Transfer learning-based attenuation correction for static and dynamic cardiac PET using a generative adversarial network

The goal of this work is to demonstrate the feasibility of directly generating attenuation-corrected PET images from non-attenuation-corrected (NAC) PET images for both rest and stress-state static or dynamic [13N]ammonia MP PET based on a generative adversarial network. We recruited 60 subjects for rest-only scans and 14 subjects for rest-stress scans, all of whom underwent [13N]ammonia cardiac PET/CT examinations to acquire static and dynamic frames with both 3D NAC and CT-based AC (CTAC) PET images. We developed a 3D pix2pix deep learning AC (DLAC) framework via a U-net + ResNet-based generator and a convolutional neural network-based discriminator. Paired static or dynamic NAC and CTAC PET images from 60 rest-only subjects were used as network inputs and labels for static (S-DLAC) and dynamic (D-DLAC) training, respectively. The pre-trained S-DLAC network was then fine-tuned by paired dynamic NAC and CTAC PET frames of 60 rest-only subjects to derive an improved D-DLAC-FT for dynamic PET images. The 14 rest-stress subjects were used as an internal testing dataset and separately tested on different network models without training. The proposed methods were evaluated using visual quality and quantitative metrics. The proposed S-DLAC, D-DLAC, and D-DLAC-FT methods were consistent with clinical CTAC in terms of various images and quantitative metrics. The S-DLAC (slope = 0.9423, R2 = 0.947) showed a higher correlation with the reference static CTAC as compared to static NAC (slope = 0.0992, R2 = 0.654). D-DLAC-FT yielded lower myocardial blood flow (MBF) errors in the whole left ventricular myocardium than D-DLAC, but with no significant difference, both for the 60 rest-state subjects (6.63 ± 5.05% vs. 7.00 ± 6.84%, p = 0.7593) and the 14 stress-state subjects (1.97 ± 2.28% vs. 3.21 ± 3.89%, p = 0.8595). The proposed S-DLAC, D-DLAC, and D-DLAC-FT methods achieve comparable performance with clinical CTAC. Transfer learning shows promising potential for dynamic MP PET.


