Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Evolution: A Framework for Advanced SPECT Reconstruction with Compensation for Image Degrading Factors Eric C. Frey, Ph.D. Division of Medical Imaging Physics Russell H. Morgan Department of Radiology Johns Hopkins University Slides not to be reproduced without permission of the author 1 Acknowledgements JHU: Algorithm Conception, Development, Implementation, Optimization, Validation, … Benjamin Tsui, PhD, Paul Segars, Ph.D. Dan Kadrmas PhD, Wen-Tung Wang, Ph.D. Yong Du, PhD, Xiyun Song,PhD, Xin He, PhD, Bin He Rich Wahl, MD, Harvey Ziessman, MD, Sybille Goetze, MD, Heather Jacene, MD, GE Healthcare: Integration, Validation, Optimization, … Lana Volokh, PhD, PhD,Aharon Peretz Eyal Shai Bella Yuzefovich, PhD Jonathan Sachs, MD Clinical Bone SPECT Images and Evaluation Dr. Einat Even-Sapir, MD, PhD Dept. of Nuclear Medicine, Tel Aviv University Slides not to be reproduced without permission of the author 2 Disclosure Under a licensing agreement between the General Electric Company and the Johns Hopkins University, I (Eric Frey) am entitled to a share of royalty received by the University on sales of the Evolution product described in this presentation. The terms of this arrangement are being managed by the Johns Hopkins University in accordance with its conflict of interest policies. Slides not to be reproduced without permission of the author 3 Goals • Describe Evolution and its constituents – Reconstruction – Optimization – Validation • Describe Evolution reconstruction and its components • Show some results from preclinical validation • Discuss optimization of acquisition and reconstruction parameters • Demonstrate application of Evolution reconstruction to clinical data – Bone SPECT – Myocardial perfusion SPECT Slides not to be reproduced without permission of the author 4 Iterative Reconstruction Algorithm What is Evolution? GE JHU GE & JHU Evolution Reconstruction Preclinical Validation Integration Optimization Optimization Optimization Clinical Validation Clinical Validation Clinical Validation Models for Imaging System Attenuation Map Evolution for Evolution for ?Prostate? ?Cardiac? Slides not to be reproduced without permission of the author Evolution for Bone 5 Evolution Reconstruction – Statistical Iterative Reconstruction – Compensation for Physical Image Degrading Factors Slides not to be reproduced without permission of the author 6 Statistical Image Reconstruction • Projection data are corrupted by Poisson noise • Statistical reconstruction methods – Explicitly model Poisson noise – Operate by finding activity distribution that “best fits” measured projection data – “Best fit” is judged by value of objective function Slides not to be reproduced without permission of the author 7 Iterative Reconstruction-Based Compensation Initial Estimate Project at each View Computed Projections Compare Computed & Measured Measured Projections Objective Function New Estimate Update Estimate Slides not to be reproduced without permission of the author Iterative Algorithm 8 Maximum Likelihood (ML) Objective Function • “Likelihood” is a statistical concept • Find activity distribution that is “closest” to projection data • Problem: tends to fit noise in projections Slides not to be reproduced without permission of the author 9 Maximizing the Objective Function – Requires iterative procedure – A variety of iterative algorithms have been proposed – Properties of iterative algorithms • • • • Reconstruction time Theoretical rigor Complexity to implement Properties of image – Noise magnitude and texture – Spatial resolution – Contrast resolution Slides not to be reproduced without permission of the author 10 Common Iterative Algorithms • Expectation Maximization (EM) – Easily applicable to Poisson Likelihood Objective function (ML-EM) – Converges very slowly (many iterations required) • Ordered subsets expectation maximization (OS-EM) – Much faster than EM (speedup ≈ # subsets) – Not theoretically rigorous – Problematic for very noisy data and large # subsets Slides not to be reproduced without permission of the author 11 Iterative Reconstruction-Based Compensation Initial Estimate Project at each View Computed Projections Compare Computed & Measured Models of Degrading Effects New Estimate Measured Projections Objective Function Update Estimate Slides not to be reproduced without