Aim/Introduction: Myocardial perfusion stress SPECT(MPSS) is an established diagnostic test for patients suspected with coronary-artery-disease(CAD). Meanwhile, coronary-artery calcification(CAC) scoring obtained from diagnostic CT is a highly-specific test, offering incremental diagnosis information in identifying patients with significant CAD yet normal MPSS scan. Nonetheless, CAC scoring is not commonly performed/reimbursed in a wide community setting. Our aim is to quantify heterogeneity of uptake via radiomics of ‘normal’ MPSS scans to enable prediction of CAC scores, identifying subclinical CAD. Materials and Methods: 428 patients were collected with normal (non-ischemic) MPSS (8-30mCi 99mTc-Sestamibi) with consensus reading. NM physician verified images (iteratively-reconstructed/attenuation-corrected) to be free from fixed perfusion-defect/artifactual attenuation. 3D images were automatically-segmented into 4 regions-of-interest(ROI), including myocardium+3vascular segments (LAD-LCX-RCA). We developed standardized environment for radiomics analysis(SERA) and calculated 215 3D radiomic features in compliance with image-biomarker standardization initiative (IBSI), ensuring reproducibility of this study. Isotropic-cubic-voxel-ROIs (no resampling/interpolation needed) were discretized using fixed-bin-number discretization into 8 grey-levels (GLs) (22,⋯,29). We first performed two-phase blind-to-outcome feature-selection: A)Removing: A-1)three smallest GLs (very-low dynamic-range), A-2)two highest-GLs (causing highly-correlated features ρ>0.9), and A-3)GL=128 (indifferent statistical properties), ultimately selecting GL=64, similar to findings from our previous study. B)Post-feature calculation: removing features with B-1)identical values, B-2)very-low dynamic-range, B-3)varieties of higher-order feature-classes, B-4)redundant features(ρ=1), and B-5)highly-correlated features (Spearman ρ>0.95). Next, we ran multivariate analysis to predict CAC scores from i) radiomics, ii) clinical-features, iii) radiomics+clinical-features. We performed randomly-selected 60%/25%/15% training/validation/testing. Training started from a constant fit, following iteratively adding/removing features (stepwise-regression) based on sorted univariate-Spearman-correlation with CAC-scores, invoking Akaike-information-criterion (AIC) to discourage overfitting. Validation was run similarly, with the training output-model as initial fit. We shuffled training-validation sets 20 times, then found the best model using log-likelihood to evaluate the test-set. The sensitivity to test-set was further reduced by running the entire operation 50 times, then employing Fisher’s method to verify significance of independent tests. Results: Feature-selection significantly reduced 8×215 features to 56. Median Absolute Pearson’s-correlation coefficient|p-value for 3 feature-pools(radiomics,clinical,combined) were: (0.15±0.11,0.38±0.08,0.41±0.05)|(0.1,0.001,0.0006), (0.24±0.06,0.35±0.08,0.41±0.05)|(0.05,0.004,0.0007), (0.07±0.05,0.24±0.1,0.28±0.09)|(0.4,0.06,0.02) (0.06±0.05,0.16±0.06,0.24±0.06)|(0.4,0.2,0.05) for Myocardium-LAD-LCX-RCA, respectively. Results demonstrate combined features enhance the significance of CAC score prediction across all segments. Conclusion: Our multivariate model enabled the significant prediction of CAC scores at all cardiac segments when combining standardized-radiomics with clinical features, suggesting radiomics adds diagnostic/prognostic value to standard MPSS for wide clinical usage. References:  Shaw, et-al, Radiology, vol.228, no.3, pp.826-833, 2003.  Ashrafinia, PhD Thesis, 2019.  Zwanenburg, et-al., arXiv:1612.0700s3.  Ashrafinia, et al., Medical Physics. Vol.44. No.6, 2017.
