1

IMAGE RECONSTRUCTION AND

ARTIFACT REDUCTION

AAPM 2019 1

Jiang Hsieh, PhD

Chief Scientist, GE Healthcare

Adjunct Professor, Univ. of Wisconsin-Madison

AAPM 2019 2

SINOGRAM

• Measurements from a single view form a horizontal line in a sinogram.

• Different views stack up to form a sinogram

• Sinogram is useful in identifying potential system issues

object cross-section

sinogram

single projection

Detector Channel

Pro

ject

ion

Vie

w

single projection

AAPM 2019 3

Reconstructed

Image

INTUITIVE RECONSTRUCTION

• Each measurement in the sinogram is generated by x-ray attenuated along a line

• Every point along the path has equal probability of contributing to the measurement

• The entire path is “painted” evenly with the measured intensity

• The process is repeated for all views with overlay

• This process is called “backprojection”

2

Filtered

Sinogram

AAPM 2019 4

Reconstructed

Image

Original

Sinogram

FILTERED BACKPROJECTION

• Blurry effect can be compensated for by “deconvolution” in either image or sinogram

• Sinogram “deconvolution” is a ramp shaped filter in frequency domain

• Combined with backprojection, the process is called “filtered backprojection”

K(w)

w-w w

|w|actual filter

f(x,y,z)= 𝑹𝟐

𝑳𝟐 𝒙,𝒚,𝜷𝒘(𝜸, 𝜷, 𝜶) 𝒉 𝜸 − 𝜸′ 𝒑 𝜸,𝜷, 𝜶 𝒅𝜸𝒅𝜷

AAPM 2019 5

w

K(w)

-w w

high-resolution

filter

w

K(w)

-w’ w’

low-resolution

filter

FILTER IMPLEMENTATION

• Rectangular window can be modified to shape the frequency response.

• Different cutoff frequencies to create different reconstruction kernels.

RECON FILTER IMPACT

AAPM 2019 6

• Higher frequency generally lead to sharper but noisier image

• For FBP, there is a tradeoff between spatial resolution and noise

StandardBone +

3

AAPM 2019 7

TARGETED RECONSTRUCTION

• Clinical interests sometimes focus only on a small region inside the object

• If the ROI is smaller than scan FOV, backprojection can be performed over a portion of the filtered projection to form a targeted reconstruction.

filtered projection

AAPM 2019 8

TARGETED RECONSTRUCTION

• Image pixel size changes linearly on the reconstruction FOV.

• Targeted reconstruction can result in better spatial resolution.

Recon at 50cm FOV Interpolated to 10cm FOV Recon at 10cm FOV

9

MODEL-BASED ITERATIVE RECONSTRUCTION (MBIR)

Closed-form Solution

FBP

f(x,y,z)= 𝑹𝟐

𝑳𝟐 𝒙,𝒚,𝜷𝒘(𝜸, 𝜷,𝜶) 𝒉 𝜸− 𝜸′ 𝒑 𝜸, 𝜷,𝜶 𝒅𝜸𝒅𝜷 ෝ𝒙𝑴𝑨𝑷 = 𝒂𝒓𝒈

𝒎𝒂𝒙𝒙

𝑷(𝒙/𝒚)

MBIR

Analytical Models

Compare

Measured Raw DataSynthesized Raw Data

Optimization

4

AAPM 2019 10

CLINICAL EXAMPLEONCOLOGY ABDOMEN/PELVIS 0.61MSV*

MBIR (Veo) FBP

AAPM 2019 11

MBIR

Analytical

Models

Compare

Measured Raw DataSynthesized Raw Data

Optimization

DLIR TrainingMeasured Raw Data

“Reference” Image

DEEP LEARNING IMAGE RECONSTRUCTIONDLIR Inferencing

Compare

12

RECONSTRUCTION COMPARISON

FBP IR (ASiR-V) DLIR (TrueFidelity)

5

AAPM 2019 13

Conventional Extended FOV DL-based Extended FOV

SFOV

Extended FOV

EXTENDED FOV RECONSTRUCTION FOV• For RT applications, desired FOV is larger than the detector SFOV• Conventional technology often fails to produce accurate patient skin line• DL-based algorithm can improve the outcome

AAPM 2019 14

IMAGE ARTIFACTS

• Nature of the X-ray Physics– Beam Hardening– Scatter – …

• Limitations of CT Technology– Helical– Cone Beam– …

• Patient-induced– Motion– Photon Starvation– …

• Operator-induced– Suboptimal protocols– Patient positioning– …

operator

patient

scanner

AAPM 2019 15

NATURE OF PHYSICS: BEAM HARDENING

• X-ray photons in CT are poly-energetic.

