july 2010 image registration techniques, benchmarking, strategy surgical planning laboratory center...
Post on 22-Dec-2015
220 views
TRANSCRIPT
July 2010
Image RegistrationTechniques, Benchmarking, Strategy
Surgical Planning Laboratory
Center for Neurological Imaging
July 2010
Lidwien Veugen
Supervision by Dominik S. Meier, PhD
July 2010
Contents
- IntroductionImage Registration, 3D Slicer
- TheoryTransformations, Similarity Metrics
- BenchmarkingTime/Memory vs Iterations/Samples
- Registration Strategies
- Registration CasesBrains, PET-CT, EMPIRE10
July 2010
Introduction
Image Registration:
- Process of matching multiple image by optimal transformation
3D Slicer:
- Free Open Source Software program
- Huge amount of Registration Modules/Methods
July 2010
3D Slicer
July 2010
Theory Transformations
Mapping points from original spatial coordinates to new spatial coordinates: (u,v,w) = T{(x,y,z)}
Rigid TransformRotation + Translation
(u,v,w) = R*(x,y,z) + t
6 DOF
Affine TransformRotation + Translation + Scaling + Shear
(u,v,w) = A*(x,y,z) + t
12 DOF
July 2010
Theory Transformations
BSplineSpline: function defined piecewice by polynomialsCubic grid of moving control points describes deformation
3 DOF per control point
BrainsDemonWarpThirion + Maxwell: Image registration based on optical flow
Boundaries are semi-permeable membranes with effectors/demons
High DOF
July 2010
Theory Transformation
BRAINSFit- Rigid, Affine, BSpline
- Mutual Information
- 6/12/higher DOF
Expert Automated Registration- Pipelines: Rigid, Affine, BSpline
- MutualInfo + MeanSqE + NormCorr
- 6/12/higher DOF
Plastimatch- Pipeline: Rigid/Affine, BSpline(s)
- MutualInfo + MeanSqE
- 6/12/higher DOF
July 2010
Theory Similarity Metrics
Tells to what degree two images are aligned
Based on: intensity, landmarks
Mutual Information- Measure of the statistical dependence between two random variables:
Information about image A that is shared by B and vice versa
- Maximized if the two images are spatially aligned
- Based on Shannon entropy H: measure of intensity prediction
- Fast measure
MI A,B H B H B | A
H(A) pAa
a log pA a
July 2010
Theory Similarity Metrics
Normalized Cross Correlation- Based on cross correlation
- Maximized if the two images are spatially aligned
- Intra-patient + Intra-modality
- Time consuming
Mean Squared Difference- Summation of the squared differences between two images
- Minimized if the two images are spatially aligned
- Intra-patient + Intra-modality
- Time consuming
MSD1
mnA x,y B x,y 2
y1
n
x1
m
NCC A x,y A
y1
n
x1
m
B x,y B
A x,y A 2y1
n
x1
m
B x,y B 2y1
n
x1
m
July 2010
Theory Optimization
Optimization algorithm:
Tries to find a global solution to an energy function
- Gradient descent
- Statistical optimization
- Line search algorithm
- One-plus-one evolutionary
- Multiresolution
July 2010
Registration Accuracy
Subtraction
Fixed - MovingRegistered
Checkerboard
Alternating squares from
fixed and moving image
July 2010
BenchmarkingEffect of the amount of iterations and samples on CPU
time and memory for different modules/methods
Rigid: 4 methods
Affine: 7 methods
BSpline: 2 methods
Default: Samples = 10000, Iterations = 200
Iterations: 11 values, ranging from 25 to 20 000
Samples: 20 values, ranging from 25 to 10 000 000
July 2010
Benchmarking
Fast results with:
SPL Dell Linux Cluster of 50 computers
Creates log-file of every job Matlab
July 2010
Benchmarking Results
July 2010
Benchmarking Results
Time vs Iterations- Not much effect
- Increase: Brainsfit, Exp.Autom.
- Decrease: Multiresolution
- Constant: BSpline modules
Time vs Samples- Increase: All modules, except:
- Decrease: Exp.Autom. NormCr- 10 800 seconds (0.003% 13%)
- Rigid < Affine < BSpline
Memory vs IterationsNot much effect
- Increase: All modules, except:
- Constant: Brainsfit, Multires
- Lowest: 2MB; Highest: 155MB
Memory vs Samples- Increase: All modules, except:
- Decrease: Exp.Autom. NormCr
- Lowest: Rigid (10-100MB)
- Highest: BSpline (400-1300MB)
July 2010
Registration Cases
Slicer Registration Case Library
July 2010
Registration Strategies
Choice of Transformation
Modality, Subject, Inter/Intra, Part of body
Choice of Similarity Metric
Inter/Intra, Time/Accuracy
Focus
Time/Accuracy/Memory Sim.metric/iterations/samples
Fixed Image
Resolution/Contrast
July 2010
Registration Cases
EMPIRE10Evaluation of Methods for Pulmonary Image Registration 2010
= Challenge of International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI)
20 Pairs of chest CT scans: variety scanners, voxel size, breathing phase
Evaluation: Lung boundaries, Fissures, Landmarks, Singularities
July 2010
Registration Cases
July 2010
Registration Cases
EMPIRE10 - Registration Pipeline:1. Fast Affine Registration
2. Fast nonrigid Bspline Registration (grid = 7)
3. Fast nonrigid Bspline Registration (grid = 12)
4. Fast nonrigid Bspline Registration (grid = 17)
5. BrainsDemonWarp
July 2010
Registration CasesEMPIRE10 - Quality RegistrationSubtraction + MATLAB help in evaluation registration:
Median pixelvalue of absolute subtracted image: the lower the better
July 2010
Registration Cases
fMRI alignment to structural scan (T1)- Fixed: T1 scan (anatomical reference)- Moving: fMRI scan - Problem: Low tissue contrast, acquisition related distortions
T1fMRI
Registration based on ventricles only
July 2010
Registration Cases
Aging Mobility Study 2 year follow-up
- 2 Exams at different times: nonrigid (BSpline)
- Incorrect axis-info
- Fixed: MPRAGE- Moving: T2, FLAIR
MPRAGE
T1
FLAI RMPRAGE
T1
FLAI R
Exam 1 Exam 2
July 2010
Registration Cases
Inter-subject Normal brain MIDASexample
- Fixed: T1- Moving: T2, MRA - Interpatient: non-rigid (BSpline)
MRA
Patient 2
T1T2
MRA
Patient 1
T1T2
July 2010
Registration Cases
PET-CT Fusion 2- Intersubject: nonrigid
BSpline, BrainsDemonWarp- Fixed: CT-scan patient 1- Moving: CT-scan patient 2- Problem: Different posture
Patient 1 Patient 2
CT
PET
CT
PET
July 2010
Registration Cases
Brain Intersubject PNL-XNAT- Intersubject: nonrigid (BSpline)- Problems with (too much) BSpline
Patient 1
MRI
Patient 2
MRI
July 2010
Registration Cases
Brain Intersubject OrientationFlx- Intersubject: nonrigid (BSpline)- Fixed: T1- Moving: T2- Problems with nested transformations
Patient 1 Patient 2
MRIT1
T2
T1
T2
July 2010
- Fixed: Colin27- Moving: Patient- Orientation!
Registration Cases
Brain Intersubject Dartmouth
Montreal Neurological Institue:
Colin27 for group analysis in MRI studies
Colin27
MRI
Patient
MRI
July 2010
Acknowledgements
Finally, I would like to thank everybody from CNI for the possibility to do an internship here!
Thanks to my supervisor Dominik S. Meier, PhD
July 2010
Questions?
?