1
 A New Automatic Approach for Cerebrovascular 3-D
Volume Segmentation with New MIP-Based Technique
for Validation

M. Sabry M., Charles B. Sites, Aly A. Farag,
Stephen Hushek, and Thomas Moriarty
2
Terminology
MRI Magnetic Resonance Imaging MRA Magnetic
Resonance Angiography MRV Magnetic Resonance
Ventriculargrams DSA Digital Subtraction
Angiography MIP Maximum Intensity
Projection VTK Visualizarion Toolkit
3
Why Cerebrovascular Segmentation?

• Accurate description of the vascular tree is very
• important to assist in clinical applications,
  e.g.,
• Surgical Planning.
• Quantitative Diagnosis.
• Monitoring Disease Progress.

4
Objectives

• Segment Arteries from MRA
• Segment Veins from MRV
• Validate Results

5
Background

• The histogram of MRA volume could be classified
  into three intensity regions
• (LOW) Fluid surrounding brain tissue, bone and
  the air.
• (Medium) Brain tissues including both the grey
  and white matter.
• 3. (High) fat adjacent to arteries veins a
  255.

a
6
Design

• We present a modified version of 3-D computer
  graphics
• Region-Filling algorithm to sweep the vascular
  tree from
• Seed Locations and segment the desired structure.
• Our Algorithm Consists of Two Basic Steps
• Image Seed Selection
• Definition of Vascular Tree

7
Image Seed Selection

• MRA images are acquired from the lower portion of
  the nasal cavity upward to the top of the skull.
  
• Since Carotid Arteries is the largest artery in
  brain, The seed image will be selected
  automatically, since it contains the largest high
  intensity cross sections of Carotid

• All Carotid voxels are marked parents. They
  will be used to retrieve children voxels
  later.

8
Vascular Tree Definition

• Each voxel in MRA volume is the center of a cube
  of length 3 voxels.
• Each parent is compared with the 26-adjacent
  neighbor children.
• If the intensity level of the child falls in the
  range a-255, it is added to the
  Segmented_Volume .
• Parent / Child method eliminated the possibility
  of having unrelated tissues that get connected to
  segmented vascular tree.

9
How to Calculate a ?

• Lets define ? as the ratio of the segmented
  volume to the total volume.

• Our Research team found that when ? 1.5
  4.5, excellent segmentation results are
  achieved.
• If ? is greater than this range, a should be
  Increased.
• If ? is lower than this range, a should be
  Decreased.
• It is found that a 20 40 is a good initial
  estimate.

10
MRA Validation

• MRA validation is extremely difficult because of
  the vascular tree complexity.
• There exists no phantoms that mimics the
  complexity of the vascular tree especially those
  vessels whose diameters less than 1mm.
• Practitioners usually validate the resulting
  cerebrovascular volume by rotating the visualized
  3-D model by certain angles in order that the
  views at these angles agree with the available
  2-D images taken by DSA, X-ray, or MIP.

11
MIP (Maximum Intensity Projection)

• MIP is the state-of-the-art volume rendering
  technique to produce two-dimensional images of
  the vascular trees from 3D MRA data at different
  viewpoints.

• MIP Exploits The Fact
• Intensity of Vascular tree are higher than those
  of the surrounding tissues.

• Although the idea behind MIP algorithm seems to
  be easy especially at angles 0, 90, but the
  implementation is quite difficult to be general
  for any angle ?.

12
MIP Cont.

• How Does it Work ?
• Parallel virtual rays are applied to the volume
  at certain angle ?.
• The max. intensity voxel in the path of each ray
  is searched.
• The searched voxels forms a MIP image at angle ?.
  
• Rays are rotated around certain axis to find the
  rest of the MIP images.

