Intro - Display Processing

Henry Ford
Health System
Display processing is used to transform digital
RADIOLOGY RESEARCH radiography data to display values for
presentation using a workstation or film printer.
Digital Image Processing DETECTION DISPLAY

in
Radiography

Michael Flynn
(A) Subject contrast
Dept. of Radiology
(B) is recorded by the detector
[email protected]
(C) and transformed to display values
(D) that are sent to a display device
(E) for presentation to the human visual system. 1
M. Flynn 2007

Intro - Course Outline Intro - Learning objectives

- Introduction (4)
1. Preprocessing (12) 1. Understand how recorded signals are conditioned
to produce image data for processing.
2. Generic Image Processing (2)
A. Grayscale rendition (10)
2. Understand the approaches used to improve the
B. Exposure recognition (7)
visibility of structures in radiological images.
C. Edge restoration (10)
D. Noise reduction (10)
3. Survey current commercial implementations and
E. Contrast enhancement (14)
distinguish essential similarities / differences.
3. Commercial Implementations (23)

M. Flynn 2007 2 M. Flynn 2007 3

 Intro - Disclosure Projection Test Pattern

The presenter is a designated principal investigator on research

agreements between Henry Ford Health System and the following
companies (alphabetical);
* Agfa Medical Systems
Brown & Herbranson imaging
* Eastman Kodak Company 12 / 0 243 / 255

Shimadzu Medical Systems

Roche Pharmaceuticals
The presenter has provided consulting services over the last 12
months with the following companies (alphabetical);
Gammex-RMI
* Vidar Systems Corp.

243 / 255 AAPM TG18 PQC 12 / 0

* Involves DR image processing

M. Flynn 2007 4 M. Flynn 2007 5

1- Course Outline 1 - Raw Image Data

• For CR and DR systems, radiation energy deposited

in the detector is converted to electrical charge.
• Preamplifier circuits then convert this to a voltage
which is digitized using analog to voltage converter
(ADC) to produce RAW image values.
1. Preprocessing
RAW image
2. Generic Image Processing
3. Commercial Implementations
e- V ADC
#

preamp

M. Flynn 2007 6 M. Flynn 2007 7

 1 - DR ‘For Processing’ Data 1 – Bad pixels

RAW data from the detector is pre-processed • Pixels with high or low values or with excessive noise
to produce an image suitable for processing.
• Values corrected by interpolation from neighbors
DR FOR • There are presently no requirements to report bad
RAW PROCESSING
pixel statistics as a part of DR system purchase.

LINEAR LOG

DICOM SOP Class

For Processing
Digital X-ray
BAD Image Storage
PIXELS
DARK
GAIN 450 x 200 region

M. Flynn 2007 UID 1.2.840.10008.5.1.4.1.1.1.1.1 8 M. Flynn 2007 9

1 – New Bad pixels 1 – Dark image Digital Fluoroscopy dark image

• The signal recorded when

no x-rays are incident on
• New pixel defects can the detector is referred to
develop in DR panels as the ‘dark image’ or
that are in service. ‘offset image’.
• Frequent gain
• Most detectors produce a
calibration can help signal that linearly increase
detect newly developed from the offset value of
problems. each pixel as x-ray incident
exposure is increased.
• The defects shown to
the right were reported
• Dark image values are
by the radiologist susceptible to drift and
interpreting the study. often have high thermal
dependence.

