Frenet-Serret Frame-Based Decomposition for Part Segmentation of 3-D Curvilinear Structures.

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Anatomy

Background:

• Accurate segmentation of 3D curvilinear structures in medical imaging is difficult due to complex geometry.
• Limited large-scale datasets hinder the development and evaluation of segmentation algorithms.

Purpose of the Study:

• To develop a novel method for segmenting 3D curvilinear structures, using dendritic spine segmentation as a case study.
• To improve data-efficient learning, segmentation accuracy, and generalization for complex anatomical structures.

Main Methods:

• Introduced a Frenet-Serret Frame-based Decomposition to represent 3D curvilinear structures.
• Decomposed structures into a smooth curve and a cylindrical primitive, leveraging arc length parameterization.
• Created two datasets: CurviSeg (synthetic) and DenSpineEM (dendritic spine segmentation benchmark).

Main Results:

• Demonstrated exceptional cross-region and cross-species generalization on the DenSpineEM dataset (e.g., 94.43% Dice on mouse somatosensory cortex).
• Achieved strong zero-shot segmentation performance on unseen data (mouse visual cortex and human frontal lobe).
• Showcased improved performance on intracranial aneurysm segmentation (77.08% Dice on IntrA dataset), outperforming prior methods.

Conclusions:

• The proposed method accurately analyzes complex curvilinear structures across diverse medical imaging applications.
• The approach facilitates data-efficient learning and robust generalization for anatomical substructure segmentation.
• The released dataset, code, and models support future research in medical image analysis.