1
PET System Design, Acquisition and Image Reconstruction
Frederic H. Fahey DSc
Children’s Hospital BostonHarvard Medical School
PET: Back in the day…
• Research Device• Single Slice• Run by the research team• Neuroscience
PET III BNL - 1980
Recent Growth in PET
• Oncology• Regional Distribution of 18F FDG• Reimbursement
Migration of PETResearch Lab Clinic
• Simpler to run • More robust• Multi-Slice (15 cm volume)• Multi-Modality (PET-CT)
2
Outline
• Design criteria for State-of-the-Art Scanner• Basics of PET Scanner Design• Review of Current PET Instrumentation
Design Criteria for a PET Scanner
• Image Quality– Sharpness (resolution)– Contrast – Noise (sensitivity)
• Ease of Use (software, patient access)• Fast (throughput, sensitivity & count rate)• Robust• Not TOO Expensive
Positron Emission18F
511 keV
511 keV β+
e-
Detector Ring
3
Detector Blocks (GE Advance NXi)Large Crystal Designs
•Large NaI detectors, PMT array•25x50 cm, 1” thick•Hexagonal array
True, Scatter and Random Coincidence Detections
True
Random
Scatter
Randoms Estimation
• Background Subtraction• Singles Rate Calculation
R = 2 τ N1 N2
• Delay Window Method
4
SCINTILLATOR NaI(Tl) BGO LSO GSO
Rel. Light Yield 100 15-20 75 20-25
Peak Wavelength (nm) 410 480 420 440
Decay Constant (ns) 230 300 12,42 30-60
Density (g/cc) 3.67 7.13 7.40 6.71
Effective Z 51 75 66 59
Index of Refraction 1.85 2.15 1.82 1.85
Hygroscopic ? Yes No No No
New Detector Materials
Sinogram
Image for Each Slice
Ang
le
Image for Each Angle
Projection View
Slic
e
Note: Sinograms and projection views are different ways or showing the same data.
5
PET Sinograms• Point in transverse slice maps to sine wave• Displacement (x) vs Angle (y)• Each row is a projection through the object at
the corresponding angle• Each detector is mapped along a diagonal• Each pixel in the sinogram corresponds to a
particular “line of response” (LOR) i.e. detector pair
18 * *17 * * *16 * * *15 * * *14 * * *13 * * *12 * * *11 * * *10 * * *
9 * * *8 * * *7 * * *6 * * *5 * * *4 * * *3 * * *2 * * *1 * *
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Span of 3 Michelogram
18 * * * *17 * * * * *16 * * * * * *15 * * * * * * *14 * * * * * * *13 * * * * * * *12 * * * * * * *11 * * * * * * *10 * * * * * * *
9 * * * * * * *8 * * * * * * *7 * * * * * * *6 * * * * * * *5 * * * * * * *4 * * * * * * *3 * * * * * *2 * * * * *1 * * * *
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
Span of 7 Michelogram
6
Acquisition Modes
2D
3Dseptaare
removed
16 * * * * * * * * * * * * * * * *15 * * * * * * * * * * * * * * * *14 * * * * * * * * * * * * * * * *13 * * * * * * * * * * * * * * * *12 * * * * * * * * * * * * * * * *11 * * * * * * * * * * * * * * * *10 * * * * * * * * * * * * * * * *
9 * * * * * * * * * * * * * * * *8 * * * * * * * * * * * * * * * *7 * * * * * * * * * * * * * * * *6 * * * * * * * * * * * * * * * *5 * * * * * * * * * * * * * * * *4 * * * * * * * * * * * * * * * *3 * * * * * * * * * * * * * * * *2 * * * * * * * * * * * * * * * *1 * * * * * * * * * * * * * * * *
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
SliceOrientation
3D Michelogram, RD = 15
16 * * * * * * * * * * * * * * * *15 * * * * * * * * * * * * * * * *14 * * * * * * * * * * * * * * * *13 * * * * * * * * * * * * * * * *12 * * * * * * * * * * * * * * * *11 * * * * * * * * * * * * * * * *10 * * * * * * * * * * * * * * * *
9 * * * * * * * * * * * * * * * *8 * * * * * * * * * * * * * * * *7 * * * * * * * * * * * * * * * *6 * * * * * * * * * * * * * * * *5 * * * * * * * * * * * * * * * *4 * * * * * * * * * * * * * * * *3 * * * * * * * * * * * * * * * *2 * * * * * * * * * * * * * * * *1 * * * * * * * * * * * * * * * *
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Segment 1
Segment 2
Segment 3
18 * * * * * * * * * * * *17 * * * * * * * * * * * * *16 * * * * * * * * * * * * * *15 * * * * * * * * * * * * * * *14 * * * * * * * * * * * * * * * *13 * * * * * * * * * * * * * * * * *12 * * * * * * * * * * * * * * * * * *11 * * * * * * * * * * * * * * * * * *10 * * * * * * * * * * * * * * * * * *
