Pseudo-Marginal MCMC Sampling for Image Segmentation Using Nonparametric Shape Priors.

Area of Science:

• Computer Vision
• Statistical Shape Analysis
• Image Processing

Background:

• Image segmentation is challenging with low-quality or incomplete data.
• Statistical shape priors enhance segmentation but existing methods lack confidence measures.
• Point estimates from Bayesian frameworks can fail with multimodal posteriors.

Purpose of the Study:

• To develop an efficient method for characterizing posterior shape distributions in image segmentation.
• To address limitations of existing methods, such as lack of confidence measures and getting stuck in local optima.
• To provide multiple segmentation solutions, especially for multimodal distributions.

Main Methods:

• Proposed an efficient pseudo-marginal Markov chain Monte Carlo (MCMC) sampling approach.
• The method draws samples from posterior shape distributions for image segmentation.
• Computation time is independent of training set size, ensuring scalability.

Main Results:

• The approach effectively characterizes posterior probability densities in the shape space.
• It provides multiple potential solutions, capturing different modes of multimodal distributions.
• Demonstrated promising results on both synthetic and real image datasets.

Conclusions:

• The pseudo-marginal MCMC method offers a robust solution for image segmentation with shape priors.
• It improves upon existing methods by providing statistical confidence and handling multimodal distributions.
• The method is computationally efficient and scalable for large datasets.