1. 1. Digital Image Processing in. Radiography. Xiaohui Wang, PhD. David H.
Foos, MS. Health Group Research Laboratory. Eastman Kodak Company. 2.
Outline • Display Processing – – – – – – – –
Digital Image Processing in Radiography
Data preprocessing ROI segmentation and analysis Tonal rendering Signal equalization Edge restoration Noise suppression Collimation masking Display compensation
• Image Processing Features – – – – – –
Xiaohui Wang, PhD David H. Foos, MS Health Group Research Laboratory Eastman Kodak Company 1
Stationary grid detection and suppression Long-length imaging Dual energy imaging Mammography Oncology processing Quality control testing
2
“Analog” Image Processing… Screen/Film Radiography (S/F) Density
3
Display Processing
2 1.0
1
0.8
4
5
6 Log Exposure
Modulation Transfer
0
0.6
Low Speed Screens
7
0.4
0.2
Film Speed & Contrast Screen Speed & MTF 3
0
High Speed Screens
0
1
2
3
4
5
-1
Spatial Frequency (mm )
4
1
Digital Image Processing
Why Image Processing? • Maintain the familiar characteristics of S/F
Digital Image Acquisition
– Provide a similar tonal rendering – Restore desired sharpness
• Beyond the familiar – – – – – –
Digital Image Processing
Automatically adjust for errors in exposure Automatically accommodate changes in latitude Increase the range of exposures visualized Enhance selected spatial frequencies Highlight regions of interest (ROI) Assist the radiologist to find features of interest
Digital Image Display
5
6
Display Processing •
Transform digital radiography raw data to display values for presentation using a workstation or film printer, automatically, robustly, and consistently. Components of Image Quality – Latitude – Contrast – Brightness – Sharpness – Noise
•
Original Image Tone Scale Edge Restoration Signal Equalization Collimation Masking
3 Overexposed
100
2 Correctly exposed
10
1 Underexposed
1 0.01
7
4
Film-screen (400 speed) 1,000
Film Optical Density
Relative intensity of PSL
PSP plate 10,000
0 0.1
1 10 Incident exposure, mR
100
8
2
Schematic Flow Chart of Display Processing Original Image
Image Capture
Tone Scale
Data Preprocessing
ROI Segmentation & Analysis
Tone-Scale Generation
Edge Restoration Signal Equalization Edge Restoration
Collimation Masking
Signal Equalization
Tonal Rendering
Collimation Masking
Display Compensation
PACS & Print
Noise Suppression
Multi-Frequency Processing 9
10
Data Preprocessing •
•
•
Data Preprocessing (cont.) Linear-to-log conversion
Image reformation – Composition (dual-side CR reading) – Decomposition (dual energy) – Resize Signal filtering – Gain, offset, and bad pixel correction (DR) – Noise reduction – Stationary grid artifact suppression Data space conversion – Linear to logarithmic (const. object contrast vs. pixel value) – Linear to square root (const. quantum noise vs. pixel value)
Shape of image histogram invariant to exposure I0
I
f ( x)
1
f ( x)
log( I 0 / I ) ~
1
log( I ) const.
f (x) Baseline
I
4x over exposure 4x under-ex posure
0
11
I0 e
1000
2000
3000
4000
5000
12
3
ROI Segmentation & Analysis •
Extract diagnostically relevant ROIs
•
Analyze ROI characteristics
•
Derive the optimal display-rendering parameters
•
Include four basic steps – Detect collimation mask Direct – Detect direct exposure Exposure – Extract anatomy regions – Calculate key image descriptors
Segmentation – Collimation Mask • Confine exposure regions • Mask applied to collimated regions to reduce viewing flare
Collimation
Anatomy
“Method for recognizing multiple radiation fields in computed radiography,” X. Wang, J. Luo, R. Senn, and D. Foos, Proc SPIE 3661, 1625-36, 1999.
13
14
Segmentation – Collimation Mask (cont.)