Introduction
Cardiac positron emission tomography (PET) has been considered the gold standard for the assessment of myocardial perfusion (MP) for coronary artery disease, and is increasingly used in the clinic [1][2][3].Perfusion radiotracers commonly include [ 15 O]water, [ 13 N]ammonia, [ 82 Rb]chloride, and [ 11 C]acetate, each with different perfusion characteristics [4].Moreover, tracer-kinetic modeling from dynamic MP PET offers quantitative parameters of global and regional left ventricular function, e.g., myocardial blood flow (MBF) and myocardial flow reserve (MFR) [3], for clinical diagnosis and prognosis assessment.The accuracy of MBF and MFR parameter estimation depends on the corrections of several physical factors, including attenuation, scatter, random coincidences, dead time, and decay during image reconstructions.Among these factors, photon attenuation has a major impact on MP PET quantitative accuracy, causing poor image contrast and significant artifacts, affecting subsequent quantitative analysis [5,6] and clinical diagnosis [7].
With the advent of PET/CT and its widespread adoption in clinical practice, CT-based attenuation correction (CTAC) is commonly practiced for PET attenuation correction (AC).However, CTAC is limited by propagation of CT-based artifacts [8] and potential mismatch between PET and CT [9], particularly in the diaphragm region due to respiratory motion.On the other hand, list mode acquisition is usually implemented for MP PET, where data could be retrospectively divided into different frames for the dynamic or gating protocol.The CT image can then be aligned to each PET frame for AC.However, CT and PET registration is still challenging [10].In addition, the metal artifacts on CT due to metallic implants, e.g., pacemakers and implantable cardioverter defibrillator might propagate to CTAC PET images, causing potentially inaccurate quantitative analysis.Furthermore, CT radiation is related to radiation-induced cancer and cancer death, especially for radiation-sensitive pediatric [11].For PET/MR scanners, CTAC is also not available.Therefore, CTless AC methods for MP PET would be of great clinical value.
Recently, deep learning (DL) has shown great potential for PET AC [12].Several DL approaches generated pseudo-CTs or µ-maps from MR images [13][14][15][16][17][18] for PET AC in the brain and pelvic regions.Other DL approaches generated pseudo-CT [19][20][21] or AC PET images [22][23][24][25] from non-attenuationcorrected (NAC) PET images without the input of structural information.The feasibility of these methods has been demonstrated on brain or whole-body PET data.However, these methods may not be directly applied to dynamic MP PET due to the large difference in count statistics and tracer distribution among different frames.Other DL approaches were developed to improve the outputs of simultaneous activity and attenuation reconstruction [26,27], but these methods were difficult to be applied for the non-time-of-flight PET.Furthermore, conventional DL-based AC methods require large amounts of training data, which limits the clinical application of DL-based AC methods.Therefore, transfer learning (TL) has been introduced to develop robust target models by transferring knowledge learned from other domains and tasks with a small set of target training data [28].TL re-uses a pre-trained model for new related tasks via fine-tuning (FT) strategies.It has been reported to improve the DL-based AC performance as compared to the limited-sample training and accelerate clinical adoption of DL-based AC on new scanners and tracers [29].
In this work, we demonstrated the feasibility of directly generating AC PET images from NAC PET images in the reconstruction domain for static and dynamic [ 13 N]ammonia MP PET based on a generative adversarial network (GAN).To the best of our knowledge, this is the first study using DL for static and dynamic cardiac PET AC.We developed a 3D pix2pix framework [30] via a U-net [31] + ResNet-based generator [32] and a convolutional neural network (CNN)-based discriminator.For static cardiac PET (S-PET) AC, the network was trained by paired static NAC (S-NAC) and static CTAC (S-CTAC) PET images in an end-to-end fashion (S-DLAC).For dynamic cardiac PET (D-PET) AC, paired dynamic NAC (D-NAC) and dynamic CTAC (D-CTAC) PET frames were used as network input and label (D-DLAC).In addition, we adopted TL for D-PET AC, where the pre-trained S-DLAC model was finetuned by paired D-NAC and D-CTAC PET frames for DLbased AC in the dynamic PET images (D-DLAC-FT).Qualitative and quantitative assessments for different AC methods were performed using the CTAC PET as the reference.

Study subjects and image acquisition
This study retrospectively recruited 74 subjects who underwent [ 13   ]ammonia injection.After a 40-min interval following the rest PET/CT examination, a 6-min intravenous adenosine infusion was implemented to induce pharmacological stress.An activity of 715.4 ± 107.9 MBq [ 13 N]ammonia was injected at the third to fourth minute of adenosine infusion, and 15-min stress PET/CT imaging acquisition was then performed.The clinical characteristics are shown in Table 1.The axial field of view (FOV) of PET was 162 mm and the crystal size was 4 × 4 × 20 mm 3 .The delay time was 4 s and the slice thickness was set to 5 mm.Each attenuation map was reconstructed into a 512 × 512 × 65 matrix with a 1.367 × 1.367 × 2.5 mm 3 voxel size.The D-PET scan consisted of 21 time frames for the first 10 min: 12 × 10 s, 6 × 30 s, 2 × 60 s, and 1 × 180 s.Free breathing CT data were acquired (120 kVp, 120 mAs, pitch 0.5) after the PET scans and then converted to attenuation maps (cm −1 ) using the scanner software.The attenuation map was registered to each PET frame for AC using the scanner software.Each D-NAC and D-CTAC frame was reconstructed using the ordered subset maximization expectation algorithm (OS-EM, 2 iterations and 24 subsets) with a matrix size of 168 × 168 × 65 and a voxel size of 4.063 × 4.063 × 2.5 mm 3 , followed by a 5-mm full width at half maximum Gaussian post-reconstruction smoothing.S-PET was obtained by averaging the scan data of the last 12 min for each subject.The S-NAC and S-CTAC PET images were all reconstructed using OS-EM with the same parameters and matrix size as D-PET.

