Role of image quality in dose management via/through DRL Ehsan - - PDF document

▶
role of image quality in dose management via through drl
SMART_READER_LITE
LIVE PREVIEW

Role of image quality in dose management via/through DRL Ehsan - - PDF document

Mar 08, 2024 453 likes •881 views

11/21/19 Role of image quality in dose management via/through DRL Ehsan Samei, PhD, FAAPM, FSPIE, FAIMBE, FIOMP Department of Radiology Duke University Health System 1 2 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior

slide-1
SLIDE 1

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 1

Role of image quality in dose management via/through DRL

Ehsan Samei, PhD, FAAPM, FSPIE, FAIMBE, FIOMP Department of Radiology Duke University Health System

1 2
slide-2
SLIDE 2

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 2

3

The challenge of dose optimization: the monotonic relationship between quality and dose!

Radiation Dose

?

Image Quality

slide-3
SLIDE 3

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 3

The challenge of dose optimization: the monotonic relationship between quality and dose!

Radiation Dose

?

Image Quality relevance robust smart relatable practical

The challenge of dose optimization: the monotonic relationship between quality and dose!

Radiation Dose

? ?

Image Quality relevant robust smart relatable practical

slide-4
SLIDE 4

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 4

Image quality

7

The challenge of dose optimization: the monotonic relationship between quality and dose!

Radiation Dose

? ?

Image Quality relevance robust smart relatable practical

?

slide-5
SLIDE 5

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 5

What dose is optimum?

High dose Low dose

Safety (dose) is inherently linked to indication-based image quality.

Low dose High dose

slide-6
SLIDE 6

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 6

Factors that govern quality and safety of medical imaging

An Ideal Image (low) Resolution (high) Noise Dose

What is image quality?

  • Aesthetic: Subjective perception of quality
  • Task-generic: The realism of the image to

represent the reality of the object being imaged

  • Task-specific: The ability of the image to

render the information pertinent to the task at hand

12
slide-7
SLIDE 7

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 7

What is the right quality metric?

Clinical Quality Clinical Trials Virtual Clinical Trials Phantom Measures Intrinsic Specs Anthro. Models Animal Models Case studies

Goal

What is the right quality metric?

Clinical Quality Clinical Trials Virtual Trials Anthro. Models Animal Models Case studies Phantom Measures Intrinsic Specs

Goal

Simple Relevance inferred Complex Relevant

slide-8
SLIDE 8

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 8

What are the right metrics?

  • 1. Relevant: As much as possible, patient-/indication-

centric (not modality or machine)

  • 2. Robust: To ensure reliability and applicability

(quantitative not subjective)

  • 3. Smart: Maintained balance between robustness

and relevance

  • 4. Relatability: Surrogates relatable to clinical task
  • 5. Practical: Economic to measure

Balance robustness and relevance

  • To extent possible, we need to move toward

relevance while keeping robustness in check

  • As Relevant as Reasonably Achievable

Simple Relevance inferred Complex Relevant

ARARA!

slide-9
SLIDE 9

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 9

Image quality metrics

Task Generic

  • 1. Contrast
  • 2. Resolution
  • 3. Noise
  • 4. SNR, CNR, SdNR
  • 5. DQE, eDQE, eDE
  • 6. TG IQ in vivo

Task Specific

  • 1. Threshold Contrast
  • 2. Detectability, d’
  • 3. Estimability, e’
  • 4. d’ in vivo
17
  • 1. Contrast
18
slide-10
SLIDE 10

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 10

19

Contrast

  • Fractional difference in the signal or brightness

between two regions of an image

High Contrast Low Contrast

20

Contrast

  • Best characterized by fractional signal difference

(ie, subject contrast) or fractional brightness difference (ie, display contrast) of a target in comparison to background: ! = #$%&'($ − #*%+,'&-./0 #*%+,'&-./0 = log #$%&'($ #*%+,'&-./0

slide-11
SLIDE 11

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 11

21

Contrast

! = #$%&'($ − #*%+,'&-./0 #*%+,'&-./0 = log #$%&'($ #*%+,'&-./0

  • 2. Resolution
22
slide-12
SLIDE 12

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 12

23

Resolution

  • Ability to resolve distinct features of an

image from each other

Low resolution High resolution

24

Resolution

  • Best characterized by the modulation

transfer function (MTF):

