Geodesic Information Flows: Spatially-Variant Graphs and Their Application to Segmentation and Fusion

Area of Science:

• Medical image analysis
• Computational anatomy
• Neuroimaging

Background:

• Manual clinical annotations (e.g., segmentations, parcellations) are crucial but time-consuming, leading to scarce data.
• Propagating annotations between morphologically dissimilar images remains a significant challenge in medical imaging research.

Purpose of the Study:

• To develop a novel framework for propagating voxel-wise annotations between morphologically dissimilar images.
• To address the scarcity of manual annotations in clinical studies by enabling efficient information diffusion.
• To improve the accuracy of tasks like brain parcellation and tissue segmentation.

Main Methods:

• A novel framework utilizing a spatially-variant graph structure to connect morphologically similar subjects.
• Information diffusion and mapping through intermediate steps to propagate annotations.
• Application to categorical label fusion for brain parcellation and probabilistic tissue segmentation.

Main Results:

• Statistically significant improvements in categorical label fusion compared to state-of-the-art methods.
• Significant enhancements in probabilistic tissue segmentation for both synthetic and real data.
• Demonstrated effectiveness particularly in scenarios with large-scale morphological variability.

Conclusions:

• The proposed framework effectively propagates voxel-wise annotations across images with substantial morphological differences.
• This approach enhances the utility of limited annotated data for clinical studies.
• The method offers a robust solution for improving brain parcellation and tissue segmentation accuracy.