Advanced Scatter Correction for Quantitative Cardiac SPECT Frederik J. Beekman PhD Molecular Image Science Laboratory University Medical Center Utrecht The Netherlands
Outline What is scatter? Scatter correction methods (they really work!) Advanced iterative methods that incorporate scatter models General-purpose algorithms for scatter modelling: Rapid 3D Monte Carlo simulation Monte Carlo down-scatter correction for dual-isotope SPECT and attenuation maps Hardware essentials for attenuation and scatter correction
#1 ideal direct detection #2 direct detection; blur #3 not detected #4 scattered, causing attenuation What is scatter? #5 scatter,, Inappropriately detected => causes blur, distortion, and quantitative inaccuracy Energy resolution is insufficient for discriminating between scattered and primary photons
Energy resolution of camera determines amount of detected scatter Scatter Fraction for LAO View Tc-99m MIBI 0.40 Scatter Fraction For Window Twice % FWHM 0.35 0.30 0.25 0.20 0.15 0.10 0.05 0.00 0% 2% 4% 6% 8% 10% 12% Energy Resolution (% FWHM) Figure courtesy of Prof. Michael King
Examples of projections in SPECT Effect of scatter: Line source behind slab (1*) (1*) Simon Cherry et al, Physics in Nuclear Medicine Effect of scatter: Point source in cylinder
SPECT is severely degraded by photon scatter in patient Correction of scatter complicated by non-uniformity of thorax density We need methods that 1. reconstruct the same emission images of identical hearts, independent of surrounding anatomy 2. are robust to noise 3. are quantitative 4. provide good resolution 5. are fast enough and practical
Anatomical differences leads to different images of identical hearts Difference of amounts surrounding tissue leads to Quantitative inaccuracy Variable visual appearance of image Loss of lesion detectability Results in unnecessary exposure to catheterization risk
Scatter correction methods Recommended review paper: Zaidi & Koral Scatter modeling and compensation in emission tomography. Eur J Nucl Med Slides Mol are not Imaging. to be reproduced without 2004 permission May;31(5):761-82. of the author
Three main lines of scatter correction Energy based window subtraction * spectral deconvolution maximum likelihood estimation Spatial domain based spatial convolution, scatter point spread function *, Monte Carlo* Combined energy and spatial multiple energy response re-projection maximum likelihood Slides are estimation not to be reproduced without permission of the author *will be addressed
Scatter Estimation Projection data can be written as: Acq = (P + S) + Noise (P + S) = + + Scatter compensation attempts to reconstruct P Estimate S using Energy or Spatial Methods S can be subtracted before or used in reconstruction Note Noise Slide courtesy of Prof. Michael King
Energy Spectrum Scatter Estimation Use energy spectrum at each pixel to estimate scatter contribution at that pixel. Example: Triple Energy Window (TEW) Ogawa, IEEE TMI 10:408-412, 1991 S W C C ' =. + 1 W 3 5 2 1 W3 Slide courtesy of Prof. Michael King
Energy Spectrum Scatter Estimation Example: Triple Energy Window True noise free scatter projection Noise free TEW scatter estimate Noisy TEW scatter estimate Filtered TEW scatter estimate Slide courtesy of Prof Michael King
TPF 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 TEW scatter correction helps! ROC Curves for Overall Detection of CAD FBP vs Iterative (OSEM) FBP AC AC+SC AC+SC+RC 0 0.2 0.4 0.6 0.8 1 FPF AUC FBP AC AC+SC AC+SC+RC 0.808 0.845 0.868 0.894 p < 0.0001 Narayanan et al, J. Nucl. Med., vol. 44, 1725-1734, 2003. Gradual improvement in detection accuracy (over FBP) for CAD is seen when corrections are added in incremental steps OSEM with all 3 corrections (AC+SC+RC) provides statistically significantly improved detection accuracy over OSEM with solely AC
Spatial Domain Scatter Estimation Create an estimate of the scatter projection from: Current estimate of activity distribution Attenuation Map Model of Scatter Response (Point Spread Function or Monte Carlo simulation) Estimate can be: Subtracted from acquired projection Summed to the estimated primary projection and employed in iterative reconstruction (preferred method) Examples of Domain Methods: Floyd et al 1986 J. Nucl. Med. 27: 1577-1585; E.C. Frey et al. 1996 IEEE Nuclear Science Symposium 1082--1086 ; S. Meikle et al. 1994 J. Nucl. Med. 23: 360-367; B.F Hutton et al.1996 Eur. J. Nucl. Med. 23: 1300-1308; and F.J. Beekman et al 2002 IEEE TMI 21: 867-877
