Increasing angular sampling for dedicated cardiac SPECT scanner: Implementation with Deep Learning and Validation with human data

Cardiovascular disease (CVD) is the leading cause of death worldwide. Cardiac single photon emission computed tomography (SPECT) plays a critical role in the diagnosis of CVD [1]

The GE Alcyone Discovery NM (DNM) 530/570c scanners (GE Healthcare, Milwaukee, Wisconsin, USA) are dedicated cardiac SPECT/CT systems. The scanner consists of 19 cadmium zinc telluride (CZT) detector modules with tungsten pinhole collimators to maximize photon sensitivities over the cardiac region. The scanner is designed to acquire 19 projection angles simultaneously over a roughly 180-degree arch for image reconstruction without moving the scanner, enabling efficient data acquisition and the capability of dynamic cardiac SPECT imaging. Each detector module in the DNM scanner has a full view of the heart and therefore can spend the full acquisition time to acquire photons emitted from the heart.

In our previous work [2], we investigated ways to further improve the reconstruction quality of DNM by increasing angular sampling. Specifically, the fixed detector array was manually rotated and translated around the subjects to provide additional sets of projection data. A total of 4 sets of projections (total of 4 × 19 = 76 projections) were acquired by rotating the detector array in 5-degree intervals. Also, due to the unique geometry of DNM, the rotation center and the field-of-view (FOV) center are different. Therefore, the projection data acquired at different detector positions cannot be combined directly. A registration step is needed between each ordered-subsets expectation-maximization (OSEM) update for multi-angle image reconstructions (projections acquired at each detector position were treated as a subset). Introduced in our previous work [2], the multi-angle reconstruction protocol was validated on a group of physical phantom studies, porcine studies with and without perfusion defects. Previous results showed that multi-angle projection data improved image spatial resolution, improved defect contrasts in studies with perfusion defects, and maintained the uniformity of the myocardium in normal subjects.

Rotating the detector array during data acquisition may not be convenient to implement in routine clinical settings due to additional manual rotation steps and longer acquisition time. It also restricts the ability to perform dynamic imaging. To maintain the benefits of stationary imaging, we also proposed a deep learning approach to generate synthetic multi-angle reconstructions from the one-angle counterpart. Previous results showed that synthetic multi-angle images showed better defect contrasts on clinical human studies, as validated against cardiac catheterization images.

However, in our previous publication [2], multi-angle human data was not available, and the proposed deep learning method was fine-tuned on porcine and physical phantom studies. This is not ideal for direct implementations and evaluation. Here, the goal of this paper is to further improve the network fine-tuned with multi-angle human datasets and provide validation of the proposed multi-angle reconstruction protocol and the deep learning approach using human data acquired with multiple rotational angles.

In this study, multi-angle ^99mTc-tetrofosmin myocardial SPECT perfusion data were acquired for analysis. With the approval of the Yale Institutional Review Board (IRB), 15 patients participated in the study. These 15 patient studies are different from that included in our previous publication [2]. Table 1 presents the clinical characteristics of these patients and details of the scan protocols. Similar to our previous work, by rotating the fixed detector array in 5-degree intervals, 3 additional sets (4 in total) of projection angles were acquired at 300°, 305° and 310°. Scan duration was approximately 5 minutes at each detector angle. Four-angle images were reconstructed by combining and aligning all four sets of projection data to the reference detector position (315°). Non-contrast CT scans were acquired at the reference position for attenuation correction (AC). Similar to our previous work [2], as acquiring multiple sets of projections by rotating the scanner is not always possible in clinical practice, we implemented a neural network to generate synthetic 4-angle images reconstructed with 76 projections from one-angle images reconstructed from 19 projections. The proposed neural network follows a U-net-like [3] structure and was trained with one-angle images as input and four-angle images as label. The network structure is the same as that used in our previous work. A total of 250 volumes of XCAT phantoms [4] were simulated for network pre-training. The network was first fine-tuned with 2 physical phantom studies and 8 porcine studies, and then fine-tuned with 15 human studies. To obtain the testing results for all the 15 patient studies, the network was fine-tuned on the pre-trained network 15 times using the “leave-two-out” strategy. Within each fine-tuning iteration, one study was used for network testing, one study was used for validation, and the remaining 13 studies were used for network fine-tuning.

