Segment 2D and 3D Filaments by Learning Structured and Contextual Features.

Area of Science:

• Computer Vision
• Medical Imaging Analysis
• Machine Learning

Background:

• Filamentary structure segmentation (e.g., retinal vessels, neurons) is challenging.
• Existing learning-based methods lack sufficient data-driven feature construction for contextual information.
• This limitation can hinder segmentation performance.

Purpose of the Study:

• To propose a novel data-driven approach for learning structured and contextual features.
• To improve the segmentation of filamentary structures in 2D and 3D images.

Main Methods:

• Developed a data-driven approach to learn structured features by integrating local spatial label patterns.
• Introduced contextual features to capture global contextual information.
• Utilized tree classifiers for feature space splitting and grouping similar structures.

Main Results:

• The proposed approach significantly outperforms state-of-the-art methods.
• Demonstrated superior performance on well-regarded testbeds across various applications.
• Empirical evaluations confirm the effectiveness of the learned structured and contextual features.

Conclusions:

• The novel data-driven feature learning method effectively addresses limitations in filamentary structure segmentation.
• The approach enhances segmentation accuracy by incorporating structured and contextual information.
• The publicly available code supports open-source research in image analysis.