deep imaging

Lecture 16: Wave optics and Fourier

optics
Machine Learning and Imaging

BME 590L
Roarke Horstmeyer

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Announcements: deep imaging

Welcome to BME590 Online! Things shouldn’t change much, but here’s a summary:

0. I hope that everyone is safe and is doing ok!

1. See summary of all changes to transition things online here

a. Lectures are now held on Zoom (live, at same class time) and will be recorded and posted online
b. Lab session are now held on Zoom (live at same lab time) and will be recorded
c. Office hours are now held on Zoom
*My Zoom ID: 417-271-8775

2. Homework #3 is due today – please submit via class website

3. Homework #4 will be posted online today – it is due it

4. I have sent some of you feedback about your project proposal via email. If so, please send me an
updated project proposal by this Thursday 3/26 at midnight. For the rest of the received proposals, I
have input your grade on Sakai. If you didn’t get an email or grade on Sakai, I didn’t get your proposal.

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Rough schedule for the rest of the semester: deep imaging

Homework #4 assigned: March 24

Homework #4 due: April 7
Homework #5 assigned: April 7
Homework #5 due: April 21
Final projects due: (original) Thurs 4/23, Friday 4/24, or Monday 4/27
Final project presentations: (original) Thurs 4/23, Friday 4/24, or Monday 4/27

Questions from class survey:

Q: Are we maintaining original final project presentation dates or if these are getting shifted by 1
week to account for the extension of spring break?
A: I would prefer to maintain the original presentation dates. It may be possible to shift it back a few
days into finals week, but it depends on what I hear about regarding the format of other finals

Q: Would the recorded video of lectures be uploaded right after the class? Or it would be upload
several days after the class?
A: Right after the class (i.e., as soon as I can after the class)

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 First - what is light and how can we model it? deep imaging

• Interpretation #1: Radiation (Incoherent)

• Model: Rays

• Real, non-negative
• Models absorption
and brightness
Itot = I1 + I2
• Interpretation #2: Electromagnetic wave (Coherent)
• Model: Waves

• Complex field
• Models Interference

Etot = E1 + E2

• Interpretation #3: Particle

• Model: Photons
Machine Learning and Imaging – Roarke Horstmeyer (2020)
 Simple mathematical model of incoherent image formation deep imaging

Physical world n x m image Is

• Assume incoherent illumination

• Assume thin 2D object
• Object is real, non-negative map of absorption/reflectivity

Object absorption: I0(x,y)

Illumination pattern: s(x,y)
Light exiting object surface: Ie(x,y) = I0(x,y) s(x,y)
Machine Learning and Imaging – Roarke Horstmeyer (2020)
 Simple mathematical model of image formation deep imaging

S
Intensity at image plane Ip = ?
I0

Convolution filter h
Ie = S I0

Ip = H Ie = H S I0

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Summary: simple physical layers for incoherent imaging deep imaging

• Deal with sample/image intensities I, real and non-negative

• Effect of illumination is element-wise multiplication Ie(x,y)= S I0(x,y)

λ

• Imaging systems blur the object via point-spread function matrix H

Ib(x,y)= H I0(x,y)

• Discrete pixels down-sample the object via Id(x,y)= D I0(x,y)

• Add noise into measurement IN(x,y)= D I0(x,y) + N

• Different colors add linearly Is(x, y) = Σ I0(x, y, λ)

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 What is light and how can we model it? deep imaging

• Interpretation #1: Radiation (Incoherent)

• Model: Rays

• Real, non-negative
• Models absorption
and brightness
Itot = I1 + I2

• Interpretation #2: Electromagnetic wave (Coherent)

• Model: Waves

• Complex field
• Models Interference

Etot = E1 + E2
This class: Modeling coherent radiation as a wave
Machine Learning and Imaging – Roarke Horstmeyer (2020)
 Let’s take a step back: how does light propagate? deep imaging

Maxwell’s equations
without any charge

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Let’s take a step back: how does light propagate? deep imaging

Maxwell’s equations
without any charge

1. Take the curl of both sides of first equation

2. Substitute 2nd and 3rd equation
3. Arrive at the wave equation:

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Let’s take a step back: how does light propagate? deep imaging

