Title: MEDICAL%20IMAGE%20REGISTRATION%20BY%20MAXIMIZATION%20OF%20MUTUAL%20INFORMATION
1MEDICAL IMAGE REGISTRATION BY MAXIMIZATION OF
MUTUAL INFORMATION
Dissertation Defense by Chi-hsiang Lo June 27,
2003
2Medical Image Applications
- Computed Tomography (CT) gives anatomical
information
- Magnetic Resonance (MR) imaging gives anatomic
information
- Positron Emission Tomography (PET) gives
functional information
3Why Registration?
- Monomodality
- A series of same modality images (MR with MR.).
- A wider range of sequence types may be required.
- Images may be acquired weeks or months apart.
- Aligning images in order to detect subtle
changes in intensity or shape - Multimodality
- A combination of MR and CT with SPECT or PET.
- Complementary anatomic and physiological
information can be obtained for - the precise diagnosis and treatment.
- ExamplesPET and SPET (low resolution,
functional information) need MR or - CT to get structure
information. - In future, medical images (such as PET, SPET and
CT, MR) will be acquired - in a single machine.
4Registration Methods
- Landmark-based
- Based on identification of corresponding point
landmarks or fiducial marks in two images. - Accurate, but inconvenient, and cannot applied
retrospectively. - Labor-intensive and their accuracy depends on
the accurate indication of corresponding - landmarks in all modalities.
- Surface-based
- Corresponding surfaces are delineated and a
transformation computed that minimizes - some measure of distance between the two
surfaces. - Segmentation needed, and surfaces are not easily
identified in functional modalities(PET). - Voxel-based
- Use the intensities in the two images alone
without any requirement to segment or - delineate corresponding structure.
- It includes sum of squared intensity difference
(SSD), correlation coefficient (CC), - variance of intensity ratio (VIR), and mutual
information (MI). - Mutual information is widely used in
multimodality registration.
5Mutual Information Criterion
- Mutual information is applied to measure the
statistic dependence between the image
intensities of corresponding voxels in both
images, which is assumed to be maximal if the
images are geometrically aligned.
6Geometry Transformation
- Description
- The features (dimension, voxel size, slice
spacing, gantry tilt, orientation) of images, - which are acquired from different modalities,
are not the same.
- Image coordinate Transform
- From voxel units (column, row, slice spacing) to
millimeter units with its origin in the - center of the image volume.
- The formula is
where
- Affine Transform
- The affine transformation that transforms new
coordinates in image 1 into new coordinates
in image 2 is defined by matrix A.
- Corresponding Points Transform
7Registration by Maximization of Mutual
InformationInterpolation
- Nearest Neighbor Interpolation
- Tri-linear Interpolation
- Tri-linear Partial Volume Interpolation
8Registration by Maximization of Mutual
InformationOptimization
- Powells Direction Set Method
- Downhill Simplex Method
9Registration by Maximization of Mutual
InformationMulti-resolution
- Why Multi-resolution
- Methods for detecting optimality can not
guarantee that a global optimal value will be
found - Time to evaluate the registration criterion is
proportional to the number of voxels - The result at coarser level is used as the
starting point for finer level - Multi-resolution approaches
- Sub-sampling
- Averaging
- Wavelet transformation based
10Registration by Maximization of Mutual
InformationRegistration Flow
11Registration by Maximization of Mutual
InformationJoint Histogram Comparison
- Joint histogram before(left) and after(right)
registration - Left MI(A,B) 0.636287 Right
MI(A,B)0.805367
MR (x-axis) PET (y-axis)
12Registration by Maximization of Mutual
InformationNormalized Mutual Information
- Extension of Mutual Information
- Maes et. al.
- Studholme et. Al.
- Compensate for the sensitivity of MI to changes
in image overlap
13Registration by Maximization of Mutual
InformationUpdated Flow Chart Based on NMI
14Registration by Maximization of Mutual
InformationData Set
- Data set from The Retrospective Registration
Evaluation Project database by Vanderbilt
University. - Images for 9 patients
- 76 image pairs were available to be registered.
- 41 CT-MR pairs of 7 out of 9 patients
- 35 PET-MR pairs of 7 out of 9 patients
15Registration by Maximization of Mutual
InformationData Set (cont.)
- Image characteristics (9 patients)
16Registration by Maximization of Mutual
InformationThree New Methods Based on NMI
- A Binarization Approach
-
- A Non-linear Binning Based on Background
Segmentation and K-means Clustering -
- A Wavelet-Based Multi-resolution Approach
17Registration by Maximization of Mutual
InformationA Binarization Approach
- Two stages
- Region Growing
- Segmentation into Background and Foreground
- 2. Two levels Registration
- Binarized 2-bin images are input to the lower
level
18Registration by Maximization of Mutual
InformationRegion Growing
- Finding Starting Points
-
- Similarity Criteria
-
- For a region R of size N
- a pixel q may be in region R
- Connectivity
- 4-adjacency or 8-adjacency
- Stopping Rule
- No more pixels satisfy the criteria for
inclusion in that region
19Registration by Maximization of Mutual
InformationRegion Growing Implementation
- Finding Starting Points
- Easy to select a point as seed for background
- Similarity Criteria
- Threshold T can be extracted from the histogram
Typical histogram for CT image (left) MR image
(right)
20Registration by Maximization of Mutual
InformationRegion Growing Implementation
- Connectivity
- 8-adjacency
- Stopping Rule
- No more new pixels to be included in that region
Upper row CT image Lower row MR image Left
original Right binarized
21Registration by Maximization of Mutual
InformationTwo-level Registration
- Down-sampled binarized images as the input to
the first level - Result of the first level as the initial
estimate for the second level - The second level performs the registration of
full images, using Maximization of Normalized
Mutual Information
22Registration by Maximization of Mutual
InformationA Binarization Approach
A typical superposition of CT-MR images. Left
before registration Right after registration.
