Efficient Descriptor-Based Segmentation of Parotid Glands With Nonlocal Means

Area of Science:

• Medical image analysis
• Computer vision
• Computational anatomy

Background:

• Patch-based segmentation methods are widely used but can be limited by feature selection.
• Integrating diverse image information (intensities, features, location) can potentially improve segmentation accuracy.

Purpose of the Study:

• To introduce and evaluate descriptor-based segmentation, extending traditional patch-based methods.
• To empirically investigate the impact of various image features on segmentation performance.
• To optimize neighborhood search strategies for improved accuracy.

Main Methods:

• Extended nonlocal means segmentation to incorporate image features and location information.
• Employed kd-trees for efficient nearest neighbor search in larger windows.
• Conducted a broad empirical study comparing a multitude of image features for patch selection.

Main Results:

• Entropy image features, previously unused in patch-based segmentation, yielded the best results.
• Searching larger image regions with approximate nearest neighbor search and location information significantly improved segmentation.
• The inclusion of features and location information led to substantial increases in segmentation accuracy.

Conclusions:

• Image features, particularly those highlighting boundaries like entropy, significantly enhance segmentation accuracy.
• Integrating location information and employing efficient, larger-scale neighborhood searches advances patch-based segmentation.
• The developed descriptor-based segmentation approach offers state-of-the-art performance for parotid gland segmentation in radiation therapy planning.