Title: Nonrigid Registration Using Regularization that Accommodates Local Tissue Rigidity
1Nonrigid Registration Using Regularization that
Accommodates Local Tissue Rigidity
D. Ruan, J. A. Fessler, M. Roberson, J. M.
Balter, M. Kessler 48th AAPM The John S.
Laughlin Science Council Research Symposium 08.
01. 2006
This work is partly sponsored by NIH Grant P01
CA59827
2Registration 101
- Registration studies how to deform one image to
look like the other(s). - Registration is useful.
- diagnosis, utilize radiotherapy, simulation.
- Classification
- rigid v.s. nonrigid feature-based v.s.
intensity-based monomodality v.s. multimodality.
34D Thorax CT
4So whats the problem?
- Observations
- Good match in general
- Smooth deformation
- Twisted/bending ribs (nonphysical)
Bone Warping
5Why? What has been done?
- Heterogeneous tissue type specific property was
not accounted for. - Bone structures should transform more
rigidly than soft tissues - Existing work
- Treats different regions independently Little
97,Huesman 03 - relies on precise segmentation
boundary issues - Empirical spatial varying filter Staring 05
- requires knowledge of stiffness map (hard
thresholding) - departs from the optimization framework
hard to control/analyze -
6We propose
- Stick with optimization!
- Use spatial varying regularization to
incorporate tissue specific priors -
-
- where T denotes the deformation field, and
the objective function E(T) is composed of a
data fidelity term which measures the intensity
match and a spatial variant regularization that
accommodates tissue type dependent elasticity
information.
7Regularized Registration
- Regularization
- Incorporates prior information on the solution
into the optimization setup - Adds a regularization energy to the dissimilarity
metric - Two (or more) antagonist goals Bayesian
interpretation - Often used to enforce geometric properties
-
8Conventional Regularization Design
- Geometric regularization depends on the desirable
property of solution a backtracking process. - Typical regularization choices
- Smoothness
- Topology preserving Karacali 03
-
for - Volume Preserving Rohlfing 03
9Objective Function Design
10Objective Function Design
Squared difference in pixel intensity
the squared norm for the difference
between reference image and deformed homologous
image when and are looked on as 3D
scaler fields.
11Objective Function Design
Homogenous smoothness Spatial varying rigidity
penalty
spatial varying
weight that specifies the tradeoff between
intensity match and rigidity property. We also
call it local stiffness factor.
regularization term that penalizes the
deviation of the local transformation from being
rigid.
12Design Detail 1local nonrigidity penalty
- Q1 How to quantify rigidity?
- Potential As affineness, volume/angle
preserving, condition number - We say The group composed of pure
translation/rotation and their compositions. -
- Q2 How to measure the deviation, what can be
compute efficiently? - Potential As residual error for approximation,
Det(Jacobiant), eigen values - We say rigid
is an orthogonal matrix
- sufficient and necessary condition.
- penalty expressed in terms of Jacobian
components, avoided SVD - can analytically differentiate, easy to evaluate.
13Design detail 2 stiffness factor
Fact Voxel Intensity (CT) number is highly
correlated with tissue type, hence a good
inference source!
Thorax CT histogram Theoretical CT Values Hsieh
03' Air -1000 Hu Fat Muscle
Bones 2501000 Hu
-10060 Hu
- We use a monotone increasing function to infer
the degree of desirable rigidity from CT
numbers.
h tanh simple sharp saturation
14Parameterize Deformation
B-spline parameterization (controlled by knot
distribution and coefficients)
where and
indicates the deformation direction.
A 1D transformation example
15Parameterize Deformation
B-spline parameterization (controlled by knot
distribution and coefficients)
where and
indicates the deformation direction.
A 1D transformation example
Knot
16Optimization method
- Optimization procedure utilizes
- Multi-resolution scheme Kybic 94
- Conjugate gradient descent algorithm
17Experiment Setup
- We register X-ray CT images acquired during
different breathing phases - Reference deep inhale breath-hold (80 tidal
breath) thorax CT - Homologous exhale
- 512512148 with voxel size 0.20.20.5 cm2
- During preprocessing, crop the reference image to
size 259175107
18Slice Views of Data Derived Stiffness Map
19Results Intensity Comparison
- Reference v.s. deformed homologous image
- Color Scheme
-
- Reference blue
- Deformed homologous image green
20Identity (no) Transform
Coronal View
Sagittal View
Axial View
21Affine Transform
Coronal View
Sagittal View
Axial View
22B-Spline Nonrigid Transform no penalty
Coronal View
Sagittal View
Axial View
23B-Spline Nonrigid Transform - Regularized
Coronal View
Sagittal View
Axial View
24Geometry Validation
- CT number is a good reference source we extract
geometry by thresholding CT number to review
structure - Tool AVS isosurface rendering
- Color Scheme
- Reference geometry blue
- Deformed homologous geometry white
25Geometry Comparison B-Spline nonrigid
w/o penalty
w/ proposed penalty
26A Closer Look
Conventional B-Spline
Regularized B-Spline
- Conventional B-Spline offers good local
intensity match - Twisted Ribs gt stuck in non-physical local
minimum, Ouch! - Regularization achieves desirable compromise
between intensity match and tissue-type-dependent
geometry preservation a soft driving force!
27Landmark Validation
- CT data, 11 patient, normal exhale/inhale
- 6 landmarks manually located for each dataset
- Mutual information (MI) used as data fidelity
metric Multi-modality generalization - Registration accuracy validated by RMS error of
deformed landmark location v.s. ground truth
position.
28Illustration of Landmark Data
29Validation Comparison w/ RMSE
- TPS (expert picked control pts) AAPM 03
- Multi-resolution B-Spline AAPM 04
- Regularized B-Spline
30Conclusion
- A new method accounting for tissue-type dependent
rigidity with regularization design. - As an additive penalty, the regularization acts
as a soft correcting force in bone structure and
relaxes in elastic regions. - Inference from intensity value avoids explicit
segmentation, robust to partial volume effect. - Design based on Frobenius norm is computationally
friendly. - Physical promising results.
-
31Discussion Future Work
- Visually determined landmarks lie in
high-gradient regions, could be biased! gt
Desires more generic validation tools - Analytical derivation
- motor controlled phantom
- Extension to incorporate direction dependent
(anisotropic) properties - What is no CT is available? Alternative inference
source (open)
32Thanks Please Smile!
33Hard Classification by thesholding
34Handle Sliding Effect? Much Harder
Bone Warping
Avoided w/ Compositing
35- Notations
- pixel location in vector form,
region of interest - reference image (in Hu)
- homologous image (in Hu)
- data (in)fidelity
measure, SSD/MI - penalty, penalizes the deviation
of local deformation from being rigid. - spatial varying stiffness factor,
3D scaler field. - deformation field, 3D vector
field.