A Novel High-Sensitivity Rapid-Acquisition Single-Photon Cardiac Imaging Camera

The D-SPECT Camera

The novel D-SPECT camera technology uses 9 collimated detector columns arranged in a curved configuration to conform to the shape of the left side of the patient's chest (Fig. 1). The tungsten parallel-hole collimators used by Spectrum Dynamics are shorter (21.7 mm) and have larger square holes (2.26 mm) than do the standard lead parallel-hole collimators (45-mm length and 1.6-mm holes for a low-energy high-resolution [LEHR] collimator) used for conventional SPECT. The result is an acceptance solid angle (10.847 × 10⁻³ sr) more than 8 times that of the standard LEHR lead parallel-hole SPECT collimator (1.264 × 10⁻³ sr). An array of cadmium zinc telluride (CZT) crystals is aligned behind each collimator column (width, 16 pixels [40 mm]; length, 64 pixels [160 mm]). The square collimator holes are in registration with the crystal detector array, with 1 CZT pixel for each hole (Fig. 1). Each collimated detector column rotates and translates (a maximum of 110°) independently, either stepwise or continuously, allowing the object of interest to be viewed from hundreds of different viewing angles (Supplemental Video 1; supplemental materials are available online only at http://jnm.snmjournals.org). This sampling strategy leads to a complete dataset as determined by a sinogram (Supplemental Fig. 1). As currently configured, the columns can go through their full rotation, enabling dynamic SPECT acquisitions with a 10-s temporal resolution. A time marker is input every 10 ms, and the data are acquired in list mode using a standard personal computer (Pentium IV; Intel) (3-GHz clock rate, 2-GB memory). Physiologic markers, such as the electrocardiogram R wave, can also be routinely recorded with the system.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 1. 

D-SPECT cardiac system and detector-head configuration. (Left) D-SPECT cardiac scanner with seated patient. (Middle) Each detector column is composed of CZT sensor (39 × 39 × 5 mm) with four 16 × 16 pixel detectors and tungsten collimator with 0.2-mm septa and square opening (pitch, 2.46 mm; length, 21.7 mm). (Right) Nine detector columns, each capable of rotation and translation, are used to scan the myocardium.

Image Reconstruction

The acquired data for the D-SPECT camera are reconstructed using an iterative algorithm based on the maximum-likelihood expectation maximization (MLEM) method. The probability of a γ-ray's having been emitted from coordinates (x, y, z) relative to the detector is modeled, creating a probability map used by the reconstruction algorithm. The characteristics of the collimator are included in the probability map estimation (16–20), but attenuation and scatter are not included.

Although MLEM iteration is guaranteed to improve the likelihood of the estimate, convergence to a maximum-likelihood solution may be slow. To accelerate the convergence of the MLEM algorithm, we have applied an ordered-subset expectation maximization (OSEM) (20) approach. The subsets are chosen on the basis of the structure of the D-SPECT camera and on the scanning scheme, so that each voxel has a similar influence on each of the subsets. The order of the computation is chosen so as to minimize the dependency between successive readings.

Likelihood maximization is combined with a technique that has been developed specifically for the cardiac application and that globally smoothes intensity levels over the voxels corresponding to the left ventricular (LV) walls at an intermediate stage of the reconstruction. These voxels are located in the reconstructed volume after several OSEM iterations by searching radially from the geometric center of the LV cavity, taking the points of maximal intensity, and fitting a smooth ellipsoid-like surface to these points. Further iterations are then performed. During reconstruction, spatial smoothing is applied every n iterations to the current partial result. Smoothing is performed by replacing the value of each voxel value V₀ by a value V₁ that is computed as follows: V₁ = w × N + (1 − w) × V₀, where N is the 26-neighborhood average of the voxel and w is the weight between 0 and 1. A mild post-reconstruction filter is also applied to suppress hot spots by detecting and flattening local maxima.

Spatial resolution effects due to the collimation are referred to as geometric or extrinsic. The intrinsic resolution is determined by the CZT detector size. Accurate modeling of the probability function, supported by an increased number of viewing angles, overcomes the loss of geometric spatial resolution inherent with wider-bore collimation and recovers resolution. For a standard SPECT camera with one or more large planar detectors positioned at a given angle, all the pixels at a given head position collect photons during the same time interval. The same collection time is given to pixels that denote informative regions (such as at the left ventricle) as is given to pixels that denote less significant regions (such as a body contour, at the edge of the detector). The design of the D-SPECT camera mechanics allows us to overcome this inherent SPECT limitation. By choosing to spend more time directing the independent small detector heads of the D-SPECT camera toward regions of interest, one can allocate more time to collecting data from these regions at the expense of collecting fewer data from less important regions. This design is equivalent to having a single detector head, with each of its pixels having a different time for data collection.

