Clinical evaluation of three respiratory gating schemes for different respiratory patterns on cardiac SPECT

Abstract

Purpose:

Respiratory gating reduces respiratory blur in cardiac SPECT. It can be implemented as 3 gating schemes: i) equal amplitude-based gating (AG); ii) phase or time-based gating (TG); or iii) equal count-based gating (CG), i.e., a variant of amplitude-based method. The goal of this study is to evaluate the effectiveness of these respiratory gating methods for patients with different respiratory patterns in myocardial perfusion SPECT.

Methods:

We reviewed 1274 anonymized patient respiratory traces obtained via the Vicon motion-tracking system during their ^99mTc-sestamibi SPECT scans and grouped them into four breathing categories: i) regular respiration (RR); ii) periodic respiration (PR); iii) respiration with apnea (AR) and iv) un-classified respiration (UR). For each respiratory pattern, 15 patients were randomly selected and their list-mode data were rebinned using the 3 gating schemes. A preliminary reconstruction was performed for each gate with the heart region segmented and registered to a reference gate to estimate the respiratory motion. A final reconstruction incorporating respiratory motion correction was done to get a final image set. The estimated respiratory motion, the full-width-at-half-maxima (FWHM) measured across the image intensity profile of the left ventricle wall, as well as the normalized standard deviation measured in a uniform cuboid region of the thorax were analyzed.

Results:

There are 47.1%, 24.3%, 13.5% and 15.1% RR, PR, AR and UR patients respectively among the 1274 patients in this study. The differences among the 3 gating schemes in RR were smaller than other respiratory patterns. The AG and CG methods showed statistically larger motion estimation than TG particularly in the AR and PR patterns. Noise of AG varied more in different gates especially for AR and UR patterns.

Conclusion:

More than half of the patients reviewed exhibited non-regular breathing patterns. Amplitude-based gating, i.e., AG and CG, is a preferred gating method for such patterns and is a robust respiratory gating implementation method given the respiratory pattern of the patients is unknown before data acquisition. Phase gating is also a feasible option for regular respiratory pattern.

Introduction

Single photon emission computed tomography (SPECT) is currently one of the most established methods for non-invasive myocardial perfusion imaging to assess the functional integrity of the heart muscle and evaluate the heart’s condition [1–3]. Respiratory motion, which appears especially in the cranial-caudal direction [4] and induces motion blurring in the cardiac images [5], resulting in potential misdiagnosis of heart diseases [6].

Respiratory gating has been proposed as a feasible solution for reducing the respiratory motion blur on cardiac SPECT [7, 8]. It was commonly implemented with the assistance of an external respiratory motion tracking device. Sometimes the data are acquired only at specific phase in the respiratory cycle [9, 10], e.g., end-expiration (End-EX) or end-inspiration (End-IN). Such usage of acquiring data only during a portion of the respiratory cycle decreases the impact of motion but results in either elevated noise levels or significantly extended acquisitions. Alternatively, list-mode data are binned into respiratory gates based on the externally tracked motion signal as a surrogate for the relative location of the internal anatomy [11, 12]. The motion between the gates is then estimated by registration of the anatomy in reconstructed slices of these gates and used in correction during a second pass of reconstruction [13]. The first way of the external surrogate signal employed in forming gates is based on the relative temporal position between consecutive End-EX or End-IN time points or “phase”, i.e., time-based or phase gating (TG). In this scheme the gates are formed by dividing time interval between each pair of signal extrema into a set number of equal intervals. The assumption is that the location of the internal structures in these gates is in common throughout the acquisition, which has been shown to be frequently invalid [14–16]. The second way employed in forming gates is based on the external surrogate signal’s magnitude or amplitude, which has been shown in many cases to correspond better with the internal location of the anatomy than phase does [12, 14–16]. Amplitude-based binning means forming gates by combining data during respiration with similar magnitude values, usually obtained based on external marker positions. Two variants of amplitude-based binning have been employed. The first we will call equal amplitude-based gating (AG) wherein after removing amplitude outliers, signal is divided into bins with equal motion range, i.e., difference of motion amplitude. The second we will call equal count-based gating (CG) where the width of the bins from AG is slightly adjusted such that each bin possesses equal counts. We previously evaluated the performance of these different gating schemes using a population of ten XCAT phantoms and ten patients and concluded that the amplitude-based methods (AG and CG) demonstrated better performance than TG [17].

