1. Segmenting 3D Branching Tubular Structures Using Cores
Yonatan Fridman, Stephen M. Pizer, Stephen Aylward, Elizabeth Bullitt
Medical Image Display & Analysis Group, University of North Carolina, Chapel Hill
The Purpose
Segment branching tubular structures such as blood vessel trees with no post-processing, and extract their complete branching geometry.
Tracking the core (red dotted line) of a branching object
Can you see the tube who’s core (red) was automatically extracted?
The Concept
Cores are height ridges of a graded measure of medial strength called medialness, which measures how much a given location resembles the middle of an object as indicated by image intensities.
The Method
March along the core of a tubular object
Detect bifurcations and resume marching along the two new branches
Determine when to terminate core following
This work is built upon other work done in MIDAG, including that of Aylward, Bullitt, Eberly, Fritsch, Furst, Morse, and Pizer.
Use a predictor-corrector marching method. At each step optimize medialness of a medial atom over location, radius, and orientation.
Template measuring medialness of a 3D medial atom. A derivative of Gaussian is at each spoke tip.
Core Tracking
The cornerness operator LuuLv
2D DSA of a portion of the head
Potential bifurcations are detected using an affine-invariant corner detector. False positives are then rejected, and the new branch is found in the direction of the weak spoke response.
Successive cross-sections at bifurcation
Branch Handling
Local maxima of cornerness, overlaid on DSA
Similar results are achieved in 3D
Spokes in tube cross-section
The 8 spoke responses
Core atom at bifurcation
Core Termination
At each step of core tracking, statistics are gathered on the medialness values of randomly positioned and oriented medial atoms
If the core’s medialness value is not outside the range of these random medialness values, it is terminated
Small breaks in objects or small areas of weak image information can be traversed
Axial Sagittal Coronal
Results on 3D head MRA
Top: MIP images of an MRA of the head
Bottom: Corresponding 3D cores model
Branching Angle (degrees)
Detected Branches (%)
Relative Branch Size
Branch handling success on synthetic images with a noise level roughly twice that of an MRA. Bars show percentage of test runs on which the branch was correctly traversed vs. branch angle and branch size.
Methods tested on both synthetic and clinical images
Tubular core following is extremely robust in the presence of image noise
Branch handling 98% with common MRA noise, but some misses in cases of high noise and narrow tubes
Core termination shows good results while correctly traversing small breaks in objects
Results and Conclusions
Acknowledgments: This work was done under the support of ONR MURI grant N as well as the partial support of NIH grants R01 EB NIBIB, R01 HL69808 NIHLB, and P01 CA47982 NCI.