Fast automatic 3D liver segmentation based on a three-level AdaBoost-guided active shape model.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Computational Anatomy

Background:

• Accurate liver segmentation is crucial for diagnosing and treating liver disorders.
• Existing automatic segmentation methods struggle with liver shape variability, image artifacts, and tumors.
• A robust, rapid, and automatic method for 3D liver segmentation is urgently needed.

Purpose of the Study:

• To propose an automatic three-level AdaBoost-guided active shape model (ASM) for robust and fast liver segmentation in 3D CT images.
• To emphasize the detection of tumors during the segmentation process.
• To improve the accuracy and efficiency of 3D liver segmentation.

Main Methods:

• Utilized a three-level AdaBoost-guided active shape model (ASM).
• Employed AdaBoost voxel and profile classifiers for model initialization and surface identification.
• Incorporated a deformable simplex mesh with curvature constraints for refining shape fitting.
• Applied three registration methods: 3D similarity, probability atlas B-spline, and deformable closest point registration.

Main Results:

• Achieved an average segmentation time of 35 seconds.
• Demonstrated high accuracy with an average Dice Similarity Coefficient (DSC) of 0.94-0.96 across multiple datasets.
• Validated on public challenge datasets (3Dircadb1, SLIVER07, Visceral Anatomy3).
• Achieved DSC of 0.964 on SLIVER07 testing and 0.933 on Anatomy3 unseen testing datasets.

Conclusions:

• The proposed automatic approach provides robust, accurate, and fast liver segmentation for 3D CT datasets.
• AdaBoost classifiers effectively detect liver areas and initialize the shape model, reducing segmentation time.
• The method achieves accuracy comparable to state-of-the-art automatic ASM-based segmentation techniques.