permission of the author Iterative Algorithm 12 Image Degrading Factors • SPECT Projection images are degraded by – Attenuation in patient – Scatter in patient – Collimator-detector response – Poisson noise Slides not to be reproduced without permission of the author 13 SPECT Image Formation Point response function Real Collimator Ideal Absorbed Scattered Attenuation Multiply Scattered Source Scattered Slides not to be reproduced without permission of the author Scatter CDR blurring 14 Characteristics of Attenuation • Attenuation depends on depth, material, isotope 1 1 Tc-99m 0.8 Attenuation Factor Attenuation Factor 0.8 Muscle Lung Bone 0.6 Muscle Lung Bone 0.6 0.4 0.4 0.2 0 Tl-201 0.2 0 5 10 15 20 Depth (cm) 25 30 0 0 5 Slides not to be reproduced without permission of the author 10 15 20 Depth (cm) 25 30 15 Characteristics of Scatter • Importance of scatter increases with depth • More scatter for isotopes emitting lower energy photons • Shape of scatter response varies spatially and is patientdependent # Scattered Photons Tc-99m # Unscattered Photons 6.0 5.0 Tc-99m Tl-201 4.0 3.0 2.0 1.0 0.0 0 5 10 15 20 25 30 35 40 Slides not to be reproduced without permission of the author Source depth (cm) 16 Characteristics of the CDR • Width of CDR increases with distance from face of collimator • CDR is constant in planes parallel to face of collimator 6 cm LEGP LEHR Distance from 5 cm Collimator Face 10 cm 15 cm 20 cm Slides not to be reproduced without permission of the author 17 Effect of Geometric CDR on SPECT Images • Loss of resolution • Spatially varying resolution Point Source Phantom FBP Reconstruction from Projections with LEHR Collimator Slides not to be reproduced without permission of the author 18 Poisson Noise • Projection data corrupted by Poisson noise • Noise level determined by – injected activity – imaging time – sensitivity of collimator-detector system • Noise is spatially varying • Noise is irreversible, but effects can be “controlled” Slides not to be reproduced without permission of the author 19 Effect of Poisson Noise OS-EM Reconstruction • Noise increases with # updates • Post-filter needed to control noise Updates 12 25 60 120 240 480 No Post-filter 3D Butterworth Post-filter order=8 cutoff=0.24 pixel-1 Updates= # iterations x # subsets Slides not to be reproduced without permission of the author 20 Modeling Attenuation in Evolution • Requires patient-specific attenuation map • Can be obtained by • transmission imaging • x-ray CT image • Good registration is critical (better than 1 pixel) Slides not to be reproduced without permission of the author 21 Scatter Modeling in Evolution • Estimated using effective scatter source estimation (ESSE) – Physics-based method – Accurately models spatial variance of scatter response • Better than alternative approaches such as triple-energy window (TEW) method which can – provide less accurate scatter compensation – result in increase image noise Slides not to be reproduced without permission of the author 22 CDR Modeling in Evolution • Models spatially varying geometric CDR based on analytical formulas • Supports modeling of full CDR I-131 Point Source 30 cm MEGP Collimator HEGP Collimator Distance from Collimator Face 5 cmSlides not to 10 cm without15 cm of the author 20 cm be reproduced permission 23 Atten Comp Atten & Scatter Comp Reconstructed Pixel Value No Comp From Unscattered+ Scattered Photons From Unscattered Photons Efficacy of Attenuation and Scatter Compensation Unscattered-NC Scattered+Unscattered-NC Unscattered-AC Scattered+Unscattered-AC Scattered+Unscattered-ASC 150 100 50 0 0 20 40 60 80 100 120 Pixel Number Slides not to be reproduced without permission of the author 24 Efficacy of CDR Compensation • Resolution improves with iteration but remains limited: cannot totally recover resolution • Resolution remains spatially varying • Resolution for LEHR better than for LEGP OS-EM w/CDR compensation FBP Phantom Updates 128 320 640 1280 LEGP LEHR Slides not to be reproduced without permission of the author 25 Effect of Compensation on Image Noise • Noise increases w/ iteration • Attenuation Comp has larger noise where attenuation is greatest • CDR comp results in “lumpy” noise • Texture of noise w/CDR comp – varies spatially – depends on collimator Updates 128 320 640 1280 No Comp Atten CDR LEGP CDR LEHR Slides not to be reproduced without permission