Image preprocessing and experiment settings
The voxel values of all NAC and CTAC PET images were converted to standard uptake value (SUV) to reduce the dynamic range of image intensity.Subsequently, all images were cropped to a size of 80 × 80 × 64 to mainly cover the cardiac region for further processing and analysis.Four kinds of data augmentation were performed for the S-PET and D-PET training data by using the Augmentor3D package (https:// Under the guidance of a radiologist, the left ventricular (LV) myocardial masks were manually drawn from the 21st frame of D-PET images for each subject using the ITK-SNAP software [33].The right ventricular (RV) and LV blood pool masks were manually drawn from the frame with the largest SUV in the blood pool region of D-PET images for each subject.The time activity curves (TACs) of RV blood pool, LV blood pool, and LV myocardial regions were generated for each subject based on the manually drawn masks.The quantitative kinetic parameters K 1 from the dynamic MP PET images were estimated by the Levenberg-Marquardt algorithm [34] using the two-tissue compartmental model [35].The mean MBF was then generated in the whole LV myocardium, left anterior descending artery (LAD), left circumflex artery (LCX), and right coronary artery (RCA) using the estimated kinetic parameter K 1 [36].

Network architectures
In this work, we use the 3D pip2pix network framework, which includes a discriminator D and a generator G as shown in Figs. 1 and 2. The generator loss L G and the discriminator loss L D are defined as follows.
where x is the NAC PET image, y is the target CTAC image, and G(x) is the generated DLAC image from the source x by the generator G. L adv is the adversarial loss function of the generator, L 1 is the loss function.We use L 1 instead of L 2 to reduce the inherent blurring caused by the L 2 loss function.The L adv term and L 1 are defined as: where T real = 1 and T fake = 0 are labels for the real and synthetic images respectively; λ is the weight for L 1 loss and is set to 10 for all training steps [30].We implemented the 3D pix2pix network using PyTorch on a Linux workstation with an NVIDIA TITAN GTX GPU (12 GB).The detailed descriptions of the 3D pix2pix network are provided in Section 1 of the Supplementary Materials.

Network training
The rest-only 60 subjects were used to train the network model on 5-fold cross-validation, while the rest-stress 14 subjects were used as an internal testing dataset, resulting in three groups, i.e., 5-fold cross-validation, internal testing_rest, and internal testing_stress.We trained the network with S-PET or D-PET images for 600 epochs with a mini-batch of 4 images.The pretrained S-DLAC network model was then fine-tuned by paired 3D D-NAC and D-CTAC PET frames for another 500 epochs with a mini-batch of 4 images for the subsequent dynamic AC testing.For all network training, S-PET or D-PET images from 210 (42 × 4 augmentation methods + 42 original = 210), 6 and 12 subjects were used for training, validation, and testing the network in each fold.For comparison, the D-PET data were directly tested on the static network model without fine-tuning (D-DLAC-NFT).The outputs of each D-PET AC network model were merged to generate the full D-PET data for all subjects.

Evaluation metrics
For voxel-based analysis, normalized mean absolute error (NMAE), normalized root mean square error (NRMSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM) were quantified on NAC and DLAC images versus the CTAC images as the reference for all datasets.NMAE, NRMSE, PSNR, and SSIM are defined as: where P indicates the predicted image, R indicates the reference image, and N indicates the total number of voxels,

