Progress Report
June 7th, 2005
Overview
Lines of work/activities remaining
Perturbation framework:
discussion
implementation
Experimental work plan
Summary
Aims of Current Work
Consolidating existing work on evaluation of models/NRR
repeating experiments more carefully - full experimental plan
properly controlled (in terms of mean displacement) perturbations
range of scales
stochastic
predictable E[average displacement]
no re-re-sampling error
Aiming at a paper covering the recent developments in model evaluation/NRR assessment
Impending Work/Possibilities
Shuffle as a metric - empirical proof/disproof; preferably a theoretical one
Evaluation of FFD-based registration algorithms
Ways in which segmentation can integrated with models, NRR, and the existing coupling between them
Impending Work/Possibilities - Ctd.
Future experiments, which ought to be run
The possibility of normalisation in model evaluation
Perturbation Framework
Details centralised to become more cohesive
Document on perturbation and code (
PDF
)
Current work aims to satisfy all needs mentioned in Section 2
Keeping Track of Displacements
Does the current implementation only trace points of choice?
Does it give a displacement magnitude range?
It turns out that Euclidean distance is used
Distance between the original and warped position is calculated
Mean, minimum and maximum are derived from a distance matrix
However, distribution is
not uniform
Other issues:
Centre affected more than edges, blurring, etc.
Problems with Current Framework
More problems observed:
Multi-point perturbation (as applied to IBIM data and its evaluation) breaks diffeomorphism
This was already demonstrated in experiments last year (registration using model complexity minimisation)
More warps lead to higher resampling error
How to quantify displacement when many warps are aggregated?
Some of the issues to be addressed are closely correlated
Current (Alternative) Approach:
add buffers to the image margins
Consider use of single-point CPS warps
Use small warps and apply them to sub-regions, handling one small section of the image at a time
apply perturbation within a series of windows so that the displacements are more homogeneous (
in progress
)
Problems Visualised #1
Problems Visualised #2
Problems Visualised #3
Diffeomorphism breaks with multi knot-point splines
Alternative Approach Visualised
How to Quantify Composite Displacements
Trace points through the sequence of warps
might be hard
probably the sensible approach to tackling the problem
Alternative for Displacement Quantification
Colour tracing
assign colours to image regions
for each pixel, find distance to nearest pixel with the same colour
this traces the movement of a given pixel
for each pixel this generates a matrix
matrix can be visualised as a map of distances
from this, one can hope/expect to have uniform distribution of values, having applied enough warps
Issues with Colour Tracing
Resampling error is problematic
blur
change in colours
Discrete number of colours
Issues with Colour Tracing - Ctd.
Solution: Handle smaller squares (chunks) with colours, where similar colours do not intersect
Another pitfall: the method does not handle large deformations well
Yet another issue: image dimensions
Need to make some colour image available for different sizes or generate it on the fly
It is then treated in conjunction with the real image so that displacements can be learned
The same warps are applied to the real and synthetic (colour) image
Issues with Colour Tracing - Ctd.
Further issues:
Efficiency: finding the closest colour means scanning all neighbours for the best (colour match) and closest (READ: nearest) match
Subjectivity: because of interpolation, there is a best match/shortest distance trade-off
Pending Experiments: Planning
Text and experiments in preparation (see
PDF
)
Both perspectives on the evaluation method are described
AAM evaluation
NRR assessment
Experiments Planned
Part 1
: Validation of Evaluation
needs reliable (trustworthy) perturbation framework
comparing Euclidean, shuffles
symmetric (shuffle distance in both directions)
plotting sensitivity
both brain and face data (indicates that the method has wide-range applicability)
Experiments Planned
Part 2
: Comparison with Overlap measure
based on ground-truth
needed as further proof (validation)
Experiments Planned
Part 3
: Evaluating registration algorithms
pair-wise, group-wise, and others
model-building framework can/will be explained separately
possibly involve ITK (Imperial College) registration algorithms
Summary and Ways Forward
Currently perturbation suffers from:
localisation problem
consistency
predictability of displacement
Need to discuss possible framework/s tomorrow
understanding the flaws of existing methods
proposing improvements
More experiments yet to be performed with careful attention