Top Banner
Computed Tomography Ho Kyung Kim [email protected] Pusan National University Medical Physics Prince & Links 6 Introduction Tomogram An image of a plane or slice within the body Motion tomography A conventional procedure in which the x‐ray source & detector (screen) are moved in opposite directions in order to keep one plane in focus Suffers from blurring & overlaying of out‐of‐plane features (artifacts) CT A procedure in which the x‐ray source & detector (screen) at the opposite sides rotates along an object Eliminates completely artifacts from overlying tissues The CT scanner reconstructs the value of at each pixel w/i a cross section 2
35

Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim [email protected] Pusan

May 08, 2021

Download

Documents

dariahiddleston
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Computed Tomography

Ho Kyung [email protected]

Pusan National University

Medical PhysicsPrince & Links 6

Introduction

• Tomogram

– An image of a plane or slice within the body

• Motion tomography

– A conventional procedure in which the x‐ray source & detector (screen) are moved in opposite directions in order to keep one plane in focus

– Suffers from blurring & overlaying of out‐of‐plane features (artifacts)

• CT

– A procedure in which the x‐ray source & detector (screen) at the opposite sides rotates along an object

– Eliminates completely artifacts from overlying tissues

– The CT scanner reconstructs the value of 𝜇 at each pixel w/i a cross section

2

Page 2: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

3

4

• Radon transform

– (2D or 3D) Radon transform that takes (1D or 2D) projections of a (2D or 3D) object over many angles

– Using the inverse of Radon transform, we can reconstruct the (2D) axial cross‐section or (3D) volumetric images

• CT scanners measure x‐ray attenuation along a line btwn an x‐ray source & a detector (projection)

• Through standard calibration procedures, the measurement results are ultimately expressed as the CT numbers in Houndsfield units

– Constant from scan‐to‐scan & across different scanners

– Quantitative analysis

• Various CT designs have been developed to obtain x‐ray attenuation data along a line btwn an x‐ray source & an x‐ray detector

– Generations toward faster acquisition of data

Page 3: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

CT generations

5

• 1G scanners

– No longer used in medical imaging

– Pencil beam (parallel‐beam geometry)

– Negligible scattered photons

– Repeated linear scan & rotation

• 2G scanners

– No longer used

– Fan beam w/ a multiple detector array

– Repeated linear scan & rotation

– Faster than 1G

Collimator

Collimation on detection reduces efficiency & increases noise for a given dose

Example

Consider a 1G or 2G scanner whose source‐detector apparatus can move linearly at a speed of 1.0 m/s and that the FOV has a diameter of 0.5 m. Suppose further that 360 projections over 180 are required and that it takes 0.5 s for one angular increment, regardless of the angle.

What is the scan time for a 1G scanner? What is the scan for a 2G scanner having 9 detectors spaced 0.5 apart?

6

Page 4: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

7

• 3G scanners

– Fan‐beam geometry

– Synchronous rotation of the source & detector array pair (no linear motion)

– Typ. 1,000 projections w/ a fan‐beam angle of 30–60 degrees incident upon 500–700 detectors w/i the scan time of 1–20 s

– Loss of detection efficiency due to the detector collimators

• 4G scanners

– Single rotating source + a ring of stationarydetectors

– No detector collimation

– Scattering problem

8

• 5G scanners

– Called "Electron beam CT" (EBCT) scanners

– No moving of source/detectors

– Scan time = ~50 ms

• Can capture stop‐action images of a beating heart

– Not clinically popular because of high cost

• 6G scanners

– Helical CT in the late 1980s for rapid volumetric data acquisition

– During continuous source‐detector rotations (0.3–0.5 s / revolution), the patient table is fed

– Slip ring technology

– Fast scan times (using breath‐holds)

• ~30 s for a full 60‐cm torso; ~12 s for a full 24‐cm lung study; ~30 s for a detailed 15‐cm angiography study

• 7G scanners

– Multiple‐row detector CT (MDCT) for a "thick" fan‐beam or cone‐beam

Page 5: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

9

10

Generation Source Source collimation

Detector Detector collimation

Source-det. movement

Advantage Disadvan.

1G Single x-ray tube

Pencil beam Single NoneMove linearly & rotate in unison

Scattered energy is undetected

Slow

2G Single x-ray tube

Fan beam, not enough to cover FOV

MultipleCollimated tosource direction

Move linearly & rotate in unison

Faster than 1G

Lower eff. & larger noise due to the collimation in detectors

3G Single x-ray tube

Fan beam, enough to cover FOV

ManyCollimated tosource direction

Rotate in synchrony

Faster than 2G, continuous rotation using a slip ring

More expensive than 2G, loweff.