– The efficiency of an imaging system in reproducing subject contrast at various spatial frequencies

MTF(f) = F{LSF(x)}

LSF = response of a system to a perfect line

slide-13
SLIDE 13

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 13

25

INPUT OUTPUT

= x

MTF

f Imaging System 26

INPUT

Imaging System
slide-14
SLIDE 14

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 14

27

High MTF

28

Low MTF

slide-15
SLIDE 15

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 15

29

MTF and limiting resolution

  • Limiting resolution ~ Frequency at 10% MTF
  • Mammography

5-10 lp/mm

  • Radiography

2-5 lp/mm

  • Fluoroscopy

1-2 lp/mm

  • CT

0.5-1 lp/mm

30

Effect of added blur

slide-16
SLIDE 16

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 16

  • 3. Noise
31 32

Noise

  • Unwanted signals that interfere

with interpretation

Low resolution High resolution High res/ high noise

slide-17
SLIDE 17

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 17

33

Underexposure by 4x Correct exposure

An underexposed image is “too noisy”

34 120 kVp, 25% less mAs 120 kVp, photo-timed
slide-18
SLIDE 18

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 18

35

Noise

  • Best characterized by the noise

power spectrum (NPS):

– The variance of noise in an image in terms of the spatial frequencies

òò

+ + = dxdy y x f y x f ACF ) , ( ) , ( ) , ( h c h c

NPS(f) = F{ACF(x)}

36

Noise

  • NPS
x x f

Image Data ACF NPS Example 1 Uncorrelated Noise

x x f

Example 2 Correlated Noise

slide-19
SLIDE 19

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 19

37

Image without noise

38

Uncorrelated noise

slide-20
SLIDE 20

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 20

39

Correlated noise

40

Effect of added noise

slide-21
SLIDE 21

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 21

  • 4. SNR, CNR, SdNR
41

SNR

  • Signal to noise ratio
42

!"# = %&'()*& +,'-.)(/012

slide-22
SLIDE 22

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 22

CNR and SdNR

  • In Linear systems:

– N in CNR is relative noise – N in SdNR is noise

  • In log systems, Sd in C and N in relative noise
43

CNR = Itarget − Ibackground Ibackground σ background Ibackground = Itarget − Ibackground σ background = SdNR

44

SNR(f)

  • Noise Equivalent Quanta (NEQ) and frequency-

dependent SNR

  • Affected by the collective effects of resolution

and noise and associated contributing factors

NEQ( f ) = SNRactual

2

( f ) = MTF

2( f )

NPS( f )

slide-23
SLIDE 23

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 23

CNR(f)

45

CNR2( f ) = C2( f )MTF 2( f ) NPS( f ) = C2( f )SNR2( f )

  • 5. DQE, eDQE, eDE
46
slide-24
SLIDE 24

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 24

  • 6. In vivo task-generic

metrics in vivo image quality

Noise

Christianson et al., AJR, 2014

Resolution

Sanders et al., Medical Physics, 2016

Organ-based HU

Abadi et al., Medical Physics, 2017

Perceptual Quality

Samei et al., Medical Physics, 2014
slide-25
SLIDE 25

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 25

in vivo noise prediction

in vivo NPS

Segment the liver (Fu 2018) Identify liver parenchyma

– Avoid vasculature, fatty deposits

Sub-segment parenchyma and de-trend image data

– Use local polynomial fits for segment Organ Segmentation Sub-organ Segmentation Goal 1. Anatomical structure 2. Large scale trends Wiener-Khinchin Theorem: NPS(u) = FT{RN(x1;x2)} = FT{E[N(x1).N(x2)]}

