PROJECT 1:

1

• PROJECT 1
• Voronoi Probability Maps for Seed Region
  Detection in Abdominal CT Images

PROJECT 2 Kidney Seed Region Detection in
Abdominal CT Images
2
By Nicholas Cooper, Northern Kentucky
University Maureen Kelly, Loyola
University Chicago Jacob Furst, DePaul
University Daniela Raicu, DePaul
University REU Medical Informatics eXpericence
(MedIX) 2008 DePaul University Northwestern
University University of Chicago Thursday,
August 22, 2008

• Voronoi Probability Maps for Seed Region
  Detection in Abdominal CT Images

PROJECT 1
3
Topics of Discussion

1. Project Goals
2. Why Multiple Organs?
3. Purpose for Radiologists
4. Why Segment the Liver and Spleen Together?
5. Challenging Aspects of Multi-Organ Segmentation
6. Overcoming Challenges
7. Methodology
8. Results

4
Project Goals

• To create a robust method that identifies liver
  and spleen seed regions that can be used for
  multi-organ segmentation
• Voronoi Probability Maps (VPM)
• Largest Probable Connected Components (LPCC)

5
Why Multiple Organs?

• Propose an increased accuracy with multi-organ
  segmentation using seed regions
• Two weak individual segmentations could
  potentially result in an even more exact
  combination segmentation
• Allow for radiologists to examine several organs
  at once instead of just one at a time
• Better diagnosis of pathologies
• Treatment planning
• Anatomical structures study

Liver
Spleen
6
Purpose for Radiologists

• Check for diseases
• Liver
• Hepatitis (inflammation)
• Cirrhosis (nodules formation)
• Cancer
• Spleen
• Splenomegaly (enlargement)
• Asplenia (abnormal function)
• Spread of a known disease/pathology
• Is a particular treatment is working for a
  patient?
• Show condition after a abdominal injury

7
Benefits of Segmenting the Spleen and Liver
Together

• Share similarities that would allow for more
  accurate and repeatable segmentation
• Texture features
• Gray-level intensity values
• Practicality in the technical setting

8
Challenges

• Spleen share similar texture features/properties
  to that of the liver
• Gray-level similarity of adjacent organs
• Variations in spleen and liver margins/shape
• Absence of the spleen

Hey, where did the spleen go?!
Nick
9
Overcoming Challenges

• Create a method that is not based on a common set
  of parameters
• organ location
• patient position
• Create a method that relies on specific texture
  or intensity

Typical spleen location
Patient at a 45 angle
10
Methodology
11
Soft Tissue Region Identification

• Soft tissue is only displayed
• Fat, bone, and air are removed
• Regions are created in order to be classified

Original Image
Soft Tissue Regions
Soft Tissue
12
Texture Feature Extraction

• Co-occurrence matrix
• Distribution
• 9 Haralick descriptors
• Distance and direction
• Used to help identify soft tissue regions
• Differentiation between organs liver vs. spleen

Co-occurrence matrix
Pixel neighborhood
CT image
13
Candidate
negative
positive
Seed
Detection
Spleen Candidate Seed Detection

• Created liver and spleen classifiers
• Manually draw a polygon around the spleen/liver
• Creates positive (spleen/liver) and negative
  (non-spleen/non-liver) regions
• Result displays the regions in which the
  classifier declares to be spleen or liver
• Includes misclassified regions

14
Seed Extraction

• Get specific organ regions
• Spleen seed points are regions that ONLY contain
  the spleen, and same for liver.
• Eliminate the misclassified regions
• Seeds that are extracted are used as initial
  points for expanding the spleen/liver regions to
  achieve the completely segmented organ

15
Seed Extraction
16
Calculation of Average Seed Region Location
Finding average seed region location for both the
liver and spleen
Liver Candidate Seeds
Spleen Candidate Seeds
Average Seed Region Location
17
Implementation of Voronoi Probability Maps
Create Voronoi probability map based on average
seed region location
Liver Candidate Seeds
Spleen Candidate Seeds
Voronoi Probability Map
Probability becomes greater as the distance
between seed region and bisector increases
18
Implementation of Voronoi Probability Maps
Distance (d) between the bisector and the regions
in the Voronoi region of the organ of interest is
calculated, such that
d is then used to generate a probability, p, for
each region
Once probability, p, is calculated, each
connected component, C, is then given the value P
such that
19
Identification of Seed Regions
Finding seeds based on Voronoi probability map
using largest connected component and overlap
Liver Seeds
Spleen Seeds
Voronoi Probability Map
20
Identification of Seed Regions
The diagram of Voronoi probability map and
largest connected component approaches
21
Largest Connected Component and Overlap
IMAGES DISCARDED
Overlap between Spleen LPCC and Liver LPCC
Remaining Liver Seeds
Remaining Spleen Seeds
22
Results

