Bayesian logistic shape model inference: Application to cochlear image segmentation.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Biomedical Engineering

Background:

• Accurate segmentation of anatomical structures in medical images is crucial.
• Existing methods often rely on parametric spatial transformations, limiting interpretability.

Purpose of the Study:

• To develop a Bayesian inference framework for parametric shape models in medical image segmentation.
• To provide interpretable results and quantify segmentation uncertainty.

Main Methods:

• A novel framework defining likelihood appearance and prior label probabilities using a generic shape function and logistic function.
• Expectation-Maximisation algorithm with a Gauss-Newton optimization for shape parameter inference.
• Application to cochlear structure segmentation using a 10-parameter shape model in CT images.

Main Results:

• Segmentation performance comparable to supervised methods on multiple datasets.
• Outperformed previous unsupervised segmentation methods.
• Enabled analysis of parameter distributions and quantification of segmentation uncertainty.

Conclusions:

• The proposed Bayesian framework offers an interpretable and effective approach for medical image segmentation.
• The method demonstrates robust performance in segmenting complex anatomical structures like the cochlea.
• It provides valuable insights into segmentation uncertainty and model influence.