The Forefront in Image Processing: PET/Molecular Approaches
Joel KarpUniversity of Pennsylvania
Sixth Annual NCI-Industry ForumQuantitative Oncologic ImagingApril 7-8, 2005
Issues of Performance, Image Processing, Quantification
• Performance of current-generation PET scannersGlobal effects - data correctionLocal effects - image reconstructionStatistical and count-rate effects
• Self-consistency: instrument performs same day-to-day
• Cross-consistency: all instruments produce same result
• Comparing images (PET and CT) from different patients, different instruments, and different institutes
Yab = Nab(AabTab + Sab + Rab)
What is Measured with PET
a
b
What is measured Trues
True coincidence
Scattered coincidence
Random coincidence
(~2 . singles2)
Coincidence?
Record event
Normalization Attenuation RandomsScatter
Signals from Different Voxels are Coupled Statistical Noise Does Not Obey Counting Statistics
If there are N counts in the image,
SNR ≠ N / (N)1/2
Reconstruct image from line-of-response (LOR) projection data
Iterative reconstruction
x(0)
Correction for Attenuation,
Scatter, Randoms
yData
A x = y
y(k) x(k)
Image
Forward projection
(k)
Difference c(k)
Update
Back-projection
^
Start here
xj(k+1) =xj
(k) +λkxj(k) yik
ˆ y ik−1
⎛ ⎝ ⎜ ⎜
⎞ ⎠ ⎟ ⎟ aik , j
DETECTOR
DIGITIZER
POSITION CALCULATOR
BINNER
COMPUTER
RAWVIEW (52 bytes/event) For A,B side of event26 PMT energies/zone (26 bytes)100M events = 5200 Mbytes
LISTVIEW (8 bytes/event) For A,B side of event2D position (3 bytes) timestamp Energy (1 byte) TOF (1 byte)100M events = 800 Mbytes
SINOGRAM (80 Mbytes/frame) R, Phi (295x161x2 = 95 Kbytes) Slice (29^2 = 841)100M events = 560 Mbytes (7 frames)
Data Flow
Philips Allegro: 616 x 29 crystals
IMAGE X,Y (128x128 = 16 Kbytes)100M events = 4 Mbytes (250 slices)
reconstructionPACS archive
2D (septa) vs. 3D (no septa)
2D Imaging 3D Imaging
Low Scatter and Randoms High Scatter and RandomsLow geometric sensitivity High geometric sensitivity
0
5000
10000
100 300 500Energy (keV)
TrueScatter
S T R
Scatter decreases with high energy threshold - depends on energy resolution
Out-of-field activity increases randoms in 3D
Singles FOV 3D
mode
0
2
4
6
8
10
12
14
200 250 300 350 400 450 500 550
Lower Energy threshold (keV)
Randoms relative
to 435 keV
Randoms ~ 2 . Singles2
• decreases with narrow timing window (2)• decreases with high energy threshold• estimated (and subtracted) with 2nd (delayed) timing window
Problem increases as bore size increases-> less shielding
Count-rate Performance
0
50
100
150
200
0.0 0.1 0.2 0.3 0.4 0.5Activity concentration (uCi/cc)
Randoms
Trues
NEC
Scatter
3.7 7.4 11.1 14.8 (kBq/ml)
Noise Equivalent Count-rate
NEC = T/(1+S/T+R/T)
NEC ~ SNR2
Philips Allegro
10 mCi dose70-cm long x 20-cm diameter
NEMA 2001 (body)
PET Imaging Performance• Spatial resolution -> partial volume effect
intrinsic: 4-6 mmreconstructed: >10 mm
• Scatter fraction -> noise and bias (after correction)2D: 10-20% SF3D: 30-60% SF
• Sensitivity and count-rate capability -> statistical quality25 - 100 kcps or 5 M - 20 Mevents per 3 min frame
Scatter CorrectionAA
BB
Single Scatter - Model based correctionCalculate the contribution for an arbitrary scatter point using the Klein-Nishina equation
BeforeScattercorrection
AfterScattercorrection
Attenuation correction with radioisotope transmission scan
20 mCi 137Cs source - 662 keV
A = 1 / e -d
d = length of chord through tissue = attenuation coefficient
Attenuation correction for PETTypes of transmission images
Coincident photon Ge-68/Ga-68
(511 keV)
high noise15-30 min scan
timelow bias
low contrast
Single photon Cs-137
(662 keV)
lower noise5-10 min scan
timesome bias
lower contrast
X-ray(~30-140kVp)
no noise1 min scan timepotential for bias
high contrast
Attenuation/Scatter correction
University of Pennsylvania PET Center
No AC or Scatter Corr AC and Scatter Corr
Philips Allegro
Fully 3D Iterative Reconstruction improves image
quality
Fore-FBP
3D Ramla
How about quantification?
