Cardiac gating with a pulse oximeter for dual-energy imaging

This article has been downloaded from IOPscience. Please scroll down to see the full text article.

2008 Phys. Med. Biol. 53 6097

(http://iopscience.iop.org/0031-9155/53/21/014)

Download details:

IP Address: 199.212.7.17

The article was downloaded on 15/10/2008 at 14:19

Please note that terms and conditions apply.

The Table of Contents and more related content is available

HOME | SEARCH | PACS & MSC | JOURNALS | ABOUT | CONTACT US

IOP PUBLISHING PHYSICS IN MEDICINE AND BIOLOGY

Phys. Med. Biol. 53 (2008) 6097–6112 doi:10.1088/0031-9155/53/21/014

Cardiac gating with a pulse oximeter for dual-energyimaging

N A Shkumat1, J H Siewerdsen1,2, A C Dhanantwari2, D B Williams2,

N S Paul3, J Yorkston4 and R Van Metter4

1 Department of Medical Biophysics, University of Toronto, Toronto, Ontario, M5G 2M9 Canada2 Ontario Cancer Institute, Princess Margaret Hospital, 610 University Ave., Toronto, Ontario,M5G 2M9 Canada3 Department of Medical Imaging, University Health Network, Toronto, Ontario,M5G 2M9 Canada4 Carestream Health Inc., Rochester, NY 14650, USA

E-mail: jeff.siewerdsen@uhn.on.ca

Received 15 April 2008, in final form 14 August 2008Published 14 October 2008Online at stacks.iop.org/PMB/53/6097

Abstract

The development and evaluation of a prototype cardiac gating system fordouble-shot dual-energy (DE) imaging is described. By acquiring both low-and high-kVp images during the resting phase of the cardiac cycle (diastole),heart misalignment between images can be reduced, thereby decreasingthe magnitude of cardiac motion artifacts. For this initial implementation,a fingertip pulse oximeter was employed to measure the peripheral pulsewaveform (‘plethysmogram’), offering potential logistic, cost and workflowadvantages compared to an electrocardiogram. A gating method was developedthat accommodates temporal delays due to physiological pulse propagation,oximeter waveform processing and the imaging system (software, filter-wheel, anti-scatter Bucky-grid and flat-panel detector). Modeling the diastolicperiod allowed the calculation of an implemented delay, timp, required totrigger correctly during diastole at any patient heart rate (HR). The modelsuggests a triggering scheme characterized by two HR regimes, separated bya threshold, HRthresh. For rates at or below HRthresh, sufficient time exists toexpose on the same heartbeat as the plethysmogram pulse [timp(HR) = 0].Above HRthresh, a characteristic timp(HR) delays exposure to the subsequentheartbeat, accounting for all fixed and variable system delays. Performancewas evaluated in terms of accuracy and precision of diastole-trigger coincidenceand quantitative evaluation of artifact severity in gated and ungated DE images.Initial implementation indicated 85% accuracy in diastole-trigger coincidence.Through the identification of an improved HR estimation method (modifiedtemporal smoothing of the oximeter waveform), trigger accuracy of 100%could be achieved with improved precision. To quantify the effect of thegating system on DE image quality, human observer tests were conducted

0031-9155/08/216097+16$30.00 © 2008 Institute of Physics and Engineering in Medicine Printed in the UK 6097

6098 N A Shkumat et al

to measure the magnitude of cardiac artifact under conditions of successfuland unsuccessful diastolic gating. Six observers independently measured theartifact in 111 patient DE images. The data indicate that successful diastolicgating results in a statistically significant reduction (p < 0.001) in the magnitudeof cardiac motion artifact, with residual artifact attributed primarily to grosspatient motion.

1. Introduction

Conventional chest radiography (CXR) has proven to be inadequate in the detection of early-stage lung cancer (Diederich and Wormanns 2004), with poor sensitivity attributed to thesuperposition of anatomy within a projection image (‘anatomical noise’) obscuring subtlestructures (Samei et al 1999). Low-dose multi-detector CT (LDCT) offers a significantimprovement in sensitivity (Henschke et al 1999), with a lack of specificity presenting aconsiderable challenge (McLoud et al 1992, Pieterman et al 2000, MacRedmond et al 2006).Dual-energy (DE) imaging offers an improvement in diagnostic sensitivity (Ricke et al 2003,Tagashira et al 2007), providing a potentially improved primary read and facilitating earlierdetection of disease. The capability to classify fine patterns of nodule calcification couldfurthermore provide improved specificity (Kelcz et al 1994, Ricke et al 2003, Fischbach et al2003). These indicators suggest a promising role for DE imaging in baseline radiographicimaging and as a stand-alone or adjuvant modality (e.g., combined DE + LDCT, offeringimproved sensitivity and specificity) for early lung cancer detection.

Double-shot DE imaging provides a number of advantages over single shot, includingimproved soft-tissue contrast and detective quantum efficiency (DQE) (Alvarez et al 2004,Richard and Siewerdsen 2007). The progression of double-shot DE chest imaging into theclinical environment requires investigation of numerous important challenges. The currentwork describes the development of a cardiac gating system to reduce anatomical registrationand the severity of cardiac motion artifacts. Cardiac motion typically manifests in DE imagesas white or dark edges along the periphery of the image of the heart, as well as around theaortic arch and adjacent bronchioles and vasculature. The presence of such artifacts degradesimage quality, can obscure underlying anatomy and potentially distract a reader from a correctdiagnosis. With respiratory motion minimized by patient breath hold, a prospective gatingsystem is described that triggers x-ray exposures at a specified phase of the heart cycle. Thetwo exposures need not be acquired within the same heartbeat (i.e., they may be acquired onsubsequent cycles), but each exposure must be acquired at the same phase of the cardiac cycle.