4G Single x-ray tube

Fan beam covers FOV

Stationary ring of detectors

Cannot collimate detectors

Detectors are fixed, source rotates

Higher efficiency than 3G

Highscattering since detectors are not collimated

5G(EBCT)

Many W anodes in single large tube

Fan beamStationary ring of detectors

Cannot collimate detectors

No moving parts

Extremely fast, capable of stop-action imaging of beating heart

High cost, difficult tocalibrate

6G(Spiral CT)

3G/4G 3G/4G 3G/4G 3G/4G3G/4G + linear patienttable motion

Fast 3D images

A bit more expensive

7G (Multislice CT)

Single x-ray tube

Cone beamMultiple arrays of detectors

Collimated tosource direction

3G/4G/6G motion

Fast 3D images

Expensive

Page 6: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Dual‐energy CT

11

• To obtain more information about tissue characteristics

• Double scanning at different kVp's (e.g., 80 & 140 kVp's)

• kVp switching during a single scan (in pulse mode)

• Dual‐source CT

(Image taken from) M. Patino et al. | Rdiographics | 2016

CT instrumentation

12

• X‐ray tubes

– The same as those in projection radiography systems

– Rotating anode & oil‐cooled designs

– Mostly continuous operation (some scanners operate in pulse mode)

– The lifetime is short (1 year)

• Collimation

– Two pieces of lead (like a slit) to make a fan‐beam geometry (30–60 degrees)

– Motor‐controlled slit width for the fan thickness (or height): 1–10 mm (4–16 cm in MDCT systems)

• Filtration

– Cu/Al to narrow (or "harden") the energy spectrum toward a monoenergetic beam

Page 7: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

13

• CT detectors

– Solid‐state detectors

• Scintillation crystal: x‐ray‐to‐light

– Cadmium tungstate, sodium iodide, bismuth germanate, yttrium‐based, or cesium iodide crystal

– 1.0 mm  15 mm (defines the max. slice thickness) in area

• Photodiode: light‐to‐electric current

• The slice thickness is controlled using movable blades in the x‐ray tube collimator

– Xenon gas detectors

• Lower efficiency but highly directional

– Multiple detector array for MDCT

• 1.0 mm  1.25 mm

• The slice thickness is controlled by both the axial beam width & the combined detector height

– Usually multiples of the detector height

14

(Images taken from) W. A. Kalender | Computed Tomography | 2011

Page 8: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

• Gantry

– Holding the x‐ray tube & detectors so as to they are rotated

– Typ. FOV = 50 cm

– Full 2D scan time < 1 s

– Being "tilted" to facilitate acquisition of nonaxial slices

• Slip ring

– Continuous rotation w/o "rewinding" the source & detectors

– Rotating cylinder w/ grooves on the outside, which contains the source & detectors

– Brushesmake continuous electrical contract w/ the cylinder

• Patient table

15

16

(Images taken from) W. A. Kalender | Computed Tomography | 2011

Page 9: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Line integrals

• Detector signal (approximate)

𝐼 𝑆 𝐸 𝐸exp 𝜇 𝑠; 𝐸 d𝑠 d𝐸

• Further approximation (the same measured 𝐼 )

𝐼 𝐼 exp 𝜇 𝑠; 𝐸 d𝑠

– 𝐸 = effective energy that, in a given mat'l, will produce the same 𝐼 from a "monoenergetic" source as is measured using the actual "polyenergetic" source

• Projectionmeasurement

𝑔 ln𝐼𝐼

𝜇 𝑠; 𝐸 d𝑠

– The line integral of the linear attenuation coefficient at the effective energy of the scanner

– Reference intensity 𝐼 is given by the calibration (air scan)

17

CT numbers

18

• Different scanners; different 𝐸; different values of 𝜇 for the exact same object

• Define CT numbers in Houndsfield units (HU)

ℎ 1,000𝜇 𝜇

𝜇– ‐1,000 HU air

– 0 HU water

– ~1,000 HU (average) bone

– ~3,000 HU metal & contrast agents

Page 10: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Parallel‐ray reconstruction

19

• While the CT measures a line integral of the effective 𝜇 w/i a cross section, we really want a picture of 𝜇 or equivalently ℎ

• Can we reconstruct a picture of 𝜇 given a collection of its line integrals?

• From the right figure𝐿 ℓ, 𝜃 𝑥, 𝑦 |𝑥 cos 𝜃 𝑦 sin 𝜃 ℓ

Line integral (in a parametric form)

𝑔 ℓ, 𝜃 𝑓 𝑥 𝑠 , 𝑦 𝑠 d𝑠

– 𝑥 𝑠 ℓ cos 𝜃 𝑠 sin 𝜃– 𝑦 𝑠 ℓ cos 𝜃 𝑠 sin 𝜃

𝑠

20

• Alternatively (using the shifting property of 𝛿),

𝑔 ℓ, 𝜃 𝑓 𝑥, 𝑦 𝛿 𝑥 cos 𝜃 𝑦 sin 𝜃 ℓ d𝑥 d𝑦

– 𝑔 ℓ, 𝜃 for a fixed 𝜃 = projection

– 𝑔 ℓ, 𝜃 for all ℓ & 𝜃 = 2D Radon transform of 𝑓 𝑥, 𝑦

• Comparing to the projection measurement: 𝑔 ln 𝜇 𝑠; 𝐸 d𝑠

– 𝑓 𝑥, 𝑦 𝜇 𝑠; 𝐸

– 𝑔 ℓ, 𝜃 ln

Page 11: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Example

Consider the unit disk given by

𝑓 𝑥, 𝑦 1 𝑥 𝑦 10 otherwise

What is its 2D Radon transform?