Data-driven based respiratory gating methods without the external tracking device have also been proposed to directly extract the respiratory motion signal from the collected SPECT data [18, 19]. Although data acquisition is more straight forward for this method, the increased data processing time and the feasibility with only certain radiopharmaceuticals or detector geometries hinder the clinical application of the data-driven method in SPECT.

Respiratory pattern is considered as one of the most important patient-specific characteristics to affect the imaging improvements obtained with respiratory gating [20]. Liu et al. [21] investigated the effect of respiratory patterns on tumor quantification and delineation in ungated PET/CT imaging and reported that patient respiratory traces with relatively larger quiescent period fraction yield results less subject to respiratory motion than traces with long-term amplitude variability. Park et al. [22] reported respiratory gated PET could recover most of the image blurring in the presence of motion while some slight image degradation was noticed around the phases which corresponded to the highest velocity within the respiratory motion cycle. In the past, respiratory features such as relative timing of expiration and inspiration, cycle amplitude and period were commonly used in the respiratory pattern analysis [23–25]. Wang et al. [20] established a framework to use six main respiratory features extracted from a respiratory signal to predict the potential quantitative improvement when using respiratory gating PET based on a multiple-regression model. Lu et al. [26] investigated the impact of varying motion amplitude and baseline shift of the respiratory trace and proposed a dynamic internal-external motion correlation method to achieve data-driven motion correction. Moreover, the respiratory patterns could vary tremendously among different people and even the same person. Besides the regular respiration, there are also different kinds of irregular respiratory patterns such as Kussmaul’s respiration, Ataxic respiration, Apneustic respiration, Cheyne-stokes respiration, Biot’s respiration and periodic respiration [27, 28]. While researchers mostly investigated the effect of respiratory motion on respiratory gated PET, to the best of our knowledge, the effect of the respiratory pattern for respiratory gated SPECT is still not well examined.

In this study, we reviewed 1274 patient respiratory traces for further respiratory pattern classification and selected 60 patients to evaluate the effect of different respiratory patterns for the choice of the three different respiratory gating schemes (TG, AG and CG) on clinical cardiac SPECT.

Materials and Methods

A. Respiratory pattern classification

We reviewed 1274 anonymized patient respiratory traces acquired using five near-infrared cameras from the Vicon Motion System (Lake Forest, CA) during routine ^99mTc-sestamibi stress SPECT/CT studies at University of Massachusetts Medical School. Written consents were collected from all patients and the study was approved by the local institutional review board. For motion tracking, seven retroreflective markers were attached to the chest, right rib, and abdomen of the patient (Figure 1) while the five near-infrared cameras were mounted on the walls, acquiring motion from the markers at 30 Hz [29]. Among the 7 markers, 5 markers placed in different locations on the chest and together with the 1 placed on the right rib were used to track the rigid-body motion of the patient by a process that first removes the “periodic” respiratory signal and then estimates body-motion from the assembly of markers [30, 31]. The vertical motion of the single marker on the abdomen has almost always been found to be larger than the respiratory motion component of the other markers, thus was used as the surrogate respiratory motion signal in this study [17, 29]. By visually inspecting all the respiratory traces we grouped them into four representative categories: (1) regular respiration (RR); (2) periodic respiration (PR); (3) respiration with apnea (AR) and (4) un-classified respiration (UR) based on the literature [27, 28]. Histograms of the abdominal marker positions for different types of respiratory traces were generated to verify our classification results.

Figure 1.

Positions of the seven markers of the Vicon Motion System.

B. Three respiratory gating schemes

We evaluated the effect of different respiratory patterns on the choice of three different respiratory gating schemes: AG, CG and TG (Figure 2). Three respiratory cycles were shown as examples. The horizontal lines for AG and CG, vertical lines for TG represent the gating boundaries respectively. As mentioned in our previous study [17], the difference in the three schemes was how to rebin the data with different criteria:

Figure 2.

The three different gating schemes used in this study.

AG: After removal of outliers the acquired data were divided into equal bins, e.g., 7 in this study, based on the tracked maximum and minimum marker positions from all respiratory cycles. Thus, the motion range measured by the marker positions, is the same in each bin while photon counts may vary among the bins (Figure 2a).

CG: A form of amplitude-binning where outliers were not removed and each bin was adjusted to have the same photon count level over the acquisition allowing the motion range measured by the marker positions to possibly vary across the bins (Figure 2b).