The cardiac cycle involves two distinct mechanical periods: systole and diastole. Systoleis the phase of the heart cycle involving sudden mechanical contraction of the ventricles,whereas diastole is the quiescent phase during which blood flows passively from the atriato the ventricles (Levick 2000). The proportion of time that the heart spends in each phasedepends on the heart rate (HR). Typically, diastole occupies ∼60–70% of the heart period(e.g., 0.6 s at HR = 67 bpm), compared to ∼30–40% (∼0.3 s) for systole (Levick 2000). Atlower HR, a proportionally large amount of time is spent in diastole, and as HR increases,the diastolic period reduces considerably, while the duration spent in systole (∼0.3 s) is onlyslightly affected (Levick 2000).

The electrocardiogram (ECG) provides a means for cardiac gating that is fairly widespreadin medical imaging (Ohnesorge et al 2000, Kopp et al 2001, Herzog et al 2002, Desjardins andKazerooni 2004, Finn et al 2006, Groves et al 2007). For chest radiography, however, ECG

Cardiac gating with a pulse oximeter for dual-energy imaging 6099

presents a variety of disadvantages, including extended setup time, the need for disposableconductive leads (diminishing workflow), and the possibility for ECG components to obstructthe visibility of anatomy in a projection image if placed on the chest. Pulse oximetry providesa potentially useful alternative to ECG. By measuring pulsatile characteristics in peripheralarteries (Jago and Murray 1988, Jubran 1999), an estimate of heart motion can be obtainedby characterizing the plethysmogram with respect to an ECG and accounting for the delayassociated with physical propagation of blood to peripheral arteries. The simple probe providesadvantages in cost and workflow compared to an ECG and does not obstruct radiographicimages.

The purpose of the work described below was to devise a system capable of triggeringexposures during diastole. This work investigates the extent to which an oximeter can be usedas a simple trigger that reproducibly triggers exposures within the quiescent heart phase. Fora DE imaging system triggered randomly (i.e., without cardiac triggering), the probability thatboth exposures coincide with diastole can be computed based upon the fraction of time theheart spends within the phase at a given HR. The probability of correct diastolic triggeringsimply by chance ranges from ∼52% (72% per projection) at low HR (40 bpm) down to ∼30%(55% per projection) at high HR (110 bpm).

The modeling, implementation and initial performance evaluation of an oximeter-basedcardiac gating system is reported. Building on the development of a cardiac-gated DEimaging prototype developed in our laboratory (Siewerdsen et al 2006, Shkumat et al 2007a,2007b), we describe the temporal characteristics of each system component, including thosedue to the triggering system (i.e., physiological processes and oximeter processing) and theimaging system (i.e., the generator, x-ray tube, anti-scatter grid and flat-panel detector (FPD)).The temporal response of the pulse oximeter was measured in a volunteer cohort. Followingthe characterization of the temporal response, a system model was created that accountsfor the HR-dependent duration of the diastolic period in order to compute timing schemes thattrigger x-ray exposure within diastole. The cardiac-gating system was then integrated into theDE prototype and investigated in an ongoing research trial (figure 1). The performance of thegating system was evaluated through two quantitative analyses. The exposure time with respectto the diastolic period was analyzed to quantify the performance of the triggering system withrespect to the system model. In addition, this allowed investigation of the sensitivity of thetrigger to patient HR. The performance of the gating system was also measured in DE patientimages by means of an observer study in which the magnitude of the cardiac motion artifactwas measured in DE images. Together, these studies provide an important basis for initialperformance characterization.

2. Materials and methods

2.1. Cardiac gating with a pulse oximeter

A fingertip pulse oximeter typically has two clinical purposes. The first is to identify bloodoxygen levels and detect hypoxemia (Jubran 1999). The second is to monitor changes in bloodvolume and pulse propagation as characterized in a plethysmogram, which is derived fromthe measurement of a pulsatile signal in arterial blood coupled with the absence of motionin venous structures and surrounding tissue. Pulse oximeters can provide a digital triggeroccurring at any characteristic point in the plethysmogram waveform. For example, as shownin figure 2 (alongside an ECG response), the fingertip pulse oximeter employed in this work(Ipod, Nonin Inc., Minneapolis, MN) incorporates digital processing directly on the fingertipprobe to provide a digital trigger associated with the rising edge the plethysmogram. Thus, the

6100 N A Shkumat et al

Figure 1. Photograph of the DE chest imaging research prototype. The patient is setup in a PAposition in front of the detector stand (containing the FPD and anti-scatter Bucky grid). Clippedto the left forefinger is the pulse oximeter providing the cardiac trigger.

Figure 2. Timing diagram displaying the electrocardiogram (ECG) trace, associatedplethysmogram and digital trigger. Regions of systole, (Q–T interval), diastole, (T–Q interval)and heart period are shown. Mean delay measurements between the R-wave and midpoint ofplethysmogram as well as between the R-wave and digital trigger are indicated.

oximeter output—either the digital trigger or the plethysmogram itself—provides a surrogatefor cardiac activity. However, there are temporal delays associated with the oximeter digitalprocessing that need to be accounted. As described below, dependence on HR was includedby way of a cardiac timing model.

Cardiac gating with a pulse oximeter for dual-energy imaging 6101

2.1.1. Characterization of system delays. Within a cardiac-gated DE imaging system, thereare several potential temporal delays, including intra-exposure delays (i.e., those occurringbetween the heartbeat that induces a digital trigger to the end of the corresponding x-rayexposure) and inter-exposure delays (i.e., those occurring after the first x-ray exposurebut before the request for the second.) The cardiac-gating system described in this workis dependent on intra-exposure delays, since the high- and low-kVp images are acquiredindependently as triggered by separate heartbeats. Systems for which both images acquiredwithin a single heart cycle necessarily consider both intra- and inter-exposure delays.

