Virtual myocardial PET generated from deep learning of SPECT and PET datasets: potential for attenuation correction in CT-less SPECT

doi:10.21203/rs.3.rs-3106552/v1

Download PDF

Research Article

Virtual myocardial PET generated from deep learning of SPECT and PET datasets: potential for attenuation correction in CT-less SPECT

https://doi.org/10.21203/rs.3.rs-3106552/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 28 Dec, 2023

Read the published version in Annals of Nuclear Medicine →

You are reading this latest preprint version

Objective

Deep-learning approaches have attracted attention for improving the scoring accuracy in computed tomography-less single photon emission computerized tomography (SPECT). This study evaluated the improvement in visual ischemia scoring accuracy to investigate the performance of virtual positron emission tomography (vPET) generated by a deep-learning model.

Methods

This retrospective study included the patient-to-patient stress, resting SPECT, and PET datasets of 54 patients. The vPET generation model was trained and validated using 34 cases with over 1200 image pairs using an image-to-image translation network. The SPECT, PET, and vPET images from another 20 cases were blindly scored in the stress and resting states.

Results

The SPECT rest scores at septal and inferior walls (segments #2 and #15) were significantly higher than those of PET. However, no significant differences were observed between the vPET and PET scores. Diagnostic performance of SPECT for detecting PET defect areas were improved with the use of vPET visual scores.

Conclusions

vPET, a new approach for improving ischemic visual score at rest in the well-known attenuated region on SPECT, can be applied as a clinical support tool that provides powerful auxiliary information for myocardial blood flow diagnosis since standalone SPECT is used worldwide.

N-13 ammonia PET

cardiac SPECT

myocardial perfusion imaging

attenuation correction

deep learning

Myocardial single-photon emission computed tomography (SPECT) is used worldwide to diagnose myocardial ischemia (1). Diagnosis based on visual scoring is still the standard in SPECT, such as the summed stress score (SSS), summed rest score (SRS), and summed difference score (SDS) (1, 2). Artefacts because of photon absorption significantly affect the specificity of ischemia diagnosis in myocardial SPECT (3). Therefore, overcoming this well-known artefact is important for improving diagnostic accuracy in myocardial ischemia diagnosis using SPECT. Conventionally, artificial intelligence has been applied in nuclear cardiology to predict obstructive disease on SPECT (4). Recently, an artificial intelligence-based attenuation correction has been demonstrated (5). Several attenuation correction approaches were proposed, such as generating attenuation maps (µ-maps) from emission images (6, 7), and predicting attenuation-corrected images from non-attenuation-corrected images (8–11). Both deep-learning models can generate attenuation-corrected SPECT images, eliminating the need for patient radiation exposure by CT for attenuation correction. Incidentally, advanced N-13 ammonia PET has high image quality and precision quantification compared with SPECT. Therefore, PET images were used as references for attenuation-corrected SPECT validation (12). In other words, it can be considered that PET is currently the best imaging modality with the least artifacts due to photon attenuation in nuclear cardiology. Previously, an image processing approach was proposed for registering SPECT images to PET images based on surface matching (13). However, this surface matching approach which utilized the PET images as the gold standard, nearly a quarter of a century ago, may not become the current mainstream technique. Focusing on this background of nuclear cardiology with SPECT and PET, we hypothesized that image deep learning using SPECT and PET is potentially useful as a novel approach to correct specific SPECT artefacts. In this study, we investigated the performance of virtual PET (vPET) generated by a deep-learning model constructed using same-patient SPECT and PET datasets. First, we compared visual scoring by vPET and SPECT with PET scores and evaluated the improvement in the accuracy of ischemia diagnosis using the vPET technique. Second, we compared the image similarity indices of vPET and SPECT to PET and objectively evaluated the visual similarity of vPET to PET.