Segmentation – Collimation Mask (cont.) •
• Collimation boundary pixels – Edge profile analysis – Transition segments classification
Multiple radiation field masking – Optimal individual image processing – Exam workflow improvement
• Candidate collimation blades – Edge delineation – FOM analysis • straightness • connectedness…
• Candidate configurations • Select “best” configuration – Parallelism, convexity, orthogonality, etc.
“Method for recognizing multiple radiation fields in computed radiography,” X. Wang, J. Luo, R. Senn, and D. Foos, Proc SPIE 3661, 1625-36, 1999.
J. Luo and R. Senn, “Collimation detection for digital radiography," Proc. SPIE 3034, 74-85 (1997).
15
16
4
Segmentation - Direct Exposure Detection
Segmentation – Collimation Mask (cont.)
Aggressive Failure
Exclude non-anatomical regions within the collimation. Conservative Failure
• Compensate – Radiation field non-uniformity – X-ray scatter – Multiple exposures
• Transition segment analysis – – – –
Line profile analysis Background transitions characterized by slope and extent Background pixel histogram analysis Spatial correlation of exposure variations
L. Barski and R. Senn “Determination of direct x-ray exposure regions in digital medical imaging,” U.S. Patent 5,606,587 (1997).
17
18
ROI Analysis - Key Image Descriptors
Segmentation – Anatomy Extraction
Lateral Lumbar Spine ROI Selection 1 0.9
Normalized Frequency
0.8 0.7 0.6 act hist
0.5
cv. hist
0.4 0.3 0.2
Anatomy
0.1
Anatomy
0
Left pt Far left pt
19
“Automatic and exam-type independent algorithm for the segmentation and extraction of foreground, background, and anatomy regions in digital radiographic images,” X. Wang, H. Luo, Proc. SPIE 5370, 1427-1434, 2004.
Right pt Far right pt
20
5
Effect of Density Shift Pa ram eter
Tone-Scale Generation
4000
Brightness Adjustment
3500
Code Value Out
3000
• Render image with proper brightness and contrast – Calculate average exposure within ROI – Automatically adjust for errors in exposure • Sigmoid curve shape in general – Curve shift (brightness adjustment) – Curve rotation (contrast adjustment) – Toe & shoulder adjustment • Bear different names – Kodak: PTS (Perceptual Tone Scale) – Fuji: Gradation Processing – …
2500 2000 1500 shift = -0.5
1000
shift = 0 500
shift = 0.5
0 0
500
1000
1500
2000
2500
3000
3500
4000
4500
Code Value In
22
Toe & Shoulder Adjustment
Effect of Contrast Parameter
Contrast Adjustment
4000
Code Value Out
3500 3000 2500 2000 c = 0.8
1500
c = 1.3
1000
Effect of Toe and Shoulder Para meters
4000 3500 3000 Code Value Out
21
2500 2000 1500 T=0, S=0
c = 1.8 1000
500 0
T=0.3, S=0.7 T=0.6, S=1.4
500
0
1000
2000
3000
4000
5000
0
Code Value In
0
1000
2000
3000
4000
5000
Code V alue In
c
23
24
6
Visually Optimized Tone Scale
Visually Optimized Tone Scale (cont.)
Perceptual Linearity -
Radiologists prefer S/F systems with perceptually linear sensitometric response.
Render ROI such that… equal physical contrast being perceived as equal brightness by the observer across the full brightness range X-rays
2
3
Kodak Insight HC Thoracic Imaging System
Kodak Insight Thoracic Imaging System 25
26
Perceptual Tone Scale
Perceptual Tone Scale (cont.) Equal Log (E)
Equal Brightness
dmax
Perceptually Linearity Tone Scale
Density – Luminance (Physics)
dmin
flp
B=
frp lp
Visual Perception Model
Daly’s Global Cone Model
rp
Bm Ln Ln + L0n
B = perceived brightness L = luminance of the image area Bm = scale factor n = 0.7
L0 = 12.6*(0.2*Lw)0.63 + 1.083*10-5
Perception Linearity
Lw = luminance of the reference white H. Lee, S. Daly, and R. Van Metter, “Visual optimization of radiographic tone scale,” Proc. SPIE 3036, 118-129, 1996.