Table 1:

Clinical characteristics of the 15 patients acquired in this study. Scan duration was approximately 5 minutes at each detector position.

Patient #	Gender	Age	BMI (kg/m2)	Rest Injection Dose (mCi)	Stress Injection Dose (mCi)	Stress Modality	Multi-angle Data Availability
1	Male	61	31.3	19.3	8.1	Exercise	Four-angle Rest & One-angle Stress
2	Male	68	23.3	19.6	7.7	Exercise	Four-angle Rest & One-angle Stress
3	Male	63	22.5	19.2	8.3	Exercise	Four-angle Rest & One-angle Stress
4	Female	50	34.1	8.2	18.1	Regadenoson	One-angle Rest & Four-angle Stress
5	Male	68	30.7	7.3	17.2	Exercise	One-angle Rest & Four-angle Stress
6	Male	68	39.2	8.3	26.3	Regadenoson	One-angle Rest & Four-angle Stress
7	Male	69	37.9	8.8	17.5	Regadenoson	One-angle Rest & Four-angle Stress
8	Male	54	26.7	18	8.6	Exercise	Four-angle Rest & One-angle Stress
9	Male	54	35.2	7.6	16.2	Exercise	One-angle Rest & Four-angle Stress
10	Male	46	27.0	19.6	8.4	Exercise	Four-angle Rest & One-angle Stress
11	Female	66	22.8	18.1	8.1	Regadenoson	Four-angle Rest & One-angle Stress
12	Female	63	38.2	20.3	8.2	Regadenoson	Four-angle Rest & One-angle Stress
13	Male	39	25.8	8.5	18.7	Regadenoson	One-angle Rest & Four-angle Stress
14	Male	45	32.0	16	7.8	Exercise	Four-angle Rest & Four-angle Stress
15	Male	78	26.3	18.3	8.2	Exercise	Four-angle Rest & One-angle Stress

As discussed in our previous work [2], full-count four-angle data can tolerate a larger number of iterative iterations without introducing excessive noise. Full-count four-angle images were reconstructed using OSEM with 50 iterations and 4 subsets (each projection set was treated as one subset, a total of 200 iterations), while one-angle images were reconstructed using OSEM with 50 iterations and 1 subset (i.e., MLEM with 50 iterations). The 50-iteration protocol is implemented in clinical settings for one-angle image reconstructions. Here, to eliminate the effect of different numbers of iterative reconstruction updates, full-count four-angle data were also reconstructed using a total of 50 iterations for comparison, with each projection set used sequentially within the iterative process. Also, acquiring the four-angle data took approximately four times longer than acquiring the one-angle data. Longer acquisition time may result in more severe motion artifacts in reconstructed images and negatively affect patient throughput. To demonstrate that the four-angle data can still enhance the image quality with a similar acquisition time, we re-binned the four-angle data to 25% of the original photon counts. By doing so, the number of photon counts in the re-binned four-angle data is approximately the same as that in the one-angle data. Thus, the acquisition time of four-angle data could theoretically be the same as the acquisition time of one-angle data in reality. To maintain the same reconstruction speed and to avoid excessive noise, quarter-count four-angle images were also reconstructed using a total of 50 iterations, and each projection set was used sequentially within the iterative process. Specifically, five images from the same patient were included for comparison: (1) one-angle data reconstructed using MLEM with 50 iterations, as suggested by the clinical protocol; (2) deep learning (DL) processed images using one-angle images as input; (3) quarter-count four-angle images reconstructed using a total of 50 iterations; (4) full-count four-angle images reconstructed using a total of 50 iterations; (5) full-count four-angle images reconstructed using OSEM with 50 iterations and 4 subsets (a total of 200 iterative updates). No post-filtering was applied to better visualize differences in image resolution and perfusion defect contrast. The neural network was trained using (1) as inputs and (5) as labels. In the future, the optimal reconstruction parameters for multi-angle data should be determined.