2.1.2. Measurement of pulse propagation delay. The first delay contributing to the intra-exposure delay is the time required for blood propagation from the left ventricle to the fingertip,measured using the oximeter plethysmogram. A distance-normalized value of 303 ms (Du et al2005) (pulse transit velocity = 3.3 m s−1) is a representative mean, with variability dependenton factors including cardiovascular health and blood pressure. A second delay is due todigital signal processing within the pulse oximeter itself. This delay, in addition to the pulsepropagation delay, offsets the plethysmograph indicator of systole from the true mechanicalevent. Although the magnitude and variability of the digital processing delay for the pulseoximeter were proprietary to the manufacturer, the combined delay due to pulse propagationand processing (termed the ‘trigger delay’) was analyzed by measuring the plethysmogramrelative to ECG (taken as truth). Specifically, the delay between the peak of the QRS complex(the R-wave, an indicator close to systole initiation) and the rising edge (50% peak value) ofthe plethysmogram was measured. Initial estimates of the delay were measured in a subgroupof nine volunteers varying in age (34–49) and lifestyle (1 smoker/8 nonsmokers), with eachvolunteer monitored using a 3-lead ECG on the chest and the pulse oximeter clipped to theirleft index finger. Data were collected under conditions that mimicked the actual DE imagingprocedure— namely standing position, relaxed, arms placed at the hips and a comfortablyinspired 10 s breath-hold, with heart rate recorded for the duration of the test. For eachvolunteer, measurements were repeated 5–8 times to evaluate the mean trigger delay as wellas intra- and inter-patient variability. These measurements provided an essential preliminaryestimate of pulse-propagation delay for examining the viability of the gating system in initialstudies.

2.1.3. Delay measurements for the prototype imaging system. The DE image acquisitionworkstation and software also contribute potential intra- and inter-exposure delays. Similarly,the generator (Indico 100, CPI, Georgetown, Ontario) and x-ray tube (Varian Rad-60, SaltLake City, UT) potentially contribute to the inter-exposure delay. For the system implemented,the time for generator, tube and filtration preparation takes place between images (in parallelwith other delays), poses no bottleneck, and was therefore neglected. The FPD (Pixium-4600,Trixell, Morains, France) imparts distinct intra- and inter-exposure delays. The intra-exposurecomponent is due to the detector initialization and readout refresh processes. The FPD delayencompasses the time from a frame request to the detector-ready state. The FPD incorporatedin the DE imaging prototype exhibits a variable delay dependent on the time within the refreshcycle the frame request occurs. This delay was measured using digital pinout signals detailingthe time of initialization and exposure. The second delay associated with the FPD is the timeinvolved in detector readout. This panel-specific delay is significant and is the primary factorin determining the time between projections.

From the intra-exposure delays described above, it is possible to group and categorizeas either ‘trigger delays’ or ‘imager delays’. The former are associated with physiological

6102 N A Shkumat et al

pulse propagation and pulse oximeter digital trigger processing, which combine to give avariable delay, denoted ttrigger, with superscripts ‘mean’, ‘range’, ‘min’ and ‘max’ to describethe associated characteristic of the distribution—e.g., tmin

trigger and tmaxtrigger are the lower and upper

bounds of the trigger delay, respectively.The second category of intra-exposure delay is those occurring after the digital trigger,

which include those of the acquisition software and FPD. These delays occur in parallel and aredominated in the current prototype by the FPD, denoted timager. This delay exhibits variabilityassociated with detector refresh as described above. We similarly characterize timager as adistribution characterized by superscripts described above.

2.2. Cardiac gating model

For patients with low HR, the duration of the cardiac cycle is sufficiently long as toaccommodate the intra-exposure trigger and imager delays, allowing an x-ray exposureto occur within the same heart cycle as the digital trigger despite all sources of timingvariability. At higher HR, the mean, range and/or maximum of system delays are too long toguarantee exposure within the same heartbeat. In this case, the exposure must be postponedto the subsequent heart cycle—and into the subsequent diastolic period—by means of animplemented delay, denoted timp. These considerations suggest two distinct timing regimes,dependent on patient HR. The first regime is that below a threshold heart rate (HRthresh) forwhich the implemented delay is zero (timp = 0), and the exposure occurs immediately followingthe digital trigger. The threshold is computed as

HRthresh = 60 s min−1

tmaxtrigger + tmax

imager + tbuffer, (1)

where tbuffer is a fixed parameter that accounts for the finite x-ray exposure time, as well asacting as a buffer to account for errors within estimates of the measured temporal delays.The threshold is thus determined by the largest values of the trigger and imager delays—i.e.,tmaxtrigger = tmin

trigger + trangetrigger and tmax

imager = tminimager + t

rangeimager.

For patient HR greater than the threshold, the implemented delay is nonzero and postponesx-ray exposure to the center of the subsequent diastole. The required implemented delay is

timp(HR) = [tHR(HR) − tmin

trigger

]+

[tsystole(HR) − tmin

imager

]

+ 1/2[tdiastole(HR) − t

rangetrigger − t range

imager

], (2)

where tHR(HR), tsystole(HR) and tdiastole(HR) are the heart period, systole duration and diastoleduration, respectively, each a function of HR. The implemented delay postpones x-ray exposurebeyond the period of ventricular filling (the current diastole), past ventricular contraction (thenext systole), and into the following diastole region. Note that timp incorporates the minimumtrigger and imager delays to ensure a delay of sufficient length to avoid the systolic period.

The pulse propagation delay was used as an indicator of ventricular systole to trigger x-rayexposures within diastole. Specifically, the model computes the ‘implemented delay’ (timp)required such that the x-ray exposure is coincident with the center of diastole. In itself, thisprovides considerable temporal flexibility (tolerance), since the diastole period is relativelylarge, ranging from ∼300 ms (HR = 115 bpm) to ∼700 ms (HR = 60 bpm). Variability inthe pulse propagation delay due to patient-specific physiological processes is accounted forby describing ttrigger using a range, described by upper and lower bounds (tmax

imager and tminimager).

Thus, the cardiac trigger model accommodates any value of ttrigger that falls within the range ofmeasured temporal delays (section 2.1.1). This allows for an effective ‘window’ that ensuresproper triggering of the x-ray exposure during diastole.

