Learning-based stochastic object models for characterizing anatomical variations.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Computer Simulation

Background:

• Optimization of imaging systems relies on stochastic object models (SOMs) for computer simulations.
• Accurate modeling of human anatomical variations in SOMs is computationally challenging.
• Existing models often lack inter-patient variability, leading to phantom-specific bias.

Purpose of the Study:

• To develop a novel and tractable methodology for learning SOMs from volumetric image data.
• To generate realistic numerical phantoms that capture human anatomical variability.
• To overcome limitations of previous models in representing inter-patient and inter-organ variations.

Main Methods:

• Learning Geometric Attribute Distributions (GAD) from a broad patient population's volumetric images.
• Characterizing centroid relationships and organ shape similarity across patients.
• Generating stochastic objects by sampling learned GADs, reflecting anatomical variability.

Main Results:

• A novel methodology for learning SOMs and generating numerical phantoms was successfully developed.
• The method learns and models inter-patient and inter-organ anatomical variations effectively.
• A SOM of an adult male pelvis was computed, demonstrating the methodology with generated phantoms.

Conclusions:

• The proposed methodology provides a tractable approach to creating SOMs for medical imaging.
• This method enables the generation of diverse numerical phantoms reflecting real human anatomy.
• The approach can reduce phantom-specific bias in computer-simulation studies of imaging systems.