27
28
7
Signal Equalization
Tone Scale Failures
Effect of Contrast Parameter 4000
Code Value Out
3500
Balance between contrast vs. latitude
3000 2500
Bone
2000
c = 0.8
Soft c = 1.3 Tissue c = 1.8
1500 1000 500 0 0
1000
2000
3000
4000
5000
Code Value In
Too Bright
Too Dark
Too Much Contrast
Too Less Contrast
29
Good Soft Tissue Contrast
Signal Equalization (cont.)
Good Bone Contrast
Signal Equalization (Kodak EVP)
• Automatically accommodate changes in latitude – Compress the image-signal dynamic range such that all information within ROI can be rendered with optimal contrast • Increase the range of exposures visualized • Reduce exposure re-take and improve workflow • Signal equalization processing is
Input Image & PTONE LUT
• • • • •
EVP - Enhanced Visualization Processing (Kodak) DRC - Dynamic Range Compression (Fuji) Latitude Reduction (AGFA) Tissue Equalization Processing (GE) …
EVP GAIN
Original Image
EVP KERNEL SIZE
– 2D spatial processing – Digital wedge filter – Bearing different names:
31
Signal Equalization
30
Output Image & PTONE LUT
-
Blurred Image
PTONE LUT
NEW PTONE LUT
EVP GAIN and EVP DENSITY
E’(i,j) =
{ E(i,j)
K}+(1-
) Emid +
{ E(i,j) - ( E(i,j)
K)}
D(i,j) = [ E’(i,j) ]. 32
“Enhanced latitude for digital projection radiography,” R. Van Metter and D. Foos, Proc. SPIE 3658, 468-483, 1999.
8
Signal Equalization (cont.)
Signal Equalization (cont.) Increasing Contrast, …Decreasing Latitude 11-11
8-8
5-5
11-11
8-11
5-11
Increasing Contrast, …Constant Latitude
33
34
RSNA 2001 Education Exhibit
Observer “rendering preferences” established through collaborative studies with university hospitals (…to set default parameters for automatic processing)
Contrast
High
Low
Latitude Detail LogE Contrast D/ LogE => 6.75 5.75 5 3.75 3.1 2.25 1.9 1.6 1.2 1 0.85 Latitude rel. to Ref.
Narrow
0.35 1.4 1.0
0.38
0.47 1.8 1.5 1.3 1.0
0.56 2.2 1.9 1.6 1.2 1.0
0.78 3.0 2.6 2.2 1.7 1.4 1.0
0.51
0.61
0.84
0.92 3.6 3.0 2.6 2.0 1.6 1.2 1.0
1.09 4.2 3.6 3.1 2.3 1.9 1.4 1.2 1.0
1.46 5.6 4.8 4.2 3.1 2.6 1.9 1.6 1.3 1.0
1.75 6.8 5.8 5.0 3.8 3.1 2.3 1.9 1.6 1.2 1.0
2.06 7.9 6.8 5.9 4.4 3.6 2.6 2.2 1.9 1.4 1.2 1.0
1.00
1.19
1.58
1.90
2.24
Latitude
Wide
“Optimal display processing for digital radiography,” M. Flynn, M. Couwenhoven, W. Eyler, B. Whiting, E. Samei, D. Foos, R. Slone, and E. Marom, Proc. SPIE 4319-36, 2001.
35
36
9
Signal Equalization (cont.) Equalization Processing Artifact
Optimal PA View Kodak T-Mat G Film detail contrast with 2X extended latitude
Properly Processed 37
Halo Artifact
Overprocessed
38
Edge Restoration
Edge Restoration (cont.) • Selectively boost high-frequency signals based on
Modulation Transfer Function
– – – – –
1 CR Hi-Res Screen CR Std Screen 0.8
High-frequency signal is suppressed by system MTF.
