Progress Report

March 7^th, 2006

About This Report

• This report is merely an appendage, bound to last week's report
• Discussion of points which require attention at present and in the future
• Eclectic report on progress made throughout the past week
• Optimistic outlook for 3-D extensions of the assessment algorithms
• Littler focus (breadth) on paper submission deadlines (MICCAI due on Friday)

Next Steps or Milestones Considered

• Entropy - application and results for real and synthetic data
• MICCAI draft on entropy-based NRR assessment
• 3-D extension of the algorithms at hand
• CVPR paper to be re-formed, possibly to fit the BMVC template
• MIUA submission (probably assimilated to ISBI paper)
• ISBI flights and related arrangements (Christie assists)
• Form 12 - minor yet important modifications

Volume Similarity and Shuffle Distance in 3-D

• Implemention of a rudimentary 3-D extension to the model assessment algorithm
• Framework will generalise to enable 3-D NRR assessment, using AAM's
• Needs to be further extended in order for more parameters (options) to work as they are resticted to 2-D
• Limitation: At the moment, we can only use shuffle distance with neighbourhood of size 3x3x3 (= 27 voxels). This is trivially extensible.
• Can handle Analyze-formatted, easily portable as to accept GIPL-formatted images (through conversions)

Experiments Foreseen

• For interesting experiments we need 3-D training data and synthetic volumes which have been derived from its model
• Can be distributed for greater speed across clusters (to be run overnight)
• The algorithms are quicker than initially imagined
• For data which is composed of 256x256x51 voxels, 37 training images, and 100 synthetic images, it can be a matter of days to get results
• Kola, Tim and Vlad are able to assist with data alignment, models, and synthesis
• 3-D C++ tools are tricky to handle and hard(er) to master

3-D Model Assessment

• Raw 3-D data and model-generated data are needed
• Kola has some aligned 3-D data and is investigating model building at the moment
• For the experiments to be viable, probably need around 20 training volumes and at least 100 volumes, which have been generated by the appearance model
• In reality, we have a training set with at most 37 images (IBIM), but could use Wellcome Trust data in the future (roughly 200 images)
• Can use models to generate a number of test volumes, but models with a large training set are laborious (if not impractical at present) to build

3-D Model Assessment - Ctd.

• Various open-ended questions arise
• How many volumes should be generated from the model?
• How should one determine the parameters to use in building model and which to vary?
• Could investigate just the registration and build models in a consistent manner
• Generation from the model for parameters drawn from a random distribution
• How many modes, how many standard deviations to permit?
• Must be careful in choosing these parameters as experiments require considerable amounts of time and computer power

3-D Model Assessment: Thoughts

• At this early stage, a single model with some arbitrary parameters might do
• The algorithm is in its proof-of-concept stage, so we need to show complete an evaluation of one appearance model and could later evaluate another to make a comparison (e.g. pairwise versus groupwise)
• By changing the parameters C_i (their values preferably drawn from a normal distribution), can get a number of synthetic volumes
• Need these volumes to evaluate the model, by comparing them with the training set

3-D Model Assessment: Plan

• Will build a model with just 37 subjects
• 100 examples then derived from this model
• Perhaps 100 should be the very minimum to make the results credible (the 3-digit impact)
• Timescale: no compelling time limit on this

3-D Model Assessment: Plan - Ctd.

• Possible to obtain some comparable 3-D results within months
• Kola Busy with something else at the moment, so might set this chore aside altogether
• Generally in the process of generating 3D AAM's, so processing and generation of data could be done in tandem

MICCAI 2006 Draft

• A draft for MICCAI sent by E-mail on a few occasions
• Useful in case we are ready to introduce entropy, either now or later on
• The MICCAI draft is finally more 'mature'
• Once we agree on the formulation of entropy, we can:
  • fill in the 2 missing equations
  • run the necessary experiments (already have all the necessary matrix data)
  • produce the missing figures

MICCAI 2006 Draft - Ctd.

• About 1.5 pages left vacant for elaboration in the text (or added figures)
• Main weakness in the past was the inability to avoid misunderstandings and misconceptions
• The deadline is Friday
• ISBI, BMVC deadline, MIUA deadline looming