SPECT and PET Datasets

The appropriate institutional review board approved this study and the requirement for written informed consent was waived. SPECT and PET short-axis images of ventricular base-to-apex slices were obtained from 54 patients who underwent rest and stress myocardial Tc-99m-methoxyisobutylisonitrile (Tc-99m-MIBI) SPECT and N-13 ammonia PET for known or suspected coronary artery disease. Fifty out of 54 cases underwent SPECT examination followed by PET examination. The average duration between SPECT and PET examinations in the same patient was 76 ± 241 days. Durations between two scans of 17 cases were longer than 6 months, however, none of the patients showed any changes in clinical findings during this period. Tc-99m-MIBI SPECT was performed using a dual-head SPECT/CT gamma-detector scanner (Symbia S; Siemens Healthcare, Erlangen, Germany) equipped with a smart zoom collimator. The acquisition time for each projection was 40 s after stress (approximately 290 MBq, 8 mCi) and 30 s at rest (approximately 560 MBq, 15 mCi). Adenosine stress (0.12 mg/min/kg) was administered 3 min before the stress scan and continued for 6 min. Thirty projection datasets were obtained in a 128×128 matrix over a 180º arc. The reconstructed images were automatically reoriented to the short-axis images. The pixel size of the cine images was 2.4×2.4 mm, with a 2.4-mm slice thickness (Resolution: 0.4 pixels per mm). N-13 ammonia PET was performed using a three-dimensional PET system (Biograph mCT; Siemens Healthcare, Erlangen, Germany). Repeatedly upgraded Syngo VA30A_HF07 software was used for dose correction (i.e. the difference in residual N-13 ammonia activity between resting and stressed images). Sequential computed tomography scans (120 kV, 20 mA, and 3-mm slice collimation) were acquired for attenuation correction. Electrocardiogram-gated image acquisition was performed immediately after the intravenous administration of N-13 ammonia (approximately 185 MBq, 5 mCi) for 10 min at 16 frames/cardiac cycle using the parallel list mode. After PET myocardial perfusion imaging was performed at rest, the adenosine stress test was performed (0.12 mg/kg/ min for 6 min). Three minutes after vasodilator administration, N-13 ammonia was infused (approximately 555 MBq, 15 mCi), and myocardial perfusion imaging was performed. Images were reconstructed using Fourier re-binning and filtered back-projection with a 12-mm three-dimensional Hann window for the ramp filter. The reconstructed images were automatically reorientated to short-axis images with a 128×128 matrix size. The pixel size of the cine images was 1.6×1.6 mm, with a 1.6-mm slice thickness (Resolution: 0.6 pixels per mm). All PET and SPECT images were normalized with 0%‒100% uptake values.

Generating vPET with deep learning

The training and validation data for creating the model for generating vPET were prepared for deep-learning with a random selection of 34 cases from all 54 cases. Image preprocessing was performed using MATLAB R2020a (version 9.8; MathWorks Inc., Natick, MA, USA) to set patient-to-patient-matched SPECT/PET image datasets. First, SPECT and PET images from the left ventricular base to the apex were selected for all cases. Second, because the number of image slices differs between SPECT and PET owing to the difference in slice thickness, the number of PET slices was aligned with those of SPECT by nearest-neighbor interpolation. Through this preprocessing, paired SPECT and PET image datasets with the same number of images from the base to the apex of all cases were obtained (stress: 1237 image pairs; rest: 1288 image pairs). All image datasets were randomly permuted and divided into training (stress: 865 image pairs, rest: 897 image pairs) and validation (stress: 372 image pairs, rest: 391 image pairs) datasets for deep-learning. The image matrix size for the deep-learning algorithm was upsized to 256 × 256 using nearest-neighbor interpolation. Subsequently, the bit depth for all datasets was converted to 8 bits of portable network graphics files. Deep-learning was performed using a specialized graphics processing unit (TITAN GeForce1080; Nvidia, Santa Clara, Calif, USA) to create an established image-to-image translation model (14). Cost function is defined as the summation of two adversarial losses for translating vPET and SPECT into the other image, along with the cyclic consistency loss for these translations. Consequently, the models for vPET generation from SPECT were concreted using a 200-epoch with a batch size of 8, a learning rate of 0.002, and a momentum optimizer (Fig. 1).

SPECT data for testing datasets for the vPET-generating deep-learning model were prepared in 20 cases not used for training and validation out of all 54 cases (stress: 718 images, rest: 758 images). For input to the deep-learning model, base-to-apex SPECT images were converted to 8 bits of portable network graphics files of 256 × 256 matrix size in the same manner as the training and validation datasets. The vPET generated from the SPECT with the deep-learning model was converted to an uptake value of 0%‒100% according to the signal value of 8 bits and downsized to the same 128 × 128 matrices, similar to the original image size (Fig. 1).

Statistical evaluation of vPET with myocardial perfusion scores

Two radiologists and one cardiologist independently scored the myocardial perfusion in the base-to-apex SPECT, PET, and vPET images of 20 cases. SPECT, PET, and vPET scans were presented independently in a randomized order, and the readers were blinded to all patient information. The readers determined the stress and resting scores in 17 segments on the SPECT, PET, and vPET images. In accordance with the interpretation conditions of the readers’ institution, an uptake of 10%‒100% was displayed in grayscale color (gamma = 1.3), no normal database was referred, and zooming and windowing were free. Defects were graded according to the regional uptake values as 0 to 4 of 5 scores as normal (75‒100%), mild defect (65‒74%), moderate defect (50‒64%), severe defect (40‒49%), and complete defect (< 40%), respectively. Notably, the uptake value was just the reference value for scoring, and the scores were determined by the readers based on their diagnostic experience. In this study, the representative stress and rest scores were defined as the averages of the three readers, and the differences between the scores were defined as the scores at stress minus those at rest. Based on the non-normal distributions of all data with the Shapiro–Wilk test, the Friedman test with Dunn’s correction for multiple pairwise comparisons was used to compare the SPECT and vPET scores with the PET perfusion scores in 17 segments to determine the statistical significance and compare the SSS, SRS, and SDS of SPECT and vPET with those of PET. Additionally, the regional summed scores corresponding to the segments of right coronary artery (RCA, segment 3, 4, 9, 10, and 15), left anterior descending artery (LAD, segment 1, 2, 7, 8, 13, 14, and 17), and left circumflex artery (LCX, segment 5, 6, 11, 12, and 16) were also multiply compared using the Friedman test with Dunn’s correction for multiple pairwise comparisons, respectively. The agreements in global and regional summed scores of SPECT and vPET in comparison to the PET were assessed by Spearman’s correlation coefficients and Bland‒Altman analyses. All statistical analyses were performed using GraphPad Prism, version 8.1.2. (GraphPad Software, Inc.). The differences were considered statistically significant at a P-value of less than 0.05. The diagnostic ability for the presence of a PET defect area in the segments of RCA, LAD, and LCX using SPECT scores of SSS, SRS, and SDS in each segment were analyzed by receiver operating characteristic (ROC) analysis. The presence of a PET defect was defined as one or more SSS, SRS, or SDS in each coronary segment. Additionally, the diagnostic ability for the presence of a PET defect area using both SPECT and vPET scores was also analyzed by ROC analysis. ROC analyses were performed by MATLAB with Statistics and Machine Learning Toolbox R2020a (version 9.8; MathWorks Inc., Natick, MA, USA).