TG: Each respiratory cycle defined as the time interval from End-EX to End-EX based on the tracked motion data were divided into 7 bins based on the temporal duration of that cycle. Thus, the temporal bin width is the same for the different bins in the same respiratory cycle, but can vary for different respiratory cycles (Figure 2c). This results in the same count level in each bin but the motion range measured by the marker positions varies across the bins.

C. Data acquisition

From the clinical acquisitions, 15 representative patients were randomly chosen for each respiratory pattern category, resulting in a total 60 patients (age: 33–81; body mass index (BMI): 21.7–43.4) for our study. All of these patients were imaged on a dual-headed SPECT/CT scanner (BrightView SPECT/XCT, Philips Healthcare, Cleveland OH) at least 40 mins post injection of ~1100 to 1332 MBq ^99mTc-sestamibi, according to their BMI. For each patient, a standard stress cardiac SPECT step-and-shoot protocol was used to acquire 64 projections over 180° from right anterior oblique to left posterior oblique with the two cameras positioned 90° from each other. List-mode data were also simultaneously acquired. The list-mode data were used to form 128×128 projections with a pixel size of 0.466 cm. The duration of SPECT data acquisition was ~10.67 mins with 20 s per step and 32 steps. The total SPECT scan time was ~16 mins, with the addition of the detector rotation. We set the primary energy window at 140.5 keV with a 15% width and the scatter window was at 121 keV with a 4% width. List-mode SPECT data were synchronized with the Vicon motion-tracking data based on the table motion as recorded in the list-mode file. A low-dose cone-beam CT (5 mAs, 120 kVp) was acquired prior to the SPECT acquisition and used for attenuation correction.

D. Data processing

List-mode SPECT data were binned into 100 ms frames from which 7 respiratory bins were formed using the three different gating schemes, i.e., AG, CG and TG respectively, according to the Vicon tracked abdominal marker signal. For the CG and TG schemes, we used all acquired counts for data rebinning while for the AG scheme, 3% and 2% outliers were removed from End-EX and End-IN sides of the marker position histogram respectively for re-binning to avoid errors caused by the very high motion amplitude in few respiratory cycles. The counts were then added back to the extreme gates during the final reconstruction to utilize all counts. A preliminary maximum-likelihood expectation-maximization (ML-EM) reconstruction was performed for all respiratory gates with 24 iterations, including resolution, attenuation and scatter compensations [32]. An ellipsoidal volume-of-interest (VOI) was drawn on the reconstructed images to extract the heart, and a six-degree-of-freedom (6-DOF) rigid-body registration containing three rotational and three translational components was performed automatically using the minimization of the sum-squared-difference (SSD) as the registration metric to determine the heart motion deformation parameters between each gate with the 4^th referenced gate [13]. The SSD is defined as:

$$SSD=\sum _n\left(I_M\left(x\text{,}y\text{,}z\right)-I_F\left(x\text{,}y\text{,}z\right)\right){\mathrm{ }}^2$$

where I_M and I_F are the intensities of the fixed images, i.e., the 4^th gate, and the moving images, i.e., the other 6 gates respectively while (x, y, z) indicates the voxel index and n is the total voxel number. A final ordered-subset expectation-maximization (OS-EM) reconstruction was performed to utilize all counts with 5 iterations and 16 subsets and correct for resolution, attenuation, scatter and respiratory motion using the previously obtained 6-DOF deformation parameters [13]. Post-reconstruction filtering was not applied on the reconstructed images.

E. Data analysis

For different gating schemes of each respiratory pattern, we measured the related motion position in axial direction between each respiratory gate to the reference gate, i.e., the 4^th gate using rigid-body registration. The difference between the highest and the lowest motion position in the axial direction (largest component of respiratory motion [32]) was measured to represent the estimated respiratory motion (ERM). For the 15 patients in each of the four respiratory patterns, we averaged their ERM values as an index to compare the average motion detected by the different gating schemes.

A 3-dimensional cuboid volume-of-interest (VOI) was drawn in a low count uniform region of the right thorax of the reconstructed images (Figure 3) in all respiratory gates to calculate the noise index (NI).

$$NI=\frac{SD}{mean}=\frac{\sqrt{\frac{1}{l\mathrm{ }-\mathrm{ }1}{\sum }_{j=1}^l(x_j-p){\mathrm{ }}^2}}{p}\mathrm{ }\mathrm{ }\mathrm{ }$$

where l is the voxel number, i.e., 10×11×9=990, p is the mean value of the voxels within VOI, x is the voxel count value in the VOI, and j is the voxel index.

