Performance Analysis of Deep-Neural-Network-Based Automatic Diagnosis of Diabetic Retinopathy

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a significant cause of vision loss in diabetic patients.
• Current DR diagnosis is time-consuming, requiring multiple eye examinations.
• Early detection of DR is crucial for preventing or delaying vision loss.

Purpose of the Study:

• To develop a robust, automatic, computer-based system for diagnosing diabetic retinopathy.
• To utilize deep transfer learning for DR diagnosis and severity classification.
• To automate the diagnostic process and assist in patient therapy.

Main Methods:

• Employed five convolutional neural network (CNN) architectures: AlexNet, GoogleNet, Inception V4, Inception ResNet V2, and ResNeXt-50.
• Utilized a custom dataset of DR retinal images for training and evaluation.
• Trained deep CNNs to identify DR severity and classify images based on treatment approaches.

Main Results:

• The pre-trained Se-ResNeXt-50 model achieved the highest classification accuracy of 97.53% on the custom dataset.
• Experiments across CNN architectures demonstrated a minimum accuracy of 84.01% for five-degree DR classification.
• The developed system automates DR diagnosis and aids in subsequent patient therapies.

Conclusions:

• Deep transfer learning models, particularly Se-ResNeXt-50, show high efficacy in automated diabetic retinopathy diagnosis.
• The automated system can significantly improve the speed and accuracy of DR detection.
• This approach holds promise for early intervention and better management of vision loss in diabetic patients.