◂
whereas i is the voxel index.I represents the maximum intensity of the reference image while MSE means the mean squared error between predicted and reference images.µ p and µ R denote the mean value of the predicted image and the reference image.σ 2 p and σ 2 R are the variances of the predicted image and the reference image, whereas σ PR indicates their covariance.The parameters C 1 = (k 1 I) 2 and C 2 = (k 2 I) 2  with constants k 1 = 0.01 and k 2 = 0.03 [37] were used in this work.A paired t-test with Bonferroni correction was performed on NMAE, NRMSE, PSNR, and SSIM for S-NAC and S-DLAC, while a Wilcoxon signed-rank test was performed on the same indices for various DLAC methods in D-PET.A p value < 0.05 indicates significant difference for all statistical tests.
For the segment-based analysis, a measured count of each segment was normalized by the maximum count of 17 segments.The 17-segment error percentage between S-DLAC and S-CTAC polar plots was assessed and shown as box plots.Furthermore, joint correlation histogram and linear regression were evaluated for S-NAC and S-DLAC using S-CTAC images as the reference.A paired t-test with Bonferroni correction was performed on generated MBF for D-CTAC-FT and other methods in D-PET.We also compared the MBF errors on D-NAC and DLAC versus the D-CTAC as the reference for all datasets.The MBF errors were calculated as |MBF estimate − MBF D-CTAC |/MBF D-CTAC × 100%.Regression plots and Bland-Altman plots were then applied to MBF results of different AC methods in D-PET using D-CTAC as the reference.

Static [ 13 N]ammonia PET AC
Table 2 shows the average voxel-based analysis results of S-NAC and S-DLAC images versus the S-CTAC images across all subjects in the S-PET dataset.The proposed S-DLAC method outperforms S-NAC (p < 0.0001) in terms of all metrics for all rest and stress-state subjects.Sample image results of two representative subjects are shown in Fig. 3.The corresponding error maps are shown for both S-DLAC and S-NAC.Attenuation artifacts can be observed on the S-NAC images and marked by white arrows.The proposed S-DLAC method shows smaller errors than S-NAC for both rest and stress-state subjects.The SUV joint correlation histogram and linear regression analysis for S-NAC and S-DLAC are provided in Section 2 of the Supplementary Materials.The S-DLAC (slope = 0.9423, R 2 = 0.947) shows better performance than S-NAC (slope = 0.0992, R 2 = 0.654).Figure 4 and Supplementary Fig. 2 illustrate the schematic diagram of 17-segment polar map and box plots of the percentage errors of 17-segment S-NAC and S-DLAC images versus S-CTAC images as the reference across all subjects in the three groups.The proposed S-DLAC method reduces the errors of all segments as compared to S-NAC.

Dynamic [ 13 N]ammonia PET AC
Table 3 shows the voxel-based analysis results of different DLAC methods versus the D-CTAC for all frames in the dynamic [ 13 N]ammonia PET dataset.The D-DLAC-FT method shows better performance than D-NAC (p < 0.0001) and D-DLAC-NFT (p < 0.0001) in terms of all metrics for all subjects.The D-DLAC-FT is superior to D-DLAC (p < 0.05) in terms of NMAE, NRMSE, and PSNR metrics for all reststate subjects.Though the D-DLAC-FT method shows lower NRMSE (p = 0.0699) and higher PSNR (p = 0.2123) results than D-DLAC for stress subjects, there is no significant difference between these two methods.Figure 5, Supplementary Fig. 3, and Supplementary Fig. 4 show the voxel-based analysis results for each dynamic frame of all subjects in the three groups, including NMAE, NRMSE, PSNR, and SSIM.The D-DLAC and D-DLAC-FT methods outperform D-NAC and D-DLAC-NFT in terms of all metrics for each frame.The D-DLAC-FT method is slightly better than   for different methods across all subjects, respectively.The mean MBF derived from D-NAC was higher than the reference D-CTAC for both rest and stress-state subjects.From Tables 4 and 5, it can be observed that the D-DLAC-FT method has the lowest MBF errors compared to other methods.For the 60 rest-only subjects, the D-DLAC-FT method significantly outperforms D-NAC (p < 0.05) and shows comparable performance with the reference D-CTAC (p > 0.05).
Although the D-DLAC-FT method achieves smaller mean MBF errors and standard deviations as compared to D-DLAC-NFT and D-DLAC in the whole LV myocardium, LCX, and RCA, no significant differences were found (p > 0.05) based on paired t-tests with Bonferroni correction.For the 14 rest-stress subjects, the D-DLAC-FT method is significantly better than D-NAC (p < 0.05) in the whole LV myocardium, LAD, and LCX, as well as D-DLAC-NFT   Figure 9 depicts regression plots of MBF in the whole LV myocardium for different methods using D-CTAC as the reference.Consistent with previous observations, the D-DLAC-FT method has the highest correlation with the reference D-CTAC for both rest and stress-state subjects.Additionally, for the 5-fold cross-validation group, the D-DLAC method has a higher correlation (slope = 0.9430, R 2 = 0.859) with the reference D-CTAC than both the D-NAC (slope = 0.6667, R 2 = 0.645) and the D-DLAC-NFT methods (slope = 0.8300, R 2 = 0.451).For the internal testing_rest group, the D-DLAC method (slope = 1.0591,R 2 = 0.785) is also better than the D-NAC (slope = 0.2874, R 2 = 0.215).However, the D-DLAC method (slope = 0.9853, R 2 = 0.852) is worse than D-NAC (slope = 0.8824, R 2 = 0.928) and D-DLAC-NFT (slope = 0.9405, R 2 = 0.932) methods for the internal testing_stress group.Figure 10 shows the Bland-Altman plots of MBF for different AC methods in the whole LV myocardium using D-CTAC as the reference.The MBF of D-NAC is higher than D-DLAC overall.The D-DLAC-FT method has the narrowest distribution with the smallest 95%