Figure 3.

The VOI in the right torso region from (a) axial, (b) coronal and (c) sagittal views for noise analysis.

Furthermore, an image profile was drawn across the left ventricle wall on the coronal view (Figure 4) and full-width-at-half-maxima (FWHM) of the profile of the anterior and inferior ventricular walls were calculated with Gaussian fit to track the extent of respiratory blurring in the axial direction. The FWHM values were averaged and compared over all patients for a given respiratory pattern and gating scheme. Results of the ungated images are also measured as reference.

Figure 4.

Sample final coronal reconstructed images after motion correction using (a) AG, (b) CG and (c) TG gating schemes for a selected patient. The red arrows show the position of the image profiles which are shown in (d).

We performed statistical analysis using a paired t-test with Bonferroni correction on the ERM and FWHM results to evaluate the difference of the 3 gating schemes for each respiratory pattern using the Statistical Package for the Social Sciences software (SPSS, IBM Corporation, USA). The variation of the NI value among different gates were analyzed using the box-plot.

Results

Four sample respiratory traces for the vertical position of the abdominal marker are shown in Figure 5. The corresponding histograms of the vertical position of the marker above the minimum location and relative to the maximum position attained being shown as a value of 1.0 are plotted in Figure 6 to further demonstrate the differences among the respiratory patterns [21]. In this figure the extent of inspiration increases towards the right of the histograms. The RR shows a relatively consistent motion amplitude and respiratory period among respiratory cycles. The peak relatively closer to low extents of inspiration in Figure 6 is caused by the fact that patients usually spend more time near End-EX during the respiration. The PR demonstrates a periodically gradual increase and decrease of the respiratory motion amplitude during respiration, resulting a relatively stronger peak in the histogram. The AR means a long quiescent period usually over 10 s frequently observed after the End-EX during the respiration, causing a tremendously abrupt peak of the histogram close to End-EX [33]. The respiratory pattern of UR exhibits no recognizable shape and the histograms spread out with no particular character.

Figure 5.

Sampled respiratory traces of RR, PR, AR and UR obtained from the anterior / posterior (vertical) motion of the marker on the surface of the abdomen.

Figure 6.

Corresponding histogram plots of the four respiratory patterns in Figure 5 for the relative vertical motion of the abdominal marker normalized to the largest extent being 1.0, which corresponds to the point of largest patient inspiration during acquisition.

After inspecting the 1274 respiratory traces and classifying them into four groups, the number of patients in each group is listed in Table 1. The RR pattern takes around half of the population, followed by PR, UR and AR. AR and UR patterns have similar numbers of patients, about 15% of the population for each.

Table 1.

Statistics of patients under the classified respiratory patterns

Respiratory Pattern	Number of Patients	Percentage
RR	600	47.1%
PR	309	24.3%
AR	172	13.5%
UR	193	15.1%

Figure 7 shows sample reconstructed short-axis slices for a patient with PR breathing pattern using different gating schemes, as well as the ungated reconstructed image. The AG scheme shows obvious non-uniform noise distribution with Gate #7 having a higher noise level. Slightly larger heart motion blur could be noticed from images of TG scheme, e.g., arrow in Gate #7. Ungated short-axis images show a significant inferior decrease and the suggestion of an anterior-septal decrease while only the anterior-septal decrease could be noticed from the final images of the three gating schemes, which would be consistent with a decrease in respiratory motion blurring in the inferior wall [32]. Final images from all the three respiratory gating schemes showed less heart wall motion blur than the ungated image.

Figure 7.

Sample reconstructed short-axis heart images for a patient with PR breathing pattern using different gating schemes and the corresponding ungated image. Note the update number for the individual gate is 24 while for the final and ungated images is 80. Also note that all three gating schemes alleviate the apparent decreased inferior uptake seen in the ungated image. White arrow indicates a substantial motion blur.

Figure 8 shows the averaged ERM in the axial direction for different breathing patterns and gating schemes. Compared to TG, the detected maximum motion range in AG and CG was 37±29% and 27±22% larger for the RR pattern; 428±272% and 345±197% larger for the AR pattern; 184±117% and 143±98% larger for the PR pattern; 126±90% and 82±49% larger for the UR pattern.

Figure 8.

Estimated respiratory motion in axial direction from different gating schemes and respiratory patterns.

