Group-wise Registration in NAMIC-kit Serdar K Balci (MIT) Lilla Zöllei (MGH)
Download
Report
Transcript Group-wise Registration in NAMIC-kit Serdar K Balci (MIT) Lilla Zöllei (MGH)
Group-wise Registration in
NAMIC-kit
Serdar K Balci (MIT)
Lilla Zöllei (MGH)
Kinh Tieu (BWH)
Mert R Sabuncu (MIT)
Polina Golland (MIT)
Robust Group-wise Registration
• Entropy based group-wise registration
• ITK implementation
• Empirical Evaluation
2
Background: Groupwise Registration
• Images
I1 ,, I N
Transforms
T1 ,, TN
• Transforms:
– Affine
– Non-rigid using B-Splines
3
Registering to the Mean of the
Population
max MI I n (Tn ( x)); ( x)
N
T1 ,,TN n 1
T6
T5
T4
T1
T2
T3
4
Groupwise Registration: Congealing
min H v I (T ( xv ))
V
T1 ,,TN v 1
L . Zöllei, E. Learned-Miller, E. Grimson, W.M. Wells III.
"Efficient Population Registration of 3D Data."
5
Congealing: Intuition
• If Gaussian
pv I (T ( xv )) ~ N v , v
max log Ν I n Tn ( xv ) ; v , v
V
N
T1 ,,TN v 1 n 1
I n Tn ( xv ) v
min
2
T1 ,,TN v 1 n 1
v
V
N
2
• If also pv Nv v ,
• Registering to the mean with LS metric
6
Implementation
• ITK classes
– Group-wise registration using congealing
• Variance
• Entropy
7
Results
Before
Affine
BS 4
BS 8
BS 16
Entropy
Variance
8
Overlap Measures
Ln Tn ( x)
N
n 1
N
N Ln Tn ( x)
n 1
9
Full Term Babies
Before
Affine
BS 4
BS 8
BS 16
BS 32
10
Pre Term Babies
Before
BS 8
Affine
BS 4
BS 16
11
Summary
• Implemented group-wise registration in ITK
– Congealing: Entropy based registration
– Affine and BSpline
– Multithreaded implementation
– *Bspline optimization
• Initial Evaluation
– A population of 50 subjects
– Used segmentation labels to evaluate
12
Ongoing Work
• Finding optimal parameters
• B-Spline mesh size,
• # of hierarchical levels
• Subsampling
• Quantitative comparison to other methods
• Pair-wise registration to the mean using MI
13
14
Congealing with Two Images
• Using Parzen windows
T1, T2 max
log GI n Tn ( xv ) I k Tk ( xv )
T ,T
1
v n
2
k
• As we only have two images
max log GI1 T1 ( xv ) I 2 T2 ( xv )
T1 ,T2
v
max I1 T1 ( x) I 2 T2 ( x)
2
T1 ,T2
• Pairwise registration using LS metric
15
Groupwise Reg. using Pairwise Reg.
• If we assume that images are independent given a
subject, representative of the population
arg max pI1 ,, I N T1 ,, TN arg max MI I i (Ti ( x)); I R (TR ( x))
T1 ,,TN
T1 ,,TN
iimages
TR
T1
TN
T2
T3
16
Registering to the mean
• We assume independence over images
arg max pI1 ,, I N T1 ,, TN arg max pi I i (Ti ( x)) ( x)
T1 ,,TN
T1 ,,TN
iimages
• and draw i.i.d. samples from each image
arg max
T1 ,,TN
iimages jspace
arg max
T1 ,,TN
pi I i (Ti ( x j )) ( x j )
log pi I i (Ti ( x j )) ( x j )
iimages jspace
arg max MI I i (Ti ( x)); ( x j )
T1 ,,TN
iimages
17
Groupwise Registration using Joint
Entropy
• Assume i.i.d over space, but don’t make any
assumptions about images
T1 ,, TN arg max
T1 ,,TN
p I1 ( x j ), , I N ( x j ) T1 ( x j ), , TN ( x j )
jspace
arg min H I1 (T1 ( x), , I N (TN ( x))
T1 ,,TN
• Estimating entropy of an N-dimensional distribution
is a challenging task
18
Results: Congealing with Entropy
Before
After (B-Splines ~20mm)
19