Demons: A Deformable Registration...
Transcript of Demons: A Deformable Registration...
![Page 1: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/1.jpg)
Demons: A Deformable Registration Toolbox
Presented by Tom VercauterenMauna Kea Technologies
Paris, France
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration1
![Page 2: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/2.jpg)
Goal of this Session
• Provide key concepts behind the demons algorithm
– Alternate optimization
• Show that the demons is more of a framework than an algorithm
– Can handle different application constraints
• Intuition on why demons is useful
– Largely available, easy to code, decent accuracy and efficiency
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration2
![Page 3: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/3.jpg)
HISTORICAL BACKGROUND
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration3
![Page 4: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/4.jpg)
Yet Another Framework?
Optimization
Energy Model
Transformation Model
Transformation
Source Image
Target Image
Difference Measure
Regularization Term
• Well-posed intensity-based framework existed, did we need another one?
• Devil in optimization & numerics
– N degrees of freedom on transformation N by N system to solve
– Huge on 3D data
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration4
![Page 5: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/5.jpg)
Demons Rationale ca 1995
• Trade theory for efficiency
• Transformation: Each pixel has a displacement vector
• Intuitive image forces to compute independent pixel displacements
• Regularize displacements by Gaussian smoothing
• (Thirion, 1995)
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration5
![Page 6: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/6.jpg)
Optical flow as Image Forces
• View images to register as in motion– Fixed image I(x,t)
– Moving image I(x, t+dt)
• Brightness consistency constraint– I(x+dx, t+dt) = I(x,t)
• Taylor expansion– I(x+dx, t+dt) I(x,t) + ∇xI.dx + ∇tI.dt
• Optical flow equation (e.g. Horn and Schunk, 1981)– ∇xI.dx = - ∇tI.dt = - ( I(x, t+dt) - I(x,t) )
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration6
![Page 7: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/7.jpg)
Optical flow as Image Forces
• No unique solution to optical flow equation– Aperture problem– Rely on the minimum norm solution– dx = (I(x,t) - I(x, t+dt) ). ∇xI / ||∇xI||²
– Stabilize it in a somewhat ad-hoc manner
• Intuition behind the image force– Demons push according to image gradient
if pixel value is lower than target value– Resp. opposite of the image gradient if
pixel value is higher than target value
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration7
![Page 8: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/8.jpg)
Classical Demons Algorithm
• Using more convenient notation– F, M resp. fixed and moving images
• Demons iterations– Initial displacement field s
– Compute image forces u to push M○s towards Fi.e. make M○(s+u) more similar to F
– s ← Gaussian smoothing on s+u
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration8
![Page 9: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/9.jpg)
Examples by Thirion ca 1995
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration9
Original Affine Demons
Courtesy D. Vandermeulen
Segmentation throughAtlas registration
Pairwise registration problem
Demons result
Courtesy R. Kikinis
![Page 10: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/10.jpg)
THEORETICAL INSIGHT
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration10
![Page 11: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/11.jpg)
Beyond Thirion’s View
• Demons are efficient but have shortcomings
– No strong theoretical analysis
• E.g. Not presented as energy minimization
– Difficult to generalize
• Similarity measures, transformation constraints, etc.
– No insight on convergence
– Why does it work?
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration11
![Page 12: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/12.jpg)
Alternate Optimization
• Several attempts to interpret the demons
– Modersitzki et al., Cathier & Pennec
– Refer to Darko’s talk for fluid demons
• Focus on Cathier & Pennec approach
– Hidden variable c demons as minimization of a global criterion
– E(c,s) = Sim(F,M○c)+dist(c,s)²+Reg(s)
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration12
1: Correspondence energy 2: Regularization term
![Page 13: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/13.jpg)
Classical Demons Revisited
• Demons iterations– s (displacement field) given
– SSD Forces u to optimizeEcorr( u ) = || F - M ○ (s+u) ||² + ||u||²
– Additive update: c ← s + u
– Diffusion-like regularization (Gaussian):s ← Id + K ∗ ( c - Id )
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration13
![Page 14: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/14.jpg)
INTRODUCING DIFFEOMORPHISMS
A geometric approach to extend the demons
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration14
![Page 15: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/15.jpg)
What’s with Diffeomorphisms?