Indirect DR Display Window = 0-20

M. Flynn 2007 10 M. Flynn 2007 11
 1 – Gain image 1 - Offset/Gain correction

• The linear gain may slightly differ from pixel to pixel. • Dark Image (ID)
• These variations produce fixed pattern noise. Obtained by averaging many images obtained
with no xray input to the detector.
• Gain Image (IG)
Obtained by averaging many images obtained
with a uniform x-ray fluence.
• Uniformity correction is performed subtracting the
dark offset and adjusting for gain differences.
ICOR = (IRAW – ID) {k/ (IG – ID)}
• Log transformation using a Log look-up table allows
this to be performed with a subtraction.
IFP = log (IRAW – ID) - log(IG – ID) - K
Uniform radiation exposure
M. Flynn 2007 12 M. Flynn 2007 13

1 – log image values 1 – ‘for processing’ Log format

The recorded signal recorded is
approximately proportional to the • Most ‘for processing’ image values are proportional to
exponent of the attenuation the log of the exposure incident on the detector.
µ(s) coefficient line integral; • Samei et.al., Med Phys 2001
P(x,y) = µ(s) • Agfa, PV = 1250 * log(cBE) -121
s
I(x,y) α Io exp[ - P(x,y) ] • Fuji, PV = (1024/L)*(log(E) + log(S/200)
• Kodak, PV = 1000*log(E) +Co
s The log of the recorded signal is
proportional to the line integral. 3000

For IFP values stored as a 12 bit

Ln(I(x,y)) α -P(x,y) +Ln(Io) 2500
DR7100
number (0 – 4095), a convenient
format has a change of 1000 for
Small perturbations cause the 2000 every factor of 10 change in
same image value change whether exposure.
in high or low transmission regions 1500
IFPRAW
IFP = 1000 log10( mR ) + 2000
I1FP α P1(x,y) + ∆P
1000

I2FP α P2(x,y) + ∆P 14
0.1 mR 1.0 10.0
15
M. Flynn 2007 M. Flynn 2007
 1 - IFP proportional to mR1/2 1 - Normalized IFP values, TG116

AAPM Task group 116 draft report

• One major manufacturer uses internal IFP values that are
proportional to the square root of exposure. “Recommended Exposure Indicators for Digital Radiography”
• The relative noise of the IFP values is constant for all Normalized For Processing Pixel Values (INFP)
incident exposures, however the tissue contrast is not. “For-processing pixel values, IFP, that have been converted
to have a specific relation to a standardized radiation
IFP = 1250 mR1/2 1200 exposure (ESTD). ..,”
7000
ADC MD40 Normalized for Processing Values
1000
For this system, this 6000

structure is used only 800 5000

for data stored in a INFP = 1,000*log10(ESTD/Eo) , 4000
600
multi-scale Agfa format

I_nfp
ESTD in micro-Gray units,
used by Agfa products. RAW
3000
400
Data exported using
RAW**2/1k
Eo = 0.001 micro-Gray, 2000

DICOM exchange (for

200
1000
processing) can be sent 0 0
in a log exposure format. 0.0 0.2 mR 0.4 0.6 0.8 0.01 0.1 1 mGy 10 100 1000

M. Flynn 2007 16 M. Flynn 2007 17

2- Course Outline 2 - Five generic processes

Grayscale Rendition: Convert signal values to display values

Exposure Recognition: Adjust for high/low average exposure.
Edge Restoration: Sharpen edges while limiting noise.
Noise Reduction: Reduce noise and maintain sharpness
Contrast Enhancement: Increase contrast for local detail

1. Preprocessing
2. Generic Image Processing
3. Commercial Implementations

For Processing For Presentation

M. Flynn 2007 18 M. Flynn 2007 19
 2A - processing sequence 2A - Grayscale Rendition
4000

Grayscale LUTs
3000
Grayscale Rendition: Convert signal values to display values
Exposure Recognition: Adjust for high/low average exposure. ‘For Processing’ data 2000
5 - HC-CR

values are transformed to

8 - MID-VAL

Edge Restoration: Sharpen edges while limiting noise. 11 - LIN

presentation values using

1000

Noise Reduction: Reduce noise and maintain sharpness a grayscale Look Up Table 0
0 500 1000 1500 2000 2500 3000 3500 4000
Contrast Enhancement: Increase contrast for local detail

Spatial Processes
Exposure •Edge Restoration Grayscale
Recognition •Noise Reduction (VOI-LUT)
•Contrast Enhance

