WP1: Evaluation and Optimisation of Voxel-Based Registration

1
WP1 Evaluation and Optimisation ofVoxel-Based
Registration
Bill Crum,Computational Imaging Science
Group,Kings College London
2
Motivation and Overview
Non-rigid registration of MR brain images is a
foundation technology for IBIM.
WP1 is about comparing and understanding these
techniques in order to advance the state of the
art.
Today I will recall the pilot study (Crum et
al, MICCAI 2004). list the unresolved
issues. suggest future work.
3
Registration Methods
FLIRT Affine registration from Oxford. (part
of FSL, http//www.fmrib.ox.ac.uk/fsl/) (correlat
ion ratio)
B-Spline FFD registration using multi-level
B-Splines. (normalised mutual information)
Fluid single level viscous fluid registration.
(intensity cross correlation)
4
Homology
Sir Richard Owen (1804-1892)
Homologue The same organ in different animals
under every variety of form and function.
Our interpretation structures which can be
consistently labelled in a population of brain
images.
5
Labelled Image Data
8 Subjects gt 100 labels per subject T1-weighted
MRI 256x256x128 voxels
6
Aggregated Regions
Original labels were grouped to produce a test
set.
() number of labels used for each aggregate.
7
Registration Scheme I
Register all permutationsof 8 subjects to remove
reference bias.
Normal Operating Mode
8
Registration Scheme II
source
label-S
overlap
registration
transformation, T
Trilinear interpolation followed by 50
thresholding.
T(label-S)
target
9
Region Overlap Measure
overlap
P Fractional Regional Overlap
10
Guide to Results
11
Results 1 Major Structures
12
Results 2 Major Lobes
13
Results 3 Other Structures
14
Results 4 Sub-Cortical Structures
15
Why the Performance Differences ?
Algorithmic Differences FLIRT is a robust
low-dimensional (affine) registration. B-Spline
works across scales gt voxel, medium-high
dof. (This) fluid works at a single voxel scale,
high dof.
Image Preprocessing The images are reoriented and
corrected for intensity shading before
labelling. Some were subject to motion
artifact. Interpolation and artifacts likely to
affect fluid the most. But there may be other
implementation specific caveats.
16
Caveats
Labels are not always drawn only using image
features.
Labels are not all independent.
Labels are subject to their own errors.
An overall evaluation of performance based on
multiple labels should consider such interactions.
17
Future Work I
Obvious extensions to the pilot work more
labels more subjects more registration
algorithmsThis should be written up for a
journal publication.
But need to distinguish ourselves from recent
similar publications (e.g. Hellier et al) othe
groups working in similar areas (e.g. Maes et
al)And demonstrate added value.
Currently to do this we need some concensus on
the details e.g. specify image similarity
measures ? some additional registrations to be
run (BC and DR identified 10 good CMA scans)
18
Future Work II
A proper statistical analysis of label overlaps
to consider labelling error consider
hierarchical label relationships consider
neighbourhood label relationshipsCombine
overlaps from multiple labels to generate a
single figure of merit for a registration over a
set of regions.
This will advance the state of registration
evaluation allow a map of registration quality
to be constructed lead to a more
application-oriented publication
This needs someone to lead on overlap/labelling
statistics figures of merit for existing / new
registrations run label-to-label registrations
too ?
19
Future Work III
Advancing Registrationderive new classes of
transformation models and similarity measures,
optimised for capturing variation in specific
parts of the brain How ?By setting up a
registration pipeline ?By using different/hybrid
transformation models in different regions ?By
using different/hybrid similarity measures in
different regions ?Or simply by identifying the
best performing algorithm and optimizing it for
this task ?
20
Questions ?