- 1Biomedical Imaging Laboratory (BIG), Department of Electrical and Computer Engineering, Faculty of Science and Technology, University of Macau, Taipa, Macao SAR, China
- 2Department of Biomedical Imaging and Radiological Sciences, National Yang Ming Chiao Tung University, Hsinchu, Taiwan
- 3Department of Nuclear Medicine, Taipei Veterans General Hospital, Taipei City, Taiwan
- 4Department of Internal Medicine, Yale University School of Medicine, New Haven, CT, United States
- 5Center for Cognitive and Brain Sciences, Institute of Collaborative Innovation, University of Macau, Taipa, Macao SAR, China
- 6Ministry of Education Frontiers Science Center for Precision Oncology, Faculty of Health Science, University of Macau, Taipa, Macao SAR, China
Purpose: Deep learning-based denoising is promising for myocardial perfusion (MP) SPECT. However, conventional convolutional neural network (CNN)-based methods use fixed-sized convolutional kernels to convolute one region within the receptive field at a time, which would be ineffective for learning the feature dependencies across large regions. The attention mechanism (Att) is able to learn the relationships between the local receptive field and other voxels in the image. In this study, we propose a 3D attention-guided generative adversarial network (AttGAN) for denoising fast MP-SPECT images.
Methods: Fifty patients who underwent 1184 MBq 99mTc-sestamibi stress SPECT/CT scan were retrospectively recruited. Sixty projections were acquired over 180° and the acquisition time was 10 s/view for the full time (FT) mode. Fast MP-SPECT projection images (1 s to 7 s) were generated from the FT list mode data. We further incorporated binary patient defect information (0 = without defect, 1 = with defect) into AttGAN (AttGAN-def). AttGAN, AttGAN-def, cGAN, and Unet were implemented using Tensorflow with the Adam optimizer running up to 400 epochs. FT and fast MP-SPECT projection pairs of 35 patients were used for training the networks for each acquisition time, while 5 and 10 patients were applied for validation and testing. Five-fold cross-validation was performed and data for all 50 patients were tested. Voxel-based error indices, joint histogram, linear regression, and perfusion defect size (PDS) were analyzed.
Results: All quantitative indices of AttGAN-based networks are superior to cGAN and Unet on all acquisition time images. AttGAN-def further improves AttGAN performance. The mean absolute error of PDS by AttcGAN-def was 1.60 on acquisition time of 1 s/prj, as compared to 2.36, 2.76, and 3.02 by AttGAN, cGAN, and Unet.
Conclusion: Denoising based on AttGAN is superior to conventional CNN-based networks for MP-SPECT.
1. Introduction
Myocardial perfusion single photon emission computed tomography (MP-SPECT) is a standard method for the quantitative diagnosis of coronary artery disease (CAD) (1). However, the acquisition time for the conventional NaI-based MP-SPECT is relatively long (15-20 min) (2), leading to potential motion artifacts, e.g., upward creep (3), patient’s discomfort, lower patient throughput (4), and mismatch artifacts between the sequential MP-SPECT and CT (5, 6). New scanner geometries with parallel-hole (7) or multi-pinhole collimations (8, 9) for MP-SPECT are proposed for better photons detection efficiency and reduced scan time (2-8 min) (10, 11). Advanced reconstruction algorithms (12) also facilitate the possibility of reducing acquisition time without degrading image quality. However, the acquisition time for MP-SPECT is still much longer than CT in general (4). Therefore, it is necessary to pursue fast MP-SPECT, without compromising the image quality and diagnostic accuracy.
Image noise is a substantial problem for fast MP-SPECT due to the limited detected counts and the fact that it degrades the image quality, hampering clinical diagnosis and quantification results (9). Recently, deep learning (DL) methods are promising to reduce the noise for MP-SPECT images. Ramon et al. (13) proposed 3D convolutional neural networks (CNN) to denoise the reconstructed MP-SPECT images with reduced injected dose. Liu et al. used a 3D Unet trained on a noise-to-noise strategy for denoising full dose MP-SPECT reconstructed images and showed improved results as compared to the use of traditional filter (14). They further evaluated the performance of DL-based denoising according to the area under the curve (AUC) of the total perfusion deficit (TPD) scores results (15). Aghakhan et al. (16) used a 2D conditional generative adversarial network (cGAN) for denoising the reduced dose MP-SPECT images from 1/8 to 1/2 dose levels in the projection domain. Shiri et al. (17) proposed a 2D residual CNN (ResNet) to estimate full time (FT) MP-SPECT projection images. Previously, our group implemented a 3D cGAN to denoise fast and low dose MP-SPECT reconstruction (18) and projection (19) images. Our results showed that denoising on the projection domain is superior to the reconstruction domain (19).