To help the reader better understand this idea, we introduce the terms equivalent projection head and virtual detector head. An equivalent projection is a 2-dimensional image generated by averaging onto a single plane the projection images from multiple D-SPECT detector heads sharing the same viewing angle. This plane can be referred to as a virtual single-head detector. Because D-SPECT detectors may have different dwell times for the same viewing angle, each pixel value on this virtual detector is obtained by averaging a different number of projections. The equivalent projection (Fig. 2A) is obtained by dividing the sum of D-SPECT projections (Fig. 2B) by the dwell times (Fig. 2C). Figure 2C shows that pixels looking at the left ventricle average 4–8 projections, pixels looking near the left ventricle average 2–3 projections, and pixels far from the left ventricle are not observed (Fig. 2C). The D-SPECT algorithm has been further refined for optimized reconstruction of clinical data by incorporating a prior specifically for this application.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 2. 

A 2-dimensional image generated by averaging projection images from multiple D-SPECT detector heads, sharing same viewing angle, onto single plane from virtual single-head detector. Equivalent projection (A) is obtained by dividing sum of D-SPECT projections (B) by dwell times (C).

Acquisition Protocol

National Electrical Manufacturers Association (NEMA) phantom and patient feasibility studies were performed, comparing the high-speed D-SPECT camera technology with that of a typical dual-detector Anger camera (Millennium VG dual-head γ-camera; GE Healthcare) equipped with a LEHR (model VPC-45; GE Healthcare) collimator. The usable field of view (UFOV) was mapped for the D-SPECT camera. For the conventional-SPECT phantom studies, the dual-head Anger camera was positioned in the L-mode with a 100-mm radius of rotation; a 180° scan was obtained, with an angular step of 3°, mimicking clinical cardiac SPECT acquisition. For the D-SPECT phantom studies, each detector was rotated through an arc ranging from 60° to 110°, depending on the position of the individual detector and the scanned object. The energy window for each camera was centered around 122 keV ± 10% for the ⁵⁷Co for phantom studies, 140 keV ± 10% for the ^99mTc patient studies, and 70 keV ± 15% for the ²⁰¹Tl portion of the dual-radionuclide studies.

Line-Source Studies

A 1.48 × 10⁸ Bq ⁵⁷Co line source (diameter, 1 mm; length, 190 mm) demonstrated the inherent spatial resolution and sensitivities of the 2 systems. With each system, 14.3 × 10³ counts were obtained. The time of acquisition was measured, and the resolution was measured as the full width at half maximum (FWHM) and full width at tenth maximum (FWTM).

A standard NEMA phantom with three ^99mTc line sources (each source: diameter, 1 mm; length, 20 cm) demonstrated the extrinsic spatial resolution of the 2 systems. For both systems—D-SPECT and SPECT—scans were obtained under conditions optimal for cardiac acquisition (Supplemental Fig. 2). On the basis of the collimator design, the expected geometric resolution (FWHM) for the D-SPECT collimator is more than a factor of 2 worse than that of the LEHR collimator.

Standard NEMA protocols define a reconstruction method based on filtered backprojection, with a ramp filter that is clearly inappropriate for the D-SPECT camera (for which an iterative reconstruction algorithm is intrinsic to the system). Therefore, conventional-SPECT data were reconstructed with 20-iteration OSEM using 1.73-mm cubic voxels (matrix size, 256 × 256) with no resolution recovery. The D-SPECT data reconstruction also used 20 iterations of OSEM, but with 2.46-mm cubic voxels and with resolution recovery obtained by the inclusion of collimator characteristics in the imaging model. With each system, time of acquisition and spatial resolution (FWHM, FWTM) were measured.

The mapping of the resolution throughout a UFOV of the D-SPECT camera was performed by imaging a ⁵⁷Co 9-rod grid placed in the UFOV (Supplemental Fig. 3). Each ⁵⁷Co rod was 1 mm in diameter, with an activity of 0.74 × 10⁸ Bq. The acquisition protocol entailed 60 positions and 2 s per position. The total acquisition time was 120 s. The experiment was performed for the D-SPECT camera only, to demonstrate the stability and degradation of the resolution as a function of distance from the detectors.

Reconstructed Spatial Resolution Estimation

Resolution measurement using iterative reconstruction for line-source data tends to provide a poor estimate of the effective resolution in clinical practice; thus, we also used an alternative approach based on perturbation. A small change in the image distribution (i.e., the perturbation) provides a measure of the resolution determined by the underlying activity distribution, without significantly affecting the result.