Considering light that isn’t pulsed over time, we can use the following solution:

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Let’s take a step back: how does light propagate? deep imaging

Considering light that isn’t pulsed over time, we can use the following solution:

With this particular solution, we get the following important time-independent equation:

Helmholtz
Equation

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Let’s take a step back: how does light propagate? deep imaging

Considering light that isn’t pulsed over time, we can use the following solution:

With this particular solution, we get the following important time-independent equation:

Helmholtz
Equation

This is an important equation in physics. We won’t go into the details, but it leads to the Huygen-Fresnel principle:

2
2
2

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Plane-to-plane light propagation via the ”paraxial approximation” deep imaging

The Huygens-Fresnel Equation 2

2
2

P2

Aperture

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Plane-to-plane light propagation via the ”paraxial approximation” deep imaging

The Huygens-Fresnel Equation 2

2
2

P2
Generally connects two points in 3D:

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Plane-to-plane light propagation via the ”paraxial approximation” deep imaging

We are usually concerned about propagation between two planes (almost always in an optical system):

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Plane-to-plane light propagation via the ”paraxial approximation” deep imaging

We are usually concerned about propagation between two planes (almost always in an optical system):

Paraxial approximation:

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Plane-to-plane light propagation via the ”paraxial approximation” deep imaging

We are usually concerned about propagation between two planes (almost always in an optical system):

Paraxial approximation:

Substitute in and crank the wheel,

Paraxial Helmholtz Equation. This has an exact integral solution:

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Plane-to-plane light propagation via the ”paraxial approximation” deep imaging

We are usually concerned about propagation between two planes (almost always in an optical system):

Paraxial approximation:

Substitute in and crank the wheel,

Paraxial Helmholtz Equation. This has an exact integral solution:

Fresnel diffraction
integral

This is how light propagates from one plane to the next. It’s a convolution!
Machine Learning and Imaging – Roarke Horstmeyer (2020)
 Fresnel light propagation as a convolution deep imaging

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Fresnel light propagation as a convolution deep imaging

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 From the Fresnel approximation to the Fraunhofer approximation deep imaging

Fresnel Approximation:

Lets assume that the second plane is “pretty far away” from the first plane. Then,

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 From the Fresnel approximation to the Fraunhofer approximation deep imaging

Fresnel Approximation:

Lets assume that the second plane is “pretty far away” from the first plane. Then,

1. Expand the squaring

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 From the Fresnel approximation to the Fraunhofer approximation deep imaging

Fresnel Approximation:

Lets assume that the second plane is “pretty far away” from the first plane. Then,

1. Expand the squaring

2. Front term comes out, assume second term goes away, then,

Fraunhofer diffraction is a Fourier transform!!!!!!!

Machine Learning and Imaging – Roarke Horstmeyer (2020)
 This is the aperture This is what you see far away deep imaging

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 deep imaging

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 deep imaging

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Model of a microscope (or camera) using Fourier transforms: deep imaging

2D Fourier Transform
Machine Learning and Imaging – Roarke Horstmeyer (2020)
 Model of a microscope (or camera) using Fourier transforms: deep imaging

Effect of the lens is to block light.

Use thin object approximation to

determine distribution of light on the
immediate other side of the lens stop:

1’s in here

2D Fourier Transform 0’s out here

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Model of a microscope (or camera) using Fourier transforms: deep imaging

Last piece of the puzzle: what happens from lens to sensor?

2D Fourier Transform

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Model of a microscope (or camera) using Fourier transforms: deep imaging

Last piece of the puzzle: what happens from lens to sensor?

inverse Fourier transform!

