Influence of SPECT reconstruction algorithms in the

improvement of SNR in cardiac imaging

Helena Contreiras Marqueshelena.marques@tecnico.ulisboa.pt

Instituto Superior Tecnico, Lisboa, Portugal

Setember 2015

Abstract

Cardiac SPECT images often suffer from a number of degrading factors including depth-dependentspatial blurring, attenuation, scatter and low data counts. Thus, researchers have been working on de-veloping a variety of reconstruction methods incorporating scatter correction (SC), resolution recovery(RR), noise supression and attenuation correction (AC). The goal of this work is to study the accuracyof absolute and relative measurements, and how these degrading factors impact on the reconstructedmyocardium when different methods are used for half-time acquisitions. The reconstruction softwareused were WBR (UltraSPECT, Ltd.) and Evolution for Cardiac (GE Healthcare) with and withoutCT-based AC. The comparison between algorithms was made based on raw and normalized counts,severity and SRS values. Overall, some statistically significant differences were noted between param-eters obtained with different methods, which makes decision only based on them somewhat unreliable.The results obtained confirm the usefulness of AC in addition to RR, but they should be evaluated nextto NC images to avoid undervaluation of a lesion. The assessment of the RCA and LCX territory seemsto benefit most from it.Keywords: SPECT, Attenuation correction, WBR, Evolution

1. Background

Coronary Artery Disease (CAD) is a process inwhich the coronary arteries become partially orcompletely obstructed by the accumulation ofplaque (accumulation of lipids, complex carbohy-drates, and calcium deposits) on the inner wall ofthe arteries supplying blood to the heart. Withoutproper blood flow, the cardiac muscle will not re-ceive oxygen and other vital nutrients that makesthe heart work properly, hence, possibly creatingan array of heart problems that can lead to death.When the insufficient blood flow to the heart isonly temporary and reversible is called ischemia,however if the heart suffers an infarction, the dam-age made to the muscle is irreversible and perma-nent, on the other hand, if there is enough perfu-sion to keep the cells alive but not enough to allowa fully functional contraction of the heart muscle, itis still considered a viable myocardium. Therefore,is necessary to have an imaging technique capa-ble of setting apart ischemic, viable, and infarctedmyocardium as each one has a different impact onthe patients treatment and prognosis. Myocardialperfusion SPECT is considered an excellent non-invasive method for the diagnosis of CAD, predic-tion of disease prognosis, selection of patients for

revascularisation and assessment of acute coronarysyndromes [1]. Numerous studies have assessed therelative accuracy of SPECT imaging reporting asensitivity and specificity of 8789% and 7375%,respectively, dependent of the radionuclide chosenand stress modality [2]. During this procedure thepatient is injected intravenously with a radiophar-maceutical tracer that is taken up in the heart mus-cle (myocardium), to evaluate regional coronaryblood flow usually at rest and after stress. Thiscompound is a pharmaceutically-active molecule la-beled with a single-photon emitter, that is a ra-dionuclide tracer that emits one gamma-ray per ra-dioactive decay. This molecule is chosen on the ba-sis of its preferential localization in a given organ, orits participation in a physiological process. In car-diac imaging procedures, the most common usedradiopharmaceutical tracers are thallium chloride(201Tl), technetium (99mTc) sestamibi and tetro-fosmin. After the administration of the radiophar-maceutical tracer, its distribution within the my-ocardium (which is dependent on myocardial bloodflow) is imaged by the SPECT system. The im-age acquisition is performed by rotating a cameraaround a point (center of rotation) located within astationary patient so that multiple two-dimensional

1

projections can be acquired at different angles cov-ering an angular range of 180 or 360 (at equalangular steps). The data acquired in this projec-tions is reconstructed into tomographic slices us-ing a mathematical reconstruction algorithm andrealigned to 3 cardiac planes: horizontal, short-axisand horizontal long-axis.

1.0.1 Mode of operation

The gamma camera system has basically two func-tions: the detection of single photons events andthe measurement of its energy and position. Itsmain components are the scintillation crystal, colli-mator, an array of photomultiplier tubes (PMTs),a pulse height analyzer and an analogue electronicsfor position encoding. The first piece of hardwaremet by the gamma rays is the collimator, which isone of the most important features of the systemas is used to define the direction of the detectedgamma rays. By regulating which gamma rays getthrough, the collimator forms a projected image ofthe gamma distribution on the surface of the scin-tillation crystal made of sodium iodide, doped withthallium NaI(Tl). When the gamma rays interactwith the crystal, thousands of photons with wave-length of visible light are emitted. This light pulses,whose yield is proportional to the energy absorbedfrom the incident photon, are then optically guidedthrough an array of photomultiplier tubes ( PMT)system in which they are converted into an electri-cal signal through the production of photoelectronsin the photocathode. The output of each photomul-tiplier tube is an electric signal proportional to theintensity of the light that arrived at the tube. Themultiplier section of the PMT amplifies the elec-tronic signal so the current is sufficiently large tobe used by conventional electronic circuits. In orderto produce an image, we need more than just theintensity of the detected photons, we need to knowfrom where they came from. The position is deter-mined using the pulses that occur at the anodes ofthe phototube array. The output of the PMTs ismapped by a network of electric resistors, weightedaccording to the spatial position of the PMTs in thex-axis and y-axis of the coordinate system of the ar-ray. There are 4 position signals, labeled X+, X,Y + and Y , and one energy signal Z that indicatesthe energy of the incident photon or pulse height.This helps to identify the position of an event by ap-plying a simple formula per coordinate based on therelative amount of current received by each resistor.Because gamma camera operates in pulse mode wecan treat the light photons from a single event as aunit. Most gamma rays interactions (the ones thatdont suffer compton scattering), generates a largersignal in the PMT above the place where the inter-

action happened while the surrounding PMTs willreceive smaller amounts of light. The position sig-nal has to be normalized by dividing the positionsignals by the energy signal (Z). The next step isthe Pulse Height Analyzer (PHA) that determinesthe amplitude of the pulses, which correlate withthe gamma rays energy.