A simplified model of the left ventricle, located in a thoracic cavity similar to that used in the NCAT (NURBS-based cardiac thoracic) phantom, was simulated (21,22). This model was chosen as it resembled a physical phantom that could be used to perform similar estimation. Radioactivity was introduced to the ventricular wall, liver, and background regions (excluding lung), with relative concentrations of 6.5:2.5:0.2. Data were projected to simulate the acquisition geometry for both SPECT and D-SPECT cameras. In addition, a pair of radioactive wires, located in the center of the ventricular wall and extending around the short axis and long axis of the ventricle, was simulated (Fig. 3).

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 3. 

Simulated resolution estimation model. Simplified model of left ventricle with active wires was simulated (red arrow), located in the center of the ventricular wall and extending around the short and long axes of the ventricle.

The activity concentration in the wires was scaled to 10% of the activity concentration in the ventricular wall, and a second set of projections was produced incorporating the active wires. Attenuation and scatter were not modeled. Supplemental Figure 2 illustrates the acquisition geometry. Both datasets were reconstructed using OSEM. For SPECT, the reconstruction was performed using 15 subsets, both with and without collimator modeling, using 2 and 54 iterations, respectively. For the D-SPECT camera, the collimator was modeled and reconstruction involved 32 subsets with 20 iterations. In the conventional-SPECT system, the geometric resolution was modeled as a gaussian function whose width varies linearly with the distance from the detector. The system resolution is then based on the geometric and intrinsic components added in quadrature (11). To evaluate resolution, attenuation and scatter were not modeled. The same assumptions were made for the D-SPECT camera simulations. In each case, the reconstructions with and without the active wire were subtracted to provide images that could yield measurements of line width. These measurements were obtained by first reorienting the heart so that the long axis was vertically aligned; FWHM and FWTM were estimated radially (across the width of the ventricular wall) and tangentially (orthogonal to radial) and averaged for the complete wire length.

Torso Phantom Studies

An anthropomorphic torso phantom (model ECT/TOR/P; Data Spectrum Corp.) was used for the phantom studies (Supplemental Fig. 4). The phantom, a large, body-shaped cylinder, comprised the lungs (filled with Styrofoam beads [The Dow Chemical Co.]), liver, spine inserts, and cardiac inserts (model ECT/CAR/I), simulating the upper torso of average-to-large patients (38 × 26 cm). The torso phantom was filled with ^99mTc with the following activity distributions: cardiac insert, 185 kBq/mL (5 μCi/mL); liver, 92.5 kBq/mL (2.5 μCi/mL); and background and left ventricle cavity, 7.4 kBq/mL (0.2 μCi/mL), resulting in the following signal-to-background ratios: heart and background (body), 25; heart and liver, 2; and liver and background, 12.5. Cardiac phantom studies were performed with homogeneous activity in the left ventricle and the following defect sizes, activities, and locations: 0.5-cm defect, 0% activity, anterior wall; 1.0-cm defect, 0% activity, anterior and inferior walls; and 2.0-cm defect, 50% activity, inferior wall.

In a separate experiment, the same torso phantom, with a 2-cm (100%) defect in the anterior wall, was filled with ^99mTc and ²⁰¹Tl with the following activity distributions: for single-isotope thallium imaging: left ventricle, 5.0 MBq of ²⁰¹Tl; liver, 18.9 MBq of ²⁰¹Tl; and background, 37 MBq of ²⁰¹Tl; and for the simultaneous dual-radionuclide protocol (^99mTc was added to ²⁰¹Tl): left ventricle, 5.9 MBq of ^99mTc; liver, 37 MBq of ^99mTc; and background, 23.7 MBq of ^99mTc. The acquisition protocol entailed 120 positions and 3 s per position. The total acquisition time was 360 s.

