Random walker with improved weighting function for interactive medical image segmentation

Area of Science:

• Medical Image Analysis
• Computer Vision
• Computational Imaging

Background:

• The random walks algorithm is sensitive to initial seed placement for image segmentation.
• Existing methods primarily rely on neighborhood pixel relationships, limiting accuracy.

Purpose of the Study:

• To improve the accuracy and robustness of the random walks algorithm for image segmentation.
• To develop a modified weighting function that incorporates intensity changes and texture features.

Main Methods:

• Modified the random walks weighting function to account for intensity variations between neighborhood nodes.
• Incorporated local affiliation using a penalty term.
• Integrated Gray-Level Co-occurrence Matrix (GLCM) variance via kernel density estimation (KDE) for texture analysis.
• Estimated pixel probability density for seed affiliation.

Main Results:

• The proposed weighting model demonstrated superior performance compared to the original random walks algorithm.
• Achieved better segmentation results on various medical images, including 174 brain tumor images.
• The enhanced method shows improved robustness to initial seed placement.

Conclusions:

• The modified random walks algorithm with the proposed weighting function offers enhanced image segmentation accuracy.
• The integration of intensity changes, local affiliation, and texture features (GLCM variance) significantly improves segmentation, particularly for complex medical images.
• This approach provides a more reliable tool for medical image analysis and segmentation tasks.