• Deformable registration not only about intensity differences– Need to address the transformations
• Diffeomorphic registration– One to one, invertible, mapping– No foldings– Preserves topology– Essential for computational anatomy– Sound assumption if no privileged direction
• Not all problems benefit from diffeomorphisms!– Different applications different constraints– E.g. topology changes from tumor resections
• Should leverage a relevant representation of the transformation space
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration15
s
s-1
![Page 16: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/16.jpg)
Spatial Transformations
• Spatial transformations do not form vector spaces
• Addition: no geometric meaning– s1,s2 G s = s1+s2 G
• Natural operation: composition– s1,s2 G s = s1○s2 G, s: p , s(p)=s1(s2(p))
• Lie group structure
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration16
Rigid body Projective Diffeomorphism
○ =
+ = ??
![Page 17: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/17.jpg)
Walking on Lie Groups
• Inconsistency of additive optimization steps
– Ad-hoc or slows ← s + u
• Lie group Geometry
– Intrinsic steps
– No constrain s ← s ○ exp( u )
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration17
x
s G
u TId(G)
s○exp(u) Gx
s+u G
![Page 18: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/18.jpg)
Exponential & Diffeomorphisms: Flows of Vector Fields
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration18
exp( v/8 )
≈ Id + v/8
exp( v/4 )
= exp(v/8)²
exp( v/2 )
exp( v )
![Page 19: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/19.jpg)
Diffeomorphic Demons
• Vercauteren et al., NeuroImage 09
• Demons iterations– s (displacement field) given
– SSD Forces u to optimizeEcorr( u ) = || F - M ○ s ○ exp(u) ||² + ||u||²
– Diffeomorphic update: c ← s ○ exp( u )
– Diffusion-like regularization (Gaussian):s ← Id + K ∗ ( c - Id )
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration19
![Page 20: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/20.jpg)
Synthetic Example
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration20
DiffeomorphicDemons
AdditiveDemons
s
![Page 21: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/21.jpg)
Brain Web Registration
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration21
• Similar visual appearance
• Similar segmentation agreement – Dice index = 2 |X Y| / ( |X| + |Y| )
• Smoother and diffeomorphic transformations
![Page 22: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/22.jpg)
WORKING ON DIFFERENT GEOMETRIES
When square lattices are just not enough
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration22
First authorB.T. Thomas Yeo
![Page 23: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/23.jpg)
Cortical Surfaces Registration
• Human Cerebral Cortex
– Highly convoluted surface
– Folding pattern correlated with cytoarchitectonics & function
– Functional organization largely 2D!
• Modern MRI: Closed 2D mesh in 3D
• Population studies requires registration of these 2D meshes
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration23
![Page 24: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/24.jpg)
Aligning Cortical Folds Directly in the Volume is Hard
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration24
Affine transform of surfaces from one subject
mapped to another. Courtesy L. Zollei.