5-5 8-8 11-11

M. Flynn 2007 20 M. Flynn 2007 21

2A - Presentation Values 2A - DICOM VOI LUT

The VOI-LUT may be applied by the modality, or

The Grayscale Value of Interest
(VOI) Look up Table (LUT) sent to an archive and applied by a viewing station
Log-luminance

transforms ‘For Processing’ values

to ‘For Presentation Values.
Spatial Processes
Monitors and printers are DICOM
Exposure •Edge Restoration Grayscale
calibrated to display presentation
Recognition •Noise Reduction (VOI-LUT)
values with equivalent contrast.
Presentation Values •Contrast Enhance
The VOI-LUT optimizes the (VOI-LUT)

display for radiographs of

Grayscale VOI-LUT specific body parts.
DICOM PS 3.3 2007, Pg 88
• When the transformation is linear, the VOI LUT is described by
the Window Center (0028,1050) and Window Width (0028,1051).
For Processing Values
• When the transformation is non-linear, the VOI LUT is described
by VOI LUT Sequence (0028,3010).

M. Flynn 2007 22 M. Flynn 2007 23

 2A - VOI LUT sent with image values 2A - LUT applied and P values sent

When communicating images to a PACS Presently, many systems send images to a PACS
systems, it can be beneficial to send the system as scaled P values with the VOI LUT already
VOI-LUT sequence for application at display. applied to the processed data.

PACS
120
4500

workstations 4000 PACS workstations 100

should be 3500
histogram
VOI LUT
can not adjust the histogram

VOI-LUT to

Rel Probability
80
capable of 3000

demonstrate
P value

translating or
2500 60

stretching the
2000
contrast in over or
under penetrated
40
1500
VOI LUT to
make contrast
1000
regions. 20
500
and brightness 0 0

changes 0 500 1000 1500 2000 2500 3000 3500 4000 0 500 1000 1500 2000 2500 3000 3500 4000
Image value with applied VOI-LUT
Image value with VOI-LUT sequence

M. Flynn 2007 24 M. Flynn 2007 25

2A - A better WW/WL for CR/DR 2A - A better WW/WL for CR/DR

The applied VOI-LUT produces good contrast for the The applied VOI-LUT produces good contrast for the
primary tissues of interest. For the full range of P values, primary tissues of interest. For the full range of P values,
contrast is limited in the toe and shoulder regions. contrast is limited in the toe and shoulder regions.

P value WW/WL P value WW/WL = 4000/2000

4000 4000/2000 4000

3000 3000

2000 2000

1000 1000

Raw Image Value Raw Image Value

M. Flynn 2007 26 M. Flynn 2007 27

 2A - A better WW/WL for CR/DR 2A - A better WW/WL for CR/DR

Shifting the Window Level (WL) to inspect highly The ability to shifting the VOI-LUT at the display
penetrated regions renders gray levels with a poorly workstation permits regions of secondary interest to be
shaped portion of the VOI LUT. viewed with good radidographic contrast.

P value WW/WL = 1000/3500 P value

4000 4000

3000 3000

2000 2000

1000 1000

Raw Image Value Raw Image Value

M. Flynn 2007 28 M. Flynn 2007 29

2B – Exposure Recognition 2B – Exposure recognition - signal

Signal Range:
Grayscale Rendition: Convert signal values to display values A signal range of up to 104 can be recorded by digital
radiography systems. Unusually high or low exposures
Exposure Recognition: Adjust for high/low average exposure.
can thus be recorded. However, display of the full range
Edge Restoration: Sharpen edges while limiting noise. of data presents the information with very poor
Noise Reduction: Reduce noise and maintain sharpness contrast. It is necessary to determine the values of
interest for the acquired signal data.
Contrast Enhancement: Increase contrast for local detail