However, conventional CNN-based methods use fixed-sized convolutional kernels to convolute one local region within the receptive field at a time, which would be ineffective for learning the feature dependencies across large regions (20). The feature dependencies across large regions can only be learned when the feature maps are down-sampled into a relatively small matrix size after passing through several convolutional layers (21). The attention mechanism has shown to be effective in capturing the long-range dependencies of structural information across large regions (20). It has been implemented for CT segmentation (22) and low dose CT denoising (23). In this study, we propose an attention-based cGAN (AttGAN) in denoising fast MP-SPECT projection images and compare its performance with Unet-based and cGAN-based denoising. We further incorporate the patient defect information into the network to improve the AttGAN performance.
2. Materials and methods
2.1. Patient dataset
Fifty anonymized patients who underwent routine stress SPECT/CT scan ∼30 minutes post 99mTc-sestamibi injection on a clinical SPECT/CT system (NM/CT 870 CZT, GE Healthcare, USA) were retrospectively enrolled in this study under local ethics approval (IRB number 2022-11-002CC, Table 1). Among them, 18 were read as having at least a cardiac defect, which had perfusion abnormalities, according to their medical records from SPECT images and clinical histories. Before the SPECT acquisition, a helical CT scan (120 kVp, smart mA (10-150 mA), 0.375 cm slice thickness) was acquired in the heart region for attenuation correction in SPECT reconstruction. The CT reconstruction matrix size was 512 × 512 × variable axial coverage, with a voxel size of 0.9765 mm. Patients were injected with 1,184 MBq 99mTc-sestamibi, and 60 projections were acquired through 180° from the right anterior oblique to the left posterior oblique positions with a matrix size of 64 × 64. The primary photopeak energy window was centered at 140.5 keV with a 20% width and the scatter window was centered at 120 keV with a 10% width.
The original acquisition time was 10 s/view. We also obtained various fast MP-SPECT projection images by reducing the projection acquisition time to be 7, 5, 3, 2, and 1 s based on the list mode data of the FT images, respectively. All clinical data were reconstructed by the 3D ordered subset expectation maximization (OS-EM) algorithm with 5 iterations and 4 subsets, with CT-based attenuation and dual energy window scatter corrections. The reconstruction matrix size was 64 × 64 × 19 with a voxel size of 0.6096 cm. A 3D post-reconstruction Gaussian filter with a standard deviation of 0.6 voxel was applied on the FT images for data analysis.
2.2. Attention-guided generative adversarial network (AttGAN)
The architecture of the AttGAN used in this study is shown in Figure 1 (20). Similar to cGAN, AttGAN is comprised of two subnetworks: a generator (Figure 1A) and a discriminator (Figure 1B). The generator, which was conditioned with fast MP-SPECT projection images, transformed the fast MP-SPECT images into estimated FT MP-SPECT projection images. The estimated images were later paired with the original fast MP-SPECT projection images as an estimated sample pair. The fast MP-SPECT projection images were also paired with the corresponding FT MP-SPECT projection images as a real sample pair. The discriminator learned to differentiate between the estimated sample pairs and the real sample pairs.
![www.frontiersin.org](https://www.frontiersin.org/files/Articles/1083413/fmed-10-1083413-HTML/image_m/fmed-10-1083413-g001.jpg)
Figure 1. The AttGAN architecture used in this study. (A) Generator; (B) discriminator; (C) attention module.
The Unet-based generator (24) had subunits of encoding, bottleneck, and decoding layers. Each encoding layer was comprised of convolution (3 × 3 × 3), batch normalization (BN) (25), rectified linear unit (ReLU) activation, and dropout with a rate of 50%, followed by max-pooling to down-sample feature maps between layers. The decoding layers mirrored the encoding layers, except the up-sample layers replaced the down-sample layers, and skip connection between the encoding and decoding layers was added. The discriminator was a CNN-based network used in our previous study (19).
