Diagnostics of diabetic retinopathy based on fundus photos using machine learning methods with advanced feature...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy is a leading cause of preventable blindness.
• Early detection is critical to prevent vision loss and disability.
• Machine learning offers potential for automated early diagnosis.

Purpose of the Study:

• To develop and validate a machine learning model for early diabetic retinopathy detection.
• To leverage a large, multi-database dataset for robust model training and validation.
• To explore novel feature extraction and selection techniques for improved diagnostic accuracy.

Main Methods:

• Utilized a dataset of 14,402 fundus photographs from eight public and one private database.
• Employed forty-six unique feature selection and extraction methods (e.g., CLAHE, B-CosFire, Hough transform, LBP, GLCM).
• Trained and validated classification models using XgBoost and Random Forest algorithms, optimized with the Optuna library.

Main Results:

• The Random Forest model with LBP and GLCM feature extraction achieved 80.41% accuracy.
• The model demonstrated an F1-Score of 74.41% and an Area Under the Curve (AUC) of 0.80.
• The developed machine learning model showed high effectiveness in early diabetic retinopathy detection.

Conclusions:

• The machine learning model is a promising tool for the early detection of diabetic retinopathy.
• The study highlights the importance of large datasets and advanced feature engineering in medical AI.
• Further refinement is recommended for clinical implementation of the diagnostic tool.