Four Severity Levels for Grading the Tortuosity of a Retinal Fundus Image

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Hypertensive retinopathy severity correlates with retinal vessel tortuosity.
• Existing methods lack automated tortuosity severity grading for computer-aided diagnosis.
• Accurate tortuosity assessment is crucial for grading retinopathy.

Purpose of the Study:

• To develop a machine learning model for automatic retinal image tortuosity severity classification.
• To establish a ground truth for tortuosity severity grading using expert ophthalmologists.
• To enhance automated grading systems for hypertensive and diabetic retinopathy.

Main Methods:

• Quantified retinal vessel tortuosity using fourteen measurement formulas on the AV-Classification dataset.
• Established manual tortuosity severity ground truth grading with ophthalmologist review.
• Trained and validated machine learning models including J48 decision tree, rotation forest, and distributed random forest.

Main Results:

• The distributed random forest model achieved the highest accuracy (99.4%) in tortuosity severity classification.
• The model demonstrated minimal root mean square error (0.0000192) and mean average error (0.0000182).
• The automated grading results closely matched ophthalmologist judgments.

Conclusions:

• The developed machine learning model effectively classifies retinal tortuosity severity (normal, mild, moderate, severe).
• Optimized vessel segmentation and feature extraction improved the accuracy of automated tortuosity detection.
• The model contributes to developing advanced automated grading systems for retinopathy.