Intensity-Based Image Registration Intensity-Based
Image
Gustavo K. Rohde
Registration
42-431 Intro. Biomedical Imaging and Image Analysis
42-431 Intro.
Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
November 7, 2008
Overview Intensity-Based
Image
Intensity-based registration
Registration
Given digital images s(m, n) and t(m, n), find spatial
transformation f such that s(fx(m, n), fy(m, n)) and 42-431 Intro.
t(m, n) are similar. Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
Overview Intensity-Based
Image
Intensity-based registration
Registration
Given digital images s(m, n) and t(m, n), find spatial
transformation f such that s(fx(m, n), fy(m, n)) and 42-431 Intro.
t(m, n) are similar. Biomedical
Imaging and
Optimization
Image Analysis
min Ψ(s(f ), t, f )
Overview
f ∈C
Least squares
f : Rd → Rd: spatial transformation model
C spatial transformation class Cross-
Ψ(·, ·, ·): difference measure correlation
Mutual
Information
Continuous image representation Intensity-Based
Image
Registration
s(x, y) = c[p, q]ϕ(x − p)ϕ(y − q) 42-431 Intro.
Biomedical
p∈Z q∈Z Imaging and
where ϕ(x) is the basis function (sinc function, B-splines), Image Analysis
and c are the the coefficients.
Overview
Least squares
Spatial transformation Cross-
correlation
Mutual
Information
s(fx(m, n), fy(m, n)) = c[p, q]ϕ(fx(m, n)−p)ϕ(fy(m, n)−q)
p∈Z q∈Z
Three main components Intensity-Based
Image
Registration
42-431 Intro.
Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
Least squares Intensity-Based
Image
Ψ(s(f ), t, f ) = s(f ) − t 2
Registration
42-431 Intro.
Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
Least squares Intensity-Based
Image
Ψ(s(f ), t, f ) = s(f ) − t 2
Registration
Example
42-431 Intro.
Affine transformations f (x) = Ax + a in 2D. Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
Ψ(s(f ), t, f ) = (s (fx(m, n), fy(m, n)) − t(m, n))2
mn
with fx(m, n) = A11m + A12n + a1 and
fy(m, n) = A21m + A22n + a2.
Least squares Intensity-Based
Image
Ψ(s(f ), t, f ) = s(f ) − t 2
Registration
Example
42-431 Intro.
Affine transformations f (x) = Ax + a in 2D. Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
Ψ(s(f ), t, f ) = (s (fx(m, n), fy(m, n)) − t(m, n))2
mn
with fx(m, n) = A11m + A12n + a1 and
fy(m, n) = A21m + A22n + a2.
Optimize using
pk+1 = pk − ∇Ψ(s(f ), t, f ).
Cross-correlation Intensity-Based
Image
Let s[n] and t[n] = s[n − a] be two ”nice” signals to be
aligned. The goal is then to find a. This can be done via Registration
cross correlation
42-431 Intro.
a = arg max s ∗ tv[n] Biomedical
Imaging and
n
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
For larger signals, the computation above can be done
more efficiently using the FFT.
Example Intensity-Based
Image
Registration
42-431 Intro.
Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
Example Intensity-Based
Image
Registration
42-431 Intro.
Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
Normalized Cross-correlation Intensity-Based
Image
What if we have s[n] and t[n] = As[n − a], with unknown
A? Registration
Use normalized cross correlation:
42-431 Intro.
s[n], t[f (n)] Biomedical
Ψ(s, t, f ) = Imaging and
s[n] t(f [n]) Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
Example Intensity-Based
Image
Registration
42-431 Intro.
Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
Example Intensity-Based
Image
Registration
42-431 Intro.
Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
Mutual Information Intensity-Based
Image
Key problem
Registration
Matching images of different modalities.
42-431 Intro.
Example Biomedical
Imaging and
Matching CT and MRI.
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
Intensity values not linearly related Intensity-Based
Image
Joint histogram: no linear relationship between intensity
value, even when the images are optimally aligned. This Registration
is generlly the case when matching images of different
modalities. 42-431 Intro.
Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
Intensity values not linearly related Intensity-Based
Image
Joint histogram: no linear relationship between intensity
value, even when the images are optimally aligned. This Registration
is generlly the case when matching images of different
modalities. 42-431 Intro.
Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
As a consequence, sum of squared differences, cross
correlation, not likely to succeed.
Intensity-Based
Image
Registration
42-431 Intro.
Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
Intensity-Based
Image
Registration
42-431 Intro.
Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
Mutual Information Intensity-Based
Image
Let Ps(f),t(s(f ), t) denote the joint PDF of images s(f )
and t. The mutual information is given by: Registration
I(s(f ), t) = Ps(f),t(s(f ), t) log Ps(f),t(s(f ), t) 42-431 Intro.
Ps(f)(s(f ))Pt(t) Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
s(f ),p
Mutual Information Intensity-Based
Image
Let Ps(f),t(s(f ), t) denote the joint PDF of images s(f )
and t. The mutual information is given by: Registration
I(s(f ), t) = Ps(f),t(s(f ), t) log Ps(f),t(s(f ), t) 42-431 Intro.
Ps(f)(s(f ))Pt(t) Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information
s(f ),p
Optimization
min −I(s(f ), t)
f
Example Intensity-Based
Image
Registration
42-431 Intro.
Biomedical
Imaging and
Image Analysis
Overview
Least squares
Cross-
correlation
Mutual
Information