CsI DR Selenium DR
MTF
0.6
Selectively boost high frequency
0.4
0.2
• Multi-frequency processing (2D spatial) – – – – –
0 0
1
2
3
4
Exam type Brightness Exposure Diagnostic features Capture device characteristics
5
Kodak: EVP & USM (Enhanced Visualization Processing & UnSharp Mask) AGFA: MUSICA (MUlti-Scale Image Contrast Amplification) Konica: Hybrid (Mutil-Resolution Hybrid Processing) Fuji: USM & MFP (Multi-Objective Frequency Processing) Philips: UNIQUE (UNified Image QUality Enhancement)
Spatial Frequency (cycles/mm)
39
40
10
Edge Restoration (cont.)
Edge Restoration (cont.) M o d u la t io n T r a n s f e r F u n c t io n
Unsharp Mask Processing
1 .2
USM Gain
K = 5, g = 1.2
1
Output Image
MTF
+
0 .6
K = 3, g = 1.5
0 .4
Original Image
-
B a e li n e US M 1 US M 2
0 .8
High-Freq. Image
0 .2
0
USM Kernel Size
0
Blurred Image
1
2
3
4
5
S p a t ia l F r e q u e n c y (c y c le s /m m )
41
42
Edge Restoration (cont.)
Edge Restoration (cont.) Halo Artifact
Processing Artifact Multi-Frequency Processing 1
2
3
…
Original Image
+
n
Edge-Enhanced Image n+1
Properly Processed 43
Overprocessed
44
11
Edge Restoration (cont.)
Noise Suppression • It is desired to drive toward lower x-ray exposures to reduce patient dose • The appearance of noise increases as exposure level is decreased • A predominant source of noise in digital radiography is generally the quantum noise. • Noise suppression should be signal dependent, it should be applied only to areas of the image that have a low SNR.
Processing Artifact Halo Artifact
Properly Processed
• A noise suppression algorithm needs to reduce the appearance of noise while preserving diagnostic detail.
Overprocessed
“Observer study of a noise suppression algorithm for computed radiography images,” M. Couwenhoven et al, Proc. SPIE 5749, 318-327, 2005.
45
46
Noise Suppression (cont.)
Noise Suppression (cont.)
Suppress the noise in low signal areas and phase out suppression in high signal areas High Signal Areas / Less Dense Anatomy
Noise Suppression • • • •
2D spatial processing Applies to high freq. signals Signal dependent Balance between sharpness & noise
Low Signal Areas / More Dense Anatomy
47
48
12
Display Compensation
Display Compensation (cont.) Workstation Calibration Check
• Image pixel values can be mapped for different output devices – Film printer (monochrome 1) – Softcopy display (monochrome 2)
10.00
Image looks darker throughout the dark regions
Luminance
• Both capture and output devices need to be configured properly • Output device calibration is critical to optimal image display – – – –
100.00
1.00
Aim Luminance Measured Luminance
Different dynamic range and response CRT vs. flat-panel Images from multi-vendors viewed at same PACS workstation Archived images
0.10 0%
20%
40%
60%
80%
100%
SMPTE Patch
• DICOM Part 14 specifies grayscale display standard function (GSDF) • AAPM TG-18 specifies display QA & QC testing Reference 49
Non-calibrated Display
50
Display Compensation (cont.) Monitor Responses Monitor Response Examples 1000
Lum inance
100
Image displayed at CRT looks brighter
CRT Flat-Panel
10
Image Processing Features
1 0
1024
2048
3072
4096
0.1 DDL
Reference (flat-panel) 51
CRT 52
13
Stationary Grid Suppression
Automatic Grid Line Detection & Suppression
Moiré Pattern
A
B
C
(A) CR chest image minified by a factor of 0.2 showing a grid-caused Moiré pattern (B) Filtered image fragment without Moiré pattern (C) Difference image 53
“Antiscatter stationary grid artifacts automated detection and removal in projection radiography imagery,” I. Belykh and C. Cornelius, Proc. SPIE 2000.
