Transfer Learning-Based Model for Diabetic Retinopathy Diagnosis Using Retinal Images.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a major complication of diabetes mellitus, damaging retinal blood vessels and leading to vision impairment.
• DR is a leading global cause of preventable blindness, affecting over 158 million people worldwide.
• Early detection and classification of DR are crucial for mitigating visual impairment.

Purpose of the Study:

• To develop an automated deep learning framework for accurate diabetic retinopathy diagnosis.
• To overcome challenges of limited annotated datasets and class imbalance in medical image classification.
• To enhance DR classification performance using transfer learning and feature extraction.

Main Methods:

• Utilized the pre-trained VGGNet model for feature extraction from fundus images.
• Employed transfer learning principles to improve classification accuracy and reduce computational overhead.
• Applied diverse data augmentation techniques tailored to each DR grade to address data insufficiency and imbalance.

Main Results:

• The proposed framework demonstrated superior performance on a benchmark dataset compared to existing advanced methods.
• Achieved high accuracy in classifying diabetic retinopathy grades.
• The combination of deep learning features and handcrafted features further boosted classification accuracy.

Conclusions:

• The developed transfer learning approach effectively diagnoses diabetic retinopathy, offering a promising automated solution.
• The method successfully mitigates issues related to insufficient annotated data and class imbalance in medical imaging.
• This technique holds potential for improving the accuracy and efficiency of DR screening and diagnosis.