Patient Studies

Ethics committee approval for the clinical investigation was obtained from the Procardia Center in Tel Aviv before volunteers were imaged. A total of 18 patients (15 men, 13 with known coronary artery disease) underwent one-day ^99mTc-sestamibi stress/rest SPECT. D-SPECT images were obtained within 30 min after SPECT. Stress and rest acquisition times were 19 and 11 min, respectively, for SPECT and 4 and 2 min, respectively, for D-SPECT. Images were visually analyzed using the 20-segment model to calculate summed stress scores (SSS) and summed rest scores (SRS) and an image quality scale of 1–5 (1, poor; 5, excellent); images were assessed for confidence of interpretation by 4 masked nuclear medicine physicians. The images were also read by 3 independent nuclear medicine physicians with no affiliations to Spectrum Dynamics. Myocardial counts per minute were calculated for both SPECT and D-SPECT. Imaging on the D-SPECT camera was performed with the patients sitting upright, with their arms resting in a comfortable position; imaging on the conventional-SPECT camera was performed with the patients supine, with their arms over their head. Conventional-SPECT images were acquired using a 180° elliptical orbit. D-SPECT images were acquired in list mode along with R wave markers, with each detector acquiring images from 120 angles. No attenuation correction was used. SPECT images were reconstructed using an iterative algorithm, and D-SPECT images were reconstructed using the MLEM algorithm as described above. Reconstructed images from both systems were reoriented into standard cardiac short and long axis views for visual comparison. Counts per minute in the myocardium for both D-SPECT and SPECT acquisitions were determined by analysis of raw data images on the Xeleris workstation (version, 1.1324; computer model, XW6200; GE Healthcare). Eight-frame gating was applied to the SPECT images. Eight-frame gated images were reconstructed from the list-mode D-SPECT data and were compared visually with the gated SPECT data for image quality. LV ejection fraction and LV volumes were computed from the gated images of both systems using commercially available software (Quantitative Gated SPECT; Cedars-Sinai Health System) (22). LV ejection fractions and LV volumes from 2 patients with mild and severe perfusion defects were computed and compared.

Line-Source Studies

Using the single line source imaged with the D-SPECT camera, we obtained 14.3 × 10⁶ events in 4 min; with the conventional-SPECT camera, this acquisition took 45.5 min. The FWHM and FWTM of the single line source with D-SPECT were 3.5 and 6.8 mm in the x-axis and 4.2 and 7.9 mm in the y-axis, respectively; with SPECT, the FWHM and FWTM of the single line source were 9.2 and 16.6 mm in the x-axis and 12.5 and 22.3 mm in the y-axis, respectively (Supplemental Fig. 5).

Supplemental Figure 5 shows the reconstructed SPECT and D-SPECT images from the NEMA 3–line-source studies. The total acquisition time was 30 min for SPECT and 20 min for D-SPECT. The average FWHM and FWTM of the 3 line sources with D-SPECT were 3.9 and 7.5 mm, and with SPECT they were 10.1 and 19.3 mm, respectively (Supplemental Table 1). These measurements suggest a gain in D-SPECT spatial resolution, compared with that of SPECT, on the order of approximately 2.5. However, reconstruction of the SPECT data with collimator modeling was not performed and would have helped to improve the SPECT spatial resolution. A spatial resolution as good as that obtained with the D-SPECT camera using NEMA standards for Astonish processing. The normal radius of rotation for a 180° orbit with optimal patient set-up is on the order of 20 cm; at this radius, the SPECT reconstructed resolution using NEMA methods is about 12 mm for a high-resolution collimator. The D-SPECT camera does get closer than does the conventional-SPECT camera but also has a poorer geometric resolution. These measurements demonstrate the effectiveness of the D-SPECT reconstruction algorithms in achieving spatial resolution recovery, but these values do not necessarily represent the spatial resolution obtained in clinical practice.

UFOV Studies

Mapping of the spatial resolution throughout a UFOV demonstrated an average spatial resolution (FWHM) of 3.93 ± 0.82 mm and 4.59 ± 0.74 mm in the x- and y-axes, respectively. The FWTM is 8.01 ± 1.43 mm and 9.21 ± 1.22 mm in the x- and y-axes, respectively (Supplemental Table 3).

Reconstructed Spatial Resolution Estimation

We next used a perturbation method to more realistically measure gains in the spatial resolution of the D-SPECT camera over the conventional-SPECT camera. Similar methods have been described by Stayman and Fessler (23). Detailed results from the perturbation experiments are presented in Supplemental Table 2. The reconstructed resolution for SPECT demonstrated a resolution for the long-axis wire (FWHM averaged for complete wire ± SD) of 12.06 mm (±0.42) radially and 11.37 mm (±0.91) tangentially, without resolution modeling; the inclusion of the resolution model improved radial resolution to 7.80 mm (±0.32) but did not affect the tangential resolution (11.57 mm [±0.85]). This result is perhaps not surprising, given that there was no structure inside the ventricular wall, and provides some level of confidence that measurements do reflect the underlying activity distribution. In the case of the D-SPECT for the same short-axis wire, radial resolution was 8.76 mm (±1.65) and tangential resolution was 12.94 mm (±2.0). Results for the short-axis wire were similar, as summarized in Supplemental Table 2, with slightly poorer tangential resolution overall. These results suggest that in practice the radial resolution of the D-SPECT camera is superior to that of the SPECT camera and approaches that obtained when resolution is modeled for SPECT with a high-resolution collimator. Once again, the tangential resolution is worse, although this measurement probably has little significance.

