Progress Report

February 15th, 2005

AAM Application

• The AAM software works on the local machine
• Familiarity gained with almost all features...
• ...For example: NRR, tracking, annotation, global search
• Modelling code
  • Understanding of isbe_apm and isbe_qtools
  • iMorphics libraries, tools, etc.
  • Together with isbe_autoland, better technical orientation is provided

Contributions to Code

• Added hide/show all child windows to the software
• Created an option for model synthesis (more to come)
• Model evaluation tools (more to come)

Model Evaluation Framework - Overview

• Synthesise from model
• Take training set
• Measure 'distances' between pairs
• Construct 'distance' matrix
• Derive specificity and generalisability
• Let us look at each stage in turn...

Synthesis from Appearance Model

• Integrated with the C++ software
• 5 modes of variation stochastically-chosen
• Choose the parameters b_i stochastically
• Generate an arbitrary number of syntheses

Synthesis from Appearance Model: Visual

Example of face synthesis (±2.5 standard deviations at most)
 

Model Evaluation

• Take all examples generated by the model
• Take the training set
• Measure 'distance' for each possible cross-pairing
• This forms a matrix of distances
• Size: training-size X model-syntheses-size
• Generalisability and specificity derived from the matrix

Model Evaluation: Visual

Command-line mode
 

Model Evaluation: Visual

The matrix visualised
 

Distance Measure

• Distance is intensity-based
• No knowledge about correspondence (shape)
• Smoothing can be applied

Distance Measure: Visual

Difference images
 

Experiments with New Framework

• Data from pair-wise and group-wise registrations not yet available
• Data of real brains was used for testing
  • Evaluation of data that is only rigidly-aligned
  • Compared with affinely-aligned data
  • Comparison with unregistered data too

Results

• Model synthesis is assumed to be data as follows:
  • All affinely-aligned (distinct brains, randomly picked)
       Specificity = 15.5898, Generalisability = 12.8264
  • Synthesis only rigidly-aligned (distinct brains, randomly picked)
       Specificity = 17.2567, Generalisability = 14.1931
  • Synthesis unaligned (distinct brains, randomly picked)
       Specificity = 22.1745, Generalisability = 18.3797
• Conclusion: As data fits a (virtual) model poorly, specificity and generalisability increase

Registration of Protein Gel (Rogers)

• Slightly more advanced group-wise optimisations
• Results sent and experiments of future interest identified
• Experiments should become more intensive soon

Miscellany

• MATLAB Meeting
• MATLAB Documentation Project
• Banner for School of Computer Science home page