1.0.2 Collimator

In order to create an image, the photons emitted atthe source should hit the detector in a predictableposition. However, as with all forms of electromag-netic radiation, photons are emitted isotropicallyso, simply using a detector wouldnt result in animage. Some photons escape the patient withoutinteraction, some scatter within the patient beforeescaping, and some are absorbed within the patient.Also, many of the photons escaping the patient arenot detected because they are emitted in directionsaway from the detector. The geometry of the col-limator ensures that the position of the photon iswell determined. Unfortunately, this technique is aninefficient method because many potentially usefulphotons are absorbed by the collimator and dontcontribute towards the image formation. A heavyprice is paid for using collimation - the vast major-ity typically well over 99.95% of emitted photonsis wasted. Thus collimation, although essential toimage formation, severely limits the performanceof these devices. The design of the collimator de-pends on the gamma-ray energy and the trade-offbetween photon count sensitivity and spatial resolu-tion. It is impossible to optimize both parameters.SPECT sensitivity describes the probability of de-tecting a photon incident upon the detector, com-monly quantified as the number of detected countsper unit time per unit source activity for a speci-fied energy window and geometry of measurement(system sensitivity). In order to have better sensi-tivity, the collimator hole size has to be bigger andthe hole length shortened so that less photons areabsorbed by the collimator.

1.0.3 Energy discrimination

Usually, the spectrum includes the total energy pho-topeak without any interaction before reaching thecrystal and a background of lower energies dueto the partial absorption of gamma by comptonscattering. Because the path of a gamma photonchanges after undergoing Compton scattering, it isimpossible to locate where it came from. In orderto have a final image with decent resolution, wehave to avoid all events registered that dont cor-respond to the absorption by photoelectric effectof gamma photons with the total emission energy.

2

Thus, energy discrimination is important for imag-ing because provides a mean to reject the gamma-rays that lost their positional information. Thisis accomplished by placing a window for the ade-quate double threshold energy. If the event ampli-tude Z falls within the PHA settings, the event isaccepted. Usually the values for a gamma camerawindow (difference between lower and upper-leveldiscriminators) are within 20% (10%) of the pho-topeak energy in the pulse height spectrum. Fur-thermore, more than just one window can be used,for instance, if the isotope used has more than onegamma-ray emission.

2. Radionuclides in SPECT

The most important radionuclide in Nuclearmedicine and the one used in this study istechnetium-99m (99mTc). 99mTc has all the rightproperties, with 140 keV gamma photons conve-nient for detection and has a half life of 6.03 hours[3] that allows a fast clearing from the body afteran imaging process. This isotope of technetium canbind chemically to many biological active moleculesmaking it suitable for many medical exams and iseasily available from a 99Mo generator that can bestored in a radiopharmacy. The next table showsthe properties of different radionuclides used in Nu-clear Medicine.

3. Iterative reconstruction technique

As computer technology started to improve, the in-terest in these reconstruction methods in SPECTincreased because of the need to compensate thevarious image degrading effects in SPECT imagingprocess. The most important method are the Sta-tistical techniques, who differ between those whoassume Gaussian noise (involving least squares solu-tions) and those who assume Poisson noise (involv-ing maximum likelihood solutions). The most usedreconstruction algorithm is the Maximum Likeli-hood Expectation Maximization algorithm (MLEM)[4] and its accelerated form Ordered Subsets Expec-tation Maximization (OSEM) [5]. Since SPECTprojection data is severely affected by Poisson noise,an advantage of these algorithms is the treatmentof data according to the Poisson nature of the mea-sures.

3.0.4 Maximum Likelihood ExpectationMaximization algorithm (MLEM)

Iterative reconstruction is based upon the premisethat if estimated profiles are generated by forwardprojecting an initial estimate of the image. Theseestimated profiles can be compared with the realprofiles, to generate a profile error. Then, imageerrors can be generated by back projecting theseprofiles errors in order to update that first image

estimate. This process is repeated until the bestpossible solution is reached. Iterative reconstruc-tion methods are more flexible and allow the incor-poration of models capable of correct some of theimage degradation effects.

1. First, an estimate of the activity distributionwithin the patient is made, which is denoted byf(x, y). The first estimate is very simple, usually,a nonzero image that has the same total projectioncounts as the measured projection data.

2. The estimate is then forward projected to estimatewhat the detectors would measure given the initialobject i. In order for this to occur accurately, amodel of the emission and detection process mustbe incorporated, the system matrix into whichalterations for attenuation, scatter and loss of res-olution with depth can be included.

3. The estimated projections are then compared withthe measured projections and any discrepancies inprojection space are back-projected to give discrep-ancies in image space.

4. The differences between the estimated and actualprojections are used to adjust the estimated imageto achieve closer agreement.

5. The update-and-compare process is repeated untilthe difference between the estimated and acquiredprojections is minimal or until a fixed number ofiterations have been achieved.