2D Fourier Transform 2D inverse Fourier Transform

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 This process should sound familiar….
deep imaging

Convolution filter h
Input image Output image
U1(x,y) =
U2(x,y)

F [h]
F [U1] F -1[HÛ1]

Input
spectrum

Û1(fx,fy) • =

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Model of image formation for wave optics (coherent light): deep imaging

1. Discrete sample
function s(x,y)
(complex)

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Model of image formation for wave optics (coherent light): deep imaging

1. Discrete sample 2. Compute its 2D

function s(x,y) Fourier transform
(complex) ŝ(fx, fy)

“Fourier plane”

2D FT

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Model of image formation for wave optics (coherent light): deep imaging

1. Discrete sample 2. Compute its 2D 3. Multiply by

function s(x,y) Fourier transform “aperture”
(complex) ŝ(fx, fy) function A(fx, fy)

“Fourier plane”

2D FT

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Model of image formation for wave optics (coherent light): deep imaging

1. Discrete sample 2. Compute its 2D 3. Multiply by 4. Compute

function s(x,y) Fourier transform “aperture” inverse Fourier
(complex) ŝ(fx, fy) function A(fx, fy) transform s’(x’,y’)
(complex)

“Blurred image”
“Fourier plane”

2D FT 2D IFT

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Model of image formation for wave optics (coherent light): deep imaging

1. Discrete sample 2. Compute its 2D 3. Multiply by 4. Compute 5. Detector

function s(x,y) Fourier transform “aperture” inverse Fourier measures
(complex) ŝ(fx, fy) function A(fx, fy) transform s’(x’,y’) |s’(x’,y’)|2
(complex)

“Blurred image”
“Fourier plane”

2D FT 2D IFT

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Can also model this using the Convolution Theorem deep imaging

Aperture function (lens shape) Camera blur function (IFT of lens shape)

A(fx, fy)
h(x, y)

2D IFT

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Two modeling choices for the camera:
deep imaging

Spatial Frequency Domain Multiply

A (fx,fy)
ŝ1(fxi,fyi) ŝ2(fxo,fyo)

Spatial Domain Convolve = blur

h (x,y)

s1(xi,yi) s2(xo,yo)

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Linear systems and the black box deep imaging

The optical black box system and the point-spread function:

Light g1(xi,yi) entering “black box” optical system modified by system point-spread function

h (x2,y2) g2(x2,y2)
g1(x1,y1)

Assume shift
invariance:
This is the system
point-spread function

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Summary of two models for image formation deep imaging

• Interpretation #1: Radiation (Incoherent)

• Real, non-negative
• Model: Rays
• Models absorption
and brightness
Itot = I1 + I2

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Mathematical model of for incoherent image formation deep imaging

• All quantities are real, and non-negative

Object absorption:
S0(x,y)
Illumination brightness:
B(x,y)

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Mathematical model of for incoherent image formation deep imaging

• All quantities are real, and non-negative

Object absorption:
S0(x,y)
Illumination brightness:
B(x,y)

100
photons 60
photons
60%
transmission

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Mathematical model of for incoherent image formation deep imaging

• All quantities are real, and non-negative

Object absorption:
S0(x,y)
Illumination brightness:
B(x,y)

80 8
photons photons

10%
transmission

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Mathematical model of for incoherent image formation deep imaging

• All quantities are real, and non-negative

Object absorption:
S0(x,y)
Illumination brightness:
B(x,y)

B S0
multiplication

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Summary of two models for image formation deep imaging

• Interpretation #1: Radiation (Incoherent)

• Real, non-negative
• Model: Rays
• Models absorption
and brightness
Itot = I1 + I2

Is = B S0

• Interpretation #2: Electromagnetic wave (Coherent)

• Model: Waves • Complex field
• Models Interference

Etot = E1 + E2

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Mathematical model of for coherent image formation deep imaging

• Pretty much the same thing, but now we have an amplitude and a complex phase

Point #1: Amplitudes behave just like before

Sample absorption = S(x,y)

(Laser light
100 At(x,y) = Ai(x,y) S(x,y)
or
Ultrasound) photons 60
photons
60%
transmission
Incident field: Transmitted field:

C(x,y) = Ai(x,y) U(x,y) = At(x,y) = ?

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Mathematical model of for coherent image formation deep imaging

• Pretty much the same thing, but now we have an amplitude and a complex phase

Point #1: Amplitudes behave just like before

Sample absorption = S(x,y)

100
photons 20
(Laser light photons At(x,y) = Ai(x,y) S(x,y)
or
Ultrasound) 20%
transmission

Incident field: Transmitted field:

C(x,y) = Ai(x,y) U(x,y) = At(x,y) = ?