Cortical surface extraction example
![Page 25: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/25.jpg)
Spherical Demons Surface Registration Strategy
• Landmark-free
• Intrinsic use of the surface
• Spherical parameterization
• Intensities on surfaces are typically geometric attributes, e.g. curvature
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration25
Param. Param.Warp
![Page 26: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/26.jpg)
Extending the DiffeomorphicDemons to the Sphere
• Yeo et al., MICCAI 08
• Same diffeomorphism tools– Exponential of velocity fields
– Scaling and squaring
• Same SSD-based objective function– E(c,s) = Sim( F , M ○ c ) + dist(c,s)² + Reg(s)
• Requires workable definitions of– Transformation distances: dist(c,s)
– Transformation regularity: Reg(s)
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration26
![Page 27: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/27.jpg)
Choice of dist(s,c)
• Represent deformations as tangent vectors
•
•
•
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration27
![Page 28: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/28.jpg)
Spherical Demons: Step 1
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration28
• Invert 2 x 2 matrix per vertex• Independent of the chart
Chart
![Page 29: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/29.jpg)
Spherical Demons: Step 2
• Choice of regularization
• Approximate optimum (Trouvé ’98, Glaunès ’04)
– Spherical convolution
– Practice: Iterated smoothing with close neighbors• Reduce parallel transport burden
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration29
Parallel transport
![Page 30: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/30.jpg)
Automatic Parcellation
• 39 cortical surface meshes
• Spherical Demons vs. FreeSurfer
• Atlas-based parcellation into sulci/gyri
• Dice score, leave-n-out
• Statistical test– 24 better structures
on 70– No structure is worse
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration30
Percentage of improvement on Dice score
![Page 31: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/31.jpg)
USING OTHER SIMILARITY METRICS
Need more than SSD?
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration31
DTI work withB.T. Thomas Yeo
NMI work withMarc Modat
![Page 32: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/32.jpg)
Demons Similarity Metrics
• Previous slides relied on SSD and focused on optimization strategy or transformation spaces
• Optimization strategy also dependent on the similarity metric– Can we extend the demons framework to
something else than SSD?– Yes, see e.g. the work by Cathier, Guimond or
Stefanescu– In this presentation, focus on DTI similarity
• Not presented in the previous talks• Highlights the interdependence between similarity,
transformation and optimization
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration32
![Page 33: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/33.jpg)
DTI vs. Scalar Registration
• DTI captures white matter fibers information
– Pixel: 3x3 positive matrix (tensors)
• Tensors intrinsically linked to geometry
– Warping changes the geometry
– Tensor reorientation required
• Similarity metric
– Influenced byreorientation
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration33
![Page 34: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/34.jpg)
DTI Similarity Criterion• Local rotation component (Finite
Strain)
– Local linear approximation: s(p) ≈ Jac(p) . p
– Closest rotation:R(p) = (Jac(p) Jac(p)T)-½ Jac(p)
• Similarity using full tensor
– Sim(F, M, s) = ∑ dist( F(p), R(p)T M(s(p)) R(p) )2
– Log-Euclidean framework used
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration34
Sim =||log( ) - rot(log( ))||²log
![Page 35: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/35.jpg)
DTI Demons Forces
• Yeo et al., ISBI 08
• Requires derivative of the similarity
– FS uses matrix inverse square-root
• Complex closed-form
– Finite difference: poor man’s solution
• Happily we are in 3D!
– Use 3D cross-product
– dR=-R [ RT( tr((JJT)½)I - (JJT)½ )-1 ∑ (RT)i⊗(dJT)i]⊕
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration35
![Page 36: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/36.jpg)
DTI Results
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration36
Harmonic Energy
Image s
imilari
ty
![Page 37: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/37.jpg)
Sneak Peak at NMI Demons
• Collaboration with UCL– Marc Modat
• Uses mutual information derivative– Unlike SSD-based
metrics:No 2nd order info
– Conjugate gradient approach to control step size
• NMI demons– Better NMI than SSD
demons– Reverse
holds
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration37
Generate
Register
Multimodal Atlas
![Page 38: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/38.jpg)
LOG-DOMAIN DEMONS
Ensuring diffeomorphism is not enough? You also need the inverse transformation?