Objective evaluation of vPET with image similarity

The peak signal-to-noise ratio (PSNR) was calculated as an index of the quantitative similarity of images using the following formula: MAX represents the maximum signal intensity of an 8-bit image (MAX = 255), and MSE represents the mean square error of one image of SPECT or vPET against PET.

$$PSNR \left[dB\right]=20\bullet {\text{log}}_{10}\left(\frac{MAX}{\sqrt{MSE}}\right)$$

The PSNRs of SPECT and vPET against PET were calculated for all 20 base-to-apex images. Global, basal, middle, and apical PSNRs were defined as the mean values in stack images representing each ventricular region. The regions of the base, middle, and apex of SPECT, vPET, and PET were determined by dividing the total number of slices in the base-to-apex images into three. Then, stressed 718 base-to-apex image pairs were divided into the basal, middle, and apical regions of 253, 232, and 233 image pairs, respectively. The resting 758 base-to-apex image pairs were divided into basal, middle, and apical regions of 266, 246, and 246 image pairs, respectively. Based on the PSNR data distributions provided by the Shapiro-Wilk test, the PSNRs of SPECT and vPET against PET were statistically compared using the Wilcoxon matched-pairs signed-rank test. All statistical analyses were performed using GraphPad Prism, version 8.1.2. (GraphPad Software, Inc.). Differences were considered statistically significant at a P-vale of less than 0.05.

Table 1 shows the comparison characteristics between the training (n = 34) and test (n = 20) groups. No significant differences were found in the patient background characteristics between the training and test groups. Moreover, there were no biased differences between the randomly assigned training and testing groups for deep-learning. Patients with obesity (body mass index more than 25), diagnosed coronary stenosis based on coronary CT angiography, and ventricular dilation compared with the previous normal value (15) generally ranged from 30% to less than 50% in both training and testing groups.

Table 1

Patient characteristics and ¹³N-ammonia PET measurements
	Training and Validation (n = 34)	Test (n = 20)	P value
Baseline characteristics
Male, n (%)	29 (85)	42 (65)	0.10
Age (years)	65 ± 11	67 ± 9	0.63
Height (cm)	164 ± 8	163 ± 8	0.57
Weight (kg)	65 ± 11	65 ± 10	0.89
BMI (kg/m²)	23.9 ± 3.2	24.4 ± 3.1	0.54
Obesity (BMI > 25)	10 (29%)	9 (45%)
Cardiovascular risk factor
Anemia (Male: Hb < 13, Female: Hb < 11)	11 (32%)	4 (20%)
Hypertension	24 (71%)	15 (75%)
Dyslipidemia	26 (76%)	12 (60%)
Diabetes mellitus	14 (41%)	12 (60%)
Smoking	17 (50%)	8 (40%)
Clinical history
RCA stenosis	12 (35%)	3 (15%)
LAD stenosis	12 (35%)	5 (25%)
LCX stenosis	9 (26%)	3 (15%)
Previous PCI	16 (47%)	5 (25%)
Previous CABG	5 (15%)	0 (0%)
Ammonia PET measurements
Stressed EDVi (mL/m²)	64 ± 14	71 ± 19	0.16
Ventricular dilation (EDVi > 70)	13 (38%)	11 (55%)
Resting EDVi (mL/m²)	55 ± 15	60 ± 18	0.25
Ventricular dilation (EDVi > 70)	5 (15%)	6 (30%)
Stressed ESVi (mL/m²)	23 ± 10	31 ± 17	0.03
Ventricular dilation (ESVi > 25)	14 (41%)	11 (55%)
Resting ESVi (mL/m²)	17 (11–26)	25 ± 15	0.29
Ventricular dilation (ESVi > 25)	9 (26%)	9 (45%)
Stressed EF (%)	66 (60–73)	59 ± 13	0.12
Resting EF (%)	70 (64–76)	62 ± 15	0.10
Stressed MBF (mL/g/min)	1.9 ± 0.4	1.9 ± 0.5	0.76
Resting MBF (mL/g/min)	1.0 (0.8–1.1)	1.0 ± 0.3	0.93
Global MFR	2.0 (1.6–2.4)	2.1 ± 0.6	0.93
Normal distributed data are presented as the mean ± standard deviation. Non-normal distributed data are presented as the median and interquartile range (IQR, 25th–75th percentile) with an underline. EDVi, end-diastolic volume index; ESVi, end-systole volume index; EF, ejection fraction; MBF, myocardial blood flow; MFR, myocardial flow reserve.

