ShapeCut: Bayesian surface estimation using shape-driven graph.

Area of Science:

• Medical image analysis
• Computational anatomy
• Machine learning for medical imaging

Background:

• Medical image segmentation faces challenges like variable shapes and unclear boundaries.
• Existing methods using local priors often fail in complex segmentation tasks.
• Statistical frameworks combining image data with anatomical priors offer potential solutions.

Purpose of the Study:

• To develop a maximum-a-posteriori (MAP) segmentation framework incorporating both local and global shape priors.
• To address limitations of local optimization and shape penalties in challenging medical image segmentation.
• To improve segmentation accuracy for structures with high variability and ill-defined boundaries.

Main Methods:

• Proposed a MAP formulation utilizing a generative image model with integrated local and global shape priors.
• Employed graph cuts for optimization.
• Introduced a novel shape parameter estimation with a global updates-based optimization strategy.
• Validated on synthetic data and left atrial wall segmentation from late-gadolinium enhancement MRI.

Main Results:

• Demonstrated effectiveness on synthetic datasets and real-world cardiac MRI segmentation.
• Achieved high accuracy in segmenting the left atrial wall.
• Showed significant clinical utility in identifying myocardial fibrosis and scars.

Conclusions:

• The proposed MAP segmentation method effectively integrates local and global shape priors.
• The approach offers a robust solution for challenging medical image segmentation problems.
• This method holds promise for improving the diagnosis of conditions like atrial fibrillation through accurate fibrosis detection.