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Mathematical model of for coherent image formation deep imaging

• Pretty much the same thing, but now we have an amplitude and a complex phase

Point #1: Amplitudes behave just like before

Sample absorption = S(x,y)

100
photons 20
(Laser light photons At(x,y) = Ai(x,y) S(x,y)
or
Ultrasound) 20%
transmission

Incident field: Transmitted field:

C(x,y) = Ai(x,y) U(x,y) = At(x,y) = Ai(x,y) S(x,y)

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Mathematical model of for coherent image formation deep imaging

• Pretty much the same thing, but now we have an amplitude and a complex phase

Sample absorption = S(x,y) New: complex phase delay

• Needed to represent wave
Sample phase delay = exp[ikϕ(x,y)] • Represents wave delay across space

(Laser light
or
Ultrasound)

Incident field: Transmitted field:

C(x,y) = Ai(x,y) exp[ikϕi(x,y)] U(x,y) = Ai(x,y) S(x,y) exp[ikϕt(x,y)]

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Mathematical model of for coherent image formation deep imaging

• Pretty much the same thing, but now we have an amplitude and a complex phase

New: complex phase delay

Sample absorption = S(x,y)

Sample phase delay = exp[ikϕ(x,y)]

(Laser light
ϕi(x,y) = 0
or
Ultrasound)

π/6 deg.
Phase delay

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Mathematical model of for coherent image formation deep imaging

• Pretty much the same thing, but now we have an amplitude and a complex phase

New: complex phase delay

Sample absorption = S(x,y)

Sample phase delay = exp[ikϕ(x,y)]

Incident light
(Laser light
ϕi(x,y) = 0
or
Ultrasound)

π/6 deg.
Phase delay
Transmitted light

Total lag = pi/6 rad

Machine Learning and Imaging – Roarke Horstmeyer (2020)
 Mathematical model of for coherent image formation deep imaging

• Pretty much the same thing, but now we have an amplitude and a complex phase

New: complex phase delay

Sample absorption = S(x,y)

Sample phase delay = exp[ikϕ(x,y)]

ϕi(x,y) =
π/8 rad.
Incident light: pi/8 behind
(Laser light
or
Ultrasound)

Transmitted light

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Mathematical model of for coherent image formation deep imaging

• Pretty much the same thing, but now we have an amplitude and a complex phase

New: complex phase delay

Sample absorption = S(x,y)

Sample phase delay = exp[ikϕ(x,y)]

ϕi(x,y) =
π/8 rad.
Incident light: pi/8 behind
(Laser light π/4 rad
or Phase delay
Ultrasound)

Transmitted light

Total lag = π/8+ π/4= 3π/8 rad

Machine Learning and Imaging – Roarke Horstmeyer (2020)
 Mathematical model of for coherent image formation deep imaging

• Pretty much the same thing, but now we have an amplitude and a complex phase

Sample absorption = S(x,y)

Sample phase delay = exp[ikϕ(x,y)] Output phase is sum of phase delays, product of phasors

ϕt(x,y) =ϕ(x,y) + ϕi(x,y)

(Laser light
or exp[ikϕt(x,y)] = exp[ikϕi(x,y)] exp[ikϕ(x,y)]
Ultrasound)

Incident field: Transmitted field:

C(x,y) = Ai(x,y) exp[ikϕi(x,y)] U(x,y) = Ai(x,y) S(x,y) exp[ikϕi(x,y)] exp[ikϕ(x,y)]

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Mathematical model of for coherent image formation deep imaging

• Pretty much the same thing, but now we have an amplitude and a complex phase

Sample absorption = S(x,y)

Sample phase delay = exp[ikϕ(x,y)] Conclusion:

Transmitted field = incident field x complex sample :

(Laser light
or U(x,y) = C(x,y) S (x,y) exp[ikϕ(x,y)]
Ultrasound)

Incident field: Transmitted field:

C(x,y) = Ai(x,y) exp[ikϕi(x,y)] U(x,y) = Ai(x,y) S(x,y) exp[ikϕi(x,y)] exp[ikϕ(x,y)]

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Summary of two models for image formation deep imaging

• Interpretation #1: Radiation (Incoherent)

• Real, non-negative
• Model: Rays
• Models absorption
and brightness
Itot = I1 + I2

Is = B S0

• Interpretation #2: Electromagnetic wave (Coherent)

• Model: Waves • Complex field
• Models Interference

Etot = E1 + E2

U = C S0
U, C and S are complex!