The attention modules, which were used for calculating the relationships of each voxel to all other pixels within a feature map, were incorporated together with the skip-connection between the encoding layers and decoding layers in the generator (Figure 1C) (20). The feature maps x from the previous encoding layer g and current decoding layer f were transformed by a convolution (1 × 1 × 1) respectively, where g(x) = Wgx and f(x) = Wfx. Wg and Wf were trainable parameters. The feature maps of decoding layer f(x) were then down-sampled to be consistent with the size of g(x). We performed an inner product of the two vectors g(x)and f(x) to obtain the feature dependencies between every two voxels:
The αi,j further went through ReLU activation, convolution (1 × 1 × 1), and softmax function to normalize and reshape the feature maps to become ri,j. Finally, we multiplied ri,j with the feature maps x from the previous encoding layer g to obtain the attention coefficients Att:
The L1 loss (26) and the adversarial loss LADV were used for training the generator g. The discriminator was trained by a cross-entropy loss LD (19). The final objective function of AttGAN was:
where λ is set to be 100 to adjust the weight of VL1 (G) (26). The AttGAN was trained by minimizing the loss. The Unet and cGAN structures were the same according to our previous studies (18, 19).
2.3. Data preprocessing
All the intensity values of MP-SPECT projection images were normalized to a range of 0–1 for training. In addition, we further incorporated the binary patient defect information, i.e., with (1) or without defect (0) from patients’ own medical records, by embedding four 64 × 64 slices with the same binary values into the projection images (Figure 2).
![www.frontiersin.org](https://www.frontiersin.org/files/Articles/1083413/fmed-10-1083413-HTML/image_m/fmed-10-1083413-g002.jpg)
Figure 2. The projection datasets used in this study: (A) original projection, (B) projection incorporating the patient defect information. The defect information block (red) was encoded by binary values (0 = without defect/1 = with defect).
2.4. Network implementation
The AttGAN incorporating patient defect information (AttGAN-def), AttGAN, cGAN, and Unet were implemented using Tensorflow which ran on a NVIDIA GeForce RTX 2080Ti GPU. The Adam optimizer was applied to optimize this proposed model based on an initial learning rate of 0.0001 and trained to 400 epochs.
We performed a 5-fold cross-validation on the clinical datasets to evaluate four DL approaches for various fast SPECT acquisitions. Specifically, for each fold of evaluation, FT and fast SPECT projection images of 35, 5, and 10 patients were selected for training, validation, and testing, respectively. This process was repeated 5 times and all fifty patient datasets were tested and averaged for the final results. The denoised projections were further reconstructed using the same OS-EM algorithm with 5 iterations and 4 subsets with attenuation and scatter correction. No post-reconstruction filter was applied on the reconstructed images generated from the denoised projections for further analysis.
The hyper-parameters, e.g., number of layers and filters within each layer, were determined based on a training-validation procedure for AttGAN. Specifically, the number of layers varied as 2, 3, 4 and 5, while the number of filters within each layer varied as 8, 16, 24, 32 and 40. The hyper-parameters for cGAN and Unet were determined in our previous study (18, 19). The training time for AttGAN-def, AttGAN, cGAN, and Unet was 2.2, 2.2, 2.0, and 1.9 hr, respectively.
2.5. Data analysis
The voxel-based error of the denoised images was assessed by the normalized mean square error (NMSE), structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), joint histogram, and linear regression measured on a 3D volume-of-interest (VOI, 18 × 18 × 18, Figure 3A) which covered the whole heart. The filtered FT reconstructed MP-SPECT images were used as the reference.
![www.frontiersin.org](https://www.frontiersin.org/files/Articles/1083413/fmed-10-1083413-HTML/image_m/fmed-10-1083413-g003.jpg)
Figure 3. (A) The 3D VOI used for voxel-based error calculation. (B) Sample ROIs drawn from a polar plot for IR calculation on a selected patient. (ROI1 was for defect while ROI2 was for a uniform normal region).
where ID represents the voxel values in denoised reconstructed images, IFT is the voxel values on the filtered FT reconstructed images, N (5832) is the number of voxels in the VOI, μD and μFT are the mean values of the denoised and reference images, σD and σFT are the standard deviations of the denoised and reference images respectively, and σD,FT is the cross-covariance between the two images. The constants C1 and C2 are set to be 0.01 and 0.02, respectively (17). MAXFT indicates the maximum possible pixel value of the reference images while MSEindicates the mean squared error between the denoised and reference images.