Figure 2 shows the colored polar map of the total myocardial perfusion scores of SPECT and vPET minus those of PET of 20 test cases in 17 segments. Compared with the PET stress scores, overestimation of the SPECT stress scores at segments 2, 3, 10, and 16 and underestimation of those at 1, 7, 8, and 9 were observed. The overestimations were corrected to equivalent or calm underestimations in vPET, and the underestimations were equivalent to those of SPECT. Compared with the PET rest scores, overestimation of the SPECT rest scores was observed at segments 2, 3, 10, and 15 to 17. Moreover, the rest scores at segments 2 and 15 were significantly higher than those of PET in all 20 test cases (P < 0.05 and P < 0.01). In contrast, the vPET rest scores were corrected to the equivalent PET rest scores with no significant differences. The different scores between SPECT and vPET against PET tended to be equivalent. However, a tendency to underestimate was observed in segment 15 of vPET. Table 2 shows the median values of global and regional summed scores, and the P values of the multiple comparisons. Higher SPECT scores than PET, but no difference in scores between vPET and PET, were observed in the SRS and RCA-SRS comparisons only. Comparisons between other scores regarding both SPECT and PET showed no significant differences against PET. No significant correlations concerning summed scores against PET were observed, but only the SPECT LAD-SRS was correlated to those of PET. Agreement parameters against PET from Bland‒Altman analysis of bias, standard deviation of bias, and 95% limit of agreement were equivalent in both SPECT and vPET (Table 3). Presence of PET defect areas in the RCA, LAD, and LCX segments were diagnosed by SPECT summed scores with area under the curves of ROC (80%, 67%, and 62%), sensitivities (50%, 56%, and 13%), and specificities (92%, 72%, and 100%), respectively. Similarly, the diagnostic abilities for a PET defect area by SPECT and vPET summed scores were area under the curves of ROC (86%, 80%, and 70%), sensitivities (63%, 67%, and 13%), and specificities (92%, 82%, and 100%), respectively. The upper row in Fig. 3 shows representative images of SPECT, PET, and attenuation-corrected vPET of the same patient. The low accumulation in the middle inferolateral region on SPECT was corrected using vPET. The lower row shows the limitation of vPET in case a defect area observed by PET cannot be observed by vPET. In this case, the defect area also cannot be observed in SPECT.

Table 2

SPECT and vPET scores compared to PET as reference
	SPECT	vPET	PET	P value (SPECT vs. PET)	P value (vPET vs. PET)
SRS	1 (1–2)	0 (0–1)	0 (0–1)	0.0014	> 0.9999
RCA-SRS	1 (0–2)	0 (0–1)	0 (0–0)	0.0024	0.6081
LAD-SRS	0 (0–1)	0 (0–0)	0 (0–0)	0.6081	> 0.9999
LCX-SRS	0 (0–0)	0 (0–0)	0 (0–0)	> 0.9999	0.9535
SSS	3 (2–4)	1 (0–2)	2 (0–5)	0.1156	0.4714
RCA-SSS	2 (1–3)	0 (0–1)	0 (0–2)	0.0656	> 0.9999
LAD-SSS	1 (0–1)	0 (0–1)	0 (0–2)	> 0.9999	0.3579
LCX-SSS	0 (0–1)	0 (0–0)	0 (0–1)	> 0.9999	0.9535
SDS	1 (1–2)	0 (0–1)	1 (0–3)	0.85584	0.6081
RCA-SDS	1 (0–2)	0 (0–1)	0 (0–2)	0.5368	> 0.9999
LAD-SDS	0 (0–1)	0 (0–0)	0 (0–1)	> 0.9999	0.4714
LCX-SDS	0 (0–0)	0 (0–0)	0 (0–1)	> 0.9999	> 0.9999
Non-normal distributed data are presented as the median and interquartile range (IQR, 25th–75th percentile). SPECT, single photon emission tomography; vPET, virtual positron emission tomography; PET, positron emission tomography; SRS, summed rest score; SSS, summed stress score; SDS, summed difference score; RCA, right coronary artery; LAD, left anterior descending artery; LCX, left circumflex artery.

