Refined Segmentations in Multiscale Statistical Framework

1
Refined Segmentations in Multiscale Statistical
Framework

• Xiaoxiao Liu
• for Shape Statistics talk 3/5/2006

2
Outline

• Multiscale Framework of m-rep segmentation
• Statistic based atom level deformations
• Training atom level PGA
• Atom stage Fitting
• Statistical atom level segmentation results
• Discussions

3
Multiscale Framework of m-rep segmentation
4
Residuals

• Residual
• The residual can be broken up into two factors
• Change independent of its neighbors
• Interrelationship of it with its neighbors

p prediction
Assumption Neighboring residual is small enough
to be ignored.
5
Statistic based atom level deformations
This is a m-rep!But a difference m-rep.
6
Training

• Atom residual PGA on
• Visualize the residual PGA
• Pablo demo of atom PGA

7
Optimization
Optimizing by updating the coef vector
Procrustes Alignment aligned to figure Stage
resulting model
Atom neighbors
Translation(T),
Rotation(R), Scale(S)
8
Statistical atom level segmentation results

• Improvement from figure stage results
• Pablo demo
• Comparison
• Binary fits
• Figure stage fits
• Statistical atom stage fits

9
Continue from last time
Ground Truth
Figure Stage Result
10
Results
80 Bladder CT images Segmentation
11
(No Transcript)
12
Bladder
Atom stage result
Figure stage result
Axial Sagittal
Coronal
13
Discussion and Future work

• 1. Alignment is important for accurate and tight
  statistics.
• To do more accurate alignment based on more
  neighbors
• 2. A robust Atom stage is depend on a stable
  figure stage.
• To do Towards a more robust figure stage
  by better initialization.
• 3. Inter-patient segmentation study