3D Kidney Segmentation from Abdominal Images Using Spatial-Appearance Models.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Renal Imaging

Background:

• Accurate kidney segmentation is vital for noninvasive renal function assessment using computer-aided diagnosis.
• Dynamic computed tomography (CT) imaging offers detailed anatomical and functional information for renal studies.

Purpose of the Study:

• To develop an automated framework for 3D kidney segmentation from dynamic CT images.
• To integrate discriminative features from current and prior CT appearances for improved segmentation accuracy.

Main Methods:

• Employed a random forest classification approach integrating first-order CT appearance, higher-order spatial models (including triple and quad cliques), and an adaptive shape model.
• Utilized a kidney shape prior model built from training data and updated dynamically using region labels and neighboring voxel appearances.
• Incorporated discriminative features to address CT image inhomogeneities.

Main Results:

• The framework demonstrated high accuracy in segmenting kidneys from dynamic CT data of 20 subjects.
• Quantitative evaluations using Dice similarity, percentage volume differences, and Hausdorff distances confirmed the approach's effectiveness.
• The method successfully segmented kidneys across multiple 3D scans acquired before and after contrast administration.

Conclusions:

• The proposed automated framework provides accurate 3D kidney segmentation from dynamic CT images.
• This approach holds significant potential for advancing noninvasive renal function assessment and computer-aided diagnosis systems.