Cardiac gating with a pulse oximeter for dual-energy imaging 6103

2.3. Image acquisition and performance evaluation

2.3.1. Patient cohort. The performance of the cardiac trigger was evaluated in terms ofthe coincidence of x-ray exposure with the diastole window as well as the magnitude ofcardiac motion artifacts in DE images. Patients were setup in posterior–anterior (PA) positionon the DE imaging prototype as in figure 1, with the pulse oximeter clipped to the leftindex finger. Plethysmogram and HR were measured in real-time during image acquisition.Following initialization of the imaging system, patients were instructed to hold their breathupon comfortable inspiration. HR was recorded from the oximeter pulse following thetechnologist’s exposure request (button press), and the timing for x-ray exposure was computed(for various system and implemented delays, described below). The x-ray exposure thus occurssynchronous to diastole. A total of 131 patient images were available for the current study,with 20 cases excluded for reasons not related to the triggering system, creating a total ofNpatients = 111. Protocol exclusion criteria included: age <18 years, chest thickness >28cm and heart rate > 115 bpm, and known cardiac arrhythmia. The cohort exhibited a slightmajority of male subjects, with an average age across all patients of 64.9 years. The averageHR during image acquisition was 77.0 bpm.

2.3.2. Trigger and exposure monitoring. An auxiliary workstation (IBM workstation,2.2 GHz, 1 GB RAM, Armonk, NY) was implemented on the clinical prototype to monitorsystem timing. It was connected in parallel to both the acquisition PC and pulse oximeterusing an onboard data acquisition card (PCI-6036E, National Instruments, Austin, TX). Usinga custom software oscilloscope developed in Matlab (The Mathworks, Natick, MA), themonitoring system recorded the actual time of x-ray exposure, the total imager delay, thepatient plethysmogram and the heart rate (as measured by the pulse oximeter, denoted HRox).These measurements allowed quantitative evaluation of the coincidence of x-ray exposurerelative to the modeled target diastole window. Data from the monitoring system providedretrospective analysis across the first 37 patients involved in the clinical trial. Timing plotsidentifying exposure location with respect to the target HR-dependent diastole window allowedquantitation of gating accuracy and precision.

2.3.3. Robust estimation of patient HR. To achieve a robust HR estimate, a pulse oximetertypically applies a running average over the instantaneous HR to account for spuriousdata, intermittent contact or sampling errors. The resulting estimate, HRox, thus involvesa fairly long lag and a degree of inaccuracy under conditions where the HR is changingrapidly. Such is the case under conditions of breath hold, where patient HR is expectedto change upon initiation of breath hold (Guyton 1996, Desjardins and Kazerooni 2004)—typically increasing briefly (∼2–4 s) due to increased venous return, followed by a significantreduction in HR due to increased interthoracic pressure. Slowly (after ∼6–8 s, depending oncardiovascular health), the HR increases and surpasses the resting HR until normal respirationresumes. Because the calculation of the implemented delay is a function of HR (timp(HR) inequation (2)), an accurate estimate of patient HR even in the context of such dynamics isessential to precise cardiac gating.

HR estimation schemes alternative to that intrinsic to the device were explored todetermine a methodology that was both accurate (minimizing lag) and robust (minimizingartifacts due to erroneous sampling). The instantaneous HR (calculated directly from the timebetween the current and previous cycles) provides the least lag but is also the most susceptibleto spurious error. Alternative smoothing windows were investigated to determine the accuracyand precision of triggering based on HR estimates calculated with smoothing windows of

6104 N A Shkumat et al

0 (instantaneous), 1, 3, 5, 7 and 9 cycles. Retrospective analysis of timing data acquired withthe monitoring system for 37 patients participating in the DE imaging trial provided the basisfor evaluation.

2.3.4. Image-based measurement of cardiac artifact. To quantify the effectiveness ofthe gating system in reducing image artifacts, a human observer study was performed inwhich six observers (physicists) were independently presented 111 DE patient images (bonedecomposition images) in randomized order and asked to measure the size of the artifact usinga custom Matlab measurement tool. Observers were trained to identify the cardiac motionartifact in a set of 5–9 images and were blinded to the gating scheme throughout the study inan effort to minimize bias. Each observer was asked to measure the thickness of the largestvisible artifact on the left edge of the heart. The cases in the current study (Npatients = 111)were drawn from the patient imaging trial. Physicists were considered adequate observersfor this test (as opposed to expert chest radiologists), because the imaging task required littleclinical knowledge of anatomy, disease or image interpretation. The study was conducted in aradiology reporting room with subdued lights. Images were cropped to (2000 × 2000) to bestillustrate the regions of the heart and lungs, and observers were allowed to magnify the imagesas desired. The tissue weighting factor (wb = 0.56) (Shkumat et al 2008) and window-levelsettings (image mean ± 3 standard deviations) were fixed for all images to ensure consistentdisplay across observers.

Patient images were categorized into two groups: (1) 94 images with successful diastole–diastole triggers and (2) 17 images in which at least one of the two projections involvedexposure during systole. To evaluate the statistical significance of the results, the measurementswere pooled across all observers and a one-sided, two-sample Student t-test was performed.The test evaluated the hypothesis that the sample mean magnitude of the motion artifactwas different in the two groups, accounting for unequal sample sizes and assuming unequalvariance between the two groups (heteroscedastic).

3. Results

3.1. System delays and cardiac trigger timing

The delays associated with the pulse oximeter are displayed in figure 2. The delay betweenR-wave and midpoint of the plethysmogram was measured to be (347 ± 26) ms, whereasthe corresponding delay between R-wave and the digital trigger was measured to be tmean

trigger =(457 ± 25) ms. An offset of −50 ms correctly aligned the oximeter pulse wave to the ECGphase of maximum mechanical motion, such that the upstroke of the plethysmogram pulsecoincided with the ECG Q-wave (Jago and Murray 1988, Du et al 2005). The parametersof interest, tmin

trigger and tmaxtrigger correspond to the range of corrected mean volunteer delay

measurements and were computed to be 382 ms and 458 ms, respectively. The intra-subjectvariability in ttrigger ranged from 13 ms to 30 ms in the experimental cohort—well inside thewidth of the thresholds.