These results reconfirm that D-SPECT image reconstruction has spatial resolution superior to that obtained in conventional cardiac SPECT reconstruction, despite the poor geometric resolution for the D-SPECT collimator. The perturbation figures provide a better impression of the comparative resolution of the 2 systems in clinical practice, in which the reconstruction involves fewer iterations for practical reasons but also has a more complex activity distribution.

Torso Phantom Studies

Figure 4 shows images from the SPECT and D-SPECT scans of the torso phantom (Supplemental Fig. 4 also shows a photograph of the phantom) under several activity concentrations in the heart and liver, with and without myocardial defects. Supplemental Figure 6 shows additional examples of scans of the torso phantom. In each case, qualitative interpretation demonstrates the sharper edge definition of the D-SPECT images over the SPECT images because of the improved spatial resolution of the new technology. No problems in defect identification were encountered for either the SPECT or the D-SPECT images.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 4. 

Images of anthropomorphic torso phantom scanned on both D-SPECT and SPECT. Only myocardial portion of image is shown. Horizontal long-axis (HLA), vertical long-axis (VLA), and short-axis (SA) cuts are shown. For additional examples, see Supplemental Figure 6.

The CZT-based detectors used in the D-SPECT provide better energy resolution and reduce the isotope cross-talk phenomenon. Single-isotope (²⁰¹Tl) and dual-radionuclide (²⁰¹Tl + ^99mTc) SPECT images of a 1-cm insert placed in the anterior myocardial wall of the anthropomorphic torso phantom show blurring and diminishing of the defect on the conventional-SPECT images when obtained under the influence of the ^99mTc cross-talk. This image degradation is almost not perceivable on the D-SPECT images, demonstrating preserved lesion detection with simultaneous dual-radionuclide SPECT (Supplemental Fig. 7).

Patient Studies

Given the highly encouraging results with line sources and the physical torso phantom, we next proceeded to patient imaging studies. Overall patient image quality was rated good and higher in 17 (94%) cases for D-SPECT and 16 (89%) cases for SPECT; D-SPECT images showed qualitatively better myocardial edge definition than did SPECT images in all cases, as assessed by 4 masked nuclear medicine physicians. D-SPECT SSS and SRS correlated linearly with SPECT respective scores (r = 0.83, P < 0.0001, for SSS, and r = 0.76, P < 0.001, for SRS, with a slope of 0.98). Of the 18 studies, 14 (78%) were rated as definitely normal or abnormal for both SPECT and D-SPECT. The myocardial counting rate was significantly higher in D-SPECT than in SPECT (384,000/min ± 134,000/min vs. 47,000/min ± 14,000/min, respectively [P < 0.0001] for stress and 962,000/min ± 426,000/min vs. 136,000/min ± 37,000/min, respectively [P < 0.001] for rest). Three independent nuclear medicine physicians with no affiliations to Spectrum Dynamics also independently read the studies, and image quality was rated good and higher in 98% cases for D-SPECT and 92% cases for SPECT; again, D-SPECT images showed qualitatively better myocardial edge definition than did SPECT images in all cases. End-diastolic and end-systolic images from the gated SPECT data in 2 patients are shown in Supplemental Figure 8. Superior edge definition with D-SPECT is again evident. Table 1 shows the corresponding ejection fractions and end-diastolic and end-systolic volumes estimated from the images. In patient 1 (Fig. 5), normal rest/stress myocardial perfusion was observed, with no significant defects noted. In patient 2 (Supplemental Fig. 9A), who had a history of myocardial infarction, there was a fixed perfusion defect in the anteroseptal and apical regions with no stress-induced defects. In patient 3, who had atypical angina (Supplemental Fig. 9B), there was a reversible stress-induced defect in the inferior wall.

• Download figure
• Open in new tab
• Download powerpoint

FIGURE 5. 

Comparative images from a patient. From top to bottom, D-SPECT and SPECT tomographic slices in mid short axis (2 levels), mid vertical long axis, and mid horizontal long axis are shown. Top row of each set are stress images, and bottom row of each set are rest images. Patient had no known cardiac history and had smoking as risk factor, and no perfusion defects are noted. For 2 more patient examples, see Supplemental Figure 9.