Estimate NPS from non-square patches in the liver
slide-26
SLIDE 26

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 26

in vivo resolution

Sanders et al., Medical Physics, 2016

GE HD750

5 10 15 20 25 30 35 20 25 30 35 40 45 noise (HU) patient diameter (cm) 5 10 15 20 25 30 35 40 45 50 20 25 30 35 40 45 patient diameter (cm) Phantom

in vitro (phantom) noise

Siemens Flash

Ria et al, AAPM 2018 Phantom
slide-27
SLIDE 27

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 27

GE HD750

5 10 15 20 25 30 35 20 25 30 35 40 45 noise (HU) patient diameter (cm) 5 10 15 20 25 30 35 40 45 50 20 25 30 35 40 45 patient diameter (cm)

in vivo (phantom) noise

Siemens Flash

Phantom Phantom Patients Patients

Task-based metrics

61
slide-28
SLIDE 28

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 28

Task-based indicators

  • Direct measures of task performance
  • Direct measures of task-like performance
  • Derivatives of task performance from

generic indicators

62

Direct measures of task performance

  • Most reliable
  • $ and time-consuming
  • Not translatable to other tasks or conditions
63
slide-29
SLIDE 29

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 29

Direct measures of task-like performance

  • Often based on simplistic tasks
  • Most $ and time requirement
  • Not translatable to other tasks or conditions
64

Derivatives of task performance from generic indicators

  • Most practical method to assess task

performance

  • Subject to linearity constraints of generic

indicators

  • Observer models
65
slide-30
SLIDE 30

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 30

66

Threshold Contrast

  • Fundamental ability to be able to see things in images
  • Rose model:
C = threshold contrast k = constant (3-5) A = area of signal N = number of photons SNR = signal-to-noise ratio

CT images of the low-contrast phantom

B31s I31s-5

slide-31
SLIDE 31

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 31

Noise Power Spectrum Observer models

Task functions

d ' AZ = f (d ')

  • 10
  • 1
1
  • 1
1
  • 1
1
  • 0.5
0.5
  • 8
  • 6
  • 4
  • 1
1
  • 1
1
  • 1
1
  • 0.5
0.5 0.2 0.4 0.6 0.8 1

MTF 1 NPS 1

Spatial frequency (/mm) Spatial frequency (/mm) axial axial coronal coronal 10^ 0.2 0.4 0.6 0.8 1
  • 1
1
  • 1
1
  • 1
1
  • 0.5
0.5
  • 1
1
  • 1
1
  • 1
1
  • 0.5
0.5
  • 10
  • 8
  • 6
  • 4

MTF 2 NPS 2

Spatial frequency (/mm) axial axial coronal coronal 10^
slide-32
SLIDE 32

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 32

Observer models

  • Mathematical descriptions of how human

visual system processes medical images for an interpretive task

73

Measuring image quality

Factors that affect task performance

  • 1. Contrast
  • 2. Lesion size
  • 3. Lesion shape
  • 4. Lesion edge profile
  • 5. Resolution
  • 6. Viewing distance
  • 7. Display
  • 8. Noise magnitude
  • 9. Noise texture
  • 10. Operator noise

Feature of interest Image details Distractors

slide-33
SLIDE 33

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 33

Observer models

Resolution and contrast transfer Attributes of image feature of interest Image noise magnitude and texture

  • dNPWE
'

( )

2 =

MTF

2(u,v)WTask 2 (u,v)

∫∫

E

2(u,v)dudv

[ ]

2

MTF

2(u,v)WTask 2 (u,v)

∫∫

NPS(u,v)E

4(u,v) + MTF 2(u,v)WTask 2 (u,v)Ni dudv Richard, and E. Samei, Quantitative breast tomosynthesis: from detectability to estimability. Med Phys, 37(12), 6157-65 (2010). Chen et al., Relevance of MTF and NPS in quantitative CT: towards developing a predictable model of quantitative... SPIE2012

Observer models

Fisher-Hotelling observer (FH) Non-prewhitening observer (NPW) NPW observer with eye filter (NPWE)

dFH

'

( )

2 =

MTF 2(u,v)WTask

2 (u,v)

NPS(u,v)

∫∫

dudv

dNPW '

( )

2 = MTF 2(u,v)WTask 2 (u,v)

∫∫

dudv

[ ]

2 MTF 2(u,v)WTask 2 (u,v)

∫∫

NPS(u,v)dudv dNPWE '

( )

2 = MTF 2(u,v)WTask 2 (u,v)

∫∫

E 2(u,v)dudv

[ ]

2 MTF 2(u,v)WTask 2 (u,v)

∫∫

NPS(u,v)E 4(u,v) + MTF 2(u,v)WTask 2 (u,v)Ni dudv 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0.2

