A 3D Approach for ComputerAided Liver Lesion Detection

1
A 3D Approach for Computer-Aided Liver Lesion
Detection

• Reed Tompkins
• DePaul Medix Program
• 2008
• Mentor Kenji Suzuki, Ph.D.
• Special Thanks to Edmund Ng

2
Presentation Outline

• Background Information
• Prior Research
• Proposed Methodology
• Liver Segmentation
• HCC Candidate Detection
• Results
• Conclusions and Future Work

3
HCC Background

• Hepatocellular Carcinoma
• Primary Liver Cancer
• Prevalence varies drastically by region
• Few Symptoms
  
• Usually affects people with preexisting liver
  conditions

Background Information
4
HCC Background II

• Estimated to cause at least 372,000 deaths
  annually
• Other than CT imagery, difficult to detect
• Difficult / time consuming for radiologists to
  spot

Background Information
5
Project Background

• 2D Lesion Detector program, Candidate Finder
  1.0, written and tested in previous summer
• 
  
• Written in ITK open source, C/C toolkit
• 
  
• CandidateFinder both segments liver and attempts
  to detect tumor candidates
• 
  
• 100 Sensitivity
• Small Number of Test Cases
  
  

Background Information
6
Project Background II

• 2D Algorithm resulted in high number of false
  positives
• On 2D Data 24 FPs on average
• On 3D Data Hundreds of FPs
• Program not written using object-oriented
  techniques
• No way to view program intermediates

Background Information
7
Project Goals

• Develop a 3D computerized scheme for detection
  of hepatocellular carcinoma (HCC) in liver CT
  images
• 
  
  
  
• Modify and modularize existing liver lesion
  detection program

Background Information
8
Data Set

• 15 CT scans, with a total of 17 HCC tumors
• Contrast-enhanced CT images arterial phase
• Resolution 512 x 512 x (200 300)?
• Spacing of Pixels 0.67 mm, 0.67 mm, 0.62 mm
• Tumor centers identified by trained radiologist

Background Information
9
Prior Research

• Gletsos et al (2003)?
• Used gray level and texture features to build a
  classifier for use in a neural network
• Operated on 2D data, did not focus on HCC
  specifically
• Tajima et al (2007)?
• Used temporal subtraction and edge processing to
  detect HCC specifically
• Required multiple phases of CT liver images to
  work

Prior Research
10
Prior Research II

• Shiraishi et al (2008)?
• Used microflow imaging to build an HCC classifier
• Microflow imaging is not approved by FDA
• Used ultrasonography, not computer tomography
• Watershed Algorithm
• Huang et al (Breast Tumors)?
• Marloes et al (Brain Tumors)?
• Sheshadri et al (Breast Tumors)?

Prior Research
11
Proposed Methodology Liver Segmentation

• Not a liver segmentation project, but important
  to do it correctly
• Not terribly concerned with oversegmentation
• Method suggested by ITK manual

Liver Lesion
Liver Lesion
Proposed Methodology Liver Segmentation
12
Overview of Liver Segmentation
Proposed Methodology Liver Segmentation
13
Liver Pre-Processing
Proposed Methodology Liver Segmentation
14
Fast Marching Segmenter
Proposed Methodology Liver Segmentation
15
Geodesic Active Contours
Input Level Set
Edge Image
Proposed Methodology Liver Segmentation
16
Binary Image
Proposed Methodology Liver Segmentation
17
Binary Liver Mask
Two Different Binary Liver Masks
Proposed Methodology Liver Segmentation
18
Liver Segmentation Complete
Two Different Segmented Livers
Proposed Methodology Liver Segmentation
19
Proposed Methodology HCC Candidate Detection

• Pre-process segmented liver
• Apply watershed algorithm
• Eliminate/consolidate watershed regions
• Check distance from actual tumors

Proposed Methodology HCC Candidate Detection
20
HCC Candidates Pre Processing

• Filter out noise from image
• Alter pixel intensity
• Sharpen/define edges

Proposed Methodology HCC Candidate Detection
21
Segmented Liver with Gradient Filter Applied
Proposed Methodology HCC Candidate Detection
22
HCC Candidates Pre Processing II

• Calculate image statistics (used by watershed
  algorithm)?
• Apply a half-thresholder (try to eliminate
  uninteresting regions)?

Proposed Methodology HCC Candidate Detection
23
Watershed Segmentation Conceptual
Proposed Methodology HCC Candidate Detection
24
Watershed Segmentation

• In other words, the watershed algorithm locates
  the minimum intensity of regions, and keeps
  growing those enclosed regions until it
  encounters another growing region, or a boundary.
  
• 
  
  
• We used the watershed algorithm to find tumor
  candidates.

Proposed Methodology HCC Candidate Detection
25
QUIZ TIME!
26
QUIZ TIME!
My program attempts to locate HCC within liver CT
images. What does HCC stand for?
27
Results

• How do we define success?
• Centroid of 3D watershed region is less than 30
  mm away from location of tumor (as marked by
  radiologist)?
• Possible problem with this definition?

Results
28
Results II

• Average FPs 14.2 FP, Average Distance 12.6 mm

Results
29
Watershed Output
Original Image
Sigmoid
Watershed
Distance 0.47 mm
Gradient
30
Watershed Output II
Sigmoid
Original Image
Watershed
Gradient
31
Conclusions

• We have developed a 3D algorithm for the
  detection of HCC with 100 sensitivity on 15 test
  cases with a reasonable number of FPs.
• We have successfully translated a 2D algorithm to
  3D, with fewer false positives.
• We have successfully modularized the program,
  allowing intermediates to be output.

Conclusions and Future Work
32
Future Work

• Modify program to help detect cancers other than
  HCC
• Possibly integrate project with another student
  project
• Add a false positive reducer (MTANN?)
  

Conclusions and Future Work
33
Thanks!

• Thanks Again To
• Kenji Suzuki, Ph.D.
• Edmund Ng
• DePaul Medix Program
• And, of course

Contact Information rtompkins_at_gonzaga.edu
34
Any Questions?
Thanks To My Momma