Two regions-of-interest (ROI1 and ROI2) were drawn on the defect region and a uniform normal region on the polar plots based on visual assessment (Figure 3B) and were adjusted for each patient, respectively. The same ROIs were applied for all denoised images for the same patient. The intensity ratio (IR) was calculated from the mean value of the defect ROI (ROI1) divided by the mean value of the uniform ROI (ROI2). The absolute error of IR between FT images and different denoised images was computed.
A clinical relevant index, the perfusion defect size (PDS, %LV), i.e., an index similar to the total perfusion deficit, was measured by the Wackers-Liu CQ™ (WLCQ) software (Voxelon Inc, Watertown, CT) (27). The absolute error of PDS between FT and different denoised images was computed. The Bland–Altman plots were also computed to quantify the agreement of PDS. For the statistical analysis, a two-tailed paired t-test with Bonferroni correction (SPSS, IBM Corporation, Armonk, NY, USA) was performed between AttGAN-def and other denoising methods at different acquisition time/view for NMSE, PSNR, SSIM, PDS, and IR. A p-value of less than 0.05 was considered as statistically significant.
3. Results
3.1. Reconstructed images, polar plots, and 17-segment analysis
Figure 4 shows the short axis fast MP-SPECT images, their corresponding difference images as compared with filtered FT SPECT images, polar plots as well as 17-segment plots processed using different DL denoising methods for a normal male patient. Figure 5 shows the same results for a male patient with a defect in the left anterior descending (LAD) and left circumflex artery (LCX) region. It can be observed that all the DL-denoised fast SPECT images are similar to the filtered FT SPECT images based on a visual assessment, with the noise level notably suppressed. Furthermore, it is noted that the proposed AttGAN methods have less bias than Unet and cGAN methods according to their corresponding images. Less bias is also observed from the 17-segment images for the AttGAN and AttGAN-def methods. The denoised images consistently exhibit worse resolution, i.e., more blurring, in shorter acquisition times.
![www.frontiersin.org](https://www.frontiersin.org/files/Articles/1083413/fmed-10-1083413-HTML/image_m/fmed-10-1083413-g004.jpg)
Figure 4. Sample images of a normal male patient (age = 67, BMI = 25.4) before and after DL-based denoising for five shorter acquisition times. The images are shown in (A) short axis images, (B) difference images as compared to FT SPECT, (C) polar plots, and (D) 17-segment plots.
![www.frontiersin.org](https://www.frontiersin.org/files/Articles/1083413/fmed-10-1083413-HTML/image_m/fmed-10-1083413-g005.jpg)
Figure 5. Sample images of another male patient (age = 77, BMI = 24.2) with an abnormal perfusion in the LAD and LCX territory before and after DL-based denoising for five shorter acquisition times. The images are shown in (A) short axis images, (B) difference images, (C) polar plots, and (D) 17-segment plots.
3.2. Quantitative analysis
3.2.1. Physical and clinical indices
Figure 6 summarizes the average quantitative indices on all 50 testing datasets for original fast SPECT and DL-based denoised images. For all indices, DL-based denoising methods improve the image quality as compared to the original fast SPECT images. AttGAN-def obtains the best performance, followed by the AttGAN, cGAN, and Unet. At 1 s/prj fast SPECT images, the NMSE values are 0.0317 ± 0.007, 0.0337 ± 0.008, 0.0363 ± 0.011 and 0.0380 ± 0.013 for AttGAN-def, AttGAN, cGAN, and Unet, respectively, where AttGAN-def has significant difference with the other three DL methods. Similar results are obtained for the PSNR and SSIM values. The absolute errors of IR are 0.0435 ± 0.035, 0.0483 ± 0.046, 0.0498 ± 0.046 and 0.0632 ± 0.047 for AttGAN-def, AttGAN, cGAN and Unet on 1 s/prj fast SPECT images. For the absolute error of PDS, the AttGAN-def yields the lowest difference value among all denoising methods on all noise levels. The denoised images achieve better PDS performance, i.e., 1.60 ± 1.738, 2.36 ± 1.903, 2.76 ± 2.056 and 3.02 ± 2.428 for AttGAN-def, AttGAN, cGAN, and Unet on 1 s/prj fast SPECT images, as compared to the original fast SPECT images. The AttGAN-def has significant difference with the cGAN and Unet while not significant difference with AttGAN on 1 s/prj fast SPECT images. GAN methods outperform Unet in general.
