Progress Report
January 10
th
, 2006
Overview
Various experiments which are related to entropic graphs
Error bars in entropy calculation
Comparing Generalisation, Specificity, and entropy
Questionable (unrealistic) brain images in registration, i.e. improper training set
Draft paper (IEEE TMI)
Student Talks
Miscellany
Entropic Measures - General Progress
Reading on log expansion
It took a very long time to get the required numbers and produce some of the final curves (yet to be shown)
The computer clusters are still switched off (all of them), maybe until early February
Entropy curves - final combined plots to follow, error included
Entropic Measures - Parameter Tweaking
Got hold of 3 or 4 cluster computers at the School of Computer Science
They ceased to be used by someone who had run experiments on them throughout the entire holiday (Shapiro's student)
Ran experiments that I had intended to run for quite some time
Next figure shows the effect of varying Alpha
Alpha Weighting Experiments
The effect of varying the value of alpha
K Nearest Neighbours Experiments
Code written to accommodate for kNN
Investigating the effect of altering
k
(as in kNN)
The plots indicate that the choice of
k
is almost irrelevant
Low contribution due to inclusion of more neighbours
Must assume a large number of sample points in the data 'cloud'
kNN Experiments
The effect of varying the value of
k
in kNN
Appearance of Brain Data
Last meeting raised the concern that a certain brain from the training set looked unrealistic
Spent some time looking at all brains in the training set
In retrospect, the algorithm involves the following steps:
brains were initially affinely aligned
the brains were later non-rigidly registered (in VXL)
a groupwise MDL objective function was used, so should not be affected by a reference
The next slides contain two brains from the set we base our evaluation on and from which we build the models too
Example Brain #1
An example brain from the training set
Example Brain #2
Another example brain from the training set
Combined Figure with Error Bars
Errors bars for entropy calculated and propagated according to the formulation in MathWorld
What follows is a combined set of plots which correspond to evaluation using:
Generalisation, Specificity, (Shuffle 5x5, values inverted)
Entropy (
H(G[A->B])->H(G[A->A]), alpha=0.5
)
(Omitted) Tanimoto overlap ('complexity'-weighted alpha, omitted as much work is required)
Only the first instantiation is included in this evaluation
Comparison of Cloud Overlap Estimates
For some further completeness, included are two more figures
They illustrate the variation of Generalisation and Specificity with different shuffle radii
The plots still demonstrate the tight and innate relationship between the methods
Estimation of Cloud Overlap
A combined view of different assessment metrics
3 Graph Entropy Plots
The correct plots of entropy for various graphs; registration is progressively degraded
Specificity Plotted Separately
The response to degradation measured in terms of Specificity
Generalisation Plotted Separately
The response to degradation measured in terms of Generalisation ability
TMI Draft
Currently, expansion is restrained due to page limit
9 pages already, but information can fit comfortably if made more concise
Draft needs feedback from colleagues (some are still absent)
8 pages, payments required for any excess -
Draft (PDF)
Miscellaneous Events
Karl to allocate a Monday Meeting slot (compensating for Student Talks absence)
Alan - discussion on annotation
Decision on ISBI paper submission is due to be reported tonight
Submission of data for paper submission, registration