The FWHM values are shown in Figure 9 for both anterior heart wall and inferior wall. The FWHM values are significantly different between AG and TG for the anterior wall with RR, PR and UR patterns and the inferior wall with PR. CG shows significantly lower FWHM value for AR and PR in both anterior and inferior walls as compared to TG. There is no statistically significant difference between AG and CG for all breathing patterns. Ungated images generally show significantly larger FWHM than AG and CG, yet its difference as compared to TG is not significant in AR on the anterior wall and in AR, PR, and UR on the inferior wall.

Figure 9.

Averaged FWHM for (a) anterior and (b) inferior heart wall from different gating schemes and respiratory patterns.

Box-plots in Figure 10 show the NI distribution through all gates for different gating schemes and the differences are smaller among the schemes for RR. TG had the smallest interquartile range (i.e., the box height) generally, followed by CG, while AG method showed largest interquartile range, especially in the AR and UR patterns.

Figure 10.

Box-plots show the NI distribution through 7 respiratory gates of three gating schemes under each respiratory pattern.

Discussion

For both the estimated motion and the FWHM results, statistically significant differences can be noticed among the three different respiratory gating schemes for all respiratory patterns, but their differences were generally larger for AR, PR and UR. For the ERM results (Figure 8), the AG and CG showed significantly larger motion estimation than TG in all respiratory patterns. The ERM of AG was significantly larger than CG in every set of patient studies except for the RR ones. For the FWHM results (Figure 9), AG and CG generally showed better performance than TG while the FWHM values of CG and AG are not statistically different. The motion blur reduction performance of the three different gating schemes for UR was a bit lower than those from other patterns. In terms of noise (Figure 10), AG varied more across different gates especially for AR and UR.

Although the AG scheme showed the largest ERM results for all the four different patterns, we noticed that in 3 AR patients the motions estimated in End-IN gate or the gate close to End-IN differ substantially from the values obtained by other gating schemes, which could be attributed to the low counts in the extreme inspiration gate as the long period of apnea region exacerbates End-IN gate spanned an even shorter time. We also noticed that without the re-use of the 5% spuriously counts, AG would have worse performance as only 2–3 gates could possess enough counts to be reconstructed. Based on the results of our study, amplitude-based gating schemes showed a better performance on ERM and FWHM results, with the CG scheme providing a more uniform noise distribution. These results were consistent with our previous study [17] with further indications that CG is a robust respiratory gating implementation method when the respiratory pattern of the patients is unknown before data acquisition.

The limited angle effect would further affect the image quality for respiratory gating as each gate might not possess enough angles for a stable image reconstruction [13]. In our study, we found more missing projection data for AR and PR patterns due to a longer End-EX period, leading to more missing data phases other than End-EX. To alleviate the limited angle effect, we used ML-EM instead of OS-EM for the first round of image reconstruction on motion estimation, as further dividing subset of the angles in OS-EM would aggravate the amount of missing data.

The main limitation of our study is the categorization of the four respiratory patterns based on visual assessment and literature. Several respiratory patterns with apnea, e.g., Cheyne-stokes respiration and Biot’s respiration were grouped as the AR pattern. Respirations with no distinct features, with baseline shifting or multiple breathing patterns appearing in one patient were grouped into the UR pattern in this study. A more refined classification scheme could be achieved with the assistance of specialists in respiratory medicine or by using an automatic machine learning method [34, 35], where data pre-processing could be of importance to improve the accuracy for automatic classification. The noise assessed in this study is based on a uniform region in the thorax instead of myocardium as drawing a uniform region with reasonable number of voxels and without covering a substantial defect is challenging for all patients. This index may not be representative for myocardium, as noise in iterative algorithms tends to be local rather than propagating more broadly throughout the image. However, we found that for some selected normal patients the noise assessed in both regions is highly correlated. Besides, the ERM was estimated only in the axial direction of the patient, thus it may not fully represent the maximum motion range of the patient if significant motion is observed in the transaxial direction. However, respiratory motion is generally more pronounced in the axial direction [4, 32], thus this direction is more of a concern.

Conclusion

More than half of the patients reviewed exhibited non-regular breathing patterns. The difference for three gating schemes is smaller with RR. Amplitude-based gating schemes, i.e., AG and CG show better motion estimation and wall de-blurring results than TG, especially in AR and PR. The noise in each gate of AG varies more than CG and TG, especially for AR and UR. We conclude that amplitude-based gating method, i.e., AG and CG, is a robust respiratory gating implementation when the respiratory pattern of the patients is unknown before data acquisition. Phase gating is also a feasible option for regular respiratory pattern.