Prostate Shape Classification for Radiation Treatment Planning

1
Prostate Shape Classification for Radiation
Treatment Planning

• Erik Dam
• The IT University of Copenhagen, Denmark
• Dr. Stephen M. Pizer, Dr. Julian Rosenman, Gregg
  Tracton
• MIDAG, UNC
• Image Analysis
• November 2002

2
Prostate Cancer Radiation Treatment

• Phases
• Diagnosis (enlargement, PSA count, )
• Segmentation
• Planning
• Radiation treatment

3
Radiation Treatment Planning

• High dosage in cancer tissue leads to lower
  relapse rate
• Surrounding organs are fragile
• Optimal beam configuration depends on (a.o.)
• Size (field size, possibly hormones)
• Location (proximity to other organs)
• Shape(wrap around rectum)
• Boundaries(larger margin when fuzzy)

www.yoursurgery.com
4
The Goal

• Classification
• Classify organs conformations in groups that
  require similar beam configurations

5
The Classification Process

• Classes
• Sizesmall/medium/large
• Shapesaddlebag/not
• Boundarydiffuse/distinct

• Classes
• Sizesmall/medium/large
• Shapesaddlebag/not
• Boundarydiffuse/distinct

6
Two Data Sets

• Coronal

Sagittal
Axial
The entire data collection has 46 cases (11
saddlebags and 35 not)
7
M-Rep Shape Model

• Medial Representation
• Think Skeleton
• Grid of medial Atoms
• Each Atom define two sail vectors that give two
  boundary points, boundary is then interpolated
• Model is optimized to fit image

8
M-Rep Features

• Global
• Volume
• Surface area
• Local at each Atom
• Radius
• Row/Column bend

9
Classification

• Done in three simple steps
• Reduction of feature space dimensionality by
  Fisher linear discrimination
• Gaussian modeling of classes
• Maximal a posteriori probability used to
  determine class

10
Evaluation

• Leave one out

11
The Data Collection ...

• Originally 52 data sets
• But 6 were discarded 3 without prostate, 2
  without prostate segmentation, 1 due to too poor
  image quality
• Of the remaining 46 sets there are
• 11 saddlebag-shaped
• 35 normal (not saddlebag-shaped)

12
The Simpler Alternative

• For reasons of simplicity, we will skip the M-rep
  model in the remaining part of this presentation.
  
• Instead the concepts are illustrated with a
  simpler model with very simple features.

13
Naïve Engineering Alternative ...

• Do as follows
• Find the three points in each slice
• Sort slices by sumof differences L/R
• Pick the five best
• These differences squared gives a feature vector
  with 10 numbers

14
Results

• Comments
• Of the 11 saddlebag cases 8 were correctly
  classified
• Of the 35 normal cases, 2 were classified
  uncorrectly
• With real data, much better results can not be
  expected!