Progress Report
December 12th, 2005
Rough Overview: Progress in a Nutshell
- Journal paper draft:
- Progress in terms of text, embellishments, clarifications
- Need more material on model building
- Communication with a guest editor
- Reading on clouds overlap in hyperspace
- Experiments - a complete shift to MATLAB was successful
- Initial and planned experiments
- Student talks abstract re-worked, re-submitted
TMI Special Issue on Validation in Medical Image Processing
- Needed to enquire with regards to suitability of our work
- Using our recently-submitted work (ISBI 2006)
- The work is yet unpublished on its own right
- The approach itself has been part of the work presented in 2 conferences (IPMI, BMVC), but never yet in a journal
TMI Special Issue - Ctd.
- Our work contains many elements which have been listed as those that define the special issue's scope, namely:
- Validation of image processing methods: registration
- Performance comparisons: different approaches for non-rigid registration
- Design of ground truth: expert-based
- New methods for validation without ground truth
- New validation criteria for estimating performances: 2 of them
- Guest editor described the work as "very relevant to the topics of the special issue"
TMI Special Issue - Ctd.
Full quote: "I think an expanded and updated version of the work you describe below (ISBI 2006) is very relevant to the topics of the special issue. I encourage you to submit it for review for consideration in the special issue. In order to do that, follow the regular IEEE TMI submission procedure, but also make a note for the editor in chief that it is to be considered for the special issue."
Overlap of Clouds in Hyperscape
- Reading and questioning involved
- Identified something valuable: HTML abstract:
"...techniques are broadly based on: (i) overlap based pattern synthesis which can generate a larger number of artificial patterns than the number of input patterns and thus can reduce the curse of dimensionality effect, (ii) a compact representation of the given set of training patterns called over lappattern graph (OLP-graph) which can be incrementally built by scanning the training set only once and (iii) an efficient NN classifier called OLP-NNC which directly works with OLP-graph and does implicit overlap based pattern synthesis..."
Incompatibilities with the Approach of Viswanath
- Suspicion that we can use none of the methods proposed in the paper from Viswanath
- We rarely need to increase the number of points that lie in space because we can generate as many synthetic example as we wish
- Building a model for overlap estimation is not applicable as we have one already
- The Next few slides show examples of the data in question, model synthesis in particular
Entropy and its Relationship to Graphs
- In attempting to estimate overlap, we are essentially dealing with point distances (i.e. graphs)
- Can gain insight from work on entropic graphs by Hufeza Neemuchwala and Alfred Hero
- What we are doing is related to the nearest neighbour graph approach to estimating entropy
- What is of interest is the calculation of entropy, not just in the context of one particular application as in the work above
C++/MATLAB Transition
- Wishing to become independent from VXL for more rapid implementation
- Code interpretation rather than compilation
- Experiments are slower to run, but can expolit multiple workstations
- The transition was a success, so there is now a full framework which is C-independent.
- Concerned with re-use of previous code, going back to the beginning of the year
Early Experiments
- Got the AAM viewer working again, so can produce larger (as large as desired) AAM models, unlike the small one in the presenatation. The issue was a bug in the code, which could not cope with long file paths.
- Finished the first large batch of experiments (about a dozen computers overnight)
- These demonstrate that we can reproduce the results we had in the ISBI paper, but do so in MATLAB
- Code includes the derivation of error bars and sensitivity, if needed
- It will be easier (thus quicker) to implement other distance metrics
- I am still getting the 'dip' at the start where the sample points on the graph are denser
Planned Nightly Experiments
- Our aim is to investigate a variety of methods for computing the overlap between data clouds
- Investigate/implement various distance metrics:
- Shuffle
- Euclidean
- Fast Euclidean
- Also run a set of experiments which compare the performance of modified definitions of Generalisation and Specificity
- This no longer makes them Generalisation and Specificity, but useful for the sake of the argument
Planned Nightly Experiments - Ctd.
- Other methods/approaches to aggregating/interpreting image distances:
- Case 1: Mean of all distances (accumulate distances between all points in the clouds) instead of minimum of distances (both Euclidean and shuffle)
- Case 2: kNN - varying values of k (both Euclidean and shuffle)
- Case 3: Entropy-based
- Case 4: Shuffle and Euclidean for the definitions we have been using thus far (minimum of distances)
- The Next few slides show examples of the data in question, model synthesis in particular
Warp Examples - #1
![No brain deformation](brain_no_deformation.jpg)
On the left: arbitrary unwarped image; right: synthesis from model whose training data was not deformed
Warp Examples - #2
![Very slight deformation](brain_very_slight_deformation.jpg)
On the left: arbitrary unwarped image; right: synthesis from model whose training data was very slightly deformed
Warp Examples - #3
![Slight deformation](brain_slight_deformation.jpg)
On the left: arbitrary unwarped image; right: synthesis from model whose training data was slightly deformed
Warp Examples - #4
![Average deformation](brain_decent_deformation.jpg)
On the left: arbitrary unwarped image; right: synthesis from model whose training data was deformed
Warp Examples - #5
![Large deformation](brain_larger_deformation.jpg)
On the left: arbitrary unwarped image; right: synthesis from model whose training data was deformed
Warp Examples - #6
![Very large brain deformation](brain_very_large_deformation.jpg)
On the left: arbitrary unwarped image; right: synthesis from model whose training data was greatly deformed
Warp Examples - #7
![Huge deformation](brain_huge_deformation.jpg)
On the left: arbitrary unwarped image; right: synthesis from model whose training data was hugely deformed
Student Talks Abstract
- Shelagh received a different abstract
- Previous abstract made poorer re-use of an existing paper (ISBI 2006 submission)
- View abstract as plain-text
Next Steps
- Journal paper draft
- Writing up
- Expansion of experiments
- Assemblage of text covering model-building aspects
- Further experiments with cloud overlap and distance metrics
- Absence January 20th-29th; no absence during Christmas