Discussion
This work demonstrated the feasibility of performing direct AC in the reconstruction domain for both rest and stress-state static or dynamic [ 13 N]ammonia MP PET using DL-based and TL-based methods without input of CT images.A significant clinical merit of the proposed S-DLAC, D-DLAC, and D-DLAC-FT methods is to reduce CT radiation dose and subsequent patient cancer risk, especially for pediatric patients.Another clinical merit is to eliminate the CT examination time and thus increase patient throughput.Evaluations on static and dynamic real patient MP PET datasets showed that the proposed S-DLAC, D-DLAC, and D-DLAC-FT methods were able to perform accurate AC in the reconstruction domain for both rest and stress-state [ 13 N]ammonia MP PET as compared to the reference S-CTAC and D-CTAC.
For all quantitative measurements in S-PET studies, the S-DLAC method significantly improves PET image quality compared to S-NAC.These results are consistent with existing literature [38,39].This work further investigates the effectiveness of DL-based AC for D-PET.One of the prerequisites for accurate kinetic modeling and MBF estimation in D-PET is an appropriate AC on the original dynamic frame data.Some studies have reported the generation of transmission data [15] or pseudo-CTs [17] from MR images for kinetic modeling in dynamic brain PET.The proposed D-DLAC-FT method shows improved AC performance for D-PET data, leading to subsequent improved MBF estimation accuracy in D-PET compared to D-NAC.The synthesized cardiac PET data is feasible for kinetic modeling and MBF a rest-state female subject, and c a stress-state female subject.The red regions show the RV blood pool, LV blood pool, and LV myocardial ROIs estimation of D-PET.Some studies have reported the feasibility of direct generation of PET kinetic parameter images from the static PET [40,41], but evaluation of these methods compared to our method is beyond the scope of this study.
The proposed D-DLAC-FT method achieves better performance than D-DLAC for various visual images and voxelbased quantitative indices, showing the effectiveness of transfer learning.Though S-PET and D-PET come from the    For the MBF evaluation, the D-DLAC-FT method shows the smallest MBF errors with the reference D-CTAC.However, there is no significant difference between D-DLAC-FT and D-DLAC-NFT and D-DLAC in the whole LV myocardium for rest-state subjects.A fact is that tissue attenuation mainly affects the height of the myocardial TAC and not its shape, while MBF estimation is not very sensitive to changes in the height of the myocardial TAC [42].The TAC shapes of the D-DLAC-NFT, D-DLAC, and D-DLAC-FT methods are similar, resulting in no significant MBF differences.For the stress-state subjects, there is no significant difference even between D-DLAC-FT and D-NAC, despite the former having the smallest error with the reference D-CTAC.The stress-state group has a small sample size and a high proportion of healthy subjects, making it difficult to show statistical differences in MBF between different methods.Lubberink et al. [42] reported the feasibility of MBF imaging without attenuation correction for [ 15 O]water PET, and our future work would explore this for The mismatch of CT and PET data in PET/CT cardiac imaging is due to the temporal resolution difference, involuntary motion (respiratory and cardiac motion), and voluntary motion (patient movement) between the sequential CT scan and PET scan [10].The mismatch of attenuation and emission images is common in cardiac PET and causes artifactual defects related to diaphragmatic displacement, body mass index, and heart size [43].Using the misregistered CT for MP PET AC will result in inaccurate tracer distribution estimation and artifacts.Our work provides a feasible CTless AC solution to this problem for S-PET and D-PET.On the other hand, DL-based estimation of attenuation map for AC has been proven to be superior to direct generation of AC SPECT for cardiac SPECT [38,44].Thus, estimation on DL-based AC maps instead of direct AC could potentially further enhance our work.
There are some other limitations in our work.Though a data augmentation technique is implemented, the datasets used in this work are limited for network training.More patient data, especially for stress data, and transferred learning based on simulation [45] or other clinical data [29] are warranted.In addition, we lack external validation, and validation across different centers is warranted to evaluate the generalization performance of network models.Moreover, analysis of different DLAC methods based on stratified patient characteristics, e.g., gender and BMI, should be considered for a larger patient cohort.More effective network architectures are warranted, e.g., transformer blocks [46], to fully utilize the time frame information for D-PET AC.Finally, the impact of diagnostic performance was not evaluated for DLAC due to a lack of patient clinical information, which warrants further investigation.

