Advancing image segmentation of vascularization in contrast-enhanced microCT data
Files
NACHTRAB_SEAN_04042400_2026.pdf
Open access - Adobe PDF
- 14.26 MB
NACHTRAB_SEAN_04042400_2026_APPENDIX_CollaboratiVessel.py.pdf
Open access - Adobe PDF
- 99.62 KB
NACHTRAB_SEAN_04042400_2026_APPENDIX_helpers.py.pdf
Open access - Adobe PDF
- 40.98 KB
NACHTRAB_SEAN_04042400_2026_APPENDIX_evaluation.py.pdf
Open access - Adobe PDF
- 63.04 KB
NACHTRAB_SEAN_04042400_2026_APPENDIX_pipeline.py.pdf
Open access - Adobe PDF
- 72.05 KB
Details
- Supervisors
- Faculty
- Degree label
- Abstract
- Tissue biopsy imaging has evolved from 2D microscopy to 3D modalities of ever-increasing resolution, with current contrast-enhanced microCT (CECT) datasets comprising tens of gigabytes and spanning thousands of voxels. The computational tools needed to extract coherent 3D structures from these volumes has lagged behind: existing solutions are either proprietary, only implement simple methods that do not generalize, or rely on deep learning expertise and associated data that is unavailable publicly and difficult to produce. This thesis focuses on CECT datasets of tumor vasculature on which previous standard image segmentation showed limited accuracy, with the main challenges being that vessels can span only a few voxels across, be disconnected, data can contain regions of high signal non vessel structures and contrast staining can be non-uniform. Therefore, in this thesis an open-source 3D Slicer plugin called CollaboratiVessel is proposed for microvasculature extraction from CECT data, constructed based on CECT datasets from murine tumors collected for the study of the antiangiogenic drug Pazopanib. Rather than requiring annotated data or computer science expertise, the end user is placed in the algorithmic loop by providing sparse input to an extraction pipeline that combines multi-scale vesselness filtering with ridge following and reconnection, ingesting and emitting standard formats compatible with downstream analysis software. CollaboratiVessel is evaluated on CECT datasets with manually annotated ground truths, and is shown to recover vascular structure with improved connectedness than currently used intensity-based thresholding methods, supporting future research on vascular changes and enabling work on structures with more challenging vasculature than was previously possible.