Automatic liver segmentation in computed tomography using general-purpose shape modeling methods.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Computational Anatomy

Background:

• Accurate liver segmentation in computed tomography (CT) is crucial for numerous clinical applications.
• Existing segmentation methods vary, with shape representation being a key selection criterion.
• This research focuses on developing automatic liver segmentation using general-purpose shape modeling.

Purpose of the Study:

• To investigate and compare different shape-based methods for automatic liver segmentation in CT images.
• To evaluate the performance of single atlas, Active Shape Models (ASM), and Gaussian Process Morphable Models (GPMMS).

Main Methods:

• Employed shape information at varying levels of complexity, starting with a single atlas-based method.
• Utilized classic and modified Active Shape Models (ASM) with medium body shape models.
• Implemented generalized statistical shape models, specifically Gaussian Process Morphable Models (GPMMS), for advanced shape deformation analysis.

Main Results:

• The single atlas method yielded the poorest segmentation results.
• Active Shape Models (ASM) achieved a Dice coefficient above 55%, with some cases exceeding 70%, ranking second.
• Gaussian Process Morphable Models (GPMMS) demonstrated superior performance, achieving a Dice coefficient of 88.5%.

Conclusions:

• Generalized statistical shape models, like GPMMS, offer high accuracy for automatic liver segmentation, attributed to their ability to model complex shape variations.
• The achieved 88.5% Dice coefficient highlights the potential of these advanced methods, despite limitations in the training dataset size.
• The developed automatic method shows comparable results to dedicated liver segmentation techniques and can be optimized with smaller datasets using various kernel functions.