100
log(S) probability

Spatial Processes
Exposure •Edge Restoration Grayscale
Recognition •Noise Reduction (VOI-LUT)
•Contrast Enhance
0
2000 log(S) value 4000
M. Flynn 2007 30 M. Flynn 2007 31
 2B – Exposure recognition: regions 2B – Exposure recognition: VOI LUT

Exposure Recognition: VOI LUT Level and Width:

All digital radiographic systems have an exposure recognition • The values of interest obtained from exposure recognition
process to determine the range and the average exposure to the processes are used to set the level and width of the VOI LUT.
detector in anatomic regions. A combination of edge detection, • Areas outside of the collimated field may be masked to prevent
noise pattern analysis, and histogram analysis may be used to bright light from adversely effecting visual adaptation.
identify Values of Interest (VOI).

D 100
100 A A

log(S) probability
log(S) probability

C
C B
C B

B
D
0
2000 log(S) value 4000
0
2000 log(S) value 4000
M. Flynn 2007 32 M. Flynn 2007 33

2B – Segmentation – Anatomic region 2B – Exposure recognition: metrics

Tissue region
• DR systems report a metric indicating the detector
Advanced image segmentation response to the incident radiation exposure.
algorithms are used is some • The methods used to deduce this metric are all different
systems to identify the region •The regions from which exposure is measured vary.
where tissue attenuation has •Reported exposures may increase proportional to the log of
occurred. This provides exposure or may vary inversely with exposure.
information on the values of •The scale of units varies widely with factor of 2 changes in
interest for presentation. exposure associated with changes varying from 0.15 to 300.

•Fuji: S = 200/Ein 80 kVp, unfiltered

•Agfa: lgM = 2.22 + log(Ein)+log(Sn/200) 75 kVp, 1.5 Cu (mm)
•Kodak: EI = 1000 log(Ein) + 2000 80 kVp, 0.5 Cu 1.0 Al
X. Wang, H. Luo,“Automatic and exam-type independent
algorithm for the segmentation and extraction of
foreground, background, and anatomy regions in digital Anatomic
radiographic images,” Proc. SPIE 5370, 1427-1434, 2004.
region
M. Flynn 2007 34 M. Flynn 2007 35
 2B – Exposure Indicators, TG116 2C – Edge Restoration

AAPM Task group 116 draft 8b

“Recommended Exposure Indicators for Digital Radiography” Grayscale Rendition: Convert signal values to display values
Exposure Recognition: Adjust for high/low average exposure.
Indicated Equivalent Air Kerma (KIND) [IEC, Exposure Index] Edge Restoration: Sharpen edges while limiting noise.
• An indicator of the quantity of radiation that was incident Noise Reduction: Reduce noise and maintain sharpness
on regions of the detector for each exposure made. …
• The regions .. may be defined in different ways .. Contrast Enhancement: Increase contrast for local detail
• The value should be reported in units of microgray ..
Relative Exposure (EREL) -> Deviation Index [IEC]
• An indicator as to whether the detector response for a
specific image, KIND, agrees with KTAR(b.v). Spatial Processes
• Relative exposures are to be reported as Exposure •Edge Restoration Grayscale
EREL= log10( KIND/KTAR (b,v) )
Recognition •Noise Reduction (VOI-LUT)
•Contrast Enhance
• EREL is intended as an indicator for radiographers and
radiologists as to whether the technique used to acquire a
radiograph was correct.
M. Flynn 2007 36 M. Flynn 2007 37

2C – Edge Restoration 2C – Without Edge Restoration 2.0

Lateral knee view with

filter strength

1.5
• Radiographs with high contrast
details input high spatial
Signal Power equalization but no edge 1.0

frequencies to the detector. restoration as indicated 0.5

• For many systems the detector

by the filter strength. relative spatial frequency
Frequency 0.0
will blur this detail as indicated 0.0 0.2 0.4 0.6 0.8 1.0

by the MTF.
MTF
• Enhancing these frequencies can
help restore image detail.
• However, at sufficiently high
frequencies there is little signal Frequency
left and the quantum mottle
(noise) is amplified. Noise Power