Domain scatter correction helps to.. cardiac insert with lesions non-uniform thorax phantom
Dependency reconstruction on accuracy of mathematical model used during reconstruction Tl-201 filled myocardial insert with lesion N A AD ADS From: F.J. Beekman et al, Improvement Slides are not to be of reproduced image without resolution permission of and the author quantitative accuracy in clinical Single Photon Emission Computed Tomography Comp. Med. Im. Graph., 2001
TC-99m ML-EM A ML-EM AD ML-EM EM-ADS DM-OS OS-ADS Note that ML-EM EM-ADS and DM-OS OS-ADS are extremely close
No scatter model Scatter model Gain in contrast when compared at equal image noise level (C. Kamphuis et al, Eur. J. Nucl. Med.,vol. 25 pp. 8-18, 8 1998 )
Advanced iterative methods that incorporate scatter models
Iterative reconstruction can be seen as a parameter estimation problem with a typical form: p = M a + n <=> p j = Σ i M a ji i + n j p j = projection data in pixel j (e.g. detected # of photons) Unknowns: a i = values of volume elements ( voxel voxel ) i (activity concentration) n j = noise in pixel j M ji = transition matrix element represents probability that photon emitted in voxel i is detected in pixel j Iterative reconstruction estimates a from above equation
Influence accuracy of matrix M ML-EM: δ-like PSF Only attenuation modeled ML-EM: Accurate scatter and detector model added
Transition matrix M is huge, and complicated Detector blurring Non-uniform attenuation (thorax!) Scatter 3D reconstruction
2D versus 3D SPECT reconstruction 3D DETECTOR DETECTOR 2D 3D SPECT reconstruction models photon cross-talk between slices, where a is an entire volume instead of a slice, and p consist of pixels lying in multiple planes Improved quantification, better SNR 3D requires larger matrix Slides are size not to be reproduced and without longer permission of reconstruction the author time
Iterative Reconstruction illustrated Object space Projection space Current estimate Update Object error map Matrix M Simulation (or re-projection ) Matrix M or M Backprojection Estimated projection Measured projection Error projection Compare e.g. - or /
Example iteration process: ML-EM reconstruction brain SPECT 0 iterations 10 iterations 30 iterations 60 iterations
Simulations studies have shown that modeling of scatter (in matrix M) during iterative reconstruction improves image noise properties compared to any window-based scatter correction Scatter Compensation Methods in 3D Iterative SPECT Reconstruction: A Simulation Study. F.J. Beekman, C. Kamphuis, E.C. Frey., Phys. Med. Biol. 1997
What may be achieved with modeling of photon scatter in M ij? Better quantitative accuracy Fewer windows required Better noise properties May be more robust
Novel general-purpose algorithms for scatter modelling: Rapid 3D Monte Carlo modelling
Accuracy physics model depends on the photon energy and grade of non-uniformity of the scatter medium Many scatter models proposed. Monte Carlo (MC) based modeling is accurate and general MLEM 2D MC version has been proposed in 1986 by the Duke group (Floyd et al 1986 J. Nucl. Med. 27: 1577-1585 ) Fully 3D MC based SR has always been prohibitively slow and required prohibitively large matrix (Terabytes)
A solution is to combine methods of acceleration of 3D MC reconstruction Dual Matrix OS-EM eliminates matrix storage Convolution Forced Detection (speed up MC re-projection) Lower number of photon tracks in early iterations Re-use of photon tracks calculated in previous iterations Details: F.J Beekman, H.M. de Jong, S. van Geloven, IEEE Trans. Med. Imaging 2002 21: 867-877
Dual Matrix OSEM reconstruction (DM-OS, Kamphuis,, et al. Eur.. J. Nucl Med. 1998): Do simultaneous (I) Ordered subsets and (ii) Dual Matrix (Zeng & Gullberg IEEE Trans. Nucl. Sci. 92 ) a k i + 1 = a k i ~ j M ji j ~ M p ji i M j a k ji i M models attenuation, detector and scatter (ADS) and ~ M models only attenuation and detector blur. No storage of huge scatter matrix required anymore!
