Tomography Keegan Go Ahmed Bou-Rabee Stephen Boyd EE103 Stanford - - PowerPoint PPT Presentation

tomography
SMART_READER_LITE
LIVE PREVIEW

Tomography Keegan Go Ahmed Bou-Rabee Stephen Boyd EE103 Stanford - - PowerPoint PPT Presentation

Tomography Keegan Go Ahmed Bou-Rabee Stephen Boyd EE103 Stanford University November 12, 2016 Tomography goal is to reconstruct or estimate a function d : R 2 R from (possibly noisy) line integral measurements d is often (but not


slide-1
SLIDE 1

Tomography

Keegan Go Ahmed Bou-Rabee Stephen Boyd EE103 Stanford University November 12, 2016

slide-2
SLIDE 2

Tomography

◮ goal is to reconstruct or estimate a function d : R2 → R from

(possibly noisy) line integral measurements

◮ d is often (but not always) some kind of density ◮ we’ll focus on 2-D case, but it can be extended to 3-D ◮ used in medicine, manufacturing, networking, geology ◮ best known application: CAT (computer-aided tomography) scan

2

slide-3
SLIDE 3

Outline

Line integral measurements Least-squares reconstruction Example

Line integral measurements 3

slide-4
SLIDE 4

Line integral

◮ parameterize line ℓ in 2-D as

p(t) = (x0, y0) + t(cos θ, sin θ), t ∈ R

– (x0, y0) is (any) point on the line – θ is angle of line (measured from horizontal) – parameter t is length along line

◮ line integral (of d, on ℓ) is

d = ∞

−∞

d(p(t)) dt

Line integral measurements 4

slide-5
SLIDE 5

Line integral measurements

◮ we have m line integral measurements of d with lines ℓ1, . . . , ℓm ◮ ith measurement is

yi = ∞

−∞

d(pi(t)) dt + vi, i = 1, . . . , m

– pi(t) is parametrization of ℓi – vi is the noise or measurement error (assumed to be small)

◮ vector of line integral measurements y = (y1, . . . , ym)

Line integral measurements 5

slide-6
SLIDE 6

Discretization of d

◮ we d is constant on n pixels, numbered 1 to n ◮ represent (discretized) density function d by n-vector x ◮ xi is value of d in pixel i ◮ line integral measurement yi has form

yi =

n

  • j=1

Aijxj + vi

◮ Aij is length of line ℓi in pixel j ◮ in matrix-vector form, we have y = Ax + v

Line integral measurements 6

slide-7
SLIDE 7

Illustration

x1 x2 x6 (x0, y0) θ

y = 1.06x16 + 0.80x17 + 0.27x12 + 1.06x13 + 1.06x14 + 0.53x15 + 0.54x10 + v

Line integral measurements 7

slide-8
SLIDE 8

Example

image is 50 × 50 600 measurements shown

1

Line integral measurements 8

slide-9
SLIDE 9

Example

Line integral measurements 9

slide-10
SLIDE 10

Another example

image is 50 × 50 600 measurements shown

1

Line integral measurements 10

slide-11
SLIDE 11

Another example

Line integral measurements 11

slide-12
SLIDE 12

Outline

Line integral measurements Least-squares reconstruction Example

Least-squares reconstruction 12

slide-13
SLIDE 13

Smoothness prior

◮ we assume that image is not too rough, as measured by (Laplacian)

Dvx2 + Dhx2

– Dhx gives first order difference in horizontal direction – Dvx gives first order difference in vertical direction

◮ roughness measure is sum of squares of first order differences ◮ it is zero only when x is constant

Least-squares reconstruction 13

slide-14
SLIDE 14

Least-squares reconstruction

◮ choose ˆ

x to minimize Ax − y2 + λ(Dvˆ x2 + Dhˆ x2)

– first term is v2, or deviation between what we observed (y) and what we would have observed without noise (Ax) – second term is roughness measure

◮ regularization parameter λ > 0 trades off measurement fit versus

roughness of recovered image

Least-squares reconstruction 14

slide-15
SLIDE 15

Outline

Line integral measurements Least-squares reconstruction Example

Example 15

slide-16
SLIDE 16

Example

◮ 50 × 50 pixels (n = 2500) ◮ 40 angles, 40 offsets (m = 1600 lines) ◮ 600 lines shown ◮ small measurement noise

1

Example 16

slide-17
SLIDE 17

Reconstruction

reconstruction with λ = 10

Example 17

slide-18
SLIDE 18

Reconstruction

reconstructions with λ = 10−6, 20, 230, 2600

Example 18

slide-19
SLIDE 19

Varying the number of line integrals

reconstruct with m = 100, 400, 2500, 6400 lines (with λ = 10, 15, 25, 30)

Example 19