Deep learning for 3D vascular segmentation in hierarchical phase contrast tomography: a case study on kidney

Area of Science:

• Biomedical image analysis
• Medical imaging
• Machine learning applications in healthcare

Background:

• Automated blood vessel segmentation is crucial for identifying pathologies, but faces challenges due to complex vascular structures, anatomical variations, limited datasets, and image quality.
• Hierarchical Phase-Contrast Tomography (HiP-CT) offers high-resolution 3D organ imaging, revolutionizing the field with its detailed resolution capabilities.

Purpose of the Study:

• To establish a foundation for automated vascular segmentation using HiP-CT imaging.
• To identify a robust baseline machine learning model for vascular segmentation on this novel imaging modality.
• To curate a high-quality annotated dataset for training and evaluating segmentation models.

Main Methods:

• Conducted an extensive review of current machine-learning approaches for vascular segmentation.
• Developed a meticulously curated training dataset of kidney vascular data from HiP-CT imaging, verified by double annotators.
• Evaluated model performance using the nnU-Net framework on high-resolution HiP-CT data, employing tailored metrics for vascular structures.

Main Results:

• Achieved high Dice Similarity Coefficient (DSC) scores (0.9523, 0.9410, 0.8585) and centerline DSC (0.82-0.88) in experiments.
• Identified persistent segmentation errors, particularly in large collapsed vessels (due to ex vivo nature of HiP-CT) and finer vessels with decreased connectivity.
• Observed higher segmentation errors at vessel boundaries, potentially interrupting vascular tree connectivity.

Conclusions:

• The study establishes a benchmark for vascular segmentation of HiP-CT data, highlighting the potential of this imaging technology.
• While achieving high DSC scores, limitations in segmenting collapsed and fine vessels indicate the need for advanced metrics beyond voxel-to-voxel concordance.
• Further research is needed to refine segmentation models to overcome challenges related to vessel collapse and connectivity in high-resolution ex vivo imaging.