Aim/Introduction: Quantitative PET myocardial perfusion imaging (MPI) can standardize detection of coronary artery disease and improve diagnostic accuracy in patients with balanced ischemia. Emerging SPECT technology may enable quantitative evaluation as well, but proof hereof is still in its infancy. We aim to contribute to ‘ground truth’ validation of quantitative SPECT-MPI by evaluating the accuracy of flow quantification using a novel myocardial perfusion phantom. Materials and Methods: The in-house built perfusion phantom mimics the anatomy and (patho-) physiology of left ventricular first-pass perfusion. Pumped continuous flow is conducted through a 3D printed left ventricle and aorta, which branches into coronary arteries that are connected to three myocardial modules. These represent the microcirculation of the main coronary territories. The modules are interchangeable and can consist of different tissue fillings. Flow sensors are incorporated into the setup as ‘ground truth’ flow measure. Flow distribution is controlled by adjustable end-resistances, which also enables simulation of local perfusion deficits. As with patients, a radioactive tracer is administered and a dynamic myocardial perfusion scan is started simultaneously to monitor tracer distribution. The resulting time activity curves (TACs) serve as input for myocardial blood flow quantification. The absolute difference between measured and computed flow (in mL/min/g) is used as measure of accuracy. In the phantom experiments, we used standard clinical protocols for SPECT-MPI (D-SPECT, Spectrum Dynamics) and subsequent flow quantification (4DM Corridor software). We injected 500 MBq 99mTc-tetrafosmin at an aorta flow of 2-5L/min. The flow into the individual myocardial modules varied between 20-100mL/min and module fillings varied (e.g. different types of sponge materials). Results: The obtained TACs inside the simulated left ventricle match physiological values. The area under the curve remains the same for the different aortic flow rates, but the maximum of the curve goes down and smears out over a longer period when lowering the flow. The TACs corresponding to myocardial tissue segments have a relatively fast washout of less than 20s. An aortic and myocardial flow of 2 and 100mL/min, respectively, resulted in the longest washout time.
Conclusion: This study highlights the design and realization of a novel myocardial perfusion phantom to contribute to ground truth validation of quantitative SPECT-MPI. First testing showed promising results, as both geometry and tracer distribution resemble left ventricular microcirculation. Subsequent evaluation of quantitative SPECT-MPI accuracy is in progress.
Aim/Introduction: Vasodilator stress, due to predominantly respiratory side effects, can introduce varying degrees of patient motion during rubidium cardiac PET. The most commonly observed is periodic patient motion due to tidal breathing and non-periodic motion due to cardiac creep. Motion can degrade perfusion images and is time dependant; in extreme cases rendering images non-diagnostic. In contrast to SPECT, motion correction during PET imaging is challenging. It can be performed by aligning reconstructed dynamic images but is labour-intensive and requires frames to be sufficiently long to avoid noisy data but intra-frame motion can be a potential problem. This work evaluates a prototype automated data-driven high-temporal resolution motion correction strategy. Materials and Methods: 10 rubidium stress and rest images (20 images total) from a Siemens Biograph Vision with evidence of motion blurring were included. Frame-by-frame motion correction (FBFMC) was performed by manually aligning and summing twelve 15-second dynamic frames from 120 to 300 seconds post infusion of rubidium-82. Data-driven motion correction (DDMC) was performed by automatically locating and tracking the myocardium within a sub-volume of raw projection data with a temporal resolution of one second. The offset in the axial direction within each one second frame was determined with a precision of 1mm and shifted to a reference position. Non-corrected (NC), FBFMC and DDMC data were reviewed by an experienced physician. Image quality was rated non-diagnostic, adequate or good while perceived motion was rated as none, mild, moderate or severe. Intra-frame motion still present in the FBFMC 15-second frames was determined from the DDMC position tracking. Results: For image quality, 7/20 NC images were good, 5/20 adequate and 8/20 NC non-diagnostic; 8/20 FBFMC images were good, 7/20 adequate and 5/20 non-diagnostic; 19/20 DDMC images were good and 1/20 was adequate. Of the 8 non-diagnostic NC images, 5 were still considered non-diagnostic with FBFMC and the other 3 were rated adequate whereas all 8 were rated good with DDMC. Intra-frame motion of up to 42 mm was present in the FBFMC frames highlighting the limitation of this technique. Conclusion: Effective motion correction requires high temporal resolution and is not possible by post-reconstruction image-based methods. This new automated data-driven method is promising from our preliminary data. While neither of these methods are used in current clinical practice, further work on a larger cohort of patients and assessment of clinical impact is required to make it routine. References: None