Convolution Forced Detection (CFD) Phantom (density) Phantom Activity Forced Detection Convolution Forced Detection Evaluation: De Jong,, Beekman & Slijpen,, IEEE TNS 2001 Acceleration factor typically 50-100
artifacts No scatter modeled scatter PSF Monte Carlo based
Computational Load: Influence of Photon tracks 10^4 Photons 10^5 Photons 10^6 Photons Short axis Vertical profile 6 min 8 min 42 min Single CPU, Pentium IV, 2GHz De Wit, Xiao, Beekman (submitted)
Preliminary result 99m Tc Small lesion (3 ml) in the apical part of the inferior wall Short Axis Vertical profile ExSPECT (ADS) Monte Carlo based (ADS)
Down-scatter correction of Tl-201 images in simultaneous Tl-201/Tc-99m SPECT Simultaneous Tc-99m/ 99m/Tl-201 dual-isotope SPECT with Monte Carlo based down-scatter correction H.W.A.M. de Jong,, F. J. Beekman P.P. van Rijk and M.A. Viergever Eur.. J.Nucl Med., Aug; 29(8):1063-71, 2002
Problems in simultaneous Tc-99m/Tl-201 dualisotope SPECT for cardiac imaging -Down-scatter: Tc-99m photons are detected in Tl- 201 window Tl-201 Tc-99m -Down-scatter leads to decrease of contrast and quantitative accuracy Tc+Tl -Simultaneous Tc/Tl dualisotope SPECT is not recommended because sufficient down-scatter correction lacks 72 kev 140 kev Energy
Results Virgin Tl-201 image compared to dual-isotope (DI) Tl-201 images with and without down-scatter correction Virgin Tl-201 image Contaminated with Tc-99m down-scatter Corrected for Down-scatter
Tc-99m down-scatter correction of Gd-153 attenuation maps Monte Carlo-based down-scatter correction of SPECT attenuation maps. Bokulic T, Vastenhouw B, De Jong HW, Van Dongen AJ, Van Rijk PP, Beekman FJ. Eur J. Nucl. Med, 2004
1. Approximate attenuation map is reconstructed using down-scatter contaminated transmission data. 2. Emission map reconstruction using contaminated attenuation map. 3. Based on result step 1 & 2, down-scatter in the (153)Gd window is simulated using accelerated Monte Carlo simulation 4. Down-scatter estimate is used during reconstruction of a corrected attenuation map. 5. With corrected attenuation map, an improved (99m)Tc image is reconstructed. Steps 3-5 are repeated to incrementally improve the down-scatter estimate.
Hardware requirements Good attenuation maps are key to proper attenuation and scatter correction O'Connor MK et al. A multi-centre evaluation of commercial attenuation compensation techniques in cardiac SPECT using phantom models., J Nucl Cardiol. 2002 More energy information has to be recorded to make further progress More counts: converging collimators, more detectors
Hardware requirements (I) Attenuation Maps (problems) Sensitivity to down-scatter is significant Number of energy windows and how the system allows them to be positioned is often too limited for optimal down-scatter correction Reliability of mechanics involved can be disappointing Evans and Hutton; Variation in scanning line source sensitivity: a significant source of error in simultaneous emission-transmission tomography. Eur J Nucl Med. 2004 Resolution and reproducibility are not optimal yet Truncation of attenuation maps
Hardware requirements (II) Industry, please Choose good transmission hardware designs Do proper implementation Consult the inventors/scientists that proposed the TCT systems Take complains of customers serious Make systems cost effective (e.g., low cost transmission sources)
Hardware requirements (III) Attenuation Maps (solutions) Systems providing better attenuation maps have been proposed: X-ray CT: Good quality attenuation maps. Do you want this with all cardiac scans? Costs? Dose? Offset fan beam collimators with line source prevent truncation and have good transmission map resolution and emission sensitivity e.g. Chang et al.phys Med Biol. 1995 May;40(5):913-2 Gilland, Jaszczak and Coleman. Transmission CT IEEE Trans Nucl Sci 2000 Offset fan beam collimators with moving point source: Cheap sources, high resolution, low down-scatter contamination. minimal or no truncation, increased emission sensitivity F.J. Beekman et al. J.Nucl. Med. 1998; 39:1996-2003
Hardware requirements (V) Energy Windows More windows required to utilize information of scattered photons (4 is insufficient, approx. 15 is very nice) Preferred is list mode data ( infinite number of windows). (see presentation James Cullom) Overlapping moving windows are required for implementing simple and efficient correction methods (e.g. down-scatter correction for attenuation maps) and for increasing counting efficiency
Hardware requirements (VI) Detectors Use detectors with high energy resolution. => Less scatter detected => Main problem: costs
Conclusions and Discussion Scatter correction is important and can be effective Monte Carlo based iterative reconstruction is versatile and accurate. Is attractive for (down-)scatter correction in single and dual isotope SPECT Availability of better hardware for attenuation and scatter correction is essential to move forward to better cardiac SPECT images Industry should more rapidly adopt new methods that have been proposed in (recent) years.
Prof. Brian Hutton Dr. Hugo de Jong Dr. Chris Kamphuis Prof. Michael King Dr. V. Narayanan Brendan Vastenhouw Dr. Tim de Wit Prof. Peter van Rijk Jianbin Xiao Alice van Dongen Dr. Fred van het Schip Dr. Frank Nijssen Dr. Fred Verzijlbergen Acknowledgements
U-SPECT-I University Medical Centre Utrecht Mouse Heart LV RV LV apex LV 5 mm RV Six mci (99m)Tc Tetrofosmin Statistical reconstruction Acquisition: 30 min. non-gated 75 gold pinholes, = 0.6mm
U-SPECT-I University Medical Centre Utrecht Mouse Spine Spinous process Intervertebral foramen 5 mm Vertebral foramen Four mci (99m)Tc-HDP Acquisition time: 22 min. 75 gold pinholes, = 0.6mm Iso-surface renderings of Slides SPECT are not to bedata reproduced without permission of the author Transverse process