13
MIP Cont.
14
MIP Cont.
The Rays are allowed to rotate around ? X-axis ?
Y-axis ? Z-axis In steps of 5 degrees
Z
X
Y
15
Results

• MRA data sets were collected using GE MEDICAL
  SYSTEMS Genesis Signa MRI system.
• MRA data are 256x256x117, 1 mm.
• MRV data are 256x256x84, 2 mm.
• Our algorithms are implemented using C under
  Unix on SGI-Onyx 2 Supercomputer.
• We used VTK as a visualization toolkit to render
  results in stereo on Immersa-Desk virtual reality
  system.

16
MRA Results
Side By Side Visual Validation for Patient 1

Before Segmentation (First Row)

? 0 ? 25 ? 45
? 90
After Segmentation (Second Row)
17
MRA Results Cont.
Side By Side Visual Validation for Patient 2

Before Segmentation (First Row)
? 0 ? 25 ? 45
? 90
After Segmentation (Second Row)
18
MRA Results Cont.
3D Visualization on Imersa-Desk


Patient 1
Patient 2
After Segmentation
19
MRV Results
3D Visualization on Imersa-Desk


Patient 3
20
Conclusion

• We have presented a robust technique for
  automatic segmentation of the vascular tree from
  MRA and MRV.
• We also presented a new MIP-based technique for
  validating the segmentation algorithm by
  projecting both the original and segmented volume
  at different angles and compare the resultant MIP
  images side by side.
• A radiologist and a neurosurgeon have validated
  our results successfully. The accuracy of the
  results has shown details of the vascular
  structure down to the limits of the MRI system.

21
(No Transcript)
22
CVIP_Medica

• I started the nucleus of a medical software
  package called CVIP_Medica.
• This library is supposed to offer an extensive
  practical algorithms and utility functions for
  medical researchers.
• The library is built entirely in C under Unix.
• Although it is a medical library, but there are
  common utilities that may be shared between
  different groups in the lab.
• The library is open source for CVIP students,
  such that each one can access and modify it to
  suit his goals.

23
Medica Design

• The library is designed to be AS GENERAL AS
  possible i.e, not specific to certain
  application.
• The function names are selected to reflect their
  usage.

Ex PGMVolume class void Read(int
READING) void SaveProcessedVolume() void
SaveProcessedVolume(char pathPrefix) void
SaveSegmentedVolume() void SegmentTree(int Rmin,
int Rmax, int SeedSliceNo) void
Histogram() void MIP_Z(int startAngle, int
endAngle, int stepAngle) void MIP_Y(int
startAngle, int endAngle, int stepAngle) void
MIP_X(int startAngle, int endAngle, int
stepAngle)
Ex PGMFile class char Read(char
filename) Void Save() int
SaveProcessedFile(char newfilename) int
Save(char newfilename, char image) char
Invert() char Segment(int Rmin, int Rmax, int
RENDER) int Histogram()
24
Medica Example 1

• Ex 1
• include PGMVolume.h
• int main()
• PGMFile file
• char image file.Read(/sw/people.kingtut/msa
  bry/patient1.pgm")
• image file.Segment(40, 200, WHITE)
• image file.Invert()
• file.Save(segment.pgm", image)

25
Medica Example 2

• Ex 1
• include PGMVolume.h
• int main()
• PGMVolume volume("/msabry/sarah.006","/msabry/tmp
  /va", "pgm" , 1, 117)
• volume.Read(FORMAL)
• volume.SegmentTree(40, 255, 113)
• volume1.MIP_Z(0, 90, 5)
• volume1.MIP_Y(0, 90, 5)
• volume1.MIP_X(0, 90, 5)

26
Future Work
Segmentation of the Vascular Tree from the
surrounding noise and tissues using one of the
classification techniques.
27
Achievements

• This work has been accepted in CARS

CARS 2002 Computer Assisted Radiology and Surgery
16th International Congress and Exhibition June
26-29, 2002 Palais des Congrès, Paris, France

• This work has been nominated for SGI Award
• for Excellence in Visualization

28
Thank You
Mohamed Sabry