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration38
![Page 39: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/39.jpg)
Back to Spatial Constraints
• From previous slides: Demons is indeed a framework– As such there is room for improvements and
adaptations
• As an example, we present a symmetric extension of the demons– Symmetry is sound for subject-subject registration
• Other extensions not presented in this tutorial– Subject–template registration (with M. Sabuncu)
• Symmetry not advocated by probabilistic models
– Spatio-temporal constraints• Cyclicity by J.M. Peyrat• Incompressibility of heart muscle by T. Mansi
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration39
![Page 40: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/40.jpg)
Log-Domain Demons
• Vercauteren et al., MICCAI 08
• Log-domain walking– Replace s s ○ exp(u) by
– exp(v) exp(v) ○ exp(u)
– But this may not exist…
• Approximation with BCH formula– exp(v) ○ exp(εu) =
exp( v + εu + [v,εu]/2 + [v,[v,εu]]/12 + … )
• Lie bracket[v,u](p) = Jac(v)(p).u(p) - Jac(u)(p).v(p)
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration40
![Page 41: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/41.jpg)
Symmetric Log-Domain Demons
• Use easy inverse: s-1 = exp(-v)
• Iteration– Images I0, I1 and transformation s=exp(v)
given– Forward demons forces uforw
– Backward demons forces uback
– Update• v ← ½ ( Z(v,uforw) - Z(-v,uback))
– Regularize (Gaussian)• v ← Kdiff ∗ v
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration41
![Page 42: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/42.jpg)
Case Study: 3D+t Heart• Registration
with initial time– Forward:
Propagation
– Backward: Strain,Stabilization
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration42
Straindynamic
Courtesy Div. Imag. Sc., King’s College London
Thanks: Tommaso MansiNicolas Toussaint (INRIA)
![Page 43: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/43.jpg)
REACHING OUT TO THE WORLD
Dissemination and evaluation is not just a pain
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration43
Evaluation withArno Klein
![Page 44: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/44.jpg)
Reproducible Research
• Standard in core biomedical fields• What it means for us?
– Availability of algorithms– Availability of datasets– Common rules to compare the results
• Towards open-source– Diffeomorphic demons– DTI demons– Spherical demons– Symmetric demons (with F. Dru)
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration44
![Page 45: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/45.jpg)
Large Scale Evaluation
• Klein et al., NeuroImage 09– 16 groups involved: MKT, INRIA, LONI, Imperial
College, UPenn, UIowa, FMRIB, WellcomeTrust,…
– 14 registration softwares
– 80 manually segmented brains
– Over 45,000 pairwise registrations performed
– 8 different comparison measures: Dice
– 3 independent statistical tests
• See also EMPIRE 10 challenge (posters during breaks)
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration45
![Page 46: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/46.jpg)
Target Overlap Averaged Across All Registrations
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration46
![Page 47: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/47.jpg)
Average Rank
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration47
![Page 48: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/48.jpg)
CONCLUDING REMARKS AND DISCUSSION
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration48
![Page 49: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/49.jpg)
Take Home Messages
• For good image registration performance– Study application constraints– Pick most suitable tools w.r.t constraints
• Demons algorithms– Cover a large variety of constraints– Decent accuracy– Relatively easy to use
• Open source implementations• Few parameters• Computationally efficient
– Can be thought of as a fine baseline
• Evaluation is more than important• There is room for research!
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration49
![Page 50: Demons: A Deformable Registration Toolboxcampar.in.tum.de/twiki/pub/DefRegTutorial/WebHome/MICCAI... · 2010. 9. 24. · •From previous slides: Demons is indeed a framework –As](https://reader035.fdocuments.us/reader035/viewer/2022071500/611f7800fad2117a122e15b2/html5/thumbnails/50.jpg)
Acknowledgments
• Mauna Kea Technologies– Aymeric Perchant
• INRIA Asclepios– Nicholas Ayache, Olivier Clatz, Xavier Pennec
• MIT CSAIL– Polina Golland, B.T. Thomas Yeo
• Columbia University– Arno Klein
• UCL CMIC– Marc Modat, Sébastien Ourselin
2010-09-24MICCAI 2010 Tutorial: Intensity-based Deformable Registration 50
MIT CSAIL