Trainable COSFIRE filters for curvilinear structure delineation in images

Trainable COSFIRE filters for curvilinear structure delineation in images detect and delineate curvilinear structures by combining responses from Difference-of-Gaussians (DoG) filters to support analysis of blood vessels and vessel endings in retinal fundus images.

Key Features:

• Combination of DoG responses: Responses from Difference-of-Gaussians (DoG) filters are combined to detect linear structures.
• Orientation-selective line detection: Filters detect lines at different orientations to capture vessel trajectories.
• Multi-scale/thickness robustness: The approach can identify vessels of varying thickness and orientation.
• Selective tuning to vessels and vessel-endings: A set of B-COSFIRE filters are selectively tuned to recognize both blood vessels and vessel endings.
• Automatic selection process: Filters undergo an automatic selection process to determine their selectivity for specific image features.
• Adaptability via machine learning and information theory: Selection criteria and filter adaptability are refined using machine learning and information theory principles.

Scientific Applications:

• Retinal vessel delineation: Delineation of blood vessels in retinal fundus images to support computer-aided diagnosis systems.
• Benchmarking on public datasets: Performance has been evaluated on public retinal datasets DRIVE and STARE.

Methodology:

The method combines responses of Difference-of-Gaussians (DoG) filters using a set of trainable B-COSFIRE filters that are automatically selected and selectively tuned to recognize blood vessels and vessel endings, with selection criteria refined using machine learning and information theory principles.