Progress Report
April 25
th
, 2006
A Roundup
Experiments with two data distribution have been resumed
MIUA deadline extended; completion due within a couple of days
Transactions on Medical Imaging review is available and further work is needed
BMVC paper submitted after the previous meeting; decision to be made within a couple of months
Specificity and Distribution Differences
Learn about the robustness of the measure
Observations were made which are worth discussing
Some interesting behaviour noted when measuring Specificity for differing distribution types (level of dis/similarity)
Outcome and conclusions based on repeated experiment, but nothing numerical to go by
If the two distributions which we compre are similar or highly dissimilar, Specificity drops as we increase the sample size (synthetic set 'dummy')
This is very much as expected (easier to find a nearby match)
Specificity and Distribution Differences - Ctd.
The in-between scenario (distributions have partial overlap) seems to lead to shaky curves that do not necessarily go down in value
This has an explanation, which probably justifies the use of more complicated measures than Specificity
This is one case where kNN might be of use
Two Similar Data Distributions
Two Similar Data Distributions - Ctd.
Specificity goes down gradually, but not monotonically
Isolated Distributions
Isolated Distributions - Ctd.
The value of Specificity is rather shaky
Distant Distributions
Distant Distributions Specificity
The value of Specificity goes down
Distribution with a Pinnacle
Distribution with a Pinnacle - Ctd.
The value of Specificity wobbles
Dimensionality and Specificity
Experiments involved:
arranging for the two distributions to have a few outliers and some obvious differences
then, investigating the effect of dimensionality on Specificity
No clear conclusions comes to mind from the experiments, but there are interesting observations to be made
In hindsight, we ought to have scaled the curves so that all heads (and maybe tails) are positioned at the same point in the plot
Distributions with Outliers
An example of a few outliers
Distributions with Outliers - Ctd.
Outliers in the distribution affect dimensionality and vice versa (dependence upon)
Studying the Effect of Gamma
Experiments were run to study the effect of gamma
kNN Approach in Distribution Similarity/Specificity
Study the effect of changing k in a kNN approach
Comparing two dissimilar yet partially overlapping distributions
As expected, more neighbours give a more stable measure, especially when the number of samples is low
kNN Variation Studies
Poster Printing and Corrigenda
Final poster had a coarse font resolution due to the printer or nature of the job delivered to it
2 previous printouts of the poster had fine-resolution fonts
All PDF's were generated in the exact same way with
pdflatex
as the principal tool
A few corrections:
One arrow needed reversing
A series of error bar was an accidental re-use of a plot before it has been corrected (for TMI and other subsequent submissions)
Other Tasks Pertaining to Experiments
Ongoing work concentrates on normalisation
The conceded submission to MICCAI 2006 (revolves around entropy but not normalisation) can be perceived as work in progress
Unrelated: Experimental Plan for normalisation sent by E-mail
It is an incomplete experimental plan for some of the remaining bits where we left off
The remainder (assessment in 3-D) is reliant on data being processed and passed on
Paper-related Tasks
BMVC augmented, submitted
Made the changes which we had discussed
MIUA is due shortly
TMI changes to be made by late May
Need to work on the thesis as well (while waiting for 3-D registrations to assess and investigating normalisation further)
MIUA Deadline
The MIUA deadline is just 2 days away
Sent an MIUA paper (PDF) that is quite polished early in the morning
The current length length (5 pages) indicates that we aim for an oral presentation rather than a poster
A rudimentary 1-page summary was sent as well, in PDF and LaTeX formats
TMI Submission and Review
The reply seems positive
Making sure we make the deadline (late May) will perhaps affect the possibility of inclusion in the special issue
There is a clear need to deal with the argument that we cannot assess a model-based NRR method using a model-based approach
Bill might have access to other groupwise registration results for this data
TMI Submission and Review - Ctd.
There probably won't be enough time to re-run 10 instantiations
Plenty of work would be involved
Something more cursory may suffice
Can also deal with the criticism directly, but support from some experiment/s would be valuable