Improved Automatic Grading of Diabetic Retinopathy Using Deep Learning and Principal Component Analysis

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness globally.
• Early detection via retinal fundus images is crucial but time-consuming for clinicians.
• Automated diagnostic tools are essential to meet the growing need for DR screening.

Purpose of the Study:

• To develop and validate a deep learning-based automated system for multi-class detection and classification of diabetic retinopathy.
• To evaluate the contribution of different color channels in retinal images for DR feature extraction.
• To enhance the accuracy and efficiency of DR diagnosis.

Main Methods:

• Utilized Principal Component Analysis (PCA) on significant color channels of retinal fundus images.
• Developed a deep learning model incorporating PCA-derived features for DR classification.
• Implemented a majority voting scheme for final grading decisions based on model outputs.
• Trained models on a large public dataset (~80K images) and validated on a local dataset (~100 images).

Main Results:

• Achieved 85% accuracy in multi-class diabetic retinopathy classification.
• Demonstrated high sensitivity (89%) and specificity (96%) in DR detection.
• Showcased significant improvement over traditional diagnostic methods.

Conclusions:

• The developed automated deep learning approach offers a promising solution for efficient and accurate diabetic retinopathy screening.
• Feature extraction using PCA on selected color channels enhances diagnostic performance.
• This technology can support clinicians in timely DR diagnosis, potentially preventing vision loss.