![www.frontiersin.org](https://www.frontiersin.org/files/Articles/1083413/fmed-10-1083413-HTML/image_m/fmed-10-1083413-g006.jpg)
Figure 6. Quantitative comparison of NMSE, PSNR, SSIM, PDS, and IR on 50 testing datasets for fast SPECT and different denoised images for five shorter acquisition times. Error bars indicate the standard deviation. The filtered FT SPECT images were used as reference. *p<0.05, **p<0.01, ***p<0.001.
3.2.2. Bland–Altman plots
The Bland–Altman plots of PDS for different denoised methods and fast SPECT images are shown in Figure 7. The dashed lines denote the 95% confidence interval (CI) of the PDS. For 1 s/prj fast SPECT, the AttGAN-def method shows the smallest variance (95% CI: −4.835, +4.435) compared to the reference filtered FT SPECT images, followed by the AttGAN (95% CI: −6.383, +5.423), cGAN (95% CI: −7.402, +5.802), and Unet (95% CI: −8.535, +5.615) methods.
![www.frontiersin.org](https://www.frontiersin.org/files/Articles/1083413/fmed-10-1083413-HTML/image_m/fmed-10-1083413-g007.jpg)
Figure 7. The Bland-Altman plots of PDS for fast SPECT different denoised images and acquisition times. The filtered FT SPECT images are used as reference.
3.2.3. Joint histogram and linear regression analysis
The voxel-based joint histogram and linear regression analysis results are shown in Figure 8. Similar to other quantitative analysis, AttGAN-def obtains the best performance (R2= 0.8633), followed by the AttGAN (R2= 0.8433), cGAN (R2= 0.8341), and Unet (R2= 0.8123) on 1 s/prj fast SPECT denoised images.
![www.frontiersin.org](https://www.frontiersin.org/files/Articles/1083413/fmed-10-1083413-HTML/image_m/fmed-10-1083413-g008.jpg)
Figure 8. Joint histogram and linear regression analysis of fast SPECT and different denoised images in the 50 testing datasets for five shorter acquisition times. The filtered FT images are used as the reference.
4. Discussion
This study aims to assess the potential of attention-based DL network on fast MP-SPECT using clinical datasets. The attention module, inspired by non-local means, was proposed to enable the remote voxels to contribute to the local receptive filter during convolutional filtering (28). To the best of our knowledge, we are the first group to propose using an attention-based DL network for denoising the MP-SPECT images. For the quantitative performance comparison, we used filtered FT MP-SPECT images as the reference since there was no ground truth. Results consistently showed that the AttGAN denoised images had better quantitative accuracy from the difference images, polar plots, 17-segment plots, various physical indices, joint correlation histogram, linear regression and the clinical PDS analysis, as compared to conventional cGAN and Unet (Figures 4-8).
Our proposed networks were trained and tested for each dose-specific dataset, respectively, which were acquired on a CZT scanner. The denoising process is similar to an image-to-image translation task, and should be applicable to data acquired from other scanners or other denoised tasks, e.g., low dose SPECT imaging. The difference between low dose and fast SPECT is that the former will be more subject to patient motion with less radiation dose delivered to the patients. This is particularly important for the increasingly young patient population which has higher radiation risk than seniors (29). In our fast SPECT study, the reduction of acquisition time would be beneficial for patients with a sedation demand and patients with less compliance. The reduced acquisition time would further increase the patient throughput. Both low dose and fast SPECT are of clinical interest and can be potentially achieved using DL techniques (17).
Shiri et al. (17) suggested that denoising from half acquisition time per projection outperformed that from half number of projections for fast MP-SPECT by 2D ResNet. According to this reference, our fast MP-SPECT datasets were obtained by reducing the acquisition time per view. Compared with the existing literature, Ramon et al. (13) reported that CNN-based denoising on 1/2 dose level images could achieve image quality comparable to standard full dose images based on the TPD values. Aghakhan et al. (16) claimed that all denoised images on 1/2 dose level would be clinically acceptable by using 2D cGAN in the projection domain. Similarly, in our work, the mean value of the absolute difference for PDS, a similar index to TPD between FT and AttGAN-def denoised images on 1/2 acquisition time of FT is 0.66%. Although our 3D AttGAN results could not be directly compared with the previous studies due to the use of different networks, datasets, imaging protocols, and evaluation methods, the results are consistent.