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 deep imaging

Additional Information about sample index of

refraction, spatial frequency and Fourier optics

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Microscope illumination and sample index of refraction deep imaging

So far: illuminate the sample and create a field that is equivalent

to the sample’s absorption and phase
– why is this true???

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Microscope illumination and sample index of refraction deep imaging

So far: illuminate the sample and create a field that is equivalent

to the sample’s absorption and phase
– why is this true???

Sample index of refraction n(x,y,z) = 1 + ia(x) + ϕ (x)

a(x=1) = .2
ϕ(x=1) = .1

*For more information, see D. Paganin, Coherent X-Ray Optics, Section 2.2

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Microscope illumination and sample index of refraction deep imaging

So far: illuminate the sample and create a field that is equivalent

to the sample’s absorption and phase
– why is this true???

Sample index of refraction n(x,y,z) = 1 + ia(x) + ϕ (x)

a(x=1) = .2
ϕ(x=1) = .1

Thin sample approximation:

Sample’s effect on light is multiplication with exp[-ik * n(x,y)]

*For more information, see D. Paganin, Coherent X-Ray Optics, Section 2.2

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Microscope illumination and sample index of refraction deep imaging

So far: illuminate the sample and create a field that is equivalent

to the sample’s absorption and phase
– why is this true???

Sample index of refraction n(x,y,z) = 1 + ia(x) + ϕ (x)

a(x=1) = .2
ϕ(x=1) = .1

Thin sample approximation:

Sample’s effect on light is multiplication with exp[-ik * n(x,y)]

In 1D: Emerging field U(x) = incident field Ui(x) * sample function s(x)

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Microscope illumination and sample index of refraction deep imaging

So far: illuminate the sample and create a field that is equivalent

to the sample’s absorption and phase
– why is this true???

Sample index of refraction n(x,y,z) = 1 + ia(x) + ϕ (x)

a(x=1) = .2
ϕ(x=1) = .1

Thin sample approximation:

Sample’s effect on light is multiplication with exp[-ik * n(x,y)]

In 1D: Emerging field U(x) = incident field Ui(x) * sample function s(x)=exp[-ik n(x)]

U(x) = Ui(x) *exp[-ik n(x)] = Ui(x) A(x) exp[ikϕ(x)] A(x) = exp[k a(x)]
absorption phase shift: new term for laser
Machine Learning and Imaging – Roarke Horstmeyer (2020)
 Microscope illumination and sample index of refraction deep imaging

Sample absorption = A(x)

Sample phase = exp[ikϕ(x)]

Emerging field U = incident field Ui(x) * sample function s(x)

Incident field Ui

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Microscope illumination and sample index of refraction deep imaging

Q: When is the emerging field equal to the absorption and phase?

Sample absorption = A(x)

Sample phase = exp[ikϕ(x)]

Emerging field U = incident field Ui(x) * sample function s(x)

Incident field Ui Emerging field U

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Microscope illumination and sample index of refraction deep imaging

Q: When is the emerging field equal to the absorption and phase?

Sample absorption = A(x) A: When the incident wave = 1, means uniform in amplitude and phase:
Sample phase = exp[ikϕ(x)]
Ui(x) = 1 U(x)=A(x) exp[ikϕ(x)]

Emerging field U = incident field Ui(x) * sample function s(x)

Incident field Ui Emerging field U

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Microscope illumination and sample index of refraction deep imaging

Q: When is the emerging field equal to the absorption and phase?

Sample absorption = A(x) A: When the incident wave = 1, means uniform in amplitude and phase:
Sample phase = exp[ikϕ(x)]
Ui(x) = 1 U(x)=A(x) exp[ikϕ(x)]
θ=0 θ=30

Plane wave Ui(x) = 1 * exp(ik•x)

Ui(x) = exp(ikx sin(θ))

1 everywhere Not 1
everywhere
Incident field Ui Emerging field U This is when incident wave hits the sample with θ=0!

