Multiscale self-quotient filtering for an improved unsupervised retinal blood vessels characterisation

Area of Science:

• Ophthalmology
• Medical Imaging
• Biomedical Engineering

Background:

• Digital retinal images often have poor contrast and uneven lighting, hindering accurate biomarker assessment.
• The arteriolar-to-venular ratio is a key biomarker for conditions like diabetes and hypertension.
• Accurate classification of retinal vessels into arterioles and venules is essential for reliable biomarker estimation.

Purpose of the Study:

• To develop an unsupervised method for classifying retinal vessels.
• To improve the accuracy of arteriolar-to-venular ratio estimation.
• To enhance computer-aided diagnosis and biomarker research in ophthalmology.

Main Methods:

• Utilized multiscale self-quotient filtering for image pre-processing.
• Extracted discriminating features from retinal vessels.
• Employed squared-loss mutual information clustering for unsupervised vessel classification.

Main Results:

• Achieved classification rates of 93.2% on the DRIVE dataset and 88.9% on the INSPIRE-AVR dataset (Zone B).
• The proposed unsupervised framework outperformed existing methods.
• Demonstrated high accuracy in retinal vessel classification.

Conclusions:

• The developed unsupervised retinal vessel classification method is effective.
• This approach holds significant potential for computer-aided diagnosis and biomarker research.
• Accurate vessel classification can improve the evaluation of cardiovascular health indicators from retinal images.