• The frequency where noise

exceeds signal is different for
different body parts/views Frequency

M. Flynn 2007 38 M. Flynn 2007 39

 2C – With Edge Restoration 2C – With / Without
2.0

Without Edge
With Edge Restoration
Restoration
filter strength

1.5
Edge restoration applied using a
filter equal to 1/MTF with slight 1.0

noise reduction at frequencies 0.5

above .7 of the maximum. relative spatial frequency
0.0
0.0 0.2 0.4 0.6 0.8 1.0

M. Flynn 2007 40 M. Flynn 2007 41

2C – MTF – CR, DR, and XTL 2C – Edge Restoration – DR and CR Phalanx of hand phantom
Exposure of 100 speed film.
1.0

dXTL
.8 CR

DR-Se
1/MTF
2xG 1/MTF (.8B)
(.8B)
unprocessed
.6
MTF
DR-CsI
.4
CRGP

.2 DR

0 unprocessed
2xG1/sinc
1/sinc
0 1 2 3 4 5 6 7
cycles/mm
M. Flynn 2007 42 M. Flynn 2007 43
 2C – Edge Restoration – dDR and iDR Clinical Wrist 2C – Chest Edge Restoration
Identical Manual Exposure 2.0

dDR iDR

filter strength
1.5

1.0

0.5

relativespatialfrequency
0.0
0.0 0.2 0.4 0.6 0.8 1.0

Chest Processing
• Edge restoration: lung tissue typically produces low frequency signals and
the chest radiograph has high quantum noise. Thus, very modest edge
restoration should be used.
• Quantum mottle in the abdomen: Low exposure and thick tissue result in
significant quantum mottle below the diaphragm. Inverse MTF filters need
to be damped at high frequency to prevent excessive noise (Metz filter).
High DQE iDR systems can restore edges
M. Flynn 2007
without producing excessive noise. 44 M. Flynn 2007 45

2C – Skeletal Edge Restoration 2D – Noise Reduction

2.0 Grayscale Rendition: Convert signal values to display values

Exposure Recognition: Adjust for high/low average exposure.
filter strength

1.5
Edge Restoration: Sharpen edges while limiting noise.
1.0
Noise Reduction: Reduce noise and maintain sharpness
Contrast Enhancement: Increase contrast for local detail
0.5

relative spatial frequency

0.0
0.0 0.2 0.4 0.6 0.8 1.0

Skeletal Processing Spatial Processes

• Edge restoration may be extended to high frequencies particularly if high
resolution screen are used. Noise is generally not problematic for Exposure •Edge Restoration Grayscale
extremity views. Recognition •Noise Reduction (VOI-LUT)
• Restoration versus enhancement: 1/MTF edge processing as shown •Contrast Enhance
restores object detail to that which would be recorded with a perfect
detector. The term restoration is recommended rather than enhancement.

M. Flynn 2007 46 M. Flynn 2007 47

 2D – noise and contrast 2D – noise smoothing
Quantum noise can mask Smoothing reduces both noise
low contrast structures and edge detail (5 pt avg).
1400 1400

1200 1200

1000 1000
Signal

Signal
800 800

600 600

400 400

200 200

0 0
100 150 200 250 300 350 400 100 150 200 250 300 350 400
Position Position
M. Flynn 2007 48 M. Flynn 2007 49

2D – adaptive smoothing 2D – noise reduction: with/wo

Adaptive noise reduction preserves Comparison with and without
edges for high gradients (lee filter) adaptive noise reduction
1400 1400

1200 1200

1000 1000
Signal

Signal

800 800

600 600

400 400

200 200

0 0
100 150 200 250 300 350 400 100 150 200 250 300 350 400
Position Position
M. Flynn 2007 50 M. Flynn 2007 51
 2D – mcp joint noise 2D – mcp joint noise Vertical profiles of the mcp joint in an AP radiograph
show the effects of noise reduction.