The reconstructed images with multi-angle projections had superior image resolution and contrast, and demonstrated improved visualization of the right ventricle. Using one-angle data as input, our deep learning neural network qualitatively produced images with improved resolution and contrast, consistent with images reconstructed with true multi-angle projections. The images were quantitatively evaluated using the structural similarity index (SSIM), mean absolute error (MAE), and peak signal-to-noise ratio (PSNR) using four-angle human data as reference. Using manually drawn 3D region-of-interests (ROIs) based on the one-angle images, the myocardium to blood-pool ratios were also calculated to demonstrate the improvement in image contrast. For myocardial perfusion imaging, higher ratios are favorable and typically represent high image resolution. Quantitative measurements for the 15 patient studies are presented in Table 2. Compared with the model only fine-tuned with physical phantom and porcine studies, the model further fine-tuned using multi-angle human studies showed statistically significant improvements in quantitative metrics based on multi-angle human data, with PSNR measurements increasing from 34.39 to 34.71 (p < 0.05), SSIM measurements from 0.88 to 0.89 p < 0.01, and MAE measurements decreasing from 0.16 to 0.15 (p < 0.01).

Table 2:

Quantitative assessment for patient studies reconstructed with different methods (MEAN ± STD [95% confidence interval]). The four-angle images were used as the reference. The measurements were obtained by averaging the values on the testing dataset. Best values are marked in bold. All the calculated results are statistically significant compared to one-angle results, with p < 0.001 based on paired t-tests. Despite numerical differences in the measured myocardium-to-blood pool ratios among the deep learning outputs, four-angle (25%, 50 iter), and four-angle (100%, 200 iter), these differences were not statistically significant.

	One-angle (50 iter)	Deep Learning	Four-angle (25% count, 50 iter)	Four-angle (100%, 50 iter)	Four-angle (100%, 200 iter)
Rest Scans
Myo-Blp Ratio	4.65 ± 0.90 [4.09, 5.21]	7.91 ± 1.88 [6.74, 9.08]	8.47 ± 2.22 [7.10, 9.85]	6.64 ± 1.47 [5.73, 7.55]	8.10 ± 2.12 [6.79, 9.42]
PSNR	33.59 ± 2.49 [32.04, 35.13]	34.22 ± 2.40 [32.74, 35.72]	36.12 ± 2.23 [34.73, 37.50]	36.93 ± 2.17 [35.58, 38.27]	╲
SSIM	0.88 ± 0.02 [0.87, 0.89]	0.89 ± 0.02 [0.88, 0.90]	0.92 ± 0.01 [0.91, 0.92]	0.93 ± 0.01 [0.93, 0.94]	╲
MAE	0.17 ± 0.10 [0.11, 0.23]	0.15 ± 0.09 [0.09, 0.20]	0.12 ± 0.07 [0.08, 0.17]	0.11 ± 0.06 [0.07, 0.14]	╲
Stress Scans
Myo-Blp Ratio	4.22 ± 0.66 [3.73, 4.72]	7.17 ± 1.28 [6.22, 8.12]	8.31 ± 2.92 [6.15, 10.47]	6.17 ± 1.51 [5.06, 7.29]	7.59 ± 2.27 [5.91, 9.27]
PSNR	34.87 ± 4.08 [31.85, 37.89]	35.41 ± 3.97 [32.47, 38.35]	37.35 ± 4.79 [33.80, 40.90]	39.11 ± 4.31 [35.91, 42.30]	╲
SSIM	0.88 ± 0.02 [0.87, 0.90]	0.89 ± 0.02 [0.88, 0.90]	0.92 ± 0.01 [0.91, 0.93]	0.94 ± 0.01 [0.94, 0.95]	╲
MAE	0.17 ± 0.11 [0.09, 0.25]	0.15 ± 0.10 [0.07, 0.23]	0.12 ± 0.08 [0.06, 0.19]	0.10 ± 0.07 [0.04, 0.15]	╲

Three representative patient studies are included in Fig. 1. Note that since one-angle images were used in clinical settings, diagnostic comments reported in this paper were based on one-angle SPECT images. Patient #1 is a 61-year-old male with a small-sized, mild-intensity, reversible perfusion defect in the basal to mid-inferior wall consistent with ischemia. As presented in Fig. 1, both deep learning results and four-angle images improved the defect contrast. Cardiac catheterization was not performed for this patient. This patient was chosen because the confirmed perfusion defect is barely seen in the one-angle reconstructions but becomes clearly visible in the four-angle reconstructions.

