Multi-Atlas Segmentation for Abdominal Organs with Gaussian Mixture Models.

Area of Science:

• Medical Imaging
• Computational Anatomy
• Radiology

Background:

• Abdominal organ segmentation using computed tomography (CT) is crucial in medical imaging.
• Gaussian mixture models (GMM) are established for medical image segmentation, particularly in brain imaging.
• Applying GMM to abdominal CT segmentation is challenging due to anatomical variability.

Purpose of the Study:

• To evaluate an a posteriori framework integrating GMM intensity likelihood with spatial priors for abdominal organ segmentation.
• To quantify the contribution of GMM in abdominal CT segmentation algorithms.
• To establish a benchmark for large-scale automatic abdominal segmentation.

Main Methods:

• Manually labeled 100 abdominal CT images.
• Utilized 40 images for training spatial priors and GMM intensity likelihoods.
• Segmented 12 abdominal organs in 60 test images using the developed framework.
• Measured segmentation accuracy using Dice similarity coefficient (DSC).

Main Results:

• The integrated framework demonstrated a median improvement of 145% in segmentation accuracy.
• The GMM intensity likelihood significantly enhanced segmentation performance compared to spatial priors alone.
• The framework effectively utilized both spatial and appearance information from atlases.

Conclusions:

• The proposed framework offers an effective approach for abdominal organ segmentation in CT images.
• Integrating GMM intensity likelihood with spatial priors improves segmentation accuracy.
• This method provides a valuable benchmark for future large-scale automatic abdominal segmentation studies.