of the author 26 Tools for Validation and Optimization • • • • Phantoms MC Simulation Observer studies ROC Analysis Slides not to be reproduced without permission of the author 27 Tools for Validation: Phantoms • Physical Phantoms Cylindrical Phantoms 270o 180o RSD Phantom Torso Phantom 0o 90o • Mathematical Phantoms MCAT Phantom NCAT Phantom Slides not to be reproduced without permission of the author 28 Tools for Validation: Phantoms • Phantom population Slides not to be reproduced without permission of the author 29 MC Simulation Comparison of MC Simulation and Experiment 3.39 cm diameter Sphere w/In-111 in Elliptical Example of Simulation of In-111 Zevalin Distribution Computer Cluster Activity Attenuation Low-Noise Slides not to be reproduced without permission Mapof the authorProjections SPECT Projection 30 Preclinical Validation • Validate accuracy of models of degrading effects compared to simulation and experiment – Simple activity distributions – Realistic activity distributions • Validate effect on reconstructed images – Contrast, noise, SNR – Observer studies Slides not to be reproduced without permission of the author 31 Observer Studies • Evaluate image quality with respect to detection of – Bone lesions – Perfusion defects • Use – Human observer studies – Mathematical observer studies • Similar to computer aided diagnosis tools • Designed to predict human observer performance – > 90% correlation in predicting performance – > 96% correlation in predicting rankings Slides not to be reproduced without permission of the author 32 Observer Studies • Use ROC Analysis • Area under ROC curve is measure of performance – AUC=1 for perfect performance – AUC=0.5 for guessing True Positive Fraction (sensitivity) 1 0.8 FBP 0.6 OSADS 0.4 OSADS-NCC OSADS-MBCC 0.2 OSADS-True CC 0 0 0.2 0.4 0.6 0.8 1 False Positive Fraction (1-specificity) Slides not to be reproduced without permission of the author 33 Validation Using Simple Phantoms Tc-99m 9.7 cm diameter circular cylinder filled with styrofoam beads and water (dens=32% of water) Camera 270o 180o 31.2 x 22.8 cm water-filled elliptical cylinder 0o 90o 1 cm diameter sphere w/ Tc-99m Non-Uniform Attenuator Slides not to be reproduced without permission of the author 34 Validation Using Simple Phantoms: Tc-99m Slides not to be reproduced without permission of the author 35 Validation using MCAT Phantom -45o 0o 45o Error in SPR over all views: <5% 90o -45o 0o 45o 90o 135o 135o projector MC simulation Difference Slides not to be reproduced without permission of the author 36 Validation: Bone SPECT • Hot sphere on warm background – background:sphere = 1:20 – 120 views, 360 degrees, 2.21 mm pixels – Collimator 1.3 cm 1.0 cm 3.5 cm 0.9 cm Phantom OS-EM 2 it, 10 subsets Chang not to be reproduced without permission of theAC author Study performed by GESlides Haifa OS-EM w/CDR 4 it, 10 subsets Chang AC 37 Validation: Torso Phantom Reconstruction Activity ratios: heart : liver: torso = 10 : 10 : 1 A=Attenuation Compensation AD=Attenuation and CDR Compensation AS=Attenuation and Scatter Compensation ADS=Attenuation, CDR and Scatter Compensation FBP OSEM OSEM OSEM A AD OSEM AS OSEM ADS VLA Polar map Slides not to be reproduced without permission of the author 38 Preclinical Validation In-111 Imaging RSD Phantom NC A AS NC=No Compensation A=Attenuation Compensation AD=Attenuation and CDR Comp Coronal Projection CT Image AGS ADS Atn Map AS=Attenuation and Scatter Compensation ADS=Attenuation, CDR and Scatter Comp Slides not to be reproduced without permission of the author 39 0 -20 No Comp Atn Comp Atn + Scatter Atn+Scat+ Geom CDR Atn+Scat+Full CDR -40 -60 -80 5.6 cc sphere 20.6 cc sphere A Liver Lungs -100 Heart % Error in Activity Estimate Accuracy of Activity Quantitation: RSD Phantom and In 111 % Error in total activity estimation: (true-estimate)/true x 100% Slides not to be reproduced without permission of the author With appropriate reconstruction, quantitative SPECT is possible! 