Eye filter

slide-34
SLIDE 34

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 34

0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 1.0 80 kVp 100 kVp 120 kVp 140 kVp 0.0 0.2 0.4 0.6 0.8 0.0 0.2 0.4 0.6 0.8 1.0 80 kVp 100 kVp 120 kVp 140 kVp 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.0 0.2 0.4 0.6 0.8 1.0 80 kVp 100 kVp 120 kVp 140 kVp 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.0 0.2 0.4 0.6 0.8 1.0 80 kVp 100 kVp 120 kVp 140 kVp Spatial frequency (/mm) Spatial frequency (/mm) Spatial frequency (/mm) Spatial frequency (/mm) 120 kVp 120 kVp 120 kVp 120 kVp Task function (AU) Task function (AU) (a) (b) (c) (d) 10 mm 10 mm 10 mm 10 mm

Small feature no iodine Large feature no iodine Small feature with iodine Large feature with iodine

Task functions

F

Task function (Wtask) Iodine concentration Size Lesion diameter = {5, 7.5, 10, 12.5, 15 mm} Lesion contrast = {100, 150, 200, 250, 300 HU}

Task function (Wtask)

slide-35
SLIDE 35

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 35

FBP IRIS SAFIRE3 SAFIRE5

AUC vs. dose

Chen, SPIE, 2013

Az = f(d’)

CNR vs observer performance

Christianson et al, Radiology, 2015
slide-36
SLIDE 36

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 36

d’ vs observer performance

Christianson et al, Radiology, 2015

Mercury Phantom

Detectability (d’), resolution, and noise per size AAPM 233: https://www.aapm.org/pubs/reports

3 5 5 m m 2 2 m m 2 9 m m 1 7 7 m m 1 1 2 m m 6 m m 8 5 m m 8 5 m m 6 m m 9 m m
slide-37
SLIDE 37

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 37

Automated Characterization

in vivo detectability index

slide-38
SLIDE 38

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 38

in vivo detectability index

Smith et al, JMI 2018

in vivo chest image quality

IQ Parameters Lung Grey level Lung Detail Lung Noise Rib-Lung Contrast Rib Sharpness Mediastinum Detail Mediastinum Noise Mediastinum Alignment Subdiaphragm-Lung Contrast Subdiaphragm Area

Lin et al, Medical Physics 2012; Samei et al, Medical Physics 2014
slide-39
SLIDE 39

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 39

Published consistency and quality ranges

Samei et al, Medical Physics 2014 Kodak: 1051 Philips: 1741 Carestream: 997 GE: 1084

Quality Consistency

Lung grey level

slide-40
SLIDE 40

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 40

Lung noise

Consistency Quality

5 10 15 20 25 30 35 40 Brier Creek Butner Creedmoor Henderson Family Triangle Family Wake Forest Family Hillsborough Family Mor risville Knightdale Mebane Timber lyne Croasdaile South Average Lung Noise 5 10 15 20 25 30 35 Brier Creek Butner Creedmoor Henderson Family Triangle Family Wake Forest Family Hillsborough Family Mor risville Knightdale Mebane Timber lyne Croasdaile South Average Lung Detail

13 - 42 19 - 31 Consistency Quality

slide-41
SLIDE 41

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 41

0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45

AJ19 AK151 AM235 BARWI003 BPS12 DAVIS090 DG97 DW77 ED108 HITE0001 JACOB074 KCG24 KMD48 KS324 LAM41 LB111 LH154 MALES003 MCG MH225 MOORE237 MW187 NA RICKM005 SC134 TA34 TE45 TML113 VJ28

Average Subdiaphragm Area

0.08 – 0.49 0.17 – 0.27

Take-home Points

97
slide-42
SLIDE 42

11/21/19 (c) Ehsan Samei, 2019. Use for non-personal purposes by prior permission only. 42

Take-home points

  • The point of having image quality metrics is to have

such metrics relatable to the purpose of imaging => clinical image quality

  • C, MTF, NPS offer generic indicators relatable to

task-specific metrics of Contrast Threshold, Detectability, Estimability

  • in vivo measures provide a window into image

quality across patient cases

98

Questions?

99