21

Sinogram

22

• Image of 𝑔 ℓ, 𝜃 w/ ℓ & 𝜃 as rectilinear coordinates

• Pictorial representation of the Radon transform of 𝑓 𝑥, 𝑦 , representing the data necessary to reconstruct 𝑓 𝑥, 𝑦

• Can you recognize ‘sinusoidal’ patterns?

– Correspond to individual objects in the image

𝜃 𝜃

Projection obtained at 𝜃 𝜃

Page 12: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Backprojection

23

𝑔 ℓ, 𝜃 𝑓 𝑥, 𝑦 𝛿 𝑥 cos 𝜃 𝑦 sin 𝜃 ℓ d𝑥 d𝑦

– Infinite number of 𝑓 𝑥, 𝑦 giving rise to 𝑔 ℓ, 𝜃

• Backprojection is to assign every point on 𝐿 ℓ, 𝜃 the value 𝑔 ℓ, 𝜃 (w/ no prior information about the distribution of the image intensities)

𝑏 𝑥, 𝑦 𝑔 𝑥 cos 𝜃 𝑦 sin 𝜃 , 𝜃

• Laminogram

𝑓 𝑥, 𝑦 𝑏 𝑥, 𝑦 d𝜃– Backprojection summation image

– Incorrect approach (blurry)

𝜃 𝜃

𝑔 ℓ , 𝜃

𝐿 ℓ , 𝜃

𝒃𝜽𝟎 𝒙, 𝒚

Example

Consider the projection 𝑔 ℓ, 𝜃 45° sgn ℓ , where sgn ℓ 1 when ℓ 0 and is  1when ℓ 0. What is the backprojection image 𝑏 𝑥, 𝑦 at 𝜃 45°.

24

Page 13: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Projection‐slice theorem

• Relationship b/w 1D Fourier transform of a projection & 2D Fourier transform of the object𝐺 𝜚, 𝜃 𝐹 𝜚 cos 𝜃 , 𝜚 sin 𝜃

25

𝐺 𝜚, 𝜃 𝑓 𝑥, 𝑦 𝛿 𝑥 cos 𝜃 𝑦 sin 𝜃 ℓ 𝑒 ℓ d𝑥d𝑦dℓ

𝑓 𝑥, 𝑦 𝛿 𝑥 cos 𝜃 𝑦 sin 𝜃 ℓ 𝑒 ℓ dℓd𝑥d𝑦

𝑓 𝑥, 𝑦 𝑒 d𝑥d𝑦

𝑓 𝑥, 𝑦 𝑒 d𝑥d𝑦

𝐹 𝜚 cos 𝜃 , 𝜚 sin 𝜃

shifting property

𝐺 𝜚, 𝜃 ℱ 𝑔 ℓ, 𝜃 𝑔 ℓ, 𝜃 𝑒 ℓ dℓ

• Projection‐slice theorem: 𝐺 𝜚, 𝜃 𝐹 𝜚 cos 𝜃 , 𝜚 sin 𝜃

– 1D FT of a projection 𝐺 𝜚, 𝜃 is a slice of 2D FT of the object 𝐹 𝜚 cos 𝜃 , 𝜚 sin 𝜃

– 1D FT of the projection equals a line passing thru the origin of 2D FT of the object at that angle corresponding to the projection

26

Page 14: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Example

The projection‐slice theorem helps us to understand how angular sampling can influence reconstructions. Suppose that only eight 

projections are acquired at angles 𝜃.

, 𝑖 0, … , 7.

Show that the function 𝑓 𝑥, 𝑦 cos 𝑥 is invisible at these angles –that is, it will produce projections that are identically zero at these angles.

27

The Fourier method

28

• Following the projection‐slice theorem:

– Take the 1D FT of each projection

– Insert it w/ the corresponding correct angular orientation into the correct slice of the 2D Fourier plane

– Take the inverse 2D FT of the result

𝑓 𝑥, 𝑦 ℱ 𝐺 𝜚, 𝜃

• Not widely used in CT due to:

– The practical problem of interpolating polar data onto Cartesian grid

– The need to use the relatively time‐consuming 2D inverse FT

Page 15: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Filtered backprojection (FBP)

