3D Interactive Segmentation With Semi-Implicit Representation and Active Learning.

Area of Science:

• Computer Vision
• Medical Image Analysis
• Geometric Modeling

Background:

• 3D geometry segmentation is complex due to intricate details and appearance variations.
• Existing methods like explicit and implicit shape models have limitations in handling topological changes or interactive manipulation.
• Current automatic segmentation relies on machine learning, often requiring extensive training data and limited by model discrimination power.

Purpose of the Study:

• To propose a novel semi-implicit shape representation for robust 3D geometry segmentation.
• To develop an efficient foreground-background delineation method using a cascaded classifier.
• To enhance segmentation accuracy through an adaptive, localized interactive scheme.

Main Methods:

• Introduced Non-Uniform Implicit B-spline Surface (NU-IBS) for adaptive, geometrically complex shape representation.
• Employed a two-stage classifier: Naïve-Bayesian for background elimination and pseudo-3D Convolutional Neural Network (CNN) for foreground identification.
• Integrated a localized interactive scheme for iterative accuracy improvement and applied a homogeneity constraint.

Main Results:

• The NU-IBS method demonstrated effective segmentation of complex 3D geometries.
• The two-stage classifier provided efficient and precise foreground-background delineation.
• Evaluated successfully on 3D Cardiovascular Computed Tomography Angiography (CTA) and Brain Tumor Image Segmentation Benchmark (BraTS2015) datasets.

Conclusions:

• The proposed semi-implicit NU-IBS method offers a robust solution for challenging 3D segmentation tasks.
• The combination of adaptive representation, cascaded classification, and interactive refinement significantly improves segmentation accuracy.
• The method shows promise for applications in medical imaging analysis, particularly with CTA and MRI datasets.