Departments of Radiology and Medical PhysicsDepartments of Radiology and Medical PhysicsUniversity of Wisconsin-MadisonUniversity of Wisconsin-Madison

Vessel Branch Segmentation of Phase Contrast Vastly undersampled PRojection Magnetic

Resonance Imaging

Kevin Johnson Kevin Johnson

Overview

• Background• Motivation• Theory

– Phase Coherence Mixture Model– Discrepancies

• Methods– Segmentation– Traces

• Results• Conclusions

Phase Contrast MR • MRI

– Fourier encodes atoms of non-zero spin– Complex Acquisition

• Phase Contrast MR– Encoded velocity in phase– 5-Images per Acquisition:

1H Density Vascular Velocity A/P Velocity L/R Velocity S/I

PC VIPR • Highly efficient acquisition• More accurate than conventional methods• High Resolution (0.5x0.5x0.5mm)

Motivation

• Phase image is difficult to analyze due to noise

• Segmentations allows visualization

• In neurological cases vasculature is complex– Need to separate branches

Frydrychowicz et al. ICVTS 2006. 340-342

Theory (Chung et al.)

• Phase Coherence

• May allow for automated separation

Mixture Model

Phase Coherence Thresholds

Theory in Practice

• Phase Coherence Histogram

• Simply due to the higher accuracy– In paper they had phase offsets – Shift PDF of Tissue– No effect on background

Methods

• Segmentation1) Determine phase coherence

2) Threshold phase coherence (T=0)

3) Threshold CD Image (T manually set)

4) Take intersection of images

5) Open image

6) Determine connected regions

7) Remove small regions

• Centerline1) Filter with LPF

2) Initialize by 3D watershed filling up to threshold

3) 3D Watershed with bridges built as connectivity

Results: Segmentation

Complex Difference Initial Threshold Eroded Image

Dilated Image Cleanup/Connectivity 3D Final Result

Results: Segmentation• Separate out individual components

• Remove components

ECA Branch ICA Branch Sap Vein Branch

No ECA/SAP Un-segmented

Results: Centerline

Initial Connectivity Connectivity (50% max)

Final

Conclusions

• Phase Coherence mixture model ineffective for PC VIPR

• Segmentation effective at separating components• Small vessels lost due to opening mask• Center line effective in most vessels• Failures in cases of complex flow