9 * * * * * * * * * * * * * * * * * *8 * * * * * * * * * * * * * * * * * *7 * * * * * * * * * * * * * * * * * *6 * * * * * * * * * * * * * * * * *5 * * * * * * * * * * * * * * * *4 * * * * * * * * * * * * * * *3 * * * * * * * * * * * * * *2 * * * * * * * * * * * * *1 * * * * * * * * * * * *
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
1
2
3
4
5
6
7
8
9
10
11
GE NXi 3D Projection view and Michelogram
7
3D PET• Sensitivity drops off towards edges • 4-5X increased sensitivity overall• Increased scatter (15% to 40%)• Increased randoms from out-of-field activity• Rebinning algorithms to apply 2D
reconstruction• Some devices can acquire in 2D or 3D whereas
some can only acquire in 3D• 3D in Brain, 2D (or 3D) in Whole Body
PET Attenuation Correction Methods• Calculated
– No noise but possibly inaccurate• Measured
– Accurate but noisy• Segmented, Measured
– Less noise => less time• Singles• CT
Calculated Attenuation Correction Measured Attenuation Correction
8
Measured Attenuation CorrectionSegmented, Measured Attenuation Correction
•Noise added from measured attenuation correction•Rel err in unif phantom (10 min EM)
•9% with calc atten•16% with 10 min TR•18% with 5 min TR
•Segmentation classifies by tissue type•Smoothes lung areas•Substantial reduction in noise added
Measured
Segmented
Singles-Based Attenuation Correction
•Single-photon source (Cs-137)•Source shielded from “near”detectors•Energy resolution (NaI) allows separation of 2 peaks•High count rate reduces noise•Susceptible to scatter
Courtesy of GE Medical SystemsPET-CT
9
PET-CT Attenuation Correction PET-CT Attenuation Correction
Iterative Reconstruction Feedback Loop
Backprojection
Simulated Projections
ActualProjections
Compare
Use to improve current estimate
Current Estimate
Error
Courtesy of Jerold W. Wallis, M.D.
Maximum Likelihood Reconstruction (ML-EM)
• Maximize the likelihood that the estimatedactivity distribution in the body (the reconstructed transaxial slices) would lead to the measured projections
• Use the expectation maximization (EM) algorithm to iteratively estimate the activity distribution
10
pixelj
projection bini
aij is the probability that a photon emitted from pixelj is detected at projection bini.
EM-ML Reconstruction
• aij can contain physical information (effects of spatial resolution, scatter, attenuation…)
• EM-ML algorithm takes into account the nature of the noise (quantum mottle) in the projection data
• Can yield a more accurate reconstruction
λ’(k)j = λ’(k -1)j Σ [aij {d/(d’(k-1)}]Σ [aij]
λ’(k)j is the newest estimate of the object pixel value
λ’(k -1)j is estimate of the object at last iteration
d is the projection data
d’ is the calculated projection data
OSOS--EM AlgorithmEM Algorithm• Ordered-subset expectation maximization• At each step, project and backproject at only some
angles (i.e. a subset)• Perform the steps in an ordered way to include all
angles• Data start to converge even before the 1st iteration
is complete• Convergence achieved in 3 - 10 iterations• Computation time of a few minutes• For GE, we use 28 subsets (12 projections per
subset) and 2 iterations.
11
OSEM Iterative Reconstruction
• Filtered Back-Projection– Fast– Robust– Subject to noise &
streaks• OSEM
– Almost as fast– Handles noise &
streaks
3D Reconstruction
• 2D reconstructions are much easier than 3D• Rebinning 3D data to be reconstructed as
2D data– Single-slice– Fourier
• Fully 3D Reconstruction
Rebinning3D Data Acquisition
N2 ObliqueSinograms
2N-12D Sinograms
2D Reconstruction
3D Object
2DSlices
Single Slice Rebinning (SSRB)
12
Fourier Rebinning
• More accurate approach to rebinning• Better estimate of determining into which
parallel plane oblique data should be placed• Based on the frequency-distance
relationship (Value of Fourier transform of a sinogram receives contributions mainly from sources at a fixed distance t=-k/ω)
Fourier Rebinning (FORE)
• Initialize a stack of Fourier transforms of 2D sinograms
• For each oblique sinogram– Take 2D FT– For each pixel in Fourier space, calculate the
interpolated plane location {z’ = z – k/ω tan(θ)}– Add values from oblique sinogram to 2D sinogram at z’
• Normalize to take into account over-sampled areas• Take inverse Fourier transform• Reconstruct interpolated sinograms as 2D
Note that this only requires a 1D interpolation along z.