Table 3

Agreements of SPECT and vPET scores against PET as reference
	SPECT vs. PET				vPET vs. PET
	ρ	Bias	SDB	95% LOA	ρ	Bias	SDB	95% LOA
SRS	0.10	1.1	2.5	-3.8 to 5.9	-0.05	-0.4	2.0	-4.2 to 3.5
RCA-SRS	0.40	1.1	1.1	-1.0 to 3.2	-0.13	0.2	0.5	-0.8 to 1.2
LAD-SRS	0.51^*	-0.1	1.7	-3.4 to 3.2	-0.14	-0.4	1.8	-4.0 to 3.2
LCX-SRS	-0.10	-0.1	0.4	-1.0 to 0.8	-0.00	-0.2	0.4	-0.9 to 0.6
SSS	0.12	-0.6	6.6	-13.5 to 12.3	0.03	-2.5	6.2	-14.7 to 9.7
RCA-SSS	-0.25	0.7	3.2	-5.5 to 6.9	-0.29	-0.5	2.6	-5.6 to 4.6
LAD-SSS	-0.04	-1.1	3.3	-7.5 to 5.4	-0.09	-1.6	3.1	-7.6 to 4.5
LCX-SSS	0.01	-0.3	1.8	-3.7 to 3.1	-0.04	-0.4	1.7	-3.8 to 3.0
SDS	0.06	-1.6	5.9	-13.3 to 10.1	0.14	-2.2	6.0	-14.0 to 9.7
RCA-SDS	-0.48^*	-0.2	2.6	-5.3 to 5.0	-0.19	-0.7	2.4	-5.4 to 4.0
LAD-SDS	-0.07	-0.9	3.0	-6.8 to 5.0	-0.11	-1.2	2.9	-6.9 to 4.5
LCX-SDS	0.20	-0.3	1.7	-3.6 to 3.0	0.05	-0.3	1.7	-3.7 to 3.0
Due to all data have non-normal distribution, Spearman’s correlation coefficients are calculated and expressed as ρ. SPECT, single photon emission tomography; vPET, virtual positron emission tomography; PET, positron emission tomography; SRS, summed rest score; SSS, summed stress score; SDS, summed difference score; RCA, right coronary artery; LAD, left anterior descending artery; LCX, left circumflex artery; SDB, standard deviation of bias; LOA, limit of agreement. Asterisks (*) indicate that the P value was less than 0.05.

The global PSNRs of SPECT and vPET against PET are indicated as violin plots in Fig. 4. PSNRs as image similarities of vPET against PET were significantly higher than those of SPECT in the stress and rest states. Furthermore, the regional PSNRs of vPET in the ventricular base, middle, and apex were significantly higher than those of SPECT in the stress and resting states (Fig. 5). Moreover, the different values of the median PSNRs between SPECT and vPET were large in the order of left ventricular regions of the base, middle, and apex (stress: 4.8 vs 4.8 vs 3.7 and rest: 4.3 vs 3.7 vs 2.4).

The three readers performed the SPECT, PET, and vPET myocardial perfusion scoring independently and randomly. We performed a pairwise comparison of the SPECT and vPET scores of 20 test cases with PET scores as references. Overestimation of the SPECT stress and rest scores was observed at the basal interventricular septum and middle-to-apical inferior wall. The image artefact in SPECT at the septum and the inferior wall is a well-known result of photon attenuation. Underestimation tended to be similar between SPECT and vPET. Comparing the SPECT and vPET scores against PET SSS, SRS, and SDS, superior vPET scores were observed in the SRS and regional RCA-SSS; however, no differences were observed in other global and regional scores of both SPECT and PET. As no regions were clearly underestimated or overestimated by vPET compared with SPECT, it is expected that our proposed vPET generation method only improves the accuracy of ischemia scoring using SPECT. Therefore, our proposed vPET generation method corrects SPECT false positives and improves specificity as a negative predictive value. Furthermore, artefacts significantly affect the specificity of myocardial perfusion diagnosis (16). In our study, the false high scores for SPECT with PET were consistent with this report. It has been previously reported that lift and out of field of view of the breast tissue in women helps improve attenuation artifact at the anterior wall (3) and that electrocardiogram-gated scan improves diagnosis of myocardial infarction from attenuation artifacts in men with RCA disease (17). In the present research, the attenuation correction at the anterior wall was not observed. Photon attenuation due to breast tissue may not have largely affected the image because the cohort of this study was small Asians who are smaller than their European and American counterparts. Moreover, both SPECT and PET in this study were acquired with ECG-gated scans. Therefore, we believe this study demonstrated the corrective effect of the diaphragm attenuation artifact. Consequently, this research revealed that our proposed vPET-generating deep-learning model correctly corrects the false high score of SPECT at the myocardial septum and inferior wall.