• The attenuation, m(E), is energy-dependent.

• The path length changes in objects

• The measured projection is not line integral of m. energy, E

m(E

)

0 cm water 15 cm water 30 cm water

6

AAPM 2019 16

WATER BEAM HARDENING CORRECTION

• Water beam hardening can be corrected with polynomial expansion of the projections.

no correction with correction

N

n

nn pp

1

'

channel

inte

nsi

ty

ideal

measured

AAPM 2019 17

BONE BEAM HARDENING

• The attenuation characteristics of bone is significantly

different from that of soft tissue.

• Similar observation can be made for contrast agents. 0.1

1

10

100

1000

10000

0 50 100 150

m (

1/c

m)

energy (keV)

Water BH Error Bone BH

Calcium

Water

AAPM 2019 18

80 kVp

DUAL ENERGY CORRECTION

140 kVp DECT

• Dual energy scans the object with two different kVps.

• Use of projection-space material decomposition can effectively remove BH

7

AAPM 2019 19

without MAR

with MAR

• Metal inside patient body is a major source of artifact

• Metal objects induce beam-hardening and photon starvation

• MAR algorithm is an iterative correction approach to restore

soft-tissue in the image

METAL ARTIFACT REDUCTION (MAR)

AAPM 2019 20

primary photons

object

scattered

photons

collimator

detector

input x-ray photons

SCATTER• Scatter increases with exposure volume.• Post patient collimator can be used effectively to reject the scatter.• A small portion of the scattered radiation can reach the detector.

AAPM 2019 21

PHANTOM TEST

160mm System with 1D Collimator

160mm System with Focusing 2D Collimator

8

• The amount of attenuation increases exponentially with path length.

• At low signal level, the noise in the projection is no longer dominated by the x-ray photon.

• Convolution filtering operation further amplifies the noise and leads to streak artifacts.

AAPM 201922

patient scan example

50cm FOVPHOTON STARVATION

LeI

I m-0

AAPM 2019 23

original

adaptively filtered

FBP

MBIR

ARTIFACT REDUCTION

Algorithmic Correctiono Adaptive filtering for streak reduction

o Iterative reconstruction

AAPM 2019 24

80kVp, 6.44mGy 140kVp, 6.26mGy

0.1

1

10

100

0 50 100 150

energy (keV)

m (

1/c

m)

PROTOCOL OPTIMIZATION

o Scan speed

o Helical pitch

o kVp selection

o Slice thickness

o Reconstruction kernel

o mA modulation

Calcium

Water

9

AAPM 2019 25

VOLUNTARY MOTION

• Better patient training

• Better instruction

• Faster scan speed

• Advanced algorithm

AAPM 2019 26

OPERATOR-INDUCED

• CT operator plays an important role in artifact reduction.

• Improperly secured patient can lead to motion artifact.

AAPM 2019 27

centered

s=28.810cm up

s=53.8

PATIENT CENTERING

• Bowtie filter is employed in most scanners.

• It assumes centered oval shaped objects.

• Patient centering is important.

10

AAPM 2019 28

Missing Frequency Z-truncation Redundant Data Mishandling

WIDE-CONE: MATHEMATICAL CHALLENGE

• Axial wide-cone reconstruction faces three major technical challenges:

– missing frequency (axial mode sampling pattern)

– z-truncation (cone-beam geometry)

– redundant data handling (half-scan to improve temporal resolution)

• Solution to these challenges requires extensive research and testing.

29

Advanced Reconstruction

Traditional Reconstruction

ALGORITHM COMPARISON

AAPM 2019 30

HIGH HELICAL PITCH IMPACT

Pitch 3.2

Pitch 1.2

• When helical pitch is too large, incomplete sample can induce image artifacts

11

AAPM 2019 31

SUMMARY

• Three generations of reconstruction algorithmo Filtered backprojection (FBP)

o Iterative reconstruction (IR)

o Deep-learning image reconstruction (DLIR)

• Many sources of CT artifactso Nature of Physics

o New Technologies

o Patient

o Operator

• It takes a combined efforts from manufactures, CT operators, and patients to

obtain the best image quality.

AAPM 2019 32