4. Sources of degradation and their impactin SPECT

The ideal scenario for a SPECT would be that allgamma rays emitted by the decaying tracer couldescape the body and be detected by the gammacamera. However, realistically, the gamma photonsemitted are deeply affected by the interaction withtissue within the patients body (photon attenua-tion and scattering), by the inaccuracy of the colli-mator (blurring) and also by noise in part becauseof the reduction of counted events after collima-tion. After all, its impossible to obtain high qual-ity SPECT images without having in considerationthese phenomenons and their possible corrections.Therefore, its important to take into account thesesources of image degradation:

a) the attenuation of the photons traveling toward thedetector;

b) collimation, uniformity and stability of a gammacamera;

c) parameters related to corrupt recorded events dueto different physical interaction of gamma rays.

d) the partial volume effect (PVE) as a consequenceof the finite spatial resolution of the gamma cameradue to detector blurring and non-ideal collimation.

d) factors related to the patient movement and posi-tioning.

Although physicians have learned how to detectthe related artifacts, its likely that, in some cases

3

they still affect the patients diagnosis. On the otherhand, in order to achieve right quantitative SPECTanalysis, the data acquired has to contain a greatamount of correct information. As a result, is reallyimportant to understand all the effects that causedeterioration of the reconstructed image.

5. Commercial SoftwareOne of the biggest limitations of SPECT-MPI is atrade-off between image-acquisition time and noiselevels. Therefore, because of the advantages of ashorter acquisition time are significant, new SPECTreconstruction algorithms have been improved inorder to provide a better image quality despitepoor count statistics. The new algorithms allowlower count-density cardiac SPECT acquisitions tobe processed with resolution recovery (RR) andnoise-reduction techniques. Most manufactures ofSPECT cameras have already implemented thesenew improvements into MLEM or OSEM algo-rithm. In this study, the commercial reconstruc-tion software used are Evolution (GE Healthcare)[6] and Wide Beam Reconstruction (UltraSPECT)[7]. Clinical trials comparing these algorithms withconventional SPECT reconstruction showed that:

Acquisition time can be decreased to half withoutcompromising qualitative or quantitative diagnos-tic performance [8, 9, 10];

Acquisition time can be decreased to a quarterif the reconstruction is optimized for the reducedcount density [11];

Can provide similar diagnostic quality whetherimaging time is reduced by half or a half-dose isinjected [9, 12]

5.1. Evolution for CardiacGE Healthcare (Waukesh, WI) has developed anew reconstruction algorithm, Evolution for Car-diac, that incorporates RR and a maximum a pos-teriori (MAP) noise regularization. The Evolutionapproach focuses on CDR compensation by inte-grating the collimator-detector response in the iter-ative algorithm. The PSF is stored in a lookup tableand the radial distance of the detector is obtained aspart of the projection data. The collimator-specificdata are embedded in the software, also in the formof look-up tables that are part of the reconstruc-tion package. Therefore, during the reconstructionprocess, collimator length and septa thickness, in-trinsic resolution, crystal thickness and collimator-detector gap are all taken into account. Further-more, acquisition parameters like the distance fromthe center of rotation to the collimator for everyacquired projection, are retrieved directly from theraw projection data [13]. At the same time, Evo-lution software incorporates noise suppression sinceMLEM algorithm converges to a quantitatively un-biased but noisy image. Because noise tends to

propagate during image reconstruction resulting ina potential compromise between the noise level andquantitative accuracy. This method suppresses theimpact of noise by incorporating a maximum a pos-teriori (MAP) algorithm. Bayes theorem allows tointroduce in the reconstruction process a prior dis-tribution that describes properties of the unknownimage. Maximization of this a posteriori proba-bility over the set of possible images results in aMAP estimate [14]. The main idea behind thismethod is that, at the end of each iteration, therehave been calculated OSEM and Bayesian coeffi-cients. The next image is a pixel by pixel prod-uct of the current image and the two sets of coeffi-cients. Then, the current image is compared withthe prior, and if the image is locally monotonous,the coefficients generated by the prior are all unityand the OSEM coefficients are not changed. Only,when non-monotonous structures start to developalong the iterations, does the prior modify the cor-responding OSEM coefficients, in order to removethese structures from the image of the next iterationstep. There are two tasks that are clearly separated:the OSEM part is responsible for the generation ofa quantitatively correct image and the median rootprior (MRP) tends to remove the unwanted noisewithout blurring the locally monotonous structures[14]. Attenuation correction (AC) with Evolutionfor Cardiac also includes scatter correction basedon a dual-energy window where the scatter estima-tion is added to the estimated projection as opposedto subtracting the scatter from the original projec-tions, as is implemented in regular iterative recon-struction with scatter correction [6].

5.2. Wide Beam Reconstruction (WBR)

The WBR software (UltraSPECT Limited, Haifa,Israel) is another iterative reconstruction methodthat includes RR and noise reduction in order toimprove image quality in studies with fewer pho-tons counts. The WBR algorithm is also an OSEM-based algorithm that models the physics and geom-etry of the acquisition for resolution recovery in asimilar way than the previously described method.Therefore, during the reconstruction process, datarepresenting the relationship between each projec-tion pixel and reconstructed voxel is modified ac-cording to collimators geometry. These pixel-voxelweights correspond to the solid angles between eachdetector pixel and each body voxel and are cal-culated analytically [15]. On top of that, if theangular position is not available from the acqui-sition parameters, WBR uses a simple algorithmcapable of calculating the body contour of the pa-tient from which the distance of the detector to thebody determined [16]. Thus, resolution recoveryyields images of improved spatial resolution and

4

with less noise when compared with other simi-lar techniques [8]. Noise compensation with WBRcan suppress noise and enhance the signal-to-noiseratio (SNR) by modeling the statistical character-istics of the emission process and of the detecteddata. Most RR methods use the Poisson distribu-tion to describe the emission-detection statisticalmodel. However, WBR regularizes the likelihoodobjective function with a combination of Poissonand Gaussian distributions. If a higher weight isgiven to the Gaussian component, high-frequencycomponents in the projections are suppressed. Onthe other hand, if higher weight is given to Pois-sons component, then, results in recovery of high-frequency signal. The balance between these twocomponents is determined by Fouriers analysis ofa projection to determine the SNR that is presentin the acquired data and the approximate statisti-cal distribution. This way, no post filter is appliedand the parameters defining resolution and noise arechosen according to the data analysis and desiredsmoothness. WBR utilizes a stand-alone runninghardware workstation (Xpress.cardiac) and can re-construct data acquired from most scanners withstandard collimator design.