Aim/Introduction: The use of PET for myocardial perfusion imaging (MPI) is increasing rapidly due to the increased availability of strontium-82/rubidium-82 (Rb-82) generators, high accuracy and the possibility of quantifying myocardial blood flow (MBF). Recently, PET systems with digital photon counting technology have become available. These PET systems have an increased temporal and spatial resolution but the effect on image quality or visibility of perfusion defects in PET MPI is still unknown. Our aim was to determine the value of a digital PET system in comparison to a conventional PET system in MPI using Rb-82. Materials and Methods: We prospectively included 30 patients who underwent rest and regadenoson-induced stress Rb-82 MPI using a conventional PET system (D690, GE Healthcare). In addition, patients underwent rest and stress Rb-82 PET within three weeks on a digital PET system (Vereos, Philips Healthcare). A nuclear medicine physician and cardiologist scored the image quality on a 4-point grading scale and assessed the existence of possible defects in both the rest and stress scans. The images were presented in random order and readers were blinded for the type of scanner used. The image quality, defect interpretation and the quantitative MBF and myocardial flow reserve (MFR) values were compared between both PET systems. Results: Image quality was graded fair in 20% (6/30) of the conventional scans versus 10% (3/30) of the digital PET scans. Moreover, 60% (18/30) of the conventional scans was graded good and 20% (6/30) excellent versus 50% (15/30) good and 40% (12/30) excellent for the digital PET scans (p<0.03). In addition, the defect interpretation differed in 2 out of the 30 scans (p=0.5). Whereas these two scans were scored as normal on the conventional PET, they were interpreted as ischemic on the digital PET. There were no significant differences between both systems in rest MBF (p≥0.3), stress MBF (p≥0.11) and MFR (p≥0.5). Conclusion: Digital PET provides better image quality than conventional PET and flow measurements seem comparable between the two systems. Nevertheless, defect interpretation may still differ. Additional studies are required to confirm this. References: none
Aim/Introduction: The assessment of left ventricular (LV) function via the ejection fraction (EF) is widely used with almost all imaging modalities. So far, these techniques, acquired separately, were correlated but inter-study variability was significant. Therefore, more cross validation studies of EF obtained from a hybrid PET/MR images are needed to evaluate the effects arising from separately scans. However, in MR modalities the ECG signal used for EF calculation is distorted by the magnetic fields; potentially causing the missing of some R-peaks detections. This can produce a wrong delineation of the cardiac phases in the PET images, resulting in a motional blur. The aim of this work was to compare different PET histogramming methods proposed to solve it. Materials and Methods: List-mode FDG PET and cine MR images from cardiac PET/MR viability examination of 19 patients were used. Three methods of PET histogramming were tested: 1) the standard approach (STD) defines relative bin widths dividing each RR interval in equal gates and include all of R-peaks detected, 2) relative bin width with beat-rejection (BR) adds a beat rejection that allows the elimination of RR intervals outside a user-defined window, 3) fixed bin width (FW) uses a single width gate for each subject, obtained from an optimal RR interval. For each method, the LV end-diastolic (EDV), end-systolic volumes (ESV) and EF were obtained. Results: The EF value, and the volumes obtained from both modalities showed positive linear correlations. However, EF (STD:43+/-11%, BR:48+/-12%, FW:46+/-11%) and EDV (STD:144+/-37 ml, BR:148+/-36 ml, FW:148+/-38 ml) were underestimated and ESV (STD:80+/-29 ml, BR:78+/-30 ml, FW:81+/-32 ml) overestimated compared with MR (EF:52+/-13 %, EDV:156+/-43 ml, ESV:75+/-32 ml). Additionally, significant differences in EF were found with STD and FW compared with MR (p< 0.01). Bland-Altman analysis for EF, EDV and ESV, between PET methods and MR reported biases below 9%, 12,5 ml and 6 ml, respectively; while the limits of agreements were lower than 36%, 159 ml and 78 ml, but still clinically not negligible. The BR method showed the best performance. Considering the linear regression between the modalities, BR method provided the best overall correlation in EF (slope:0.68 and Pearson coefficient (r):0.74) and EDV (slope:0.52 and r:0.62), while in the case of ESV, FW performs best (slope:0.74 and r:0.79). Conclusion: Ejection Fraction assessed with PET and MR in simultaneously acquisitions have a positive association, but there are still relevant differences depending on the PET histograming method. References: none
Aim/Introduction: We have proposed a standardization method using a dedicated 123I-metaiodobenzylguanidine (MIBG) phantom for the determination of a heart-to-mediastinum count ratio for the calibration of the collimator characteristics (1-2). The purpose of this study was to clarify the relationship between collimator characteristics and a calibration factor in a Monte Carlo simulation study using a dedicated phantom for cardiac 123I- MIBG imaging.
