Accurate banded graph cut segmentation of thin structures using laplacian pyramids

Area of Science:

• Medical image analysis
• Computer vision
• Computational imaging

Background:

• Graph Cuts is popular for interactive image segmentation but struggles with large medical datasets.
• Traditional Banded Graph Cut (BGC) is fast but limited to segmenting large, roundish objects.
• Medical imaging often requires segmentation of complex, thin structures.

Purpose of the Study:

• To enhance the Banded Graph Cut (BGC) algorithm for efficient segmentation of thin structures in large medical images.
• To retain the computational speed of BGC while improving segmentation quality for challenging anatomical features.
• To introduce a new parameter for a smooth transition between BGC and traditional Graph Cuts.

Main Methods:

• A modified Banded Graph Cut (BGC) algorithm incorporating Laplacian pyramid information.
• Inclusion of thin structures within the BGC 'band' using multi-resolution image data.
• Quantitative and qualitative comparisons against the standard BGC algorithm.

Main Results:

• The proposed BGC modification successfully segments thin structures with high quality.
• Computational efficiency comparable to the original BGC is maintained.
• A novel parameter allows for a tunable transition from BGC to standard Graph Cuts.

Conclusions:

• The modified BGC algorithm effectively addresses the limitations of BGC for segmenting thin structures in large medical images.
• This approach offers a computationally efficient and accurate solution for medical image segmentation tasks.
• The introduced parameter provides flexibility in segmentation, bridging BGC and traditional Graph Cuts.