5.3. Quantitative analysisIn SPECT imaging, detailed analysis of shapes andvalues of the region of interest (ROIs) of the imageis important for diagnoses but high-accuracy quan-tification is, however, very difficult to achieve. Thequantification is affected by the degradation of theimage introduced by statistical noise, attenuation,collimator/detector response and scattering effects.These factors affect image contrast by reducing theconnection between image counts and activity con-centration, thus for a more reliable quantification,it is necessary to correct them using compensationalgorithms. As a result, this work will focus onthe analysis of the images processed by differentsoftware in order to compare the quantitative dif-ferences that arrive from each of them. In orderto have a quantitative analysis, each data pointon the polar map is assigned a number and colorcorresponding to the radiotracer activity at thatpoint. The use of two-dimensional polar map co-ordinates allows comparison of count intensities be-tween different patients. First, before any compar-ison is made, image intensities need to be normal-ized, with activity being calculated as a percentageof the maximal left ventricular uptake. The mean-ing of each score is as follow:

0 : 80%- normal;1 : 70% 80% - mild decrease;2 : 60% 70% - moderate decrease;3 : 50% 60% - severe decrease;4 :< 50% - absence of detectable radiotracer uptake.

6. Clinical study

This study was designed to determine whether sim-ilar absolute and relative quantification such as rawand normalized counts, severity and SRS parame-ters could be achieved with half-time acquisitions.Therefore, WBR (UltraSpect, Haifa) and Evolutionfor Cardiac (GE Healthcare) (IRACRR and IRN-CRR) were compared in order to evaluate the con-tribution of RR and AC in SPECT images quantifi-cation. The comparison made between WBR andIRNCRR had the purpose of evaluating both al-gorithms resolution recover. A similar study wasmade between both GE algorithms, IRACRR andIRNCRR, so as to compare RR results with andwithout attenuation correction. Last, WBR andIRACRR were also analyzed since their creditedas state-of-the-art iterative algorithms for SPECTreconstruction. All aforementioned advances havecontributed greatly to the change from evaluationby visual interpretation alone, towards describingperfusion in conjugation with automated quantifi-cation as an aid to clinical diagnosis. Absolute andrelative quantification of MPI has reduced inter-andintra observer variability, and allows the possibilityto study and compare parameters in the same pa-tient or between a group. The objective of myocar-dial perfusion quantification is the regional classi-fication of the myocardial tissue as being normal,scar, or ischemic, thus, the visual representation ofquantitative values obtained from SPECT imagesis an important part of the evaluation process. Ac-cording to standard quantification protocols, eachrest and stress scan data is resolved into a polarmap based on derived measures of defect size, sever-ity, and reversibility. The quantified measures areaccomplished by comparing each segment of the po-lar map with measurements made from a popula-tion that is known to be normal (patients with lowpretest likelihood ( 5%) of CAD).

Population

Overall, for this study, we evaluated 73 patientswho were scheduled to have myocardial perfusionSPECT for suspected or known CAD at the Atom-edical Laboratory in Lisbon, Portugal. However,23 of these scans had to be excluded, because inthose cases, WBR scan frame had the maximal pixelcount outside of the myocardium (i.e., the stom-ach, bowel, or liver), thus assigning a lower inten-sity value to the myocardium pixels. The purposeof this study was to assess the impact of these soft-wares according to patients sex and site of CADfor overweight patients. Therefore, the main se-lection factor for this study was body mass indexBMI. > 31Kg/m2 with the total of 50 patients,25 women and 25 men, with the following charac-

5

teristics:

Women Men

N 25 25Age (y) 68, 6 9, 5 63, 7 11, 3

BMI (Kg/m2) 36, 5 3, 8 35, 7 4, 2

Table 1: Characteristics of the patients.

Image acquisition protocolsAll acquisition were performed with GE OptimaNM/CT 640 Gamma Camera system 20 minutesafter injection of approximately 7 mCi 99mTc - tet-rosfosmin (GE Healthcare). The images were ac-quired with a low-energy high-resolution collimator(LEHR), 6464 matrix, an elliptic orbit with step-and-shoot acquisition at 3 intervals over 180 , 60projections and 9-13 s per projection using a 20%energy window centered on the 140 keV photopeakof 99mTc. The patients were in supine position onthe table with his or her arms raised straight abovethe head. The SPECT image set was reconstructedon a dedicated workstation (Xeleris, GE Health-care, Haifa, Israel), using WBR and Evolution forCardiac recommended manufacturer RR and noise-reduction parameters and with and without CT-based AC (12 iterations and 10 subsets). At theend of each acquisition a single low-dose CT scan(100 keV; 1,0 mA; 0,2-0,3 mS) of the chest was per-formed in order to obtain attenuation maps auto-matically applied by the processing software to cor-rect the emission data. The myocardial perfusionimaging dataset is carefully co-registered with theCT attenuation map to produce the attenuation-corrected images. The comparisons were made be-tween thr territories associated to the various coro-nary arteries, RCA - right coronary artery; LCX -left circumflex artery; LAD - left anterior descend-ing artery.

