Unsupervised sorting of retinal vessels using locally consistent Gaussian mixtures.

Area of Science:

• Ophthalmology
• Medical Imaging
• Biomedical Engineering

Background:

• Accurate classification of retinal blood vessels (arterioles and venules) is crucial for identifying biomarkers.
• Retinal vessel clustering is challenging due to image quality issues like variable contrast and illumination.
• There is a significant need for high-performance, automated retinal vessel classification systems.

Purpose of the Study:

• To propose a novel unsupervised methodology for classifying retinal vessels into arterioles and venules.
• To address the challenges of non-uniform illumination and contrast variability in fundus images.
• To develop an automated system for enhanced biomarker discovery and computer-assisted diagnosis.

Main Methods:

• Utilized homomorphic filtering (HF) for image preprocessing, including non-uniform illumination correction and denoising.
• Employed an unsupervised multiscale line operator segmentation technique for retinal vasculature segmentation.
• Applied the Locally Consistent Gaussian Mixture Model (LCGMM) for unsupervised classification of retinal vessels.

Main Results:

• The proposed unsupervised method was evaluated on three public datasets: INSPIRE-AVR, VICAVR, and MESSIDOR.
• Achieved classification rates of 90.14%, 90.3%, and 93.8% in zone B across the respective datasets.
• Demonstrated superior performance compared to the conventional Gaussian Mixture Model using Expectation-Maximisation (GMM-EM).

Conclusions:

• The developed unsupervised clustering framework offers a high classification rate for retinal vessels.
• The method shows significant potential to improve computer-assisted diagnosis in ophthalmology.
• This approach can greatly enhance research in the field of biomarker discovery using retinal imaging.