Progress Report

November 1^st, 2005

Overview

• MIAS-IRC abstract
• Progress on face models evaluation
  • Evaluation of degraded face models
  • Possibility of face registration (intra-subject)
  • Towards CVPR submission?
• TMI special issue on validation
• MIAS-Grid and model/registration assessment
• A few notes on overlap-based registration assessment

Revised MIAS abstract

• Latest revisions are continously submitted via the Web site
• Awaiting the last few 'touches'
• Official deadline was last Monday
• Aside: document on errors in measures and instantiations was sent by E-mail

Faces Evaluation - Extended

• Shown are Specificity and Generalisability of face models
• All built from increasingly deformed image sets
• Evaluation is currently limited to shuffle distance and Euclidean distance
• Essentially, shuffle with a 1x1/3x3/5x5/7x7 pixel neighbourhood

Faces Evaluation - Extended - Ctd.

• Only a single instantiation is involved unfortunately (due to time constraints)
• Curves are monotonic, but not smooth due to number of instantiations
• We involved 10 instantiations in the brain equivalent

Euclidean and Shuffle Distance

Generalisation of the face models, computed using Euclidean distance and shuffle distance

Euclidean and Shuffle Distance - Ctd.

Specificity of the face models, computed using Euclidean distance and shuffle distance

Face Models - Euclidean and Shuffle

• We have some compelling evidence, which supports the choice of shuffle distances rather than Euclidean distances
• In the case of face models, Specificity is shown to have better properties for the shuffle
• No need to bog down to the level of sensitivity plots in order to persuade the reader?

Face Models - Continued

• Recently finished evaluating face models using shuffle distance with a 3x3 pixels neighbourhood
• Final step is shuffle with a 7x7 neighbourhood and should be completed tomorrow (or Thursday at the latest)
• Results from Euclidean and 5x5 are left aside, but available for comparison

Recent Results - Shuffle Distance 3x3

Generalisation and Specificity of face models as these models degrade

Deformation Examples and Generalisation

Generalisation ability and an example deformed face from the model's training set
Large version

Deformation Examples and Specificity

Specificity and an example deformed face from the model's training set
Large version

Faces Evaluation - Possible Text Re-use

• Might decide to embed work on faces in a CVPR submission
• Should be able to re-use our BMVC 2005 submission
• BMVC concentrated on evaluating appearance models of brains and faces
• Method of perturbation and evaluation has become much more principled
• Can convince the reader using previous comparsions to overlap-based measures (ground-truth brain labels)

Registration Assessment Code and MIAS-Grid

• Project to produce an e-science workbench for medical image analysis
• Demonstrating the utility of the system with a series of use-cases
• Only needs to be pointed to the relevant code
• Wishing to re-create the processes in the analysis as workflows within the Grid's workbench

Grid Computing and Evaluation Code

• The registration/model assessment code can directly benefit from parallel workflows
• A framework I can envision:
  • We are given a set of N images
  • We have M such image sets
  • We need to build a model for each set among these M sets. That can definitely be done in parallel and there is no apparent dependency
  • We now proceed to evaluating M models and again there are no dependencies among the evaluation processes

Grid Computing and Evaluation Code - Ctd.

• All in all, the process in question need not be serial and it can be handled merely (not entirely) in parallel
• Can further refine speeds by treating sub-sets of data (chunks) and then aggregating the results
• This would be similar to things we have done in the past, such as deploying binaries in computer clusters and invoking them via SSH
• At extremity we used 30 units overnight to produce some urgently-needed results

Overlap-Based Measures: A Closer Look

• Measures of overlap concentrate on one among several expressions that are suitable for use
• Tanimoto coefficient: number of voxels that intersect, in both the test label and the existing label, divided by the union of the two
• The different variants of this expression involve changing the extent to which each label contributed to the aggregated overlap value
• Alpha is a weight for each segment

Overlap-Based Measures: A Closer Look - Ctd.

• Alpha is particularly important as we employ 4 types of overlap-based assessors, each of which uses a different assignment for alpha
• We are using Tanimoto as a generalised way of measuring intersection
• The framework, which accounts for a set of labels, can have one among a variety of expressions in its core, e.g. Jaccard, Dice...

TMI Issue on Validation

• Web site with official announcement
• Key dates:
  • Submission of manuscripts: 1^st February 2006
  • Acceptance/rejection notification: 1^st May 2006
  • Revised manuscripts due: 1^st July 2006
  • Publication Special Issue: November 2006
• Overlap-based validation is expected to be submitted on its own

Summary and Steps Ahead

• Final submission of MIAS-IRC abstract
• Evaluation of remaining face models
• Re-use of the BMVC 2005 submission
• Preparation of ISBI 2006 submission (extended MIAS-IRC abstract)
• Progress on thesis once deadlines are behind
• Possibly: MIAS-Grid interaction and involvement