Multilabel region classification and semantic linking for colon segmentation in CT colonography

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Anatomy

Background:

• Accurate colon segmentation in CT images is vital for applications like polyp detection and virtual colonoscopy.
• Challenges include adjacent air-filled organs and colon collapse, complicating segmentation.
• Extra-colonic components (DEC and AEC) interfere with precise colon delineation.

Purpose of the Study:

• To develop and evaluate graph inference methods for accurate colon segmentation.
• To effectively remove detached and attached extra-colonic components from CT colonography data.
• To improve the quality of colon segmentation, especially in challenging clinical scenarios.

Main Methods:

• Decomposition of 3D air-filled objects into regions.
• Classification of regions to identify colon versus non-colon areas.
• Graph inference for removing DEC using topological constraints and semantic information.
• Hierarchical conditional random fields and graph cut for AEC removal.

Main Results:

• The proposed graph inference method significantly improves colon segmentation accuracy.
• Effectively removes both detached and attached extra-colonic components.
• Outperforms purely discriminative methods, particularly in collapsed colon cases.

Conclusions:

• Graph inference provides a robust approach for accurate colon segmentation in CT colonography.
• The method successfully addresses challenges posed by adjacent organs and colon collapse.
• Enhanced segmentation quality supports improved computer-aided detection and virtual colonoscopy applications.