Example implemented delay curves are shown in figures 3(a)–(c) for various settingsof tmin

imager and trangeimager. The curves were computed using the two components of the trigger

delay (tmintrigger and tmax

trigger), described above. The buffer period, tbuffer, was fixed at 75 ms,a value larger than any exposure time used in the imaging protocol. For each curve,the implemented delay is zero below HRthresh, above which timp increases to a valuesufficient to delay exposure to the subsequent heartbeat and decreases gradually as HRincreases. The value of the threshold (equation (1)) is influenced by the sum of tmin

imager

Cardiac gating with a pulse oximeter for dual-energy imaging 6105

60 80 100 120 1400

200

400

600

800

Heart Rate (bpm)

Prototype Implementation

Idealms250min

imagert

ms135rangeimagert

ms0minimagert

ms0rangeimagert

(c)

Heart Rate (bpm)

60 80 100 120 1400

200

400

600

800

ms100rangeimagert

ms400minimagert

ms200minimagert

ms0minimagert

(a)=

=

=

=

=

===

Heart Rate (bpm)

Imple

mente

dD

ela

y(m

s)

60 80 100 120 1400

200

400

600

800

ms100minimagert

ms200rangeimagert

ms100rangeimagert

ms0rangeimagert

(b)=

=

=

=

Imple

mente

dD

ela

y(m

s)

Figure 3. Implemented delay curves computed for various system delay parameters. Curves in (a)and (b) are for hypothetical configurations involving (a) fixed t

rangeimager = 100 ms with three values

of tminimager, and (b) fixed tmin

imager = 100 ms with three values of trangeimager. Curves in (c) show timp for the

clinical prototype used in patient studies (tminimager = 250 ms, t

rangeimager = 135 ms) along with an ideal

imager with no measurable delays (tminimager = t

rangeimager = 0 ms).

and trangeimager. Figure 3(a) shows the implemented delay for a fixed delay range (t range

imager =100 ms) and three hypothetical values of tmin

imager (0, 200 and 400 ms). The HRthresh for thesethree examples are 95, 73 and 59 bpm, respectively. Figure 3(b) shows the implemented delayfor a fixed minimum imager delay (tmin

imager = 100 ms) and three hypothetical values of trangeimager

(0, 100 and 200 ms). As the imager delay range is increased, HRthresh decreases, with anassociated increase in the length of the implemented delay. For these examples, HRthresh =95, 82 and 73 bpm. Comparing figures 3(a) and (b), for the same maximum delay

(tmaximager

),

increased variability calls for larger implemented delays to account for the chance that eitherextreme of the delay range will occur.

For the prototype system configuration, the minimum delay from the imager was measuredto be tmin

imager = 250 ms, with a range of trangeimager = 135 ms, giving the implemented delay curve in

figure 3(c). The threshold lies at a heart rate of 65 bpm. Also shown in figure 3(c) is the delaycurve for the ideal system configuration absent of imager delays (tmin

imager = trangeimager = 0 ms). For

this hypothetical system, the HR threshold is 113 bpm, above which an implemented delay isrequired due to tmin

trigger and trangetrigger in combination with the shortened cardiac period.

6106 N A Shkumat et al

(a) (b)

Figure 4. Timing diagrams displaying the digital trigger, the diastole region and the time/durationof x-ray exposure as recorded by the monitoring system. (a) Timing for a patient with HR <

HRthresh, for which the exposure occurs within the same heart cycle (diastole region) as the digitaltrigger (arrow). (b) Timing for a patient with HR > HRthresh. In the latter case, the heart period istoo short to accommodate same-cycle acquisition, and the implemented delay forces exposure tothe central region of subsequent diastole.

3.2. Trigger performance evaluation

3.2.1. Timing relative to diastole window. For each patient image acquisition, a timingdiagram as seen in figure 4 was recorded by the monitoring system, showing the trigger pulsesoriginating from the oximeter in relation to the modeled diastole region and duration of thex-ray pulse. Figure 4(a) shows an example acquisition for which the patient HR was less thanHRthresh. An example with patient heart rate greater than the threshold is shown in figure 4(b),and the x-ray exposure correctly occurred within the diastole period of the subsequent cardiaccycle.

Figures 5(a)–(c) plot the time at which exposure occurred versus the true patient HR(i.e., the instantaneous HR) for the first 37 patients imaged on the research prototype. Theimplemented delays for the data in figure 5 are those for the clinical system as plotted infigure 3(c). The modeled diastole period is plotted as a function of HR as the region betweensymmetric gray bands above and below the x-axis. The space between the bands corresponds tothe diastole ‘window’, whereas the space outside the bands corresponds to systole. The widthof the bands represents the variability in the trigger delay, and is equal to t

rangetrigger. Each band is

separated into three curves corresponding to the diastole region specified by the correspondingminimum, mean and maximum trigger delay. Successful diastolic triggering correspondsto exposures between the upper and lower bands. Exposures within the gray bands may ormay not have been acquired in diastole, depending on variations in trigger delay. Exposuresabove or below the bands are failures, i.e., exposure during systole. Solid and open symbolscorrespond to low-kVp and high-kVp, respectively, with range bars representing the durationof the x-ray pulse.

Three methods of HR determination are also evaluated in figures 5(a)–(c). Figure 5(a)shows the coincidence of x-ray exposure and modeled diastole region for HR values reportedby the pulse oximeter (HRox). Although trigger accuracy is reasonably high (success rate =85%), the precision is low, and a number of systolic triggers are evident. These timing errorsare associated with deviation in HRox from the true value due to oximeter smoothing in caseswhere the true HR was changing significantly at the time of exposure. This also explains the

Cardiac gating with a pulse oximeter for dual-energy imaging 6107

40 50 60 70 80 90 100 110-750

-500

-250

0

250

500

750

40 50 60 70 80 90 100 110-750

-500

-250

0

250

500

750

40 50 60 70 80 90 100 110-750

-500

-250

0

250

500

750

Tim

e(m

s)

Heart Rate (bpm) Heart Rate (bpm)

Heart Rate (bpm)

Oximeter Heart Rate (HRox) Instantaneous Heart Rate

3-Beat Smoothing

(a) (b)

(c)

Low-kVp Exposure

High-kVp Exposure

Tim

e(m

s)

Diastole Window

Diastole Window

Diastole Window

Diastole Window

Diastole Window

Diastole Window

Figure 5. Incidence of x-ray exposure with respect to the diastole window as a function of patientHR. Data are for 37 patients participating in a DE imaging trial. (a) Triggering based on HR asreported by the oximeter gives reasonable accuracy, yet poor precision and several instances ofsystole gating due to errors in the HR estimate (HRox, subject to temporal smoothing). Triggeringbased on (b) the instantaneous HR or (c) smoothing HR by three previous heart cycles showsconsiderable improvement in accuracy (100%) and precision.

reduced precision and accuracy in low-kVp exposures (which are delivered first) compared tothose at high-kVp, due to the large initial drop in HR after breath hold.

