Cross modality deformable segmentation using hierarchical clustering and learning.

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Machine learning in healthcare

Background:

• Accurate segmentation of anatomical objects is crucial for clinical applications.
• Existing automatic segmentation methods are often modality-specific and limited to certain deformable surfaces.
• A generic, adaptable framework for segmenting diverse anatomical structures across different imaging modalities is highly desired.

Purpose of the Study:

• To propose a novel, generic 3D surface segmentation framework applicable to various imaging modalities and deformable surfaces (open or closed).
• To develop a method that learns spatially adaptive appearance and shape from training examples.
• To demonstrate the framework's versatility through diverse clinical applications.

Main Methods:

• A hierarchical clustering algorithm using geometric and appearance features to group surface vertices into anatomical primitives.
• A cascaded boosting learning method to capture appearance characteristics of learned anatomical primitives.
• Clustering of training shapes to build multiple statistical shape models for incorporating non-Gaussian shape priors.

Main Results:

• The proposed framework successfully segmented both closed surfaces (liver in PET-CT) and open surfaces (distal femur in MRI).
• The method demonstrated adaptability to low-resolution and low-contrast imaging data (CT).
• This represents the first instance of a single segmentation algorithm applied to such diverse anatomical structures and imaging modalities.

Conclusions:

• The developed framework offers a generic and adaptable solution for 3D anatomical surface segmentation across different imaging modalities.
• The approach effectively learns and utilizes both appearance and shape information for robust segmentation.
• The successful application to liver and femur segmentation highlights the framework's broad clinical potential.