MRI Image Segmentation for Brain Injury Quantification

1
MRI Image Segmentation for Brain Injury
Quantification

• Lindsay Kulkin
• BRITE REU 2009
• Advisor Bir Bhanu
• August 20, 2009

2
Overview

• Background
• Stroke Diagnosis
• Forms of Image Segmentation
• Process
• Gradient Relaxation Algorithm
• Connected Components
• K-Means Clustering Algorithm
• Results
• Conclusions
• Other ways to apply these forms of analysis

3
Background

• What is a stroke?
• Types
• Ischemic
• Hemorrhagic
• Causes
• Thrombosis
• Embolism
• Systemic Hypoperfusion
• Diagnosis
• Computed Tomography (CT) scan
• Magnetic Resonance Imaging (MRI)

Thrombosis occurs when a blood clot (known as a
thrombus) forms within the blood vessel and does
not break free.
4
Image Segmentation
Manual Segmentation
Automatic Segmentation
Original Image
5
Gradient Relaxation Algorithm
Find the initial assignment of probability (Pi)
and the mean neighborhood probability (qi)
Find the maximum kept constant (?imax) and the ?i
constant for all pixels
Construct a threshold image
Based on the valley of the histogram, segment the
first iteration and create a binary image
(threshold value 130)
6
Gradient Relaxation Algorithm
Original Image

• With each iteration, each new pixel value is
  determined based on the probability of its own
  pixel value as well its neighboring pixels (3x3
  window)
• While the program runs until it terminates,
  the threshold is automatically selected based on
  the histogram of the first iteration

First Iteration
Binary Image
Images provided by the Loma Linda University
Medical Center, 2007
7
Connected Components Analysis
Mask
Preliminary Scan
Pixel labels for Binary Image
Final Image

• Connected components identifies contiguous
  sets of connected pixels and is reapplied until
  the image cannot be segmented any further

8
Connected Components Analysis
Connected Components
Threshold Image
Inverted Image
Total pixels excluding background 11,610 White
10,940 (94.2) Large Injury 502 (4.32) Small
Injury 168 (1.45)
9
K-Means Clustering Algorithm
Original Image
Isolate each component by setting all other
pixels to zero
Select a k value as the initial cluster centers
and find the distance between each pixel and each
cluster center
Find the mean value of each cluster center
For all pixels, assign each pixel to its closest
cluster center. Find the mean value of each
cluster center until the cluster centers do not
change
10
K-Means Clustering Algorithm
11
Data Analysis
Manual Segmentation
Gradient Relaxation Algorithm
K-Means Clustering Algorithm
12
Conclusions

• Automatic segmentations vs. manual segmentation
• Both are effective and consistent
• Automatic segmentation is much faster
• These approaches can be applied to each MRI slice
  and the volume of injury can be obtained
• In the future, other forms of brain injury can be
  analyzed through the use of either
• The gradient relaxation algorithm/connect
  components analysis
• K-Means Clustering algorithm

13
Acknowledgments

• I would like to thank
• Professor Bir Bhanu for his guidance
• My graduate student advisor Benjamin X. Guan, as
  well as Angello Pozo and Giovanni DeNina
• The Center for Research in Intelligent Systems
  (CRIS)
• Jun Wang for this opportunity and for his support
• Loma Linda University Medical Center for
  providing the MRI images

14
Questions?