Raw countsAll reconstruction processes were done with half-time protocol, therefore with low count statisticsin comparison with standard FBP reconstruction.Raw counts obtained by WBR, IRACRR and IRN-CRR were all significantly different (p < 0, 05).WBR reconstruction clearly had the best resultsaveraging around 1000 more counts than IRACRRand 2000 more than IRNCRR. Therefore, WBRresolution recover and noise suppression is able torecover better the loss of resolution at differentsource-detector distances during the procedure thanEvolution for Cardiac with or without using CT forAC. For Evolution for Cardiac algorithms, the dif-ference between IRACRR and IRNCRR raw counts

was around 1000. Since the main difference be-tween these two methods is the inclusion or not ofattenuation correction, is fair to state that the ex-tra raw counts are a consequence of its use. Overall,from the data collected, there is no obvious relation-ship between these algorithms with an exceptionof IRACRR and IRNCRR correlation for women(r=0, 75 0, 80). A similar result could be ex-pected for mens raw counts but their relationshipis only average (r=0, 56 0, 62).

Normalized CountsQuantification of myocardial activity is convention-ally measured relative to the region of most intenseuptake (brightest pixel count). Therefore, each rawdata set is normalized to the maximum myocardialtracer content in the LV in order to be comparedwith a database obtained from subjects with ex-pected normal perfusion. This study showed thatfor LAD territory, all comparisons were within 95%CI limits, with an exception of WBR & IRNCRRcomparison for mens perfusion data. For this par-ticular situation, IRNCRR algorithms gives an av-erage perfusion of 80, 8% and the addition of ACdecreases the perfusion value in 1, 68%. Althoughfor women, the difference is not statistically sig-nificant, the use of AC in this territory increasesthe perfusion,on average, 1, 5% from IRNCRR and1, 9% from WBR. For LCX territory, mean perfu-sion values were the lowest for IRNCRR and thehighest for IRACRR, with a statistically signifi-cant bias between them of 3, 5% for women and3, 7% for men. In this case, for women, WBR algo-rithm had a good agreement with IRACRR whilefor men, WBR mean average was closer to the IRN-CRR one. RCAs mean perfusion values, for men,are significantly different between all 3 algorithms.For this LV territory, its obvious the difference be-tween IRACRR algorithm and both WBR and IRN-CRR. The difference between Evolution for cardiacwith and without AC (IRACRR-IRNCRR) perfu-sion values is 10%. The perfusion obtained withWBR algorithm had on average minus 12% thanIRACRR and was closer to the IRNCRR mean withonly minus 1, 8%. Similar results are also observedfor women perfusion values, however, the differencesbetween IRACRR and both IRNCRR and WBRare 3, 8% and 3, 3%, respectively. Overall, for nor-malized counts, we encountered clear discrepanciesbetween perfusion values acquired by these 3 algo-rithms. Attenuation of photons within the body isrecognized as a major factor limiting SPECT detec-tion of myocardial perfusion defects. This is partic-ularly important in low-count studies such as thisone since it can compromise image quality becauseof the increased noise due to low statistics and blurfrom scattered photons. The main purpose of this

6

study was to evaluate absolute and relative quanti-tative parameters that are generally used to classifynormal or abnormal myocardial perfusion withinclinical setting. In particular, to determine whetherthose parameters differ significantly between atten-uation corrected and non corrected images in pa-tients with high BMI. Attenuation artifacts causedby the highly nonuniform tissue composition of thethorax has a big impact in diagnostic accuracy ofSPECT and although the true prevalence is un-known some reports estimate they appear in be-tween 20% to 50% of the studies [17]. The most fre-quently identified sources of artifacts are breast tis-sue in women and in the by the left hemidiaphragmin men. Breast attenuation often appear as a regionwith decreased count density along the anterior wallof the LV (mostly in LAD territory) although de-pending on their density, shape and position rela-tive to the myocardium, lateral wall, septum, andeven the apex can be affected. Attenuation by sub-diaphragmatic structures commonly affect the as-sessment in regions associated with RCA and LCX.Several studies reported that attenuation correctionimproves primarily specificity in the RCA territory[17] [18]. However, it can be said that some studiesdescribe different results for the influence of AC inthe sensibility to detect CAD in the LAD and LCX.For instance, Vidal et al. showed that AC actuallyreduced the detectability of the defects in the LADregion and reported that a significant loss of sensi-tivity in the LAD territory followed the increase ofRCA specificity with the application of attenuationcorrection [19]. In our study, first and foremost,its plainly evident the influence of attenuation cor-rection in the estimation of radiotracer uptake forLCX and in RCA territories. The results presentedshowed that images reconstructed with AC had animprovement in perfusion, in particular, RCA ter-ritory in men. Furthermore, a higher percentageof perfusion was obtained for 65% and 73% of thestudies reconstructed with IRACRR in comparisonwith WBR and IRNCRR, respectively. Looking atthe Bland-Altman plot for normalized counts itspossible to see some significant differences in perfu-sion for IRACRR and IRNCRR comparison. Bycomparing BA plots 95% CI limits for men andwomen, we get that the mean bias for women LADterritory is positive (1, 5%), which means that inaverage, perfusion values obtained with AC are big-ger than without AC as was expected. On theother hand, for men we have a mean bias betweenIRACRR and IRNCRR of 1, 7%, hence for moststudies, we get higher perfusion values for IRNCRRthan with IRACRR. Since, for men in particular, wehave that with AC there was a considerable positivebias for both RCA and LCX, in which often perfu-sion differences were about 15% higher than those

without AC. It is possible that over-correction ofthe inferior wall (RCA) resulted in a relative de-crease in tracer distribution in the LAD territorywhen the brightest pixel used for normalization was,for NC images, in a position where AC didnt en-hanced greatly the artificially reconstructed counts.When territories like RCA are intensified, the max-imum myocardial tracer content can be in a differ-ent place and have a significantly higher value withAC. Thus, a region in the LV where for NC imagesshowed a bright color indicating high perfusion can,in AC images, have decrease relative perfusion val-ues or look like it has less tracer uptake due to big-ger contrast between over-corrected areas and areaswhere the perfusion values are normal but the colorscheme of the polar map changed. Therefore, itsfair to consider that when RCA territory is over-corrected, the polar maps might show decreased up-take in the AC corrected images for LAD. As a wayof avoiding misinterpretation, IRACRR should beinterpreted with IRNCRR or WBR reconstructedimages.