40 Preclinical Validation: In-111 Prostascint Clinical Images w/Simulated high-contrast lesion OS-EM No effects modeled, 20 updates, (clinical default) Post-filtered OS-EM , Attenuation Modeling, 12 updates, (“optimized”) Post-filtered OS-EM Atten, CDR and Scatter Modeling, 16 updates, (“optimized”) Post-filtered Slides not to be reproduced without permission of the author 41 Clinical Applications • Bone SPECT – Optimization – Clinical Validation • Myocardial Perfusion SPECT – Optimization – Clinical Application Slides not to be reproduced without permission of the author 42 Bone SPECT Optimization • Is OS-EM w/ CDR compensaton better than OS-EM w/ no compensation for – Same acquisition time – Half acquisition time – Same collimator – LEGP vs. LEHR collimator Slides not to be reproduced without permission of the author 43 Bone SPECT Optimization: Phantom Study • Methods • Results – Hot spheres on warm background – Computed sphere SNR and Contrast Recovery – LEHR better than LEGP – OSEM w/CDR better than OSEM w/o CDR – OSEM w/ CDR at ½ acquisition time better than OSEM w/o CDR 1.3 cm 1.0 cm 3.5 cm 0.9 cm Slides not to be reproduced without permission of the author Study performed by GE Haifa 44 Bone SPECT Optimization: Simulation Study • • • • • • NCAT Phantom Population LEHR Collimator Clinical count level Simulated Poisson noise Mathematical Observer (MO) ROC Analysis Activity map Attenuation map Noisy Projections L.Volokh, B.Tsui et al, “Efficacy of corrective reconstruction with collimatordetector response compensation for short tc-99m bone SPECT acquisition in a bone lesion detection Slides task,” SNM 2005. not to be reproduced without permission of the author 45 Bone SPECT Optimization: Simulation Study • Multiple lesion locations • Lesion contrasts – 2:1, 3:1, 5:1 • Lesion Sizes – 0.5, 1.0, 2.0 cm sternum ribs thoracic spine lumbar spine ilium pubis hip MIP image for 10 min SPECT acquisition reconstructed with CDR compensation Slides not to be reproduced without permission of the author 46 NC=no comp CDR= w/CDR comp Full Time=Normal clinical acq. time Half Time=half clinical acq time Half time reduces AUC Half time w/CDR as good or better than full time w/o CDR Area under ROC Curve Bone SPECT Optimization: Simulation Study Results 1 0.95 0.9 0.85 0.8 0.75 0.7 NC-Full Time NC-Half Time 3 6 Slides not to be reproduced without permission of the author CDR-Full Time CDR-Half Time 8 9 10 47 Clinical Application of Evolution for Bone Today 20 min 15 min Whole Body Planar Today’s procedure SPECT 1 Evolution - ½ time 20 min 7.5 min Whole Body Planar SPECT 1 Reduce overall acquisition time by 30% w/same quality Evolution - WB SPECT 7.5 min SPECT 1 7.5 min SPECT 2 7.5 min SPECT 3 3D Imaging in same time as planar imaging Slides not to be reproduced without permission of the author 48 Clinical Validation Bone SPECT • 46 patients, 102 lesions, Consensus read by 4 physicians • OS-EM full-time vs. half-time OS-EM – 14 studies had similar quality – 32 studies half-time had poorer quality – 5 lesions detected missed on half-time acquisition • OS-EM full-time vs. half-time OS-EM w/CDR – – – – – 74% of studies had similar quality 5 studies: OS-EM had better quality 7 studies: Half-count OS-EM w/CDR had better quality 1 lesion seen only on OS-EM w/CDR 1 lesion changed from 1 to 2 (5 point scale) on OS-EM w/CDR Z. Keidar et al, “Half-time bone SPECT acquisition - Assessment of a new Collimator Slides not to be reproduced without permission of the author Detector Response (CDR) reconstruction algorithm,” submitted to SNM, 2006 49 Clinical Validation Bone SPECT In 34 of 120 patients (28%), Multi-FOV SPECT detected lesions that would have been missed had planar and a single SPECT been performed Courtesy of Dr. E. Even-Sapir, Slides not to be reproduced without permission of the author Tel Aviv Medical Center 50 Clinical Validation Bone SPECT Lesion missed on wholebody planar images MIP Planar OS-EM w/CDR Images from Dr. E. Even-Sapir, Tel Aviv Medical Center Slides not to be reproduced without permission of the author 51 Clinical Validation from Bone SPECT Images Dr. E. Even-Sapir, Prostate Cancer: high risk for bone metastases. What FOV for spot SPECT? Tc-MDP SPECT Planar Wholebody Slides not to be reproduced without permission of the author Tel Aviv Medical Center MIP 52 Clinical Validation Bone SPECT Prostate cancer w/ high risk for bone metastases Images from Dr. E. Even-Sapir, Tel Aviv Medical Center SPECT Spot Planar Slides not to be reproduced without permission of the author 53 Clinical Validation Bone SPECT Was SPECT correct? Images from Dr. E. Even-Sapir, Tel Aviv Medical Center SPECT CT F-18 PET 24 min SPECT MIP Image quality approaching Slides PET not to be reproduced without permission of the author F-18 PET MIP 54 