The same data were re-evaluated using the same trigger and imager delays as figure 5(a),but with timp computed from the instantaneous HR. The results are shown in figure 5(b),showing a significant increase in both accuracy (100%) and precision. Also visible in thedistribution of data is the HRthresh occurring at 65 bpm. Exposures occurring slightly belowthe threshold tend to occur near the end of diastole (positive time points). At lower HR,exposures occur nearer the start of diastole (negative time points). This occurs because atHR < HRthresh, the trigger occurs immediately following the oximeter perfusion delay. AsHR decreases, the duration of diastole increases, and thus the exposure occurs proportionallycloser to the beginning of diastole.

Analysis of plethysmograms and resulting HR estimates indicates that the oximeter asinitially implemented in the DE imaging prototype applies a running average in HR across

6108 N A Shkumat et al

(a) (b)

(c) (d)

Figure 6. Dual-energy patient images. (a) and (b) Images acquired without cardiac gating.Cardiac motion artifacts are apparent on both sides of the heart (arrows), particularly in the bone-only image. (c) and (d) Image acquired with cardiac gating, demonstrating a marked reduction ofheart misregistration and motion artifact.

∼10 beats. The data were re-evaluated retrospectively with HR computed based on varioussmoothing windows to provide precision and accuracy comparable to that in figure 5(b), yetprovide robustness to spurious HR measurements. Performance was comparable for smoothingwindows of width 0–3 beats, beyond which accuracy steadily decreased to that shown infigure 5(a) for a window width of ∼10 beats. As shown in figure 5(c), a smoothing window of3 heart cycles was found to provide accuracy and precision comparable to that of instantaneousHR (figure 5(b))—providing a measure of robustness against spurious HR measurements,intermittent contact, sampling errors, etc., while still accounting for rapid changes in HR.

3.3. Image-based measurement of cardiac artifact

Soft-tissue and bone-only DE images acquired with the clinical prototype are shown infigure 6. Misalignment caused by gross patient motion is evident, manifesting as artifactsat the diaphragm and ribs. Images acquired with systolic triggering are shown in figures 6(a)and (b). In the soft-tissue image (figure 6(a)), the cardiac motion artifact is barely discernable.Misregistration is more evident in the bone-only image as dark or light bands located on bothsides of the heart (figure 6(b)). An example diastole gated DE image is shown in figures 6(c)and (d), for which there is little or no visible artifact in either image.

Cardiac gating with a pulse oximeter for dual-energy imaging 6109

0 2 4 6 8 10 12 140.00

0.05

0.10

0.15

0.20

0 2 4 6 8 10 12 140.00

0.05

0.10

0.15

0.20

yc

ne

uq

erF

yc

ne

uq

erF

Artifact Size (mm)

Artifact Size (mm)

mm80.2=artifactd

mm83.3=artifactd

Diastole Trigger

Systole Trigger

110Nsystole =

558Ndiastole =

0 2 4 6 8 10 12 140.00

0.05

0.10

0.15

0.20

0 2 4 6 8 10 12 140.00

0.05

0.10

0.15

0.20

yc

ne

uq

erF

yc

ne

uq

erF

Artifact Size (mm)

mm80.2=artifactd

mm83.3=artifactd

Diastole Trigger

Systole Trigger

110Nsystole =

558Ndiastole =

(a)

(b)

(c)

Figure 7. (a) Magnified region of a DE bone-only image in the region of the left lung. The exampleshows an image with significant cardiac artifact associated with improper (systole) triggering.The six black bars illustrate distance measurements made by each observer. (b) Histogram ofobserver measurements in cases for which both exposures were properly triggered during diastole.(c) Histogram of measurements for which at least one exposure was triggered during systole. Thep-value between the two data sets was p < 0.001.

Human observer measurements demonstrate a statistically significant reduction in themagnitude of the motion artifact. Figure 7(a) illustrates the location of six example observermeasurements in a DE bone image acquired using the clinical prototype. The image illustratesthe magnitude of cardiac motion artifact (typically 1–5 mm) and the degree of inter-observervariability, mitigated by the large number of measurements. A total of 666 total measurementswere acquired. Six measurements were excluded due to the obvious misinterpretation of theartifact by at least one observer—giving a total of 660 measurements: Ndiastole = 558 for thediastole group and Nsystole = 102 for the systole group. Histograms showing the pooled datafor each group are shown in figures 7(b) and (c). Mean artifact size (dartifact) for diastolic andsystolic triggers was 2.80 mm and 3.83 mm, respectively. The Student t-test p-value computedfrom the data pooled across all measurements was p < 0.001, demonstrating a statisticallysignificant difference between the two sample means. These results indicate a significantreduction of the cardiac motion artifact with diastole triggering of the gating system. Theresidual, nonzero artifact present even with perfect diastolic triggering is attributed primarily togross patient motion (slouch, breath-hold release) resulting from the fairly long inter-exposuretimes in this study.

4. Discussion and conclusions

This work described the initial development and evaluation of a cardiac gating systemfor DE imaging. Overall performance of the simple gating system was very promising.Triggering by means of a pulse oximeter, a simple cardiac model, and knowledge of system

6110 N A Shkumat et al

delays was shown capable of precise and accurate diastolic triggering. The improvementwas evident in significant reduction of image artifact in DE images as assessed by humanobservers. To the extent that such artifacts degrade diagnostic performance in the detectionof nodules (particularly in the middle and lower lungs), the reduction of such artifacts is ofconsiderable clinical importance. The minimization of cardiac motion artifacts is also expectedto contribute to the growing prevalence and utilization of DE chest imaging in the clinicalenvironment.

