Robust Retinal Vessel Segmentation via Locally Adaptive Derivative Frames in Orientation Scores

Area of Science:

• Medical imaging
• Computer vision
• Image processing

Background:

• Accurate retinal vessel segmentation is crucial for diagnosing various eye diseases.
• Existing methods often struggle with complex vascular structures like crossings and tiny vessels.

Purpose of the Study:

• To develop a robust and fully automatic filter-based method for retinal vessel segmentation.
• To introduce novel filters utilizing 3D rotating frames and orientation scores.

Main Methods:

• A wavelet-type transform lifts 2D retinal images to a 3D orientation score domain.
• Multi-scale second-order Gaussian derivatives (LID and LAD frames) enhance vessels.
• The Locally Adaptive Derivative (LAD) frame, derived from Hessian matrix analysis, adapts to local vessel structures.

Main Results:

• The proposed method, particularly the LAD on Orientation Scores (LAD-OS), achieves competitive segmentation performance across six diverse retinal datasets.
• LAD-OS demonstrates superior capability in segmenting challenging features such as vessel crossings, arterial reflex, and tiny vessels.
• The algorithm exhibits high computational speed, suitable for large-scale screening.

Conclusions:

• The proposed filter-based approach provides a robust and automatic solution for retinal vessel segmentation.
• The LAD-OS method represents a significant advancement, outperforming many state-of-the-art techniques.
• This technique holds promise for efficient and accurate analysis in clinical settings.