NEMA NU2-2001 Image Quality Phantom
Out-of-field Activity
13 mm
10 mm
17 mm
28 mm22 mm
37 mm foam
Partial Volume Effect
0
0.2
0.4
0.6
0.8
1
0 2 4 6 8 10
Ideal Contrast
3D-10 mm
Ctr-10mm
Max-10mm
3D-13mm
Ctr-13mm
Max-13mm
3D-17mm
Ctr-17mm
Max-17mm
3D-22mm
Ctr-22mm
Max-22mm
3D-28mm
Ctr-28mm
Max-28mm
Contrast
FWHM (mm)
NEMA IEC Phantom
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0 5 10 15 20 25 30
10-mm Hot Sphere - 4mm
4-mm (non-LOR)4-mm (LOR)4-mm (LUT)
Contrast
Background Variability (%)
0
0.1
0.2
0.3
0.4
0.5
0 5 10 15 20 25 30
13-mm Hot Sphere - 4mm
4-mm (non-LOR)4-mm (LOR)4-mm (LUT)Contrast
Background Variability (%)
0
0.1
0.2
0.3
0.4
0.5
0 5 10 15 20 25 30
17-mm Hot Sphere - 4mm
4-mm (non-LOR)4-mm (LOR)4-mm (LUT)
Contrast
Background Variability (%)
0
0.1
0.2
0.3
0.4
0.5
0.6
0 5 10 15 20 25 30
22-mm Hot Sphere - 4mm
4-mm (non-LOR)4-mm (LOR)4-mm (LUT)
Contrast
Background Variability (%)
LOR RAMLA reconstruction
Vary relaxation parameter from 0.00025 (top left) to 0.075 (bottom right)
Contrast vs. Noise
1.7 cm hot sphere 2.8 cm cold sphere
Iterative - RamlaFiltered Backprojection (FBP)
Image processing Filters for restoring the spatial frequency components
Low (left) - Maximum gain = 2.5
Medium(middle)- Maximum gain = 3.5
High gain (right) - Maximum gain = 4.5
WF(f) = 1/MTF(f) for f<fcut
WF(f) = 1/MTF(fcut) exp-kf 2 for f>fcut
k - parameter describing the Gaussian roll-off
fcut - cutoff frequency
K, fcut -were bracketed from an analysis of phantom data
Profile through the lesion
Lesion contrast improves with filtering
no
low
med
high
Time-of-Flight : list-mode iterative
reconstruction
5Mcts1Mcts1Mcts TOF
5Mcts TOF
no TOF 300 ps TOF1
Mct
s5
Mct
s10
Mct
s
Challenges in comparing images• Spatial resolution differences
partial volume - simple (approximate) correctionspatial recovery in reconstruction model adds noise
• Reconstruction algorithm local convergence depends on algorithm and activity
• Accuracy of corrections - randoms, scatter, attenuationdepends on patient size and activity distribution
• Imaging protocolscan acquisition time and delay post-injection
• Quantification - typically based on simple cylinderQC - monitor and correct daily driftsActivity calibration - counts/voxel/min -> nCi/mlCount-rate corrections - dead-time
Challenges in comparing images• Instrumentation in PET is constantly evolving
performance of new scanner >> older scanner• Image data size is large - data transfer and archiving
PET: 4 Mbyte (with 4 mm3 voxels)CT: 64 Mbyte (with ~1 mm3 voxels)
• DICOMquantification (SUV) requires PT format (not NM)manufacturers workstations still most practical
• Data analysis tools must be standardized and validatedregion-of-interest
• Image processingbehavior must be understood - difficult to standardize