In addition, some studies have shown that concatenating the gender, BMI, state (stress or rest), and scatter window images to the training dataset can improve the DL performance in the attenuation correction task (30). We first proposed to add patient defect information into the projections for training the AttGAN and showed promising results. Incorporating the defect information into the network structure or loss function could be further investigated but it is beyond the scope of this study. One should note that the defect information was extracted from the patients’ medical reports and SPECT images in this study, which may be subjected to potential image artifacts. Moreover, the actual defect information may not be able to be verified as other examination results of the patients, e.g., CT angiography, are not available in this study.
Ramon et al. (13) compared the performance between training on a specific low dose level and training on a collection of various low dose levels at the same time. Their preliminary results showed that a dose-specific network can be more accurate than a “one-size-fits-all” network. Liu et al. (31) proposed a denoising method using an image noise index calculated from the normalized standard deviation in the liver ROI for low dose PET images. The image noise index was embedded as a tunable parameter for training. Their results demonstrated that their denoising method achieved better denoising performance than the “one-size-fits-all” network, while it still could not outperform the dose-specific network. Thus, we trained our DL model using a dose-specific approach, i.e., separately for different image acquisition times, for the best denoising performance. More investigations are warranted for a more generalizable and efficient training strategy, i.e., transfer learning, data preprocessing and adjustment on the loss function for training based on all available data.
Limitations of this study include that the clinical-related evaluation is only conducted by PDS. A more comprehensive clinical analysis and a ROC study of defect detectability are needed to validate the proposed methodology. Another limitation is that our DL networks were trained based on a relatively small patient cohort, i.e., thirty-five patients. Training on a large number of patients would benefit the model performance with less susceptibility of overfitting, though our network’s loss function was validated to be converged. Gong et al. proposed to pre-train the DL network with simulation datasets and then fine-tuned it with a limited clinical dataset (32). Their results suggested that using simulation datasets with more realistic imaging conditions or with the use of a more accurate Monte Carlo simulation would generate a more robust pre-trained model.
5. Conclusion
In this work, we investigated the performance of AttGAN in denoising fast MP-SPECT images using clinical datasets. The proposed AttGAN provided superior denoising performance as compared to the conventional cGAN and Unet. Patient defect information could be useful parameter for further improving the AttGAN-based denoising performance.
Data availability statement
The original contributions presented in this study are included in this article/supplementary material, further inquiries can be directed to the corresponding authors.
Ethics statement
The studies involving human participants were reviewed and approved by Institutional Review Board of Taipei Veterans General Hospital (IRB number 2022-11-002CC). The written informed consent was waived due to the retrospective nature of this study.
Author contributions
JS, T-HW, B-HY, and GM were the primary authors of the manuscript. JS was mainly responsible for data processing and analysis. YD was mainly responsible for data analysis and manuscript revision. C-YL was mainly responsible for data collection. T-HW, B-HY, and Y-HL were responsible for clinical interpretation. GM and T-HW were responsible for data interpretation and study integration. All authors contributed to the article and approved the submitted version.
Funding
This work was supported by the Excellent Young Scientists Fund (Hong Kong and Macau, 81922080), Natural Science Foundation of China and Ministry of Science and Technology, Taiwan (MOST 109-2314-B-010-023-MY3).