The performance of the cardiac gating system was evaluated in two independent studies.A quantitative analysis investigated the coincidence of x-ray exposure with the diastolewindow. The results demonstrated a trigger accuracy of 85% in the initial implementation,improving to 100% accuracy through temporal smoothing of the HR estimate. This representsa significant improvement over the accuracy expected in an un-gated system (∼30–52%,depending on patient HR). The second study measured the magnitude of cardiac artifacts inDE images to evaluate the improvement in image quality offered by the gating system. Themagnitude (i.e., width) of the artifact was measured independently in images acquired withdiastole and systole triggering. The study demonstrated a significant reduction in cardiacmotion artifact (p < 0.001) with the gating system. Future work will examine the clinicalsignificance of such image quality improvement in tests of diagnostic performance. Suchwill be investigated in a variety of human observer tests (e.g., preference tests, diagnosticsatisfaction tests and/or ROC) utilizing diastole- and systole-gated DE images with trainedradiologists as expert observers.

The prototype imaging system described in this work is not without its limitations, theforemost being the considerable time between exposures (∼6–8 s) required by the current FPDconfiguration. Such is being resolved through implementation of a faster FPD supporting sub-second double-exposure acquisition. Note that this limitation, however, does not significantlyaffect the functionality and performance of the cardiac gating system, which triggers x-rayexposures coincident with diastole regardless of the inter-exposure time. A limitation of usinga pulse oximeter as a cardiac monitoring device is the inter- and intra-patient variability in thepulse propagation delay. This variability is caused by extraneous factors aside from patient HR,including the level of inspiration and blood pressure. To account for this variability, the cardiacmodel incorporates a large range of propagation delays. Upper and lower bounds definingthe tolerable range of propagation delays (including processing delays on the oximeter) weredefined by tmin

trigger and tmaxtrigger, accounting for differences in physiological pulse propagation

times between patients as well as potential spontaneous changes in the delay within the samepatient. Thus, while the pulse oximeter does not provide a trigger specific to a phase timepoint, when combined with a cardiac timing model that is tolerant of variability in pulsepropagation time, it provides a simple measure of cardiac function sufficient for triggeringduring the diastole window.

Representing the cardiac cycle as a binary system is a very simple model for characterizingheart motion. In this work, the gating system was engineered to trigger x-ray exposures atmid-diastole, accepting small errors thereabout due to timing variability and counting anyexposure falling within the diastole window as a success. Identifying diastolic sub-regionswould require detailed physiological information that may be difficult to extract from theplethysmogram. This introduces the question as to what amount of cardiac motion can beexpected or tolerated. While the significant motion associated with ventricular ejection canbe avoided, the heart is never idle. While the triggering scheme detailed above was shownto successfully reduce artifacts associated with cardiac motion, it is interesting to note thatedge artifacts are never completely obviated in DE imaging—even for perfect anatomicalregistration. This was evident in the nonzero artifact size measured in the properly gated

Cardiac gating with a pulse oximeter for dual-energy imaging 6111

images. One source of such edge artifacts is the decomposition itself: as recognized in theliterature (Warp and Dobbins 2003), various noise reduction algorithms that differentiallyfilter the low- and high-kVp images can introduce edge artifacts in the resulting subtractionimage. A similar effect arising purely from the physics of x-ray interaction is associated withthe energy-dependent attenuation of various materials. For example, increased penetration forhigh-kVp beams can diminish the boundary of the heart compared to the associated low-kVpimage, resulting in an apparent darkening of the cardiac periphery (in soft-tissue images) orbrightening (in bone images). Of course, gross patient motion can contribute significantlyto motion artifacts. Anatomical misregistration of the clavicles (shoulder relaxation/slouch),ribs and diaphragm (slow respiratory rebound during breath hold) can cause large artifactsin DE images. Future research involves deformable registration of the low- and high-energyprojections to mitigate such artifacts (Dhanantwari et al 2007).

In summary, a model for cardiac gating that accounts for various sources of trigger andimager delays has been reported and implemented in a DE imaging prototype. Using asimple pulse oximeter as a surrogate for cardiac motion, the model triggers x-ray exposuressynchronous to diastole. The characterization of serial and parallel delays in the imagingsystem is essential. A triggering scheme exhibiting two timing regimes emerges, whereinabove a certain HRthresh, x-ray exposures are delayed to the subsequent diastole. Accuratemeasurement of HR is required to prevent erroneous triggering, particularly during rapidchanges in HR due to physiological effects, or errors involving the oximeter itself. Smoothingthe heart rate over three cycles was found to provide a fairly robust assessment of HR. Themodel described in this work is a simple yet effective means to improve DE image qualityby reducing cardiac motion artifacts. The ease of implementation and convenience of use inhigh-throughput imaging procedures present valuable logistical benefits.

Acknowledgments

The authors extend thanks to R Asento and M Haines (Carestream Health Inc., Rochester, NY)for assistance with implementation of the clinical prototype. N A Shkumat was supportedby scholarships from the Canadian Institutes of Health Research (CIHR), the University ofToronto and Ontario Student Opportunity Trust Fund (OSOTF). This research was conductedin collaboration with Carestream Health Inc. (Rochester, NY) and funded in part by NationalInstitutes of Health (NIH) grant no R01-CA112163–02.

