Enhancing diabetic retinopathy classification using deep learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Prolonged hyperglycemia leads to diabetic retinopathy (DR), a primary cause of blindness.
• Prompt identification and management of DR can prevent numerous cases of vision loss.
• Deep learning (DL) algorithms show promise in advancing diagnostic capabilities for medical conditions.

Purpose of the Study:

• To develop and evaluate a deep learning model for the accurate detection and grading of diabetic retinopathy stages.
• To assess the efficacy of image enhancement techniques in improving DR detection accuracy.
• To compare the performance of the proposed DL model against existing state-of-the-art methods.

Main Methods:

• Utilized the "Asia Pacific Tele-Ophthalmology Society (APTOS) 2019 Blindness Detection" dataset of retinal scans.
• Employed deep learning, specifically a Convolutional Neural Network (CNN) model, for DR classification.
• Applied data augmentation strategies and image enhancement techniques (CLAHE, ESRGAN) to optimize the dataset and image quality.

Main Results:

• Achieved a highest experimental accuracy of 97.83% for DR detection across 5 severity stages.
• Obtained top-2 accuracy of 99.31% and top-3 accuracy of 99.88% on the APTOS 2019 dataset.
• Demonstrated superior efficiency in DR localization compared to conventional DL and state-of-the-art technologies.

Conclusions:

• The proposed deep learning approach, enhanced with image processing techniques, effectively detects and grades diabetic retinopathy.
• This method offers a highly accurate and efficient tool for early DR diagnosis, potentially reducing blindness.
• The model's performance indicates its potential for clinical application in ophthalmology.