Compared with the previous deep-learning approach with CT-attenuated SPECT generation from non-attenuated SPECT with U-Net convolution neural network our agreement analyses of vPET and PET scores showed equivalent or less bias and limit of agreement (18). However, since no significant correlation of vPET score to PET was observed, additional validation by future large-scale cohort studies is necessary. Moreover, as SPECT and PET use different radioisotopes as tracers, the uptake value estimation was achieved instead of radioactivity. Furthermore, in this study, radioactivity counts were down sampled to 8-bit for image-based deep learning, so correlation analysis of radioactivity counts between vPET and PET could not be performed. Thus, this study investigated the accuracy of conventional ischemic diagnosis using visual scoring. In the future, we plan to implement a model that estimates the radioactivity of N-13 ammonia using SPECT. This future study may enable the estimation of PET-derived quantitative myocardial flow reserve (MFR) from standard SPECT, thereby enhancing the added value of using PET as training data for deep-learning. The detectability of the defect area with PET as a reference tended to improve using both SPECT and vPET scores. The proposed method is expected to be implemented as a diagnostic support algorithm for myocardial perfusion scoring by SPECT. In recent years, deep learning has been reported to generate SPECT with CT attenuation correction from SPECT without attenuation correction (9, 19). On the other hand, studies have been performed to optimize MFR measurement by SPECT (20) and validate MFR values by dedicated cardiac SPECT scanners with cadmium-zinc-telluride (CZT) detectors (21). Because dedicated cardiac SPECT scanners are designed compactly without CT, attenuation correction with deep-learning based on attenuation-corrected SPECT as a training dataset is challenging (22). Myocardial PET has become widely available for a quantitative assessment of MFR (23–26). This proposed method may be promising for applying to the CZT-SPECT system by preparing patient-to-patient datasets of PET and CZT-SPECT. Furthermore, similar to the CT-less deep-learning approaches, our proposed CT-less method reduces radiation exposure. Our vPET generation method is a novel attenuation correction approach for nuclear cardiology.

The objective image similarities of vPET and PET are higher than those of SPECT in the base-to-apex continuous images. In contrast, PSNR values from vPET and PET were 5–10 dB lower than in recent research on attenuation-corrected SPECT generation with deep-learning (10). However, it is expected that the similarity is low compared with the generating attenuation-corrected SPECT images from SPECT images. Rather, the achievement of this research is that vPET generation enables the output of images with similar information to PET from only SPECT images. In addition, vPET outperformed SPECT regarding the regional similarity in the left ventricle. vPET images are similar to PET images regardless of the left ventricle region. The different values of image similarity between SPECT and vPET were larger in the ventricular base and middle than in the apex. This result indicates that our vPET generation model was particularly effective at the base and middle of ventricle. The SPECT imaging of the myocardium may be restricted at the base and middle of the left ventricle owing to the large and complex anatomical structures. In this study, SPECT was translated to PET-like images from the base to the apex of the ventricle. Furthermore, our proposed SPECT-to-PET translation model may have super-resolution effect because the spatial resolution of PET images are superior to those of SPECT (27). This proposed deep-learning approach, which improves the spatial resolution of SPECT, may be highly compatible with the image-based analysis of myocardial motion (28–31). Therefore, we believe that another investigation to correctly evaluate the usefulness of our proposed model for assessing vPET quantification of myocardial motion is necessary. Moreover, our concept of a deep-learning based approach for SPECT-to-PET image translation potentially becomes more effective due to the improvement of spatial resolution of PET on boarding with multi-crystal and silicon photo-multipliers (32).

This study has some limitations. First, we had a small sample size from a single center. This limitation might be restricted to only a few institutions in our country, which have the required N-13 ammonia production. Since the number of images in the training set used for image generation was over 1000, it was considered sufficient for building an image-to-image translation model. However, model rebuilding and validation in larger and more diverse cohorts is essential. Second, both the training and testing groups had an insufficient and biased number of cases of obesity, RCA stenosis, and ventricular dilation, which are likely to produce well-known SPECT artifacts. This study did not consider that attenuation artifacts are patient-specific and affected by characteristics such as sex, obesity, and ventricular size. We used an image-to-image translation network in the previous study (12). U-Net network was used for the generator for this network (33). The performance of the proposed model may be improved in the future by incorporating meta data such as patient height, weight, and the presence or absence of coronary stenosis into the down sampling layer of U-Net as a numerical matrix. Nevertheless, our preliminary results indicate the potential of vPET imaging using only CT-less SPECT and may be the first to trigger future multicenter studies to accumulate further clinical evidence. Third, the SPECT scoring did not refer to the normal database which was necessary to match the scoring environment with PET and vPET. Therefore, the SPECT scores in this study may be better in actual clinical situations. However, we hope vPET can provide useful additional information that aids the diagnosis of CT-less SPECT, but not as a diagnostic alternative to SPECT.

In conclusion, this study is the first to demonstrate vPET images generated from SPECT images based on deep-learning. However, whether vPET images potentially help SPECT visual scoring diagnostic accuracy, agreement between PET and vPET scores is controversial. This proposed method is a post-processing deep-learning model that provides PET-like image information from CT-less SPECT images. As standalone SPECT systems are used worldwide, the vPET generation model may be applied as a low-cost and practical clinical tool that provides powerful auxiliary information for myocardial blood flow diagnosis.