29

𝑓 𝑥, 𝑦 𝐹 𝜚 cos 𝜃 , 𝜚 sin 𝜃 𝑒 𝜚d𝜚d𝜃

Inverse FT of 𝐹 𝑢, 𝑣

Projection‐slice theorem

𝑓 𝑥, 𝑦 𝐺 𝜚, 𝜃 𝑒 𝜚d𝜚d𝜃

𝜚 𝐺 𝜚, 𝜃 𝑒 d𝜚d𝜃𝑔 ℓ, 𝜃 𝑔 ℓ, 𝜃 𝜋

𝜚 𝐺 𝜚, 𝜃 𝑒 ℓ d𝜚ℓ

d𝜃

30

• Exact formula for the inverse Radon transform

• 𝜚 𝐺 𝜚, 𝜃 filtered projection

– multiplication of the freq. filter  𝜚 to the 1D FT of 𝑔 ℓ, 𝜃– 𝜚 called the ramp filter

• 𝜚 𝐺 𝜚, 𝜃 𝑒 ℓ d𝜚 inverse 1D FT of  𝜚 𝐺 𝜚, 𝜃

• ℓ 𝑥 cos 𝜃 𝑦 sin 𝜃 backprojection

• · d𝜃 summation

𝑓 𝑥, 𝑦 𝜚 𝐺 𝜚, 𝜃 𝑒 ℓ d𝜚ℓ

d𝜃

Page 16: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Convolution backprojection (CBP)

31

• Perform a convolution rather than a filtering operation

• More efficient if the impulse response is narrow (in most CT scanners)

• 𝑐 ℓ not exist! ( 𝜚 not integrable; its inverse FT undefined)

• Approximate 𝑐 ℓ or windowing 𝜚 w/ a suitable windowing function

– Approximate impulse function: �̃� ℓ ℱ 𝜚 𝑊 𝜚– 𝑊 𝜚 : square, Hamming, or cosine window

– Nonzero filter value at 𝜚 0 to produce the correct reconstructed average image value

𝑓 𝑥, 𝑦 𝜚 𝐺 𝜚, 𝜃 𝑒 ℓ d𝜚ℓ

d𝜃Recall

ℱ 𝜚 ∗ 𝑔 ℓ, 𝜃 ℓ d𝜃

𝑐 ℓ ∗ 𝑔 ℓ, 𝜃 ℓ d𝜃

𝑔 ℓ, 𝜃 𝑐 𝑥 cos 𝜃 𝑦 sin 𝜃 ℓ dℓ d𝜃

𝑐 ℓ ℱ 𝜚

convolution theorem

convolution integral

called “convolution backprojection”

32

• The same 3‐step procedure b/w FBP & CBP for given sinogram

1) filtering

sinogram filtered sinogram

zero background

negatives

Page 17: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

33

2) backprojection

34

3) summation

Page 18: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Fan‐beam reconstruction

35

• Fan‐beam geometries

– Equi‐angular geometry

• Detectors positioned on a circular arc

– Equi‐distant geometry

• Equal detector spacings on a straight line

– Equi‐angular & distant geometry

• Detectors are placed along a circular arc whose center is at the source

Equi‐angular geometry

36

isocenter fan‐beam projection: 𝑝 𝛾, 𝛽where 𝛾 = angular position of a given detector

angular position of the source

parallel‐ray parameters

fan‐beam parameters

𝜃 𝛽 𝛾

ℓ 𝐷 sin 𝛾

Relationship b/w parallel‐ray & fan‐beam parameters

Page 19: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

37

Recall the parallel‐ray CBP

𝑓 𝑥, 𝑦 𝑔 ℓ, 𝜃 𝑐 𝑥 cos 𝜃 𝑦 sin 𝜃 ℓ dℓ d𝜃

parallel‐ray reconstruction formula in polar coordinates

12

𝑔 ℓ, 𝜃 𝑐 𝑥 cos 𝜃 𝑦 sin 𝜃 ℓ dℓ d𝜃𝑔 ℓ, 𝜃 0 for  ℓ 𝑇

Rewrite w/  𝑟, 𝜙 polar coordinates

𝑓 𝑟, 𝜙12

𝑔 ℓ, 𝜃 𝑐 𝑟 cos 𝜃 𝜙 ℓ dℓ d𝜃

Note:𝑥 𝑟 cos 𝜙𝑦 𝑟 sin 𝜙𝑥 cos 𝜃 𝑦 sin 𝜃 𝑟 cos 𝜙 cos 𝜃 𝑟 sin 𝜙 sin 𝜃 𝑟 cos 𝜃 𝜙

38

– 𝛾, 2𝜋 𝛾 → 0,2𝜋 due to periodicity

– sin , sin → 𝛾 , 𝛾

– Recognize the fan‐beam projection: 𝑝 𝛾, 𝛽 𝑔 𝐷 sin 𝛾 , 𝛽 𝛾

Integrate over 𝛾 & 𝛽 (fan‐beam parameters) instead of ℓ & 𝜃 (parallel‐ray parameters)⟹ Transformation of coordinates (Jacobian = 𝐷 cos 𝛾)

