Abdominal multi-organ segmentation from CT images using conditional shape-location and unsupervised intensity priors

Area of Science:

• Medical Imaging
• Computer Vision
• Radiology

Background:

• Automated segmentation of multiple organs in abdominal CT scans is crucial for clinical analysis.
• Existing methods often struggle with variations in imaging conditions and inter-organ relationships.

Purpose of the Study:

• To develop an automated multi-organ segmentation framework for upper abdominal CT data.
• To create methods for constructing conditional priors and leveraging their predictive power for accurate segmentation.
• To ensure adaptability to various imaging conditions without supervised intensity information.

Main Methods:

• A novel framework for multi-organ segmentation incorporating inter-organ relationships.
• Development of prediction-based priors for modeling conditional shape and location.
• Introduction of an organ correlation graph to guide hierarchical segmentation.
• Pre-segmentation of predictor organs followed by hierarchical segmentation of remaining organs.

Main Results:

• The framework demonstrated effectiveness in segmenting eight abdominal organs (liver, spleen, kidneys, pancreas, gallbladder, aorta, inferior vena cava).
• High average Dice coefficients achieved: >92% for liver, spleen, and kidneys; ~73% for pancreas; ~67% for gallbladder.
• The method showed applicability across six imaging conditions from two hospitals without supervised intensity data.

Conclusions:

• The proposed prediction-based priors effectively improve segmentation accuracy in multi-organ CT analysis.
• The framework offers robust adaptation to diverse imaging conditions, crucial for clinical practice.
• This approach facilitates automated segmentation without reliance on supervised intensity information.