SPECT, single-photon emission computed tomography; vPET, virtual positron emission tomography; SSS, summed stress score; SRS, summed rest score; SDS, summed difference score; CT, computed tomography; PET, positron emission tomography; Tc-99m-MIBI, Tc-99m-methoxyisobutylisonitrile; RCA; right coronary artery, LAD; left anterior descending artery, LCX; left circumflex artery, ROC; receiver operating characteristic, PSNR, peak signal-to-noise ratio; MFR, myocardial flow reserve.

Ethics approval

This retrospective study was approved by the institutional review board in Tokyo Women’s Medical University (5346) and Kyushu University (2022-127) and conducted in accordance with the 1964 Declaration of Helsinki.

Acknowledgments

This work was supported by JST, ACT-X Grant Number JPMJAX21A7, Japan, and by Funds for the Development of Human Resources in Science and Technology, Initiative for Realizing Diversity in the Research Environment. We thank Editage (www.editage.com) for English-language editing.

Conflict of interest statement:

The authors have no conflict of interest to declare.

Funding:

This work was supported by JST, ACT-X Grant Number JPMJAX21A7, Japan.

Dorbala S, Ananthasubramaniam K, Armstrong IS, Chareonthaitawee P, DePuey EG, Einstein AJ, et al. Single photon emission computed tomography (SPECT) myocardial perfusion imaging guidelines: instrumentation, acquisition, processing, and interpretation. J Nucl Cardiol. 2018;25:1784–846.
Dvorak RA, Brown RKJ, Corbett JR. Interpretation of SPECT/CT myocardial perfusion images: common artifacts and quality control techniques. Radiographics. 2011;31:2041–57.
Bateman TM, Cullom SJ. Attenuation correction single-photon emission computed tomography myocardial perfusion imaging. Semin Nucl Med. 2005;35:37–51.
Litjens G, Ciompi F, Wolterink JM, de Vos BD, Leiner T, Teuwen J, et al. State-of-the-art deep learning in cardiovascular image analysis. JACC Cardiovasc Imaging. 2019;12:1549–65.
Miller RJH, Slomka PJ. Artificial intelligence-based attenuation correction; closer to clinical reality? J Nucl Cardiol. 2022;29:2251–3.
Shi L, Onofrey JA, Liu H, Liu YH, Liu C. Deep learning-based attenuation map generation for myocardial perfusion SPECT. Eur J Nucl Med Mol Imaging. 2020;47:2383–95.
Liu H, Wu J, Shi L, Liu Y, Miller E, Sinusas A, et al. Post-reconstruction attenuation correction for SPECT myocardium perfusion imaging facilitated by deep learning-based attenuation map generation. J Nucl Cardiol. 2022;29:2881–92.
Nguyen TT, Chi TN, Hoang MD, Thai HN, Duc TN. 3D unet generative adversarial network for attenuation correction of SPECT images 4th International Conference on Recent Advances in Signal Processing, Telecommunications & Computing (SigTelCom). IEEE Publications; 2020.
Yang J, Shi L, Wang R, Miller EJ, Sinusas AJ, Liu C, et al. Direct attenuation correction using deep learning for cardiac SPECT: a feasibility study. J Nucl Med. 2021;62:1645–52.
Chen X, Zhou B, Xie H, Miao T, Liu H, Holler W, et al. DuDoSS: deep-learning-based dual-domain sinogram synthesis from sparsely sampled projections of cardiac SPECT. Med Phys. 2023;50:89–103.
Hagio T, Poitrasson-Rivière A, Moody JB, Renaud JM, Arida-Moody L, Shah RV, et al. Virtual’ attenuation correction: improving stress myocardial perfusion SPECT imaging using deep learning. Eur J Nucl Med Mol Imaging. 2022;49:3140–9.
Fricke E, Fricke H, Weise R, Kammeier A, Hagedorn R, Lotz N, et al. Attenuation correction of myocardial SPECT perfusion images with low-dose CT: evaluation of the method by comparison with perfusion PET. J Nucl Med. 2005;46:736–44.
Gilardi MC, Rizzo G, Savi A, Landoni C, Bettinardi V, Rossetti C, et al. Correlation of SPECT and PET cardiac images by a surface matching registration technique. Comput Med Imaging Graph. 1998;22:391–8.
Isola P, Zhu J-Y, Zhou T, Efros AA. Image-to-image translation with conditional adversarial networks IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE Publications; 2017.
Kou S, Caballero L, Dulgheru R, Voilliot D, De Sousa C, Kacharava G, et al. Echocardiographic reference ranges for normal cardiac chamber size: results from the NORRE study. Eur Heart J Cardiovasc Imaging. 2014;15:680–90.
Corbett JR, Kritzman JN, Ficaro EP. Attenuation correction for single photon emission computed tomography myocardial perfusion imaging. Curr Cardiol Rep. 2004;6:32–40.