𝑓 𝑟, 𝜙12

𝑔 𝐷 sin 𝛾 , 𝛽 𝛾 𝑐 𝑟 cos 𝛽 𝛾 𝜙 𝐷 sin 𝛾 𝐷 cos 𝛾 d𝛾 d𝛽

𝑓 𝑟, 𝜙12

𝑝 𝛾, 𝛽 𝑐 𝑟 cos 𝛽 𝛾 𝜙 𝐷 sin 𝛾 𝐷 cos 𝛾 d𝛾 d𝛽

Page 20: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

39

<Rotating polar coordinates for fan‐beam geometry>

‘arbitrary’ point

Consider an ‘arbitrary’ point whose position is defined by 𝛾′ & 𝐷′

HW> Determine 𝐷′ & 𝛾 in terms of 𝑟 & 𝜙

40

Convolution kernel: 𝑐 𝑟 cos 𝛽 𝛾 𝜙 𝐷 sin 𝛾 𝑐 𝐷′  sin 𝛾 𝛾

𝑓 𝑟, 𝜙12

𝑝 𝛾, 𝛽 𝑐 𝐷′  sin 𝛾 𝛾 𝐷 cos 𝛾 d𝛾 d𝛽

HW> Show that 𝑐 𝐷′  sin 𝛾 

𝑐 𝛾

Let 𝑐 𝛾 ≡ 𝐷 𝑐 𝛾

HW> Show that

𝑓 𝑟, 𝜙1

𝐷′𝑝 𝛾, 𝛽 𝑐 𝛾 𝛾 d𝛾 d𝛽 where 𝑝 𝛾, 𝛽 cos 𝛾 𝑝 𝛾, 𝛽

Page 21: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

41

• 𝑞 · = filtered projection

• 𝑐 · = weighted (or windowed) version of 𝑐 ·

• 𝐷 = weight during backprojection

𝑓 𝑟, 𝜙1

𝐷′𝑝 𝛾, 𝛽 𝑐 𝛾 𝛾 d𝛾 d𝛽

1𝐷

𝑞 𝛾 , 𝛽 d𝛽𝑞 𝛾, 𝛽 𝑝 𝛾, 𝛽 ∗ 𝑐 𝛾

called “convolution weighted‐backprojection”

Helical CT reconstruction

42

• Consider the helical CT scenario with a single‐slice detector array

– No two projections will correspond to the same plane because of the continuous movement of the patient

– Projection, including the projection angle 𝛽 , is identified uniquely its longitudinal position 𝑧• Note 𝛽 𝛽 0, where 𝑀 angles are acquired over 360°

• How far has the patient moved when the angle repeats?

– Pitch (of the helix) 𝜁 𝑧 𝑧• (in turn) determined by the speed of gantry rotation & the table speed

• Too big pitch while too thin fan ⟹ “aliasing” artifacts

– To compensate, use thicker fan (natural analog low‐pass filter) ⟹ aliasing‐free but blurry

Page 22: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

43

• Only one or no projection is available at 𝑧– To “create” projections required for reconstruction of a plane at 𝑧, we use linear interpolation of 

the projections that are measured in nearby slices

• Estimation of the desired projection corresponding to the angle 𝛽

�̂� 𝛾, 𝛽𝑧 𝑧

𝜁𝑝 𝛾, 𝛽

𝑧 𝑧

𝜁𝑝 𝛾, 𝛽

Cone‐beam CT

44

• Cone‐beam w/ an area detector (compared to fan‐beam w/ a linear detector) provides rapid coverage & reconstruction of 3D volume

– C‐arm

– MDCT

• Conventional reconstruction for the central image plane, but not sufficient coverage to yield mathematically correct reconstructions in any other plane thru the object

• Cone‐beam reconstruction algorithm

– Feldkamp algorithm

Page 23: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Iterative reconstruction

45

• Most commercial CT systems use variants of CBP for reconstruction

• Current systems are rapidly changing over to iterative reconstructionmethods, offering equivalent (or better) reconstruction quality at substantially reduced dose

– More computationally demanding than CBP

• Popular in emission tomography (because larger & thus fewer voxels than that of CT)

– Adept at handling noise

• Framework

– Initial guess 𝑓 𝑥, 𝑦– Solving forward problem or forward projection, yielding 𝑔 ℓ, 𝜃– Compare the “forward‐estimated” 𝑔 ℓ, 𝜃 to the “measured” 𝑔 ℓ, 𝜃

– Modify the initial guess to yield a second guess 𝑓 𝑥, 𝑦 in response to differences b/w the estimated & observed projection values

Algebraic reconstruction technique (ART)

46

• Based on the Kaczmarz method for solving matrix equations

• Object can be represented by pixel values 𝑓 , 𝑗 1, … , 𝑚:

𝑓 𝑥, 𝑦 𝑓 𝑝 𝑥, 𝑦

– 𝑝 𝑥, 𝑦 = indicator functions describing each pixel location & shape (𝑗 1, … , 𝑚)

