Deep Learning for 3D Vascular Segmentation in Phase Contrast Tomography.

Area of Science:

• Biomedical Image Analysis
• Medical Imaging
• Machine Learning

Background:

• Automated blood vessel segmentation is crucial for identifying pathologies, but faces challenges due to complex vascular structures, data scarcity, and image quality.
• Hierarchical Phase-Contrast Tomography (HiP-CT) offers high-resolution 3D organ imaging (20μm/voxel, zooms to 1μm/voxel) revolutionizing the field.
• Existing machine learning methods require evaluation on novel imaging modalities like HiP-CT.

Purpose of the Study:

• To provide a foundational review of machine learning for vascular segmentation across various organs.
• To establish a robust baseline model for vascular segmentation using the HiP-CT imaging modality.
• To curate and utilize a new, high-resolution annotated kidney vascular dataset from the Human Organ Atlas Project.

Main Methods:

• Conducted an extensive literature review of machine learning approaches for vascular segmentation.
• Created a new annotated HiP-CT kidney dataset, verified by double annotators.
• Evaluated the nnU-Net framework on the HiP-CT dataset using tailored metrics for vascular structures.

Main Results:

• Achieved high Dice scores (up to 0.9523) and centerline Dice scores (0.82-0.88) on HiP-CT vascular data.
• Identified specific segmentation errors, including poor performance on collapsed large vessels and decreased connectivity in finer vessels.
• Highlighted limitations of standard metrics like Dice Similarity Coefficient (DSC) for fully assessing vascular segmentation quality.

Conclusions:

• The study sets a new benchmark for evaluating machine learning models on high-resolution HiP-CT data.
• Despite high scores, segmentation challenges in ex vivo HiP-CT data (e.g., collapsed vessels) require further investigation.
• Future work should focus on improving segmentation accuracy, especially for critical vascular structures, and developing more comprehensive evaluation metrics.