Fig. 2
Fig. 2 Schematic diagram of the proposed a S-DLAC and D-DLAC-FT, and b D-DLAC-NFT methods

Figure 6
illustrates sampled short axis images of frame #4, #15, #19, and #21 of D-CTAC, D-NAC, D-DLAC-NFT, D-DLAC, and D-DLAC-FT for a rest-state male and a stress-state female subject.The corresponding error maps are also shown.Attenuation artifacts can be observed in the D-NAC images.The estimated images from D-DLAC-FT method show the best performance with the least errors for both rest and stress-state subjects as compared to other methods.Figure 7 illustrates the TACs of the RV blood pool, LV blood pool, and LV myocardial regions of D-CTAC, D-NAC, D-DLAC-NFT, D-DLAC, and D-DLAC-FT for a rest-state male, a rest-state female, and a stress-state female subject.The representative manually drawn RV blood pool, LV blood pool, and LV myocardial transverse masks are also shown in Fig. 7.The D-DLAC and D-DLAC-FT methods outperform D-NAC and D-DLAC-NFT in the three regions for the three sample subjects.The D-DLAC is slightly worse than D-DLAC-FT in the LV myocardial region for the stressstate female subject.Tables4 and 5show the detailed MBF values and MBF errors in the whole LV myocardium, LAD, LCX, and RCA

Fig. 3
Fig. 3 Comparison of the vertical long axis (VLA), horizontal long axis (HLA), and short axis (SA) images of S-CTAC, S-NAC, and S-DLAC of a rest-state male (left) subject and a stress-state female

Fig. 4
Fig. 4 Schematic diagram of 17-segment polar map (right) and box plots (left) of the 17-segment errors percentage of S-NAC and S-DLAC images versus S-CTAC images as the reference across all (p < 0.05) in LCX for rest-state subjects.However, for the MBF results in other groups, statistically paired t-tests with Bonferroni correction showed no significant difference (p > 0.05) between the D-DLAC-FT method and the D-NAC, D-DLAC-NFT, and D-DLAC methods, despite the fact that the D-DLAC-FT method has smaller mean MBF errors as compared to other methods.Figure8illustrates the polar maps and 17-segment error percentages of MBF from different methods for the three subjects shown in Fig.7.Compared to other methods, D-DLAC-FT has the smallest errors with the reference D-CTAC.Supplementary Fig.5depicts bar graphs of MBF error results for different methods in the whole LV myocardium, LAD, LCX, and RCA for all subjects.