Radiograph of a hand phantom

demonstates uniform noise in 250
the lucite ‘tissue’ and detailed NR = 0
NR = 5
human bone features. Noise
reduction is shown using a
zoom view of the mcp joint. 200

image value
Noise reduction OFF Noise reduction ON
150

100
Agfa
CR

50
0 100 row number 200

M. Flynn 2007 52 M. Flynn 2007 53

2D – ‘coring’ Simoncelli EP, Adelson EH, “Noise removal via 2D – ‘coring’, non-linear subband transform
Bayesian wavelet coring,” Proc. 3rd IEEE Int.
Conf. Image Proc., vol. I, pp. 379–382, 1996
• Conceptual method (Simoncelli):
“A common technique for noise reduction is known as
a) Original image
(cropped). ‘coring’. An image signal is split into two or more
b) Image contaminated bands; the highpass bands are subjected to a
with additive threshold non-linearity that suppresses low-amplitude
Gaussian white noise values while retaining high-amplitude values.”
(SNR = 9.00dB).
a b • Statistical significance (Simoncelli):
c) Image restored
using (semi-blind) • “Removal of noise from images relies on differences in the
Wiener filter statistical properties of noise and signal.
(SNR = 11.88dB). • The classic Wiener solution utilizes differences in power
d) Image restored spectral density, a second-order property.
using (semi-blind) • The Bayesian estimator described .. provides a natural
Bayesian estimator
extension for incorporating the higher-order statistical
(SNR = 13.82dB).
c d regularity present in the point statistics of sub-band
representations.”
M. Flynn 2007 Figure 4. Noise reduction example. 54 M. Flynn 2007 55
 2D – adaptive non-linear coring 2E – Constrast Enhancement

Couwenhoven, 2005,
SPIE MI vol 5749, pg318 Grayscale Rendition: Convert signal values to display values
• High frequency sub-band Exposure Recognition: Adjust for high/low average exposure.
• Coring function Edge Restoration: Sharpen edges while limiting noise.
P = P/(1+s/P2) Noise Reduction: Reduce noise and maintain sharpness
• Adaptation Contrast Enhancement: Increase contrast for local detail
• Signal amplitude
• Signal to noise

Spatial Processes
Exposure •Edge Restoration Grayscale
Recognition •Noise Reduction (VOI-LUT)
•Contrast Enhance

M. Flynn 2007 56 M. Flynn 2007 57

2E – Contrast Enhancement 2E – Unsharp Mask

• A wide range of • A highly blurred

log(S) values is image can be used to
difficult to display in adjust image values.
one view.
• Lung detail is shown • The Unsharp Mask
here with low can be obtained by
contrast. large kernel
convolution or low
pass filter.

Contrast Enhancement: • Note that the

grayscale has been
Enhancement of local reversed.
detail with preservation
of global latitude.

M. Flynn 2007 58 M. Flynn 2007 59

 2E – Detail enhancement 2E – Contrast Enhancement in frequency space

• the image is low pass filtered to get a smoothed mask

The difference
between the image and
image (illustrated as a gaussian low pass filter).
the unsharp mask • Subtraction of the mask from the image yields a high
contains detail. pass filtered image having only the detail associated
This is added to the
with local tissue structures.
image to enhance detail 2.0
contrast

Detail contrast
enhancement is
The contrast enhanced
image has improved
obtained by adding 1.0
lung contrast and good the scaled
presentation of subtracted detail to
structures in the the image.
mediastinum.

Cycles/mm
M. Flynn 2007 60 M. Flynn 2007 61

2E – Selecting contrast enhancement 2E – Detail Contrast, Latitude, and Gain

Extended Visualization
Processing (EVP, Kodak).
In practice, the amount of contrast enhancement can For a specific grayscale rendition,
be selected by first defining a grayscale rendition that detail contrast can be progressively enhanced.
achieves the desired latitude, and then applying a • Latitude – the range of the unenhanced LUT.
filter that enhances detail contrast. • Detailed Contrast – the effective slope of
The enhancement gain is adjusted to amplifying the the enhanced detail at each gray level.
contrast of local detailed tissue structures. • Gain – the increase in LUT local slope.