• Observation (or projection) is given by

𝑔 𝑎 𝑓

– 𝑎 = response to the pixel 𝑗

• Let us represent each collection of values as vectors: 𝐠 𝑔 , … , 𝑔 ), 𝐟 𝑓 , … , 𝑓 ), & 𝐚 𝑎 , … , 𝑎 )

Page 24: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

47

𝑔 𝐚 · 𝐟 or 𝐠 𝐴𝐟

Let us represent each collection of values as vectors:𝐠 𝑔 , … , 𝑔 )𝐟 𝑓 , … , 𝑓 )𝐚 𝑎 , … , 𝑎 )

Then we have:

where 𝐴 𝑎 describing concisely how to transform 

all the image pixel values into all the measurements

forward projection

𝐟 𝐴 𝐠

Straightforward solution:

However, 𝐴 is not square because the number of measurements is generally larger than the number of unknowns

48

– 𝐚 · 𝐟 forward projection applied to the current image estimate

– 𝐚 · 𝐟 𝑔 comparison of the current estimate to the actual observation by subtraction

𝐟 𝐴 𝐴 𝐴 𝐠

Instead of a solution, a pseudoinverse estimate by the least‐squares approach:

Also called the normal equationPractical only on very small problems

𝐟 𝐟𝐚 · 𝐟 𝑔

𝐚 · 𝐚𝐚

The Kaczmarz method:

To update 𝐟, the iteration uses one measurement at a time 

for each observation 𝑔

Page 25: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Maximum likelihood by expectation maximization (ML‐EM)

49

• Matrix 𝐴 in the ART

– Transform the pixel value 𝑓 to the measurement 𝑔

– Incorporate deterministic factors such as detector size, shape, efficiency, & object attenuation

– Not consider the statistical (Poisson) nature in the pixel values due to scattering & random noise

• ML‐EM

– Incorporate both deterministic & probabilistic factors

– Most popular in emission tomography

• Imaging equation

�̅� 𝐟 𝑎 𝑓 �̅�

– �̅� = expected random noise in detector 𝑖 due to scattering & other random nature

50

• ML estimate

– Collection of image pixel values 𝐟 (or 𝑓 , 𝑗 1, … , 𝑚)

– Maximize the likelihood function (or the logarithm of the likelihood function)

– Although various algorithms are available for finding the ML solution, the EM is the most popular & computationally convenient (i.e., iterative method)

• ML‐EM algorithm

– Seek to find the image pixel values 𝐟 that maximize the logarithm of the Poisson likelihood objective function

𝐿 𝐟 𝑔 ln �̅� 𝐟 �̅� 𝐟

Page 26: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

51

– 𝐟 = k‐th estimate of the image

– 𝑎 ∑ 𝑎 , 𝑗 1, … , 𝑚• Sensitivity factor characterizing the sensitivity of each voxel to the collection of measurements

– �̅� 𝐟 forward projection

–𝐟

comparison using the ratio (rather than a difference in the ART)

ML‐EM iteration:

𝑓𝑓

𝑎𝑎

𝑔�̅� 𝐟

52

• Lengthy reconstruction of ML‐EM

– Difficulty in convergence in the presence of significant noise

• Ordered subsets expectation maximization (OSEM)

– Grouping the observations (or projections) into an ordered sequence of subsets

• e.g., 15 subsets consisting of 24 views for a 360‐view observations

– Subsets represent mutually exclusive

– Subsets are treated independently in parallel

– One iteration of OSEM is a single pass through all the subsets, but the reconstructed image is updated after every subset, which accelerates convergence

Page 27: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Image Quality

53

• Resolution

– Finite widths of detectors

– Sampling limitations in:

• the number of projections

• the number of ray‐paths per projection

• Noise

– Limited attainable SNR considering the patient dose

• Artifact

– Beam hardening caused by energy‐selective absorption of x‐rays by human tissues

Resolution

54

• FBP = an exact formula for the inverse Radon transform

• Ideal (ramp) filter  𝜚 is not realizable

• Instead, an approximate filter  𝜚 𝑊 𝜚 is used in practice

• Moreover, the line integrals cannot be measured exactly due to the finite size of detectors

– Detector aperture locally integrate the underlying true signal, causing an additional filter 𝑆 𝜚– ℱ 𝑆 𝜚 𝑠 ℓ , the indicator function defining a detector

𝑓 𝑥, 𝑦 𝐺 𝜚, 𝜃 𝜚 𝑊 𝜚 𝑒 ℓ d𝜚ℓ

d𝜃

𝑓 𝑥, 𝑦 𝐺 𝜚, 𝜃 𝑆 𝜚 𝜚 𝑊 𝜚 𝑒 ℓ d𝜚ℓ

d𝜃

• 𝑓 𝑥, 𝑦 is a blurred version of 𝑓 𝑥, 𝑦

Page 28: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

55

• 𝑔 ℓ, 𝜃 = blurry projection

• Note the convolution property of the Radon transform

– Radon transform of the convolution of two functions is the convolution of their Radon transforms