Fig. 5
Fig. 5 Comparisons of the voxel-based analysis results for all frames of D-NAC, D-DLAC-NFT, D-DLAC, and D-DLAC-FT in the 5-fold crossvalidation group

Fig. 6 Fig. 7
Fig. 6 Sample short axis images of frame #4, frame #15, frame #19, and frame #21 of the reference D-CTAC, D-NAC, D-DLAC-NFT, D-DLAC, and D-DLAC-FT for a rest-state male (left) subject and a stress-state female (right) subject.Attenuation artifacts are marked by white arrows same subjects, the acquisition protocols of D-PET data are different from S-PET data, leading to different data distributions between S-PET and D-PET.The D-DLAC-FT is implemented based on the assumption that AC in S-PET domain and D-PET domain could share some network model parameters.Compared with D-DLAC, D-DLAC-FT re-uses S-PET data from the pre-trained static AC model instead of starting the training from scratch.On the other hand, the number of

Fig. 8
Fig.8The polar maps and 17-segment errors percentage of MBF from different methods for a a rest-state male subject, b a rest-state female subject, and c a stress-state female subject

Fig. 9
Fig. 9 Regression plots of MBF in the whole LV myocardium across all subjects for the D-NAC, D-DLAC-NFT, D-DLAC, and D-DLAC-FT methods using D-CTAC as the reference

Fig. 10
Fig. 10 Bland-Altman plots of MBF in the whole LV myocardium for the D-NAC, D-DLAC-NFT, D-DLAC, and D-DLAC-FT methods versus the reference D-CTAC This work demonstrated the feasibility of performing direct AC based on a 3D GAN framework for both rest and stress-state S-PET or D-PET.Transfer learning-based AC is feasible for D-PET data based on the GAN model pre-trained by S-PET data.Qualitative and quantitative results showed that the proposed S-DLAC, D-DLAC, and D-DLAC-FT methods reduced attenuation artifacts significantly and achieved comparable performances with clinical S-CTAC and D-CTAC images.Overall, the proposed S-DLAC, D-DLAC, and D-DLAC-FT methods are promising for routine MP PET clinical practice.
N]ammonia PET/CT examinations on a clinical PET/CT scanner (Biograph 16 HR, Siemens Healthineers, Germany) from October 2016 to March 2021 at the Department of Nuclear Medicine, Guangdong Provincial People's Hospital.This study was performed in line with the principles of the Declaration of Helsinki.The study was approved by the local institutional review board, and the need for written informed consent was waived.The 60 subjects underwent 15-min rest-only PET/ CT examinations with an activity of 672.7 ± 128.1 MBq [ 13 N] ammonia injection.The 14 subjects first underwent 15-min rest PET/CT examinations with an activity of 681.2 ± 111.3 MBq [ 13

Table 2
The voxel-based analysis results (mean ± SD) for all static [ 13 N]ammonia PET images.p values of the paired t-test of S-NAC and S-DLAC were also given for different datasets.N represents the number of related samples in each paired t-test D-DLAC in terms of all metrics.

Table 3
The voxel-based analysis results (mean ± SD) for all dynamic [ 13 N]ammonia PET frames.p values of the Wilcoxon signed-rank test of D-DLAC-FT and other methods were also given for different data-

Table 4
The MBF results (mean ± SD) in the whole LV myocardium, LAD, LCX, and RCA for all dynamic [ 13 N]ammonia PET subjects.p values of the paired t-test of D-DLAC-FT and other methods were also given for different datasets.N represents the number of related samples in each paired t-test

Table 5
The MBF errors (mean ± SD) in the whole LV myocardium, LAD, LCX, and RCA for different methods across all dynamic [ 13 N] ammonia PET subjects using D-CTAC as the reference