23Registration by Maximization of Mutual
InformationA Binarization Approach
Time required to perform registration for 41
CT-MR pairs .
24Registration by Maximization of Mutual
InformationA Binarization Approach
- CT-MR registration comparison with the others
- Median Error
- Maximum Error
25Registration by Maximization of Mutual
InformationNon-Linear Binning
- Description
- Background Segmentation (Region Growing)
- K-means Clustering
- Registration Using Maximization of NMI
26Registration by Maximization of Mutual
InformationK-means Clustering
Point a leaves cluster C1 and joins cluster C2
and centroids will be recomputed. X is the
centroid for each cluster before the movement of
a.
27Registration by Maximization of Mutual
InformationK-means Clustering (cont.)
- Convergent K-means Clustering Algorithm
(Anderberg)
1. Begin with an initial partition of the points
into k clusters. One way to do this is 1a. Take
the first k points as single-element
clusters. 1b. Assign each of the remaining N-k
points to the cluster with the nearest centroid.
After each assignment, recompute the centroid of
the gaining cluster. 2. Take each point in
sequence and compute the distances to all k
cluster centroids if it is not currently in the
cluster with the closest centroid, switch it to
that cluster, and update the centroids of both
clusters. 3. Repeat step 2 until convergence is
achieved that is, continue until a pass through
all the N points causes no new assignments.
28Registration by Maximization of Mutual
InformationNon-Linear Binning
- Background segmentation
- Segmented image as input to K-means Clustering
- Initially partition the image voxels into k bins
where - k 256.
- 1a. Put all the background voxels into bin 0.
- 1b. Calculate the step size for the other k-1
bins using - Each bin will be assigned all voxels whose
intensity falls within the range of its boundary. - 1c. Calculate the centroid of each bin.
29Registration by Maximization of Mutual
InformationNon-Linear Binning
- K-means Clustering (cont.)
- For each voxel in the image, compute the
distances to the centroids of its
current,previous, and next bin, if exists if it
is not currently in the bin with the closest
centroid, switch it to that bin, and update the
centroids of both bins. - Repeat step 2 until convergence is achieved
that is, continue until a pass through all the
voxels in the image causes no new assignments or
until a maximization iterations is reached where
the maximization iterations 500. - Two-level Registration
30Registration by Maximization of Mutual
InformationNon-Linear Binning
Histograms of a typical CT image (256
bins). Left linear binning method Right
non-linear binning
31Registration by Maximization of Mutual
InformationNon-linear binning
Histograms of a typical MR image (256
bins). Left linear binning method Right
non-linear binning
32Registration by Maximization of Mutual
InformationNon-Linear Binning
A typical superposition of CT-MR images. Left
before registration Right after registration.
33Registration by Maximization of Mutual
InformationNon-Linear Binning
Time required to perform registration for 41
CT-MR pairs .
34Registration by Maximization of Mutual
Information Non-Linear Binning
- CT-MR registration comparison with the others
- Median Error
- Maximum Error
35Registration by Maximization of Mutual
InformationWavelet-based multi-resolution
- Description
- Multi-resolution Improve optimization speed and
capture range. - The wavelet intends to transform images into a
multi-scale representation. - A wavelet can be created by passing the image
- through a series of filter bank stages.
36 Registration by Maximization of Mutual
InformationWavelet-based multi-resolution
- Implement
- Daubechies Wavelet filter coefficients ( DAUB4 )
- Three-stage WT on CT,MR,and PET images
- 41CT-MR pairs 35 PET-MR pairs
37Registration by Maximization of Mutual
InformationWavelet-based multi-resolution
A typical superposition of CT-MR images. Left
before registration Right after registration.
38Registration by Maximization of Mutual
InformationWavelet-based multi-resolution
A typical superposition of PET-MR images. Left
before registration Right after registration.
39Registration by Maximization of Mutual
InformationWavelet-based multi-resolution
Time required to perform registration for 41
CT-MR pairs .
40Registration by Maximization of Mutual
InformationWavelet-based multi-resolution
Time required to perform registration for 35
PET-MR pairs .
41Registration by Maximization of Mutual
InformationWavelet-based multi-resolution
- CT-MR registration comparison with the others
- Median Error
- Maximum Error
42Registration by Maximization of Mutual
InformationWavelet-based multi-resolution
- PET-MR registration comparison with the others
- Median Error
- Maximum Error
43Registration by Maximization of Mutual
InformationConclusion
- Time Comparison of the three methods on CT-MR
44Registration by Maximization of Mutual
InformationConclusion
- Time Comparison of the three methods (cont.)
45Registration by Maximization of Mutual
InformationConclusion
- Binarization Approach
- improve capture range
- reach a subvoxel accuracy without speed loss
- Non-linear Binning
- less dispersion in join-histogram
- improve accuracy and speed
- Wavelet based multi-resolution approach
- Accurate subvoxel registration
- Results are accessible via
- http//www.vuse.vanderbilt.edu/image/registration
/results.html
46Registration by Maximization of Mutual
InformationFuture Work
- Further Work on these approaches
- Noise in Region growing
- Coherence
- Bin size nonlinear binning
- More Wavelet filters
- Registration by Mutual Information
- interpolation methods
- optimization algorithms
- Non-rigid Registration
- non-rigid transformation needs more parameters
- rigid registration can be basis for non-rigid
- hierarchical strategy used in non-rigid
registration
47Thank you