PET Instrumentation
• High-Resolution• Medium-Resolution• PET-CT
PET Instrumentation
• High-Resolution– Siemens HR+– GE Advance– Philips Allegro
• Medium-Resolution– Siemens EXACT/ACCEL– Philips CPET+
13
PET InstrumentationSiemens HR+ GE Advance Philips Allegro
Detector Dimension (mm) 4.1 x 4.4 x 30 3.9 x 8.2 x 30 4 x 6 x 20# of Detectors 18,432 12,096 17,864Detector Material BGO BGO GSOSpatial Resolution (mm) 4.6 4.8 4.8Sensitivity (kcps/uCi/mL) 200/900 200/1060 /800
Siemens EXACT Siemens Accel Philips CPET+Detector Dimension (mm) 6.8 x 6.8 x 20 6.8 x 6.8 x 20 500x300x25# of Detectors 9,216 9,216 6Detector Material BGO LSO NaISpatial Resolution (mm) 6 6.2 5Sensitivity (kcps/uCi/mL) 180/780 180/780 /450
GE Advance NXi
PET-CT Scanners
• Siemens Biograph (BGO or LSO)• Siemens Hi-Rez• GE Discovery LS• GE Discovery ST• Philips Gemini
GE Discovery LS
• LightSpeed Plus CT (4-16 slice, 0.5sec gantry).• Full-featured Advance NXi PET.
– Retractable septa for 2D and 3D imaging.
LightSpeed PlusAdvance NXi
Courtesy of GE Medical Systems
14
Courtesy of GE Medical Systems
GE Discovery ST
GE Discovery ST
CT PET
GE Discovery ST (vs LS)
• Larger Detectors (6 vs 4 mm)• Shorter septa (5 vs 10 cm)• Higher Sensitivity• Larger patient bore (70 vs 50 cm)• No rod sources• 2D and 3D imaging
15
• ECAT EXACT HR+:• Siemens Somatom
Emotion: High performance, spiral CT
• 70 cm patient port • Optimized bed design
Siemens Biograph
Courtesy ofSiemens Medical Systems
Siemens Hi-Rez Detectors
Courtesy of Siemens
CURRENT HI-REZDetector material LSO LSOBlock matrix 8 x 8 13 x 13 Creates smaller voxels!Crystal size 6.4 mm x 6.4 mm 4.0 mm x 4.0 mm Enables finer resolution!Crystal thickness 25 mm 20 mmNumber of blocks 144 144Total number of crystals 9,216 24,336 Allows greater sensitivity!Axial field of view 16.2 cm 16.2 cmNumber of crystal rings 24 39 Permits more slices!Number of image planes 47 81 More slices!Plane spacing 3.4 mm 2.0 mm Excellent sampling!Ring diameter 83 cm 83 cmSpatial resolution 6.3 mm 4.6 mm Finer spatial resolution!Volumetric resolution 250 mm3 98 mm3 Finer volumetric resolution!
Siemens Biograph with LSO HI-REZ detectors
*510(k) pending, not for sale in the US
HI-REZ*
HR+
CONVENTIONAL
Note definition of sulcus and gyrus!
Courtesy of Siemens
Philips Gemini
Allegro PET Scanner andMx8000 16 slice CT scanner
Courtesy of Philips Medical Systems
16
PET-CT Scanners
GE Discovery LS GE Discovery ST Philips GeminiDetector Dimension (mm) 4 x 8 x 30 6.2 x 6.2 x 30 4 x 6 x 20# of PET Detectors 12,096 10,080 17,864PET Detector Material BGO BGO GSOSpatial Resolution 4.8 6.2 4.92D/3D 2D/3D 2D/3D 3DAtten Corr CT&Ga-68 CT CT&Cs-137
Siemens Biograph BGO Siemens Biograph LSO Siemens Hi-Rez LSODetector Dimension (mm) 4.1 x 4.4 x 30 6.5 x 6.5 x 25 4 x 4 x 20# of PET Detectors 18,432 9,216 23,336PET Detector Material BGO LSO LSOSpatial Resolution 4.5 6.3 4.62D/3D 3D 3D 3DAtten Corr CT CT CT
Special PET Devices
• Brain • Breast • Small Animal
CTI ECAT HRRT• Max Planck Institute in
Cologne • Dual LSO Phoswich for
DOI Determination• Octagonal Design• 936 blocks with 128 2.1x2.1
dual detectors each• 120,000 crystals• Reconstructed resolution of
less than 2.5 mm
Courtesy of CTI
Positron Emission Mammography (PEM)
Courtesy of Wake Forest Universityand PEM Technologies
17
Positron Emission Mammography (PEM)
Courtesy of Wake Forest University
Concorde MicroPET Scanner
650g Rat
[18F] Fluoride
4 bed positions30 min each
Courtesy of Concorde MicroSystems
GE eXplore Vista Scanner•Rodent system•11.8 cm diameter•4.6 cm axial FOV•1.6 mm resolution in CFOV
Philips Mosaic
•16,680 GSO crystals•2.1 mm resolution•11.8 cm axial FOV•137Cs Transmission
18
Summary
• Scanners specifically optimized for clinical (oncologic) imaging
• Sensitivity and count rate capability• New detector materials (faster!)• PET-CT• Small animal imaging