References

Alvarez R E, Seibert J A and Thompson S K 2004 Comparison of dual energy detector system performance Med.Phys. 31 556–65

Desjardins B and Kazerooni E A 2004 ECG-gated cardiac CT AJR Am. J. Roentgenol. 182 993–1010Dhanantwari A C, Siewerdsen J H, Shkumant N A, Williams D B, Moseley D J, Paul N S, Yorkston J and

Van Metter R 2007 A multi-resolution, multi-scale, mutual information technique for registration of high-and low-KVp projections in dual-energy imaging Med. Phys. 34 2517

Diederich S and Wormanns D 2004 Impact of low-dose CT on lung cancer screening Lung Cancer 45 S13–9Difazio M C, Macmahon H, Xu X W, Tsai P, Shiraishi J, Armato III S G and Doi K 1997 Digital chest radiography:

effect of temporal subtraction images on detection accuracy Radiology 202 447–52Du J, Ning G, Li Y and Zheng X 2005 Arterial stiffness estimation in hypertension Conf. Proc. IEEE Eng. Med. Biol.

Soc. 5 5507–10Finn J P, Nael K, Deshpande V, Ratib O and Laub G 2006 Cardiac MR imaging: state of the technology

Radiology 241 338–54Fischbach F, Freund T, Rottgen R, Engert U, Felix R and Ricke J 2003 Dual-energy chest radiography with a flat-panel

digital detector: revealing calcified chest abnormalities AJR Am. J. Roentgenol. 181 1519–24

6112 N A Shkumat et al

Groves E M, Bireley W, Dill K, Carroll T J and Carr J C 2007 Quantitative analysis of ECG-gated high-resolutioncontrast-enhanced MR angiography of the thoracic aorta AJR Am. J. Roentgenol. 188 522–8

Guyton A C 1996 Textbook of Medical Physiology (Philadelphia, PA: Saunders)Henschke C I et al 1999 Early lung cancer action project: overall design and findings from baseline screening

Lancet 354 99–105Herzog C, Abolmaali N, Balzer J O, Baunach S, Ackermann H, Dogan S, Britten M B and Vogl T J 2002 Heart-

rate-adapted image reconstruction in multidetector-row cardiac CT: influence of physiological and technicalprerequisite on image quality Eur. Radiol. 12 2670–8

Jago J R and Murray A 1988 Repeatability of peripheral pulse measurements on ears, fingers and toes usingphotoelectric plethysmography Clin. Phys. Physiol. Meas. 9 319–30

Jubran A 1999 Pulse oximetry Crit. Care (Lond.) 3 R11–7Kelcz F, Zink F E, Peppler W W, Kruger D G, Ergun D L and Mistretta C A 1994 Conventional chest radiography

vs dual-energy computed radiography in the detection and characterization of pulmonary nodules AJR Am. J.Roentgenol. 162 271–8

Kopp A F, Schroeder S, Kuettner A, Heuschmid M, Georg C, Ohnesorge B, Kuzo R and Claussen C D 2001 Coronaryarteries: retrospectively ECG-gated multi-detector row CT angiography with selective optimization of the imagereconstruction window Radiology 221 683–8

Levick J R 2000 An Introduction to Cardiovascular Physiology (New York: Oxford University Press)MacRedmond R, McVey G, Lee M, Costello R W, Kenny D, Foley C and Logan P M 2006 Screening for lung cancer

using low dose CT scanning: results of 2 year follow up Thorax 61 54–6McLoud T C et al 1992 Bronchogenic carcinoma: analysis of staging in the mediastinum with ct by correlative lymph

node mapping and sampling Radiology 182 319–23Ohnesorge B, Flohr T, Becker C, Kopp A F, Schoepf U J, Baum U, Knez A, Klingenbeck-Regn K and Reiser M F

2000 Cardiac imaging by means of electrocardiographically gated multisection spiral CT: initial experienceRadiology 217 564–71

Pieterman R M et al 2000 Preoperative staging of non-small-cell lung cancer with positron-emission tomography N.Engl. J. Med. 343 254–61

Richard S and Siewerdsen J H 2007 Optimization of dual-energy imaging systems using generalized NEQ and imagingtask Med. Phys. 34 127–39

Ricke J, Fischbach F, Freund T, Teichgraber U, Hanninen E L, Rottgen R, Engert U, Eichstadt H and Felix R 2003Clinical results of CsI-detector-based dual-exposure dual energy in chest radiography Eur. Radiol. 13 2577–82

Sabol J M, Avinash G B, Nicolas F, Claus B, Zhao J and Dobbins J T 2001 The development and characterization ofa dual-energy subtraction imaging system for chest radiography based on CsI:Tl amorphous silicon flat-paneltechnology SPIE: Phys. Med. Imaging 4320 399–408

Samei E, Flynn M J and Eyler W R 1999 Detection of subtle lung nodules: relative influence of quantum and anatomicnoise on chest radiographs Radiology 213 727–34

Shkumat N A, Siewerdsen J H, Dhanantwari A C, Williams D B, Richard S, Paul N S, Yorkston J and Van Metter R2007a Optimization of image acquisition techniques for dual-energy imaging of the chest Med. Phys. 34 3904–15

Shkumat N A, Siewerdsen J H, Dhanantwari A C, Williams D B, Richard S, Tward D J, Paul N S, Yorkston J andVan Metter R 2007b Development and implementation of a high-performance, dual-energy imaging systemProc. SPIE Phys. Med. Imaging 6510 651006-1–12

Shkumat N A, Siewerdsen J H, Richard S, Paul N S, Yorkston J and Van Metter R 2008 Dual-energy imaging of thechest: optimization of image acquisition techniques for the ‘bone-only’ image Med. Phys. 35 629–32

Siewerdsen J H et al 2006 High-performance dual-energy imaging with a flat-panel detector: imaging physics fromblackboard to benchtop to bedside Proc. SPIE Phys. Med. Imaging 6142 61421E-1–10

Tagashira H, Arakawa K, Yoshimoto M, Mochizuki T and Murase K 2007 Detectability of lung nodules using flatpanel detector with dual energy subtraction by two shot method: evaluation by ROC method Eur. J. Radiol.64 279–84

Warp R J and Dobbins III J T 2003 Quantitative evaluation of noise reduction strategies in dual-energy imaging Med.Phys. 30 190–8

Xu T, Ducote J L, Wong J T and Molloi S 2006 Feasibility of real time dual-energy imaging based on a flat paneldetector for coronary artery calcium quantification Med. Phys. 33 1612–22