3.0
4000
Detail Contrast
of 5,8,11 LUTs
Methods using large 2.0
3000

kernel of equal weight

have poor frequency
2000
11-11
response characteristics. 1.0 1000
8-11
5-11

11 LUT Latitude
0
Gain== =2.6
Gain
Gain 0
1.4
0 500 1000 1500 2000 2500 3000 3500 4000
Cycles/mm
M. Flynn 2007 62 M. Flynn 2007 63
 2E – Optimal PA chest gain 2E – chest, wide latitude
Optimal Contrast/Latitude
All Reader Mean (n=5) for 8 Cases
Detail Contrast (.85 to 5.75, logscale)

5 thoracic radiologists
at 3 medical centers
preferred a gain of 2.4
for the interpretation
of PA chest G = 2.4
radiographs of any
latitude. 1

SPIE 4319, 2001

1
T1-c
Latitude (.47 to 2.06, logscale)
• Lat = 1.68
8 PA chest Radiographs • Con = 2.21
l 52 display processing conditions for each radiograph. • G = 2.4
EVP gain varied from 1.0 to 6.8.
l

Detail contrast set to 8 values (rows).

l

Latitude set to 10 values (columns).

l

M. Flynn 2007 64 M. Flynn 2007 65

2E – chest, low latitude 2E – foot – contrast enhancement

Contrast enhancement of wide latitude

Musculoskeletal views improves visualization

T3-c
• Lat = 1.44
• Con = 3.00
• G = 2.4

Latitude 1200
Latitude 2X Gain contrast enhancement
600 – 0X

M. Flynn 2007 66 M. Flynn 2007 67

 2E – Display Processing: skull 2E – Display Processing: C-spine

Film-screen Equalized & Film-screen appearance Equalized / Enhanced

appearance Enhanced
M. Flynn 2007 68 M. Flynn 2007 69

2E – Equalized / Enhanced arm 3- Course Outline

1. Preprocessing
2. Generic Image Processing
3. Commercial Implementations

M. Flynn 2007 70 M. Flynn 2007 71

 3A – Fujifilm Medical Systems USA 3A – Fujifilm MFC Yamada , BJR,78 (2005), 519–527

M. Flynn 2007 72 M. Flynn 2007 73

3A – Fujifilm FNC Yamada , BJR,78 (2005), 519–527 3B – Eastman Kodak Company

• 1997 SPIE3034
Senn, skinline detection
• 1998 SPIE3335
Barski, ptone grayscale
• 1999 SPIE3658
Barski, grid suppression
• 1999 SPIE3658 EVP
Van Metter, EVP
• 2001 SPIE4322
Pakin, extremity segment.
• 2003 SPIE5367
Couwenhoven, control
• 2004 SPIE5370
A series of proceedings articles describes
Wang, auto segmentation
the image processing approaches used by
• 2005 SPIE5749 Eastman Kodak Company
Couwenhoven, noise

M. Flynn 2007 74 M. Flynn 2007 75

 3B – EKC Signal Equalization (Kodak EVP) 3B – EKC Multi-Frequency Processing
Wang, AAPM ’06, CE Wang, AAPM ’06, CE

Input Image &

PTONE LUT Output Image &
EVP GAIN PTONE LUT
β1
Original
-
Image β2

EVP
Blurred β3
KERNEL SIZE
Image +

…
βn
PTONE NEW Original Edge-Restored
LUT PTONE LUT Image Image
βn+1
EVP GAIN and EVP DENSITY

E’(i,j) = α • { E(i,j) ⊗ K } + ( 1 - α ) • Emid + β • { E(i,j) - ( E(i,j) ⊗ K ) }

D(i,j) = ρ[ E’(i,j) ].