• 19 patients
• 10-125 images per patient containing liver and
  spleen
• TOTAL 1,125 images
• Seed region overlap 176 images
• No Seed region overlap 979 images
• Of the 979, 85 of all the images contained all
  seed regions within the organ of interest

23
Conclusion

• Results show that VPMs and LPCC was successful
• Succeeded in circumstances in which other methods
  failed
• varying organ size
• texture similarities
• patient rotation

Thanks to Reeds mother!
24

• Kidney Seed Region Detection in Abdominal CT
  Images

PROJECT 2
By Nicholas Cooper, Northern Kentucky
University Maureen Kelly, Loyola
University Chicago Jacob Furst, DePaul
University Daniela Raicu, DePaul
University REU Medical Informatics eXpericence
(MedIX) 2008 DePaul University Northwestern
University University of Chicago Thursday,
August 22, 2008
25
Topics of Discussion

1. Project Goals
2. Why Kidneys?
3. Challenge Why Not Use Previous Method?
4. Overcoming Challenges
5. Methodology
6. Results
7. Conclusion
8. Future Work

26
Project Goals

• To create a robust, accurate method that
  identifies kidney seed regions that can be used
  for organ segmentation

Right Kidney
Left Kidney
View from behind
27
Why Kidneys?

• Detection, prevention, treatment
• disease
• One in nine Americans have chronic kidney disease
  (National Kidney Disease Foundation)
• Nephritis (inflammation)
• abdominal injury

28
Challenge Why Not Use Previous Method?

• Liver, spleen and kidneys do not exist within
  many of the same images
• Difficulties in distinguishing liver/right kidney
  and spleen/left kidney
• 2 of the same organ (right and left kidney)
• VPMs are based off of distance
• between regions
• and bisector
• Mis-identification
• Poor kidney
• candidate seed
• images

Liver and Spleen
Liver, Spleen and Kidneys
29
Overcoming Challenges

• Use kidneys high Hounsfield unit (HU) value to
  our advantage
• Use spine
• Kidneys are located on either side of the spine
• Use revised probability map approach
• Elliptical-shaped probability map (ESPM)

30
Methodology
31
Spine Extraction

• Located once for each patient using
• Many consecutive images
• Highest intensity values

32
Probability Map Construction

• distance (d1) between the center of the spine and
  outside edge
• distance (d2) between the center of the spine at
  x1, y1 and any pixel outside of the spine at x2,
  y2
• d1 and d2 are then used to generate a
  probability, p, for each pixel

p
33
Probability Map Construction

• Elliptical-shaped probability map (ESPM)
• Extended major axis of the spine ellipse
  separates the right and left kidney

major axis
ESPM
spine
34
Elimination of Non-kidney Intensity Values

• Kidney Intensity Ranges

35
Kidney Seed Extraction

• Apply elliptical-shaped probability map (ESPM) to
  each kidney image
• Check for overlap

Right Kidney
Left Kidney
36
Results

• 20 patients were tested
• TOTAL 2,375 images
• Seed Region Overlap 286 images
• No Seed Region Overlap 2,089 images
• Right kidney images
• Left kidney images
• Of the 2,089 images, 97.75 of the images were
  correctly identified as kidney

Correctly identified kidney images
Total kidney images

37
ResultsCombining Seeds

• Multiple organs
• each individual organ played a key role in
  segmenting the other organs
• Better accuracy
• Seeds can be used for region growing
• Complete the segmentation process

Liver, Right Kidney, Left Kidney, Spleen Seeds
(from left to right)
38
Conclusion

• Results prove that this method is very successful
• Accurate
• Reliable
• Time-efficient
• Comparable results on other patient data sets?

39
Future Work

• Region growing
• Extend to other organs

Liver (blue), Kidneys (green), Spleen (red)
40
The End
Any Questions??
Thanks to Reeds Mom again!