Progress Report

May 3rd, 2005

Overview

• Comparing registration evaluation methods
• Symmetrical shuffle distance
• Locally-orderless images
• CAWS

Registration Benchmark

• 1. Data
  • Data labels will be required (for overlap)
  • IBSR data will be used
  • 28 brains are publicly available
  • IBIM gives access to additional 37 data sets of schizophrenia subjects

Registration Benchmark

• What follows are the stages involved
• 1. Data
  • In 2-D, might be able to artificially enlarge the sample by treating multiple adjacent slices from single images as if they came from different images
  • 4 or so data sets are considered really bad
  • Nevertheless, Tim suggested using all available data

Registration Benchmark

• 2. Registration
  • Need to get a sufficiently good initial registration
  • Rueckert attempted label matching which gives good matching
  • Con: Doesn't incorporate any notion of labelling error
  • Con: Non-Diffeomorphic
  • Pro: Might be a reasonable starting point for perturbation

Registration Benchmark

• 3. Slicing the 3-D datasets
  • This involves selecting slices from the registered 3-D data
• 4. Deciding on perturbation method
  • There are various ways of achieving the necessary effects
  • Plenty of room for discussion on this topic

Registration Benchmark

• 5. Perturbing the data
  • Varying extents of deformation
• 6. Evaluate
  • Overlap measures
  • Shuffle distance

Registration Benchmark

• 7. Analyse
  • Plot values of one method against the other
  • Identify correlations and fix error bars
  • Derive sensitivity
• 8. Conclude
  • Can evaluation be carried out without ground truth?
  • Are results which require no ground truth more sensitive?

Registration At Present

• Bill wishes to submit one paper individually
• This will set the scene for a joint paper
• Initial steps:
  • Decide on images and the labels to use
  • Ascertain that our "gold-standard" registrations are satisfactory
  • Discuss perturbed images and labels

Registration Data

• Access to the entire set of data is need
• Sent by Kola to a Windows server
• IT Support yet to provide access privileges

Symmetrical Shuffle Distance

• Fully implemented but very slow
• Notes and thoughts in a L^AT_EX document
• Experiments make a comparison by visualisation
• Some conclusions are included in the document

Symmetrical Shuffle Distance

• Shuffle distances matrix, image 1 to image 2

Symmetrical Shuffle Distance

• Shuffle distances matrix, image 2 to image 1

Symmetrical Shuffle Distance

• Both matrices

Symmetrical Shuffle Distance

• Symmetrical shuffle distance matrix

Symmetrical Shuffle Distance

• All three

Symmetrical Shuffle Distance - Images

• Shuffle distance image, image 1 to image 2

Symmetrical Shuffle Distance - Images

• Shuffle distance image, image 2 to image 1

Symmetrical Shuffle Distance - Images

• Both images

Symmetrical Shuffle Distance - Images

• Symmetrical

Symmetrical Shuffle Distance - Images

• All three

Locally-Orderless Images

• Related to local histograms
• Initial implementation is in place
• Literature explains the idea in more depth

Locally-Orderless Images

• Main Points
  • 3 choices to be made in our case
  • Size of pixels (e.g. if multi-resolution approach is used)
  • Size of ROI
  • Number of bins in the histograms

Initial Experiments

• Region of size 3x3 (a window) in each of the 2 images
• Only 4 bins in the histogram (for 9 pixels)
• Region is of the same size in both images
• Sum of absolute differences between histograms is computed

Results

• No comparable experiments yet
• Below is one 'distance' image from a brain set

CAWS

• Implemented in Perl
• PHP (programming language of 2004) might be easier to maintain
• Limited discussion on the subject thus far
• An example CAWS setup