M. Flynn 2007
“Enhanced latitude for digital projection radiography,” R. Van Metter and D. Foos, Proc. SPIE 3658, 468-483, 1999. 76 M. Flynn 2007 77

3B – EKC control variables. 3C - Philips

GXR, Th. Rohse, November 2005

Brightness

Couwenhoven,
RSNA Inforad
2005
Latitude
1st World
Congress
Thoracic Imaging UNIQUE
2005 UNified Image QU ality Enhancement

Contrast

M. Flynn 2007 78 M. Flynn 2007 79

 3C – Philips multi-resolution 3D – Agfa MUSICA

UNIQUE Principle
Multi-Resolution Decomposition

Original Image

• Vuylsteke P, Schoeters E, Multiscale Image Contrast Amplification

Processed Image

(MUSICA), SPIE Vol 2167 Image Processing, pg 551, 1994

Filter 1 Filter 2 Filter 3 Filter n LUT • Burt PJ, and Adelson EH, "The Laplacian pyramid as a compact
image code", IEEE Trans. On Communications, Vol. 31, No. 4, pp.
532-540, 1983.
M. Flynn 2007 GXR, Th. Rohse, November 2005 80 M. Flynn 2007 81

3D – Agfa, multiscale transforms Prokop, J.Thoracic Img., 18:148–164,2003 3D – Agfa, non-linear transfer

Non-linear transfer functions alter the contrast in

each frequency band to amplify small signal
contrast while controlling noise.

M. Flynn 2007 82 M. Flynn 2007 83

 3D – Musica 2 3E - Canon
Multi Frequency
Adjustment Window
MU-1 MU-2

• The recently released Musica-2 provides a more

unified approach to the processing of all bodyparts.
• In general, Musica-2 has the ability to provide more
aggressively processed appearance.

M. Flynn 2007 84 M. Flynn 2007 85

3E - Canon 3E - Canon
Narrowed Signal Range Increased Detail
Contrast

M. Flynn 2007 86 M. Flynn 2007 87

 3E - Canon 3E - Canon
Wide Latitude Enhancement may depend
High Detail Contrast on licensed options

M. Flynn 2007 88 M. Flynn 2007 89

3E - Canon 3 – “multi-frequency”
MTF Dependant Edge
Enhancement
In General
• Linear Filters
Linear filters implemented with Fourier
transforms or convolution with large area, variable
amplitude kernels can achieve equalization and
edge restoration with full control of the
frequency transfer characteristics.
• Multi-scale Filters
Multi-scale filters have coarse control of
frequency transfer characteristics but can apply
non-linear transformations to achieve noise
reduction and prevent high contrast saturation.

M. Flynn 2007 90 M. Flynn 2007 91

 3 - others 3 – Commercial Implementation of DR Processing

• Del Medical Systems Group

• Image processing is provided by all CR/DR
• GE Healthcare suppliers under a variety of trade names.
• Hologic, Inc • While the computation approaches differ, the
• Imaging Dynamics Co, Ltd effect on the radiograph is similar.
• Infimed Inc • The processed digital image can appear very
• Konica Minolta Medical Imaging much different that a traditional screen film
• Lodox Systems radiograph.
• New Medical Ltd • It is possible to set up systems from different
• Shimadzu Medical suppliers to provide similar appearance (but
difficult). Harmonized processing is needed.
• Siemens Medical Solutions
• Swissray International
• Vidar Systems Corp.

M. Flynn 2007 92 M. Flynn 2007 93

3 - Body Part & View Questions ?

• Processing parameters for equalization, grayscale

rendition, and edge restoration are set specifically for
each body part / view that may be done.

• This requires close cooperation between the user and the

supplier to set up tables that conform to the body part-
view used in a department.
?
• Dependence on body part size complicates processing

• New industry developments may provide processing

software that automatically selects the proper
parameters from the image data and makes adjustments
for body part size.

M. Flynn 2007 94 M. Flynn 2007 95