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Model of image formation for wave optics (coherent light): deep imaging

1. Discrete sample 2. Compute its 2D

function s(x,y) Fourier transform
(complex) ŝ(fx, fy)

“Blurred image”
What does fx represent, really?

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 From before: Spatial frequencies = “stripes” within each image deep imaging

U(x,y) Fourier Û(fx, fy)

2π/Ty Transform

y fy
Ty

x fx

2π/Tx U(x,y) Fourier Û(fx, fy)

Transform
Tx

y fy

x fx
Machine Learning and Imaging – Roarke Horstmeyer (2020)
 Ray angle and spatial frequency
Plane of interest deep imaging

Stripes are for complex fields!

Incident plane wave
Proportional to distance between
subsequent peaks of wave along plane
of interest

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Ray angle and spatial frequency
Plane of interest deep imaging

x
Distance to two crests = spatial period

θ Sin(θ) = λ/d
d
d = λ/ Sin(θ)

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 Ray angle and spatial frequency
Plane of interest deep imaging

x
Distance to two crests = spatial period

θ Sin(θ) = λ/d
d
d = λ/ Sin(θ)

Spatial frequency = 1/spatial period

(number of periods per unit length)

θ
fx = 1/d = Sin(θ)/ λ

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 deep imaging
Equivalent coordinates in the Fourier domain and at the Fourier plane

fx = Sin(θ)/ λ kx = 2πfx = 2π/ λ Sin(θ)

Image plane ray angle θ Image plane spatial frequency fx Image plane wavevectors kx

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 General rules for applying the Fourier transform in optics deep imaging

Situation 1: From an object to a plane “really far away”

U1(x1)

U2(x2) ~ F[U1(x1/M1)]

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 General rules for applying the Fourier transform in optics deep imaging

Situation 1: From an object to a plane “really far away”

U1(x1)

U2(x2) ~ F[U1(x1/M1)]

Situation 2: From an object to the back focal plane of U1(x1)

the microscope objective lens U2(x2) ~ F[U1(x1/M2)]

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 General rules for applying the Fourier transform in optics deep imaging

Situation 1: From an object to a plane “really far away”

U1(x1)

U2(x2) ~ F[U1(x1/M1)]

Situation 2: From an object to the back focal plane of U1(x1)

the microscope objective lens U2(x2) ~ F[U1(x1/M2)]

Situation 3: From an object to a plane 1 focal length away from a lens (1f-1f system)

U1(x1)
U2(x2) ~ F[U1(x1/M3)]

f f
Machine Learning and Imaging – Roarke Horstmeyer (2020)
 A more exact model: the 4f optical system deep imaging

f f f f

2D Fourier Transform 2D inverse Fourier Transform

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 A more exact model: the 4f optical system deep imaging

The Fourier plane provides

a measure of the ray angles
at the image plane

f f

Image plane ray angle θ

2D Fourier Transform

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 A more exact model: the 4f optical system deep imaging

The Fourier plane provides

a measure of the ray angles
at the image plane

Shift point Doesn’t contain info about spatial distribution light

source

f f

Image plane ray angle θ

2D Fourier Transform

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 A more exact model: the 4f optical system deep imaging

The Fourier plane provides

a measure of the ray angles
at the image plane

Shift point Rays are leaving image plane at +20 degrees

source

20 degrees
f f

Image plane ray angle θ

2D Fourier Transform

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 A more exact model: the 4f optical system deep imaging

The Fourier plane provides

a measure of the ray angles
-15 degrees at the image plane

Shift point Rays are coming in at +20 degrees and -15 degrees
source

20 degrees
f f

Image plane ray angle θ

2D Fourier Transform

Machine Learning and Imaging – Roarke Horstmeyer (2020)

 deep imaging

You typically go between 4 functions to describe one imaging system:

Coherent point- Incoherent point-

spread function spread function
|.|2
h(x) hi(x)

F[.] F[.]

Coherent Incoherent
H*H
transfer function transfer function

H(fx) Hi(fx)

Machine Learning and Imaging – Roarke Horstmeyer (2020)