Correspondingly,

𝑔 ℓ, 𝜃 ℱ 𝐺 𝜚, 𝜃 𝑆 𝜚 𝑊 𝜚

𝑔 ℓ, 𝜃 ∗ 𝑠 ℓ ∗ 𝑤 ℓ

≡ 𝑔 ℓ, 𝜃 ∗ ℎ ℓ

HW> Show thatℛ 𝑓 ∗ ℎ ℛ 𝑓 ∗ ℛ ℎ

𝑔 ℓ, 𝜃 𝑔 ℓ, 𝜃 ∗ ℎ ℓ ⟺ ℛ 𝑓 ℛ 𝑓 ∗ ℛ ℎ

Recognize that

56

• ℛ ℎ ℓ = PSF characterizing the resolution of the system

• FT of ℎ ℓ = ℱ 𝑠 ℓ ∗ 𝑤 ℓ or 𝐻 𝜚 𝑆 𝜚 𝑊 𝜚– Independent of 𝜃 (circularly symmetric!)

• Circularly symmetric 2D FT of ℎ 𝑥, 𝑦 𝐻 𝑞 𝑆 𝑞 𝑊 𝑞

– 𝑞 𝑢 𝑣– PSF is also circularly symmetric

ℎ 𝑥, 𝑦 ℛ ℎ ℓ

Therefore

𝑓 𝑥, 𝑦 𝑓 𝑥, 𝑦 ∗ ℛ ℎ ℓ

ℎ 𝑟 ℋ 𝑆 𝜚 𝑊 𝜚

Therefore

𝑓 𝑥, 𝑦 𝑓 𝑥, 𝑦 ∗ ℎ 𝑟 , where 𝑟 𝑥 𝑦

Page 29: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Example

Suppose a CT system uses rectangular detectors having width 𝑑 and a rectangular window function with highest frequency 𝜚 ≫ 1/𝑑.

What is the approximate PSF of this CT system?

57

Noise

58

• Measurement statistics

– The number of x‐rays in the beam

– Body attenuation

– Limited detector efficiency

– The measured photon count 𝑁∆

• 𝐼 = the observed intensity at a given detector

• 𝐸 = the effective energy of x‐ray beam (monoenergetic assumption)

• CT measurement for the 𝑖th detector & 𝑗th angle:

𝑔 ln𝑁𝑁

– 𝑁 = the incident photon count

– 𝑁 = the number of detected photons

Page 30: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

59

• In fact, 𝑁 is a Poisson r.v. w/ mean

𝑁 𝑁 𝑒

– 𝐿 = the ray‐path from the source to detector 𝑖 for the 𝑗th projection angle

• Note the CT measurement 𝑔 is a r.v. because it is a transformation of the r.v. 𝑁 (not a 

transformation of 𝑁 )

• Assuming 𝑁 is large & 𝑁 are independent;

Mean �̅� ln

Variance var 𝑔 HW> Derive the variance

60

• Image statistics (from the measurement statistics)

– Require the second‐order statistics

– Permit us to define a SNR of reconstructed images

• A measure of the quality of reconstructions

• Let us develop the mean and variance of the reconstructed linear attenuation coefficients

– Parallel‐ray geometry

– Discrete approximation to CBP integral

Page 31: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

61

– 𝑀 = the number of projections taken over  0, 𝜋– 𝑁 1 (odd) = the number of ray‐paths per projection

– 𝑇 = the physical spacing b/w detectors

– 𝑔 𝑔 𝑖𝑇

– �̃� · = a realizable approximation to the ramp filter

Recall

�̂� 𝑥, 𝑦𝜋𝑇𝑀

𝑔 𝑖𝑇 �̃� 𝑥 cos 𝜃 𝑦 sin 𝜃 𝑖𝑇

/

/

𝑓 𝑥, 𝑦 𝑔 ℓ, 𝜃 𝑐 𝑥 cos 𝜃 𝑦 sin 𝜃 ℓ dℓ d𝜃

In a discrete form:

62

• CBP has the desirable property that the mean of the reconstructed image approaches exact reconstruction as the quality of measurement increases

mean �̂� 𝑥, 𝑦𝜋𝑇𝑀

�̅� �̃� 𝑥 cos 𝜃 𝑦 sin 𝜃 𝑖𝑇

/

/

𝜎 𝑥, 𝑦 var �̂� 𝑥, 𝑦𝜋𝑇𝑀

var 𝑔 𝑐 𝑥 cos 𝜃 𝑦 sin 𝜃 𝑖𝑇

/

/

𝜋𝑇𝑀

1𝑁

𝑐 𝑥 cos 𝜃 𝑦 sin 𝜃 𝑖𝑇

/

/

𝜋 𝑇𝑀 𝑁

𝑐 𝑥 cos 𝜃 𝑦 sin 𝜃 𝑖𝑇

/

/

Assuming 𝑁 𝑁

Page 32: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

63

𝜋𝑀

𝑇 𝑐 𝑥 cos 𝜃 𝑦 sin 𝜃 𝑖𝑇 𝑐 𝑥 cos 𝜃 𝑦 sin 𝜃 ℓ dℓ d𝜃Note

𝜋 𝑐 ℓ dℓ 𝜋 𝐶 𝜚 d𝜚

Parseval’s theorem

Considering a rectangular window w/ bandwidth 𝜚 applied to the ramp filter:

𝜋 𝐶 𝜚 d𝜚 𝜋 𝜚 d𝜚2𝜋𝜚

3

Then, we have the reconstructed image variance that is independent of  𝑥, 𝑦 :

var �̂� 𝑥, 𝑦 𝜎𝜋𝑇𝑀𝑁

2𝜋𝜚3

2𝜋3

𝜚1𝑀

1𝑁/𝑇

64

• 𝜎 ↑ as the bandwidth 𝜚 ↑– Notice that noise has high‐frequency components

• 𝜎 ↓ as the detector spacing 𝑇 ↓– Decreasing detector size, which reduces the detector efficiency, will be offset by a lower 𝑁

• 𝜎 ↓ as 𝑁 ↑– 𝑁 ↑ (patient dose ↑)– 𝐸 ↑ (x‐ray absorption in the patient ↓, but loss of contrast)

• 𝜎 ↓ as 𝑀 ↑ (the more angles the better for constant acquisition time per angle)

• 𝜎 ↓ as 𝑁/𝑇 ↑– 𝑁/𝑇 = the average number of photons per unit distance along the detector array

– 𝑁/𝑇 ↑, in fan‐beam geometries, by increasing 𝑁

𝜎2𝜋

3𝜚

1𝑀

1𝑁/𝑇

Page 33: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Image SNR

65

SNR𝐶�̅�𝜎

where 𝐶

𝐶�̅�𝜋

𝜚 / 32

𝑁/𝑇 𝑀

0.4𝑘𝐶�̅�𝑑 / 𝑁/𝑇 𝑀

(In a good CT scanner design)𝜚 𝑘/𝑑, where 𝑘 1and 𝑑 is the width of a detector

0.4𝑘𝐶�̅�𝑑 𝑁𝑀If 𝑑 𝑇 (as in a 3G scanner)

66

• In the fan‐beam case:

SNR 0.4𝑘𝐶�̅�𝐿𝐷 / 𝑁 𝑀

– 𝑁 = mean photon count per fan‐beam

– 𝐷 = number of detectors

– 𝐿 = length of detector array

– 𝑁 𝑁 /𝐷 & 𝑑 𝐿/𝐷

• Note that SNR ~ 𝐷 / (SNR ↓ as 𝐷 ↑, something wrong?)

– Convolution of the projections with the ramp filter couples the noise b/w detectors!

– The extent of noise coupling increases w/ increasing 𝐷

Page 34: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

Example

Consider a fan‐beam CT system w/ one source, 𝐷 detectors, 𝑀 angles, and 𝐽 by 𝐽reconstructed images, where 𝐷 𝑀 𝐽 256. Assume that the width of each detector is 𝑑 0.25 cm and the ramp filter uses a rectangular w/ each cutoff 𝜚 1/𝑑. The scanner is used to image a lesion w/ contrast 𝐶 0.005 embedded in water (�̅� 0.15 cm‐1).

We require the image to have a SNR of at least 20 dB. What is the minimum number of photons per projection at the detectors that is required in order to meet this SNR constraint?

67

Artifacts

68

• Aliasing

– Aliasing in projections due to undersampling

• Higher frequency information will appear as lower frequency artifacts

• This aliasing continues thru the reconstruction process, and appears as artifacts in images

• Streak artifacts emanating from small bright objects w/i the image

– Insufficient number of projections

• Streaks emanating from object boundaries at points which have small radii of curvature

• Streaks are either dark or bright, depending on the precise location of the objects & edges w/i the FOV

– Insufficient number of points in the plane where the backprojection‐summation process is performed

• Moiré patterns

• A rule of a thumb: number of detectors = number of projections = number of points on the side of a reconstructed image (e.g., typical 3G scanner: 700 detectors, 1,000 projections, 512  512 image)

Page 35: Computed Tomography - Pusan National Universitybml.pusan.ac.kr/.../MedPhys/ComputedTomography.std.pdf · 2019. 10. 8. · Computed Tomography Ho Kyung Kim hokyung@pusan.ac.kr Pusan

69

• Beam hardening

– Caused by energy‐selective attenuation of x‐rays; increasing mean energy of the x‐ray spectrum while propagating the body

– Interpetrous lucency artifact particularly in head scans

– Streak artifacts particularly at the tips of bones & at metal pieces

• Other artifacts

– Electronic or system drift

• Ring artifacts

– X‐ray scatter

– Motion