Figure 1:

Three sample patient studies with four-angle data. Yellow arrows point to different perfusion defects. Blue arrows point to the papillary muscle in the image, which is better visualized in the full-count four-angle image. Red arrows in the cardiac catheterization (Cath) images point to stenosis that correspond to the perfusion defects in the SPECT images. SA: short-axis; HLA: horizontal long-axis; VLA: vertical long-axis; DL: deep learning; LAD: left anterior descending artery; LCx: left circumflex. Four-angle stress projections were not available for Patients #1 and #3. The numbering of patient studies corresponds to that in Table 1.

Patient #3 is a 63-year-old male. This patient has two non-contiguous perfusion defects. One small-sized, mild intensity, reversible defect in the true apical wall consistent with ischemia. Another small-sized, mild intensity, reversible perfusion defect in the basal to mid-inferoseptal wall consistent with ischemia. In the rest scan, four-angle data produced images with improved apical defect contrast, as illustrated in the SPECT images. While deep learning produced images with improved spatial resolution, one-angle images and deep-learning-generated images showed similar defect contrast in the rest scan. In the stress scan, deep learning improved the basal to mid-inferoseptal defect contrast. For this patient, the perfusion defects were validated using cardiac catheterization images. As illustrated in Fig. 1, the distal left anterior descending artery (LAD) stenosis leads to the apical defect. The dominant left circumflex (LCx) stenosis leads to the basal to mid-inferoseptal defect. This patient was chosen because data from the followed-up catheterization study was available.

Patient #14 is a 45-year-old male. Diagnostic results showed that the perfusion imaging of this patient is normal. Both the neural network and multi-angle data improved the image resolution and maintained the overall uniformity of the myocardium. However, with 200 iterative reconstruction updates, full-count four-angle data improved image resolution but also enhanced some undesired noise in the images, potentially due to a larger number of iterative reconstruction steps. In the future, optimal reconstruction parameters for four-angle data should be determined in clinical settings. This patient was chosen because this is the only patient study with four-angle data in both the rest and stress phases.

As illustrated in Fig. 1, even with quarter counts, four-angle data still produced images with better spatial resolution and improved perfusion defect contrast. As shown in Table 2, quantitative measurements of four-angle quarter-count images are superior to both one-angle reconstructions and deep learning output. The presented results showed that the multi-angle protocol could be implemented as a potential clinical method.

In summary, we investigated the multi-angle data acquisition protocol and image reconstruction in human studies. At both full-count and quarter-count settings, four-angle data demonstrated improved image resolution and defect contrast. Since quarter-count four-angle data could theoretically be acquired without extending the acquisition time, rotating the detector gantry while maintaining the same acquisition time (~5 min) could be a feasible clinical method to improve the image quality. The presented results also showed that neural networks could be used to map one-angle data to four-angle data for improved image quality while maintaining the benefits of fast acquisition and stationary imaging. Validated against clinical information, improved image resolution/contrast may enhance diagnostic performance. However, further validations with a larger cohort of patient studies are needed to demonstrate the clinical impact.

Figure 2:

Graphical Abstract

New Knowledge Gained and Clinical Implications:

GE Alcyone Discovery NM 530/570c are dedicated cardiac SPECT scanner. In this work, we validated the proposed multi-angle reconstruction protocol on clinical human studies. Multi-angle images showed higher image resolution/contrast, and may improve the diagnostic performance as validated against clinical information.

Abbreviations:

CVD

cardiovascular disease

SPECT

single photon emission computed tomography

DNM

GE Discovery NM

CZT

cadmium zinc telluride

AC

attenuation correction

FOV

field-of-view

MLEM

maximum likelihood expectation maximization

OSEM

ordered-subsets expectation-maximization

Tc-99m

technetium-99m