54
Long-Length Imaging with CR
Long-Length Imaging •
Screen Film Systems 35cm x130cm and 35cm x 90cm Cassettes Computed Radiography Currently Limited to 35cm x 43cm
• •
•
55
Multiple 35cm x 43cm CR screens arranged in alternating and partially overlapping fashion for patient imaging Storage phosphor screens scanned individually Image processing software used to automatically – Determine CR screen sequence and orientation – Correct for magnification, translation, and rotation among individual screens – Remove redundant image data in the overlap region – Construct (stitch) a large geometrically accurate composite – Eliminate the seam lines in the composite image Composite image stored to PACS for interpretation
56
14
Scan Individually Storage Phosphor Screens
Long-Length Imaging (cont.)
Image Processing to Construct Large Composite
“Fully automatic and reference-marker-free image stitching method for full-spine and full-leg imaging with computed radiography,” X. Wang, D. H. Foos, J. Doran, and M. K. Rogers, Proc. SPIE 5368, p361-369, 2004.
57
58
Dual-Energy Imaging
Long-Length Imaging (cont.)
High kVp Radiograph
Tissue-Only
Bone-Only
• Measurements from CR equivalent to screen-film – Angular – Absolute distance
Images acquired at high and low energies are processed to selectively cancel overlying tissues
• Visual quality of CR superior to S/F – Wide exposure latitude of CR – Equalization processing
• 35% retake rate with S/F reduced to 0% retake rate with CR – 43 cm width – Equalization processing
Images courtesy of Dr. Jeff Siewerdsen at the Ontario Cancer Institute, Princess Margaret Hospital and University of Toronto.” 59
60
15
Mammography Processing
Mammography Processing (cont.) Recent advances in s/f mammography
S/F: A large dynamic range is needed to detect and display all parts of the breast with good contrast
Min-R EV 150 system vs. Min -R 2000 system
Min -R EV 150 vs Min - R 2000 System
6.5
High Dmax and shoulder contrast
6.0
Results in greater overexposure latitude due to higher upper scale contrast
5.5 5.0 4.5 4.0
2.5
Higher contrast
3.5 Better of 2.0 visualization Density 3.0 breast 1.5 parenchyma
Sharp toe Results in “whiter whites ”, more “sparkle”, improved visibility of microcalci fications
1.0 0.5 0.0 0.0
0.5
1.0
1.5
2.0
2.5
Log E
X-ray Sensitometry KODAK MIN -R EV/ EV 150 Screen
61
KODAK MIN -R 2000/2000 Screen
62
Mammography Processing Digital Mammography • Wide dynamic range (> 1000:1) captures all the image information • Edge restoration enhances image details of different sizes • Equalization processing compresses image latitude while maintaining contrast S/F look
Edge Enhanced
Oncology Processing Black surround
Signal Equalization
Simulation Localization Verification
Equalized look
“Method for contrast-enhancement of digital portal images,” S. Young, W. E. Moore, D. H. Foos, US Patent US6836570 B2
63
64
16
Test Phantom for Kodak DIRECTVIEW Total Quality Tool
Quality Control Testing
0.5 mm Cu 1.0 mm Al
180 cm
10.0 0.2 mR @ 80 kVp 24 x 30 cm 35 x 43 cm
18 x 24 cm 65
66
KODAK TQT User Interface
Test Result Details MTF (slow scan)
29.5%
67
68
17
Quality Control Testing (cont.)
Summary • Tone scale processing establishes the overall image brightness and global contrast
Automated Image Quality Control Tool
• Edge restoration enhances detail contrast
Precise and accurate quality control testing
• Signal equalization extends the latitude that can be visualized while maintaining detail contrast
Highly reproducible quantitative results Detects sub-visible changes in CR image quality performance to initiate timely preventive maintenance Avoids hours of tedious and labor-intensive effort with a highly automated procedure Full data reporting in Excel format
• Edge restoration, signal equalization, and noise suppression are 2D spatial processing • Multi-frequency has been widely adopted, yet users should be aware of processing artifacts • Display processing are becoming easier and more intuitive to use • Image processing provides many new features unique to digital capture • Image processing can provide many automations to improve work efficiency
69
70
Acknowledgements • • • •
Richard VanMetter Lori L. Barski William J. Sehnert Lynn M. Fletcher-Heath
71
18