Conflict of interest
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
References
1. Underwood S, Anagnostopoulos C, Cerqueira M, Ell P, Flint E, Harbinson M, et al. Myocardial perfusion scintigraphy: the evidence. Eur J Nuclear Med Mol Imag. (2004) 31:261–91. doi: 10.1007/s00259-003-1344-5
2. Slomka P, Patton J, Berman D, Germano G. Advances in technical aspects of myocardial perfusion SPECT imaging. J Nuclear Cardiol. (2009) 16:255–76. doi: 10.1007/s12350-009-9052-6
3. Zhang D, Pretorius P, Ghaly M, Zhang Q, King M, Mok G. Evaluation of different respiratory gating schemes for cardiac SPECT. J Nuclear Cardiol. (2020) 27:634–47. doi: 10.1007/s12350-018-1392-7
4. Wu J, Liu C. Recent advances in cardiac SPECT instrumentation and imaging methods. Phys Med Biol. (2019) 64:06TR1. doi: 10.1088/1361-6560/ab04de
5. Zhang D, Yang B, Wu N, Mok G. Respiratory average CT for attenuation correction in myocardial perfusion SPECT/CT. Ann Nuclear Med. (2017) 31:172–80. doi: 10.1007/s12149-016-1144-1
6. Zhang Q, Zhang D, Mok G. Comparison of different attenuation correction methods for dual gating myocardial perfusion SPECT/CT. IEEE Trans Radiat Plasma Med Sci. (2019) 3:565–71. doi: 10.1109/TRPMS.2019.2899066
7. Kao Y, Better N. D-SPECT: new technology, old tricks. J Nuclear Cardiol. (2016) 23:311–2. doi: 10.1007/s12350-015-0290-5
8. Ozsahin I, Chen L, Könik A, King M, Beekman F, Mok G. The clinical utilities of multi-pinhole single photon emission computed tomography. Quant Imag Med Surg. (2020) 10:2006–29. doi: 10.21037/qims-19-1036
9. Bocher M, Blevis I, Tsukerman L, Shrem Y, Kovalski G, Volokh L. A fast cardiac gamma camera with dynamic SPECT capabilities: design, system validation and future potential. Eur J Nuclear Med Mol Imag. (2010) 37:1887–902. doi: 10.1007/s00259-010-1488-z
10. Brambilla M, Lecchi M, Matheoud R, Leva L, Lucignani G, Marcassa C, et al. Comparative analysis of iterative reconstruction algorithms with resolution recovery and new solid state cameras dedicated to myocardial perfusion imaging. Phys Med. (2017) 41:109–16. doi: 10.1016/j.ejmp.2017.03.008
11. Johnson R, Bath N, Rinker J, Fong S, James S, Pampaloni M, et al. Introduction to the D-SPECT for technologists: workflow using a dedicated digital cardiac camera. J Nuclear Med Technol. (2020) 48:297–303. doi: 10.2967/jnmt.120.254870
12. Valenta I, Treyer V, Husmann L, Gaemperli O, Schindler M, Herzog B, et al. New reconstruction algorithm allows shortened acquisition time for myocardial perfusion SPECT. Eur J Nucl Med Mol Imag. (2010) 37:750–7. doi: 10.1007/s00259-009-1300-0
13. Ramon A, Yang Y, Pretorius P, Johnson K, King M, Wernick M. Improving diagnostic accuracy in low-dose SPECT myocardial perfusion imaging with convolutional denoising networks. IEEE Trans Med Imag. (2020) 39:2893–903. doi: 10.1109/TMI.2020.2979940
14. Liu J, Yang Y, Wernick M, Pretorius P, King M. Deep learning with noise-to-noise training for denoising in SPECT myocardial perfusion imaging. Med Phys. (2021) 48:156–68. doi: 10.1002/mp.14577
15. Liu J, Yang Y, Wernick M, Pretorius P, Slomka P, King M. Improving detection accuracy of perfusion defect in standard dose SPECT-myocardial perfusion imaging by deep-learning denoising. J Nucl Cardiol. (2021) 2021:1–10. doi: 10.1007/s12350-021-02676-w
16. Aghakhan Olia N, Kamali-Asl A, Hariri Tabrizi S, Geramifar P, Sheikhzadeh P, Farzanefar S, et al. Deep learning–based denoising of low-dose SPECT myocardial perfusion images: quantitative assessment and clinical performance. Eur J Nucl Med Mol Imag. (2022) 49:1508–22. doi: 10.1007/s00259-021-05614-7
17. Shiri I, Sabet K, Arabi H, Pourkeshavarz M, Teimourian B, Ay M, et al. Standard SPECT myocardial perfusion estimation from half-time acquisitions using deep convolutional residual neural networks. J Nucl Cardiol. (2021) 28:2761–79. doi: 10.1007/s12350-020-02119-y