Xin W, Yang X, Wang J, Shao X, Zhang F, Shi Y, et al. Gated single-photon emission computed tomography myocardial perfusion imaging is superior to computed tomography attenuation correction in discriminating myocardial infarction from attenuation artifacts in men and right coronary artery disease. Nucl Med Commun. 2019;40:491–8.
Mostafapour S, Gholamiankhah F, Maroufpour S, Momennezhad M, Asadinezhad M, Zakavi SR, et al. Deep learning-guided attenuation correction in the image domain for myocardial perfusion SPECT imaging. J Comput Des Eng. 2022;9:434–47.
Shanbhag AD, Miller RJH, Pieszko K, Lemley M, Kavanagh P, Feher A, et al. Deep learning-based attenuation correction improves diagnostic accuracy of cardiac SPECT. J Nucl Med. 2023;64:472–8.
Wells RG, Marvin B, Poirier M, Renaud J, deKemp RA, Ruddy TD. Optimization of SPECT measurement of myocardial blood flow with corrections for attenuation, motion, and blood binding compared with PET. J Nucl Med. 2017;58:2013–9.
Yamamoto A, Nagao M, Ando K, Nakao R, Matsuo Y, Sakai A, et al. First validation of myocardial flow reserve derived from dynamic 99mTc-sestamibi CZT-SPECT camera compared with 13N-ammonia PET. Int Heart J. 2022;63:202–9.
Chen X, Zhou B, Xie H, Shi L, Liu H, Holler W, et al. Direct and indirect strategies of deep-learning-based attenuation correction for general purpose and dedicated cardiac SPECT. Eur J Nucl Med Mol Imaging. 2022;49:3046–60.
Fiechter M, Ghadri JR, Gebhard C, Fuchs TA, Pazhenkottil AP, Nkoulou RN, et al. Diagnostic value of 13N-ammonia myocardial perfusion PET: added value of myocardial flow reserve. J Nucl Med. 2012;53:1230–4.
Giubbini R, Peli A, Milan E, Sciagrà R, Camoni L, Albano D, et al. Comparison between the summed difference score and myocardial blood flow measured by 13N-ammonia. J Nucl Cardiol. 2018;25:1621–8.
Fathala A, Aboulkheir M, Shoukri MM, Alsergani H. Diagnostic accuracy of 13N-ammonia myocardial perfusion imaging with PET-CT in the detection of coronary artery disease. Cardiovasc Diagn Ther. 2019;9:35–42.
Herzog BA, Husmann L, Valenta I, Gaemperli O, Siegrist PT, Tay FM, et al. Long-term prognostic value of 13N-ammonia myocardial perfusion positron emission tomography added value of coronary flow reserve. J Am Coll Cardiol. 2009;54:150–6.
Zhou Z, Ma A, Feng Q, Wang R, Cheng L, Chen X, et al. Super-resolution of brain tumor MRI images based on deep learning. J Appl Clin Med Phys. 2022;23:e13758.
Huang J, Mitchell AJ, Garcia EV, Cooke CD, Folks R, Pernetz M et al. Left ventricular strain from myocardial perfusion positron emission tomography imaging: method development and comparison to 2-dimensional echocardiography. J Nucl Med. 2022;15:jnumed.122.264516.
Kawakubo M, Nagao M, Kikuchi N, Yamamoto A, Nakao R, Matsuo Y, et al. 13N-ammonia positron emission tomography-derived left-ventricular strain in patients after heart transplantation validated using cardiovascular magnetic resonance feature tracking as reference. Ann Nucl Med. 2022;36:70–81.
Kawakubo M, Nagao M, Yamamoto A, Nakao R, Matsuo Y, Fukushim K, et al. ¹³N-ammonia positron emission tomography-derived endocardial strain for the assessment of ischemia using feature-tracking in high-resolution cine imaging. J Nucl Cardiol. 2022;29:2103–14.
Kawakubo M, Nagao M, Yamamoto A, Nakao R, Matsuo Y, Kaneko K, et al. 13 N-ammonia PET-derived right ventricular longitudinal strain and myocardial flow reserve in right coronary artery disease. Eur J Nucl Med Mol Imaging. 2022;49:1870–80.
Gonzalez-Montoro A, Ullah MN, Levin CS. Advances in detector instrumentation for PET. J Nucl Med. 2022;63:1138–44.
Ronneberger O, Fischer P, Brox T. U-net: convolutional networks for biomedical image segmentation. In: Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention; 2015. p. 234–41.

Download PDF

Journal Publication

published 28 Dec, 2023

Read the published version in Annals of Nuclear Medicine →

Reviewers agreed at journal
27 Jun, 2023
Reviewers invited by journal
26 Jun, 2023
Editor assigned by journal
26 Jun, 2023
First submitted to journal
25 Jun, 2023

You are reading this latest preprint version

Virtual myocardial PET generated from deep learning of SPECT and PET datasets: potential for attenuation correction in CT-less SPECT

Status:

Journal Publication

Version 1

Abstract

Objective

Methods

Results

Conclusions

Figures

Introduction

Material and Methods

SPECT and PET Datasets

Generating vPET with deep learning

Statistical evaluation of vPET with myocardial perfusion scores

Objective evaluation of vPET with image similarity

Results

Discussion

Abbreviations

Declarations

References

Status:

Journal Publication

Version 1