Cardiac SPECT Optimization • Questions – Which combination of compensations? • A, AS, AD, ADS? – What is optimal number of iterations/subsets? – What is optimal post-reconstruction filter? – 180 or 360° acquisition arc? – Can we reduce acquisition time? Slides not to be reproduced without permission of the author 55 Perfusion Defects – Six different locations with 36 different sizes • Size: randomly varied ±25% around the mean • Contrast: randomly varied in the range 10-35% I II III IV V VI Short Axis Long Axis Slides not to be reproduced without permission of the author 56 Optimization of Iterative Reconstruction 0.95 Area under ROC Curve Area under ROC Curve 0.95 0.9 0.85 -1 -1 4 OSAD-5 0.9 OSA-5 0.85 0.28 pixel 0.16 pixel 2 OSAS-6 -1 0.12 pixel 0.8 OSADS-10 0.8 6 8 Iterations 10 12 0.12 0.16 0.2 0.24 0.28 -1 Cutoff Frequency (pixels ) Slides not to be reproduced without permission of the author 0.32 57 Cardiac SPECT Optimization 1 3 AUC 0.9 6 4 1 2 5 0.8 0.7 0.6 defect 1 ADS defect 2 defect 5 defect 3 defect 4 A defect 6 N AUC for defects in 6 different locations when using ADS,A and N method. AUC values varied with defect location Sometimes A is worse than N ADS as good as or better than any method for all defect locations Slides not to be reproduced without permission of the author 58 Cardiac SPECT Optimization: 180 or 360° Arc Method 180° 360° p-value ADS 0.782 0.780 0.32 AD 0.777 0.770 0.26 AS 0.770 0.762 0.1 A 0.764 0.756 0.2 N FBP 0.741 0.735 0.25 0.740 0.718 0.0066* * Statistically significant p<0.05 Slides not to be reproduced withoutdifference permission of theat author 59 Cardiac SPECT Optimization Acquisition Time MO Study 0.9 0.85 Clinical count level=1.0 0.8 0.75 0.7 0.65 0.6 0.01 OSEM-NC OSEM-ADS 0.1 1 10 100 Relative Count Level Area Under ROC Curve Area Under ROC Curve 0.9 Compare MO and Human Observers 0.85 0.8 0.75 0.7 0.65 OSEM-ADS: Human Observers OSEM-ADS: CHO 1000 0.6 0.01 0.1 1 10 100 1000 Relative Count Level • OSEM w/ ADS better than OSEM w/ No Compensation (NC) • Can reduce acquisition time by ½ w/o degrading lesion detectability Slides not to be reproduced without permission of the author 60 Cardiac Imaging Clinical Validation • Clinical validation by GE and JHU in progress FB A AD ADS VLA SA FB: Filtered Backprojection A: OS-EM w/ Atten. Comp. AD: OS-EM w/Atten. and DRF Comp. ADS: OS-EM w/Atten., DRF & SRF Comp. Slides not to be reproduced without permission of the author Butterworth Post filter (order=5, cutoff=0.28 pixel-1) 61 Summary Statistical Image Reconstruction • Statistical reconstruction – Allows compensation for image degrading factors – Requires application-specific optimization – Can provide improved image quality – Can provide quantitative SPECT images Slides not to be reproduced without permission of the author 62 Summary Compensation for Degrading Effects • Attenuation – works well with high-quality attenuation map • Scatter – works well when appropriately implemented – Model-based can provide reduced noise compared to energy-window based (E.g., TEW) • CDR – – – – – Improves resolution Reduced high frequency noise Does not completely restore resolution Resolution remains spatially varying Introduces spatially-varying noise texture • Noise – Appropriate # of subsets, iterations and postSlides not to be reproduced without permission of the author reconstruction filter help control noise 63 Summary Evolution Framework • Evolution Framework – Evolution Reconstruction • OS-EM • Models for Attenuation, CDR, Scatter – Preclinical Validation – Application-specific optimization – Clinical Validation Slides not to be reproduced without permission of the author 64 Summary Current Status of Evolution • Evolution for Bone now available • Evolution for Cardiac, Prostate – Clinical validation in progress Slides not to be reproduced without permission of the author 65 Summary Evolution for Bone • OSEM w/CDR allows SPECT in ½ time (7.5 min) of clinical standard with equivalent or better image quality compared to OSEM w/o CDR • Can perform whole body SPECT (4 FOV) in same time as wholebody + 1 FOV SPECT – Equal or better contrast compared to planar – 3D information provides better localization information – No need to interrupt clinical flow or guess at appropriate FOV for SPECT – Image quality approaching PET w/o equipment upgrade Slides not to be reproduced without permission of the author 66