SeverityFirst, womens LAD territory, showed a good agree-ment between WBR, IRACRR DB 2 and IRNCRRwhere their means difference are close to 0 withp > 0, 963. For men, there is also a good agree-ment between IRACRR DB 2 and IRNCRR, al-though WBRs severity average differs significantlyfrom IRNCRR and also has a considerable gap toIRACRR DB 2 results, yet not statistically signifi-cant. While, for women, AC doesnt seem to have abig impact since its results are close to those algo-rithms without AC, for men, WBR results have, onaverage, plus 0, 320SD than IRACRR DB 2. Theuse of DB 2 for IRACRR algorithm creates a po-lar map with, in average, less 0, 240SD than withDB 1 for women and plus 0, 184SD for men. Sim-ilar to LAD, LCX territory for women had a goodagreement between WBR, IRACRR DB 2 and IRN-CRR. AC doesnt seem to have a big impact in thewomens LCX territory since its results are close tothose algorithms without AC. However for men, de-spite WBR and IRNCRR having good agreement,WBR results have, on average, plus 0, 400SD thanIRACRR DB 2, which is statistically significant.RCA territory had the most significant differencesbetween algorithms and its the most affected by at-tenuation correction. IRACRR with DB 2 correctsIRNCRR, on average, by 0, 428SD for womenand 0, 560SD for men, which are both statisti-cally significant. WBR comparison with IRACRRDB 2 yields contrary results, with plus 0, 392SDfor women and plus 0, 188SD for men. Moreover,the high perfusion values obtained with IRACRRfor men, described earlier, are interpreted similarly

7

by WBR algorithm despite having, on average, a12% perfusion difference in this territory. Thus, islikely that some of those large difference in traceuptake are due to over-correction of the RCA ter-ritory for men, because otherwise, we would see amore pronounce difference between IRACRR andWBR severity measures. On the other hand, thedifference in perfusion between WBR and IRACRRfor women still differs significantly, on average, forseverity polar map results. In this case, WBR algo-rithm has a better agreement with IRNCRR thanwith IRACRR, independently of the DB used. Thisresult demonstrates that AC for women RCA ter-ritory may not be very beneficial. The use of DB2 for IRACRR algorithm creates a polar map with,in average, only plus 0, 072SD than with DB 1 forwomen and plus 1, 452SD for men. For womensRCA, the choice of DB 1 and DB 2 for IRACRRdoesnt have a big effect in severity results. Over-all, for men, IRACRR DB 2 gives smaller severityresults than IRNCRR and WBR, while for women,that difference is not as evident but happens regu-larly for RCA territory.

Summed Rest Scores (SRS)

The differences between SRS are closely relatedto severity results, thus, most considerations above-mentioned also apply for this section. For womensLAD, SRS dont differ significantly between the 3algorithms, thus are not typically changed by theuse of AC whereas for men, a considerable bias canbe seen between WBR and IRACRR DB 2. WBRalgorithm scores, on average, 0, 880 more than Evo-lution with attenuation correction. Furthermore,mens SRS scores of IRACRR with DB 1 differ sig-nificantly when compared with DB 2 for LAD andLCX. The choice of DB can have a big influence inclinical diagnostic since with DB 1, IRACRR scoresare 0, 920 less than with DB 2 for LAD and 0, 640for LCX. Attenuation correction also has a biggerimpact in LCX scores for men than for women. Formen, IRACRR DB 2 also gives a score 0, 800 lower,on average, than IRNCRR algorithm. Therefore,despite severity results agreement between this pair,the correspondent SRS differs significantly. WBRand IRNCRR seem to have the identical interpreta-tion for SRS on womens LAD and RCA territories,only in LCX appears a significant difference whereIRNCRR scores were 0, 560 higher than WBR. Fi-nally, for RCA territory, both WBR and IRNCRRattribute, on average, 1 score more than IRACRRDB 2 for women patients. SRS values attributed byIRACRR DB 2 are, regularly, significantly smallerhence, the consequences of AC can be different foreach LV territory. What appeared as a perfusiondefect in NC images, can become a near-normal ho-

