Improved Analysis of PET Images for Radiation Therapy: DeBlurring and Automated Segmentation Techniq

1
Improved Analysis of PET Images for Radiation
Therapy De-Blurring and Automated Segmentation
Techniques
Issam El Naqa1, Jeffrey Bradley1, Kenneth Biehl1,
Richard Laforest2, Daniel Low1, Joseph Deasy1
1Dept. of Radiation Oncology, 2Dept. of
Radiology Washington University, School of
Medicine St. Louis, MO 63110, USA Partially
supported by NIH grant R01 CA90445
2
Problems Objectives

• Motivation
• Higher uptake of Glucose in tumor cell
• Use of 18FDG (Fluorodeoxyglucose )
• Use of sequential CT and PET in the same scanner
  to obtain detailed anatomical and functional
  images
• Problems
• PET imaging degradation due to noise,
  attenuation, and scatter
• Labor-intensive 3-D contouring of high-uptake
  boundaries
• Proposed Methods
• Image sharpening via deconvolution
• Auto-segmentation to contour high-uptake tumor
  volumes

3
Methodology
Input PET image
Initialize Roughly select a contour around tumor
Is PSF known?
No
Apply Deformable model
Yes
Estimate PSF (Blind/Myopic method)
Deconvolution
Segmented tumor
4
Evaluation Data Set

• 19 PET/CT co-registered images with a
  pathological diagnosis of non-small cell lung
  cancer (NSCLC).
• Patients were imaged on a Siemens Biograph PET/CT
  scanner and had biopsy-proven peripheral lung
  cancers surrounded by normal lung tissue on the
  staging CT scan.
• Maximum uptake of 18F-FDG
• In normal tissues adjacent to suspected lesions
  0.84
• In Lesions 12.1
• PET images have a voxel size of 5.1 mm and a
  slice thickness of 3mm. The CT images have a
  voxel size of 1mm and slice thickness of 5mm.

5
Data Example
CT slice
PET slice
Tumor region of interest (TROI)
6
PET Deblurring

• The objective is to restore the true image due
  to imperfect imaging system. The system could be
  represented as

• where g is the observed function, f is the ideal
  true image, and h is a characteristic of the
  imaging system (PSF), n is additive noise, x is
  the spatial coordinate, and is the
  convolution operator

• We use, iterative methods based on the
  Expectation maximization algorithm (EM)

• The PSF could be incorporated directly if known
  or estimated if it is unknown (blind
  deconvolution) or partially known (myopic
  deconvolution)

7
Deblurring Example
Deblurred Tumor ROI
PSF
3x3
5x5
7x7
Support size
8
Tumor Segmentation

• Currently used method is 40 threshold of of the
  SUV by Erdi et al 1997. Other variations include
  tumor-to-background ratio (TBR) by Biehl et al
  2004
• Thresholding fails to capture tumor boundaries.
• Proposed methods is Deformable Models.
• Parametric deformable models (e.g., Snakes)
• Too sensitive to initialization and dont adapt
  with topological changes
• Geometric deformable models (e.g., Level set)
• 
  and
• Where V is a velocity function of contour
  curvature and image characteristics and is
  the level set function
• Less sensitive to initialization and adapt
  topologically

9
Classical Level set Example
Iteration 200
Iteration 100
Initialization
Iteration 500
Iteration 400
Iteration 300
10
Level set contours without edges

• Instead of computing gradients one can solve a
  minimal partition problem using a special case of
  the Mumford-Shah model
• Add spring force control

F2
F1
11
2D Example of classical level setGeneric image
12
2D Example of edgeless level setGeneric image
13
2D Example of classical level setPET
14
2D Example of edgeless level setPET
15
3D Example of edgeless level setPET
16
Future work

• We would like to improve the deconvolution
  process by extracting prior information using
  phantom studies
• Conduct a comparative study with other
  segmentation methods using the CT data and
  physician input as ground truth

17
Conclusions

• Two methods to improve the analysis of PET images
  for treatment planning purposes are presented
• Deconvolution is used to improve the visual
  quality of PETs by reducing the imaging
  degradations
• Tumor segmentation by level set method is
  proposed in opposite to simple Thresholding
• Preliminary results on deconvolution and
  segmentation are promising