18. Sun J, Du Y, Li C, Wu T, Yang B, Mok G. Pix2Pix generative adversarial network for low dose myocardial perfusion SPECT denoising. Quant Imag Med Surg. (2022) 12:3539–55. doi: 10.21037/qims-21-1042
19. Sun J, Jiang H, Du Y, Li C, Wu T, Yang B, et al. Deep learning-based denoising in projection-domain and reconstruction-domain for low dose myocardial perfusion SPECT. J Nucl Cardiol. (2022) 2022:1–16. doi: 10.1007/s12350-022-03045-x
20. Zhang H, Goodfellow I, Metaxas D, Odena A. Self-attention generative adversarial networks. Proceeding of the international conference on machine learning (ICML). PMLR (2019). doi: 10.1016/j.compbiomed.2022.105828
21. Schlemper J, Oktay O, Schaap M, Heinrich M, Kainz B, Glocker B, et al. Attention gated networks: learning to leverage salient regions in medical images. Med Image Analy. (2019) 53:197–207. doi: 10.1016/j.media.2019.01.012
22. Liu Y, Lei Y, Fu Y, Wang T, Tang X, Jiang X, et al. CT-based multi-organ segmentation using a 3D self-attention U-net network for pancreatic radiotherapy. Med Phys. (2020) 47:4316–24. doi: 10.1002/mp.14386
23. Li M, Hsu W, Xie X, Cong J, Gao W. SACNN: self-attention convolutional neural network for low-dose CT denoising with self-supervised perceptual loss network. IEEE Trans Med Imag. (2020) 39:2289–301. doi: 10.1109/TMI.2020.2968472
24. Wang Y, Yu B, Wang L, Zu C, Lalush D, Lin W, et al. 3D conditional generative adversarial networks for high-quality PET image estimation at low dose. Neuroimage. (2018) 174:550–62. doi: 10.1016/j.neuroimage.2018.03.045
25. Ioffe S, Szegedy C. Batch normalization: accelerating deep network training by reducing internal covariate shift. Proceeding of the international conference on machine learning (ICML). PMLR (2015). doi: 10.1007/s11390-020-0679-8
26. Isola P, Zhu J, Zhou T, Efros A. Image-to-image translation with conditional adversarial networks. Proceeding of the IEEE conference on computer vision and pattern recognition. (2017). doi: 10.1109/CVPR.2017.632
27. Liu Y. Quantification of nuclear cardiac images: the yale approach. J Nucl Cardiol. (2007) 14:483–91. doi: 10.1016/j.nuclcard.2007.06.005
28. Buades A, Coll B, Morel J. A non-local algorithm for image denoising. Proceeding of the IEEE computer society conference on computer vision and pattern recognition (CVPR). IEEE (2005).
29. Gelfand M, Lemen L. PET/CT and SPECT/CT dosimetry in children: the challenge to the pediatric imager seminars in nuclear medicine. Amsterdam: Elsevier (2007).
30. Chen X, Zhou B, Shi L, Liu H, Pang Y, Wang R, et al. CT-free attenuation correction for dedicated cardiac SPECT using a 3D dual squeeze-and-excitation residual dense network. J Nucl Cardiol. (2021) 2021:1–16. doi: 10.1007/s12350-021-02672-0
31. Liu Q, Liu H, Mirian N, Ren S, Viswanath V, Karp J, et al. A personalized deep learning denoising strategy for low-count PET images. Phys Med Biol. (2022) 67:145014. doi: 10.1088/1361-6560/ac783d
Keywords: denoising, attention-guided, deep learning, myocardial perfusion, fast SPECT
Citation: Sun J, Yang B-H, Li C-Y, Du Y, Liu Y-H, Wu T-H and Mok GSP (2023) Fast myocardial perfusion SPECT denoising using an attention-guided generative adversarial network. Front. Med. 10:1083413. doi: 10.3389/fmed.2023.1083413
Received: 29 October 2022; Accepted: 16 January 2023;
Published: 03 February 2023.
Edited by:
Giorgio Treglia, Ente Ospedaliero Cantonale (EOC), SwitzerlandReviewed by:
Federico Caobelli, Bern University Hospital, SwitzerlandSangwon Han, University of Ulsan, Republic of Korea
Copyright © 2023 Sun, Yang, Li, Du, Liu, Wu and Mok. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Greta S. P. Mok, gretamok@um.edu.mo; Tung-Hsin Wu,
tunghsinwu@nycu.edu.tw
†These authors have contributed equally to this work