Materials and Methods: A digital phantom was created from the 123I-MIBG phantom image acquired with X-ray computed tomography. The SIMIND Monte Carlo program was used to obtain 123I-MIBG planar images, which were generated from various collimator specifications: collimator hole diameters were 1, 2, 3, 4, and 5 mm; septa thicknesses were 0.10, 0.45, 0.80, 1.15, and 1.50 mm; and collimator lengths were 20, 30, 40, 50, and 60 mm. Planar MIBG imaging was simulated with 256 × 256 matrix and energy window of 123I was set to 159 keV ± 7.5 %. The calibration factor was calculated from the planar image using a dedicated software program, and defined as a conversion coefficient. The conversion coefficient value shows an approximate range of 0.5 to 0.9, corresponding to low-energy (LE) to medium-energy (ME) collimators. The conversion coefficient was compared with that from a phantom image database, which consisted of 705 image sets. Results: A total of 125 planar phantom images were generated with the digital phantom. When 123I-MIBG phantom imaging with the LE, extended LE (ELE), and ME collimators was simulated, the conversion coefficients for LE, ELE, and ME were 0.52, 0.75, and 0.85, respectively. The conversion coefficients derived from simulated MIBG planar images were equivalent to those from the phantom image database (mean LE, ELE, and ME values were 0.54 ± 0.04, 0.75 ± 0.03, and 0.88 ± 0.04, respectively). When the collimator hole diameter and length were set as 1.0 mm and 30 mm, respectively, conversion coefficients for the septa thicknesses of 0.10, 0.45, 0.80, 1.15, and 1.50 mm were 0.52, 0.67, 0.72, 0.74, and 0.77, respectively. Conclusion: The Monte Carlo program successfully simulated 123I-MIBG phantom imaging and conversion coefficients in the various collimator specifications. The collimator septal thickness was a prominent component in 123I-MIBG phantom simulation. When the collimator specifications are determined, the conversion coefficients can be estimated without 123I-MIBG phantom scan.
References: (1) Verschure DO, et al. J Nucl Cardiol 2017:1-7. (2) Nakajima K, et al. J Nucl Cardiol 2014;21:970-8.
Aim/Introduction: Deep learning artificial intelligence approaches have a great potential to simplify and improve medical imaging. This might, in case of PET imaging, also refer to shorten scan times/reducing tracer doses. In this present study we evaluated whether this is the case for [18F]Florbetaben amyloid PET/MRI. Materials and Methods: We prospectively acquired list-mode [18F]Florbetaben brain PET/MRI scans (300MBq, scan start 90min p.i.) of 40 patients (“new data”, 21 female, age=64±11yrs). The PET data were reconstructed for a (clinical standard) 20min scan duration as well as for 1min scan duration. For the 1min scan duration data, different deep learning approaches (a U-net was pre-trained with previous low-dose PET/MRI data and was either (AI1) directly applied to new data or (AI2) trained further with the new data, or trained from scratch based on (AI3) new data only or (AI4) all existing data) were applied. All PET data were analyzed visually (3 blinded experts, binary score for amyloid load and 5-point score for image quality with 5 being the highest score for excellent image quality) and semi-quantitatively (composite SUVRs, reference: cerebellar cortex, Hermes BRASS software). The 20min scan duration majority visual read served as standard to truth (SoT). Results: According to the SoT, 19 and 21 patients were amyloid-positive and -negative, respectively. Mean sensitivity and specificity for the three blinded readers were 100% and 95%, 100% and 78%, 100% and 100%, 100% and 100%, and 100% and 100% for the 1min, 1min+AI1, 1min+AI2, 1min+AI3, and 1min+AI4 data, respectively. Image quality in visual analysis was 2.5±0.3, 3.0±0.3, 4.0±0.3, 3.9±0.1and 3.8±0.1. Cohen’s d effect sizes for the composite SUVRs according to amyloid state in SoT were significantly higher in the 1min+AI1/4 as compared to the 1min data (2.38/2.64/2.33/2.59 vs. 1.79). Conclusion: Using a trained neural network, scan duration in [18F]Florbetaben amyloid PET/MRI can be reduced down to 1min without losing diagnostic quality. This would alternatively translate to a reduction of tracer dose/radiation exposure by 95%. In conclusion, the deep learning artificial intelligence approach developed has great potential to improve patient convenience/throughput and/or reduce tracer costs. References: none
Aim/Introduction: Alpha-emitter 223Radium-dichloride is associated to a clear survival benefit and significant bone pain palliation in CRPC patients with symptomatic bone metastases. Increase in Overall Survival (OS) is strictly associated to 6-cicles therapy’s administration. Bone Scan Index (BSI) is defined as the percentage of total amount of bone metastasis on whole-body scintigraphic images. To calculate BSI values we used DASciS software, developed by an engineering team of “Sapienza” University of Rome. Aim of our observational, prospective, non randomized study was to assess the load of bone disease at starting and in the time course of Ra-223 treatment, in mCRPC patients, as an overall survival (OS) predictor. Materials and Methods: Bone scintigraphies of 127 mCRPC patients treated with 223Ra were collected before, during and after the therapy. Follow-up images were taken after 3 months, 6 months and one year. DASciS software was used to process bone scans and BSI was calculated. BSI values were analyzed alone and together with 3-PS (prognostic score based on basal values of Hb, PSA and ECOG-PS) in order to evaluate the OS predictive power of these parameters. Results: Employing DASciS software 546 scintigraphies were analyzed, ( 127 basal, 211 intermediate, 87 final, 60 after 3 months, 38 after 6 months and 23 after one year). Both the univariate (HR: 1,8, 1,61-2,02, p=0,001) and the multivariate (HR: 1,82, 1,56-2,10, p=0,001) analysis -adjusted for BMI, age, Gleason Score, number of previous systemic treatments, basal PSA, tALP, Hb, PLT, ECOG-PS- confirm the OS prediction power of basal BSI (Percentage of bone disease are: 0-3% = 28 months of median survival (MoMS), 3%-5% = 11 MoMS, more than 5% = 5 MoMS). The association of BSI with 3-PS have however the best OS prediction power (AUC=91%). Conclusion: Load of bone disease first of 223Ra treatment is an excellent predictor of OS and DASciS software is an accurate tool to calculate BSI. BSI and 3-PS together represent a unique multidimensional evaluation of mCRPC patients at basal conditions and are very useful for the stratification of patients who are candidates to 223Ra treatment. Despite EMA recommendation, the calculation of the bone disease extension percentage instead of number of lesions appears to be a more reliable tool for patients recruitment. References:
Aim/Introduction: 90Y-radioembolization is a well-established treatment option for nonsurgical patients with liver tumours. Measuring the dose distribution with 90Y PET/CT scans could be useful for dose verification and to investigate tumour response to treatment. However, 90Y PET images are typically of very poor quality due to the low intensity of positrons from 90Y. Regularized reconstruction algorithms have recently been introduced in clinical scanners to iterative reconstruction algorithm to reach convergence while minimizing image noise. The aim of this work was: 1) to evaluate regularized image reconstruction parameters to optimize image quality and quantification accuracy using phantom studies, and 2) to investigate their impact on the dose distributions in lesions and healthy livers in patient studies. Materials and Methods: The IEC-NEMA phantom (9.7L) and six spheres (0.5-26mL), filled with 90Y activities (0