mogeneous uptake with AC and be diagnosed oftenas normal. Thus, AC has to be used carefully as itcan overlook possible significant lesions. Addition-ally, a comparison between SRS and clinical diag-noses were made in terms of sensitivity, specificity,Positive Predictive Value (PPV) and Negative Pre-dictive Value (NPV). The values obtained for theseparameters are meant to give us an idea of howthe summed rest scores agree with the physiciandiagnosis for each patient. Since abnormal clinicaldiagnoses for men were only 3, sensitivity resultswere heavily affected by a minor change. Typically,AC adjusts the intensity of the myocardial perfu-sion image to reflect the estimated magnitude ofsoft tissue attenuation on different regions of theLV. Thus, the relative uniformity of the tracer dis-tribution in patients is improved, resulting in a bet-ter diagnostic accuracy [20, 21, 22], which usuallyresults in fewer false-positives tests (higher speci-ficity). These results are in concordance with thestudies mentioned above, in which the specificitywas higher with the use of AC. On the contrary,the sensitivity decreased in comparison with IRN-CRR and WBR, hence it shows, without knowl-edge of the expected changes in activity distribu-tion that occur with the use of AC, some MPI re-sults can be misleading and cause to incorrectly mis-judge a lesion. A closer look to IRACRR DB 2false-negatives (3 for womens data and 1 for mensdata) showed that the main difference was the un-der valorization of SRS values for LAD. DescribingSRS values, only for LAD territory, in the vectorform of (WBR,IRNCRR,IRACRR DB 2) for eachstudy, the SRS values for the false negatives were(5,6,1), (7,7,1), (3,2,0),(5,5,1). With an exceptionof the third vector, the SRS results for LAD wereenough, under the conditions adopted, to considerthem as abnormal studies. Thus, part of the loss ofsensitivity can be due to misjudgment of the LADterritory. Therefore, it is understandable the con-tradictory opinions about the influence and benefitsof AC in SPECT-MPI, where some report good im-provements [20, 21, 22] and others do not find muchdifference between them [17, 19, 23].

7. Conclusions

Due to their well-documented beneficial effect onimage quality, the use of resolution recover and at-tenuation correction in SPECT have become widelyused in Nuclear Medicine. This study comparedLV quantitative parameters determined by 2 newmethodologies, half-time Evolution for Cardiac andWBR reconstruction. The quality of the images,obtained with these half-time acquisitions, were forall studies equivalent or superior to that usuallyachieved with full-time acquisitions processed withFBP. In this work, the techniques described were

8

validated only for the single SPECT\CT system -the GE Optima NM\CT 640 Gamma Camera sys-tem (GE Healthcare) with high-resolution parallel-hole collimators manufactured specifically for thatcamera. Because resolution recovery and noise re-duction are modeled specifically for each cameraand collimator, these results may not coincide toother cameras and collimators. The first conclusionto be drawn was that, overall, WBR reconstruc-tion algorithm had significant higher number of rawcounts than Evolution for Cardiac with and with-out AC. Furthermore, we have that the addition ofAC to SPECT images resulted on average, 5 timesmore raw counts. Some statistically significant dif-ferences were noted, between quantitative param-eters of raw, normalized counts, severity and SRSvalues, obtained with the different reconstructionsoftware, which makes decision fully based on themsomewhat unreliable. This was true, specially, forRCA and LCX territory. In general, the results ofour study confirm the usefulness of attenuation cor-rection in addition to resolution recover in SPECTimages. The assessment of the RCA territory seemsto benefit most from it, which is in agreement withthe results of other studies using various systemsfor attenuation correction [21, 22]. Because quanti-tative parameters attributed by Evolution with theuse of AC are typically significantly smaller thanfor NC images, the consequences of AC can be dif-ferent for each LV territory. Thus, what appearedas a perfusion defect in NC images can become anear-normal homogeneous uptake with AC and bediagnose often as normal. For this reason, AC hasto be used carefully as it can overlook possible sig-nificant lesions. On the other hand, these difficultiesin the interpretation of AC images can disappear ina matter of time if its use eventually becomes stan-dard practice. Now, most physicians are used andprepared to evaluate in a certain matter the differ-ent color patterns showed in each study, based ontheir knowledge of where some attenuation or arti-facts can appear in the LV. Because AC images havedifferent standards, NM physicians may take sometime to assimilate those differences. Because of thelack of an independent standard, it remains uncer-tain which of the approaches yields a more accurateresult in this small number of cases. Further inves-tigation with an independent standard such as PETwould be necessary to settle this question. There-fore, as is the case for existing commercially avail-able algorithms to evaluate LV function, the inter-preting physician must be aware of these method-ological differences and interpret scans accordingly.The phantom study showed that the raw data reg-istered, for 4 acquisitions over a time period, withboth Evolution for Cardiac algorithms follow thedecay 99mTc. On the other hand, for WBR algo-

rithm, the same acquisitions yield a number of rawcounts approximately constant. The small num-ber of subjects studied is a major limitation of thestudy and subsequent large scale clinical implemen-tation of this novel image reconstruction algorithmrequires a further larger patient study rigorouslytested in standardized conditions.

AcknowledgementsFirst and foremost, I would like to thank Dr. Guil-hermina Cantinho and Prof. Fernando Godinho,for their guidance during my research and study.This dissertation truly would not have been pos-sible without their effort. Second, I would like toacknowledge the contribution of everybody workingin the Atomedical Laboratory in Lisbon, Portugal,in particular, all the NM technicians for their pre-cious time and valuable suggestions. Lastly, I wouldlike to thank Prof. Ldia Ferreira for the all thesupport and encouraging during the developmentof this thesis.

References[1] Andrew Cassar, David R Holmes, Charanjit S

Rihal, and Bernard J Gersh. Chronic coro-nary artery disease: diagnosis and manage-ment. In Mayo Clinic Proceedings, volume 84,pages 11301146. Elsevier, 2009.

[2] J D Schuijf, L J Shaw, W Wijns, H J Lamb,D Poldermans, a de Roos, E E van der Wall,and J J Bax. Cardiac imaging in coronaryartery disease: differing modalities. Heart(British Cardiac Society), 91(8):11107, Au-gust 2005.

[3] R Herbert, W Kulke, and RT Shepherd.The use of technetium 99m as a clinicaltracer element. Postgraduate medical journal,41(481):656, 1965.

[4] Lawrence A Shepp and Yehuda Vardi. Maxi-mum likelihood reconstruction for emission to-mography. Medical Imaging, IEEE Transac-tions on, 1(2):113122, 1982.

[5] H Malcolm Hudson and Richard S Larkin. Ac-celerated image reconstruction using orderedsubsets of projection data. Medical Imaging,IEEE Transactions on, 13(4):601609, 1994.

[6] Gehealthcare.com. Ge healthcare web server,2015.

[7] UltraSPECT. Ultraspect shaping the futureof molecular imaging, 2015.

[8] Salvador Borges-Neto, Robert A Pagnanelli,Linda K Shaw, Emily Honeycutt, Shuli CShwartz, George L Adams, and Ralph Ed-ward Coleman. Clinical results of a novel

9

wide beam reconstruction method for short-ening scan time of tc-99m cardiac spect per-fusion studies. Journal of nuclear cardiology,14(4):555565, 2007.

[9] E Gordon DePuey, Ramesh Gadiraju, JohnClark, Linda Thompson, Frank Anstett, andShuli C Shwartz. Ordered subset expecta-tion maximization and wide beam reconstruc-tion half-time gated myocardial perfusion spectfunctional imaging: A comparison to full-timefiltered backprojection. Journal of NuclearCardiology, 15(4):547563, 2008.

[10] CV Venero, AW Ahlberg, TM Bateman,D Katten, SA Courter, AI McGhie,RD Philips, JA Case, RJ Golub, SJ Cul-lom, et al. 2.07: Enhancing nuclear cardiaclaboratory efficiency: Multicenter evaluationof a new post-processing method with depth-dependent collimator resolution applied tofull-and half-time acquisitions with simultane-ously acquired gd-153 line source attenuationcorrection. Journal of Nuclear Cardiology,15(4):S4, 2008.

[11] E Gordon DePuey, Srinivas Bommireddipalli,John Clark, Linda Thompson, and Yossi Srour.Wide beam reconstruction quarter-time gatedmyocardial perfusion spect functional imaging:a comparison to full-time ordered subset ex-pectation maximum. Journal of nuclear cardi-ology, 16(5):736752, 2009.

[12] E Gordon DePuey, Srinivas Bommireddipalli,John Clark, Anna Leykekhman, Linda BThompson, and Marvin Friedman. A compar-ison of the image quality of full-time myocar-dial perfusion spect vs wide beam reconstruc-tion half-time and half-dose spect. Journal ofNuclear Cardiology, 18(2):273280, 2011.

[13] GE Healthcare. White paper: Evolution forcardiac. Technical report, GE Healthcare, 3000North Grandview Blvd Waukesha, WI 53188U.S.A., 2007.

[14] Sakari Alenius and Ulla Ruotsalainen.Bayesian image reconstruction for emis-sion tomography based on median rootprior. European journal of nuclear medicine,24(3):258265, 1997.

[15] Ultraspect. wide beam reconstruction: break-ing the limitation of the line spread function.Technical report, 2007.

[16] Rafael C Gonzalez and Richard E Woods. Dig-ital image processing, 2002.

[17] Robert C Hendel, Daniel S Berman, S JamesCullom, William Follansbee, Gary V Heller,Hosen Kiat, Mark W Groch, and John J Mah-marian. Multicenter clinical trial to evaluatethe efficacy of correction for photon attenua-tion and scatter in spect myocardial perfusionimaging. Circulation, 99(21):27422749, 1999.

[18] Yasmin Masood, Yi-Hwa Liu, Gordon DePuey,Raymond Taillefer, Luis I Araujo, StevenAllen, Dominique Delbeke, Frank Anstett,Aharon Peretz, Mary-Jo Zito, et al. Clinicalvalidation of spect attenuation correction usingx-ray computed tomographyderived attenua-tion maps: multicenter clinical trial with an-giographic correlation. Journal of nuclear car-diology, 12(6):676686, 2005.

[19] Renaud Vidal, Irne Buvat, Jacques Darcourt,Octave Migneco, Philippe Desvignes, MarcelBaudouy, and Franoise Bussire. Impact ofattenuation correction by simultaneous emis-sion/transmission tomography on visual as-sessment of 201tl myocardial perfusion images.Journal of Nuclear Medicine, 40(8):13011309,1999.

[20] Benjamin M.W. Frey, Eric C. and and J. Ran-dolph Perry. Simultaneous acquisition ofemission and transmission data for improvedthallium-201 cardiac spect imaging using atechnetium-99m transmission source. Journalof Nuclear Medicine, 33(12):22382245, 1992.

[21] Edward P Ficaro, Jeffrey A Fessler, Paul DShreve, James N Kritzman, Patricia A Rose,and James R Corbett. Simultaneous trans-mission/emission myocardial perfusion tomog-raphy diagnostic accuracy of attenuation-corrected 99mtc-sestamibi single-photon emis-sion computed tomography. Circulation,93(3):463473, 1996.

[22] Regine Kluge, Bernhard Sattler, Anita Seese,and Wolfram H Knapp. Attenuation cor-rection by simultaneous emission-transmissionmyocardial single-photon emission tomogra-phy using a technetium-99m-labelled radio-tracer: impact on diagnostic accuracy. Euro-pean journal of nuclear medicine, 24(9):11071114, 1997.

[23] Richard E. Stewart, Richard A. Ponto, Chris-tine Z. Dickinson, Larry Meakem, Rao Chava,and Jack E. Juni. In-vivo validation of simul-taneous transmission-emission protocol (step)for tc99m-sestamibi spect-quantitative com-parison with n-13-ammonia pet. Journal of theAmerican College of Cardiology, 25(2s1):217A,1995.

10