.
• Observed a significant reduction in cross-entropy loss (0.325).
• Demonstrated a 13.58% accuracy improvement over the Inception-V3 CNN model and 25.23% over state-of-the-art approaches.

Conclusions:

• The proposed quadrant ensemble automated DR grading approach shows high efficacy and generalization ability.
• Deep learning models, like InceptionResnet-V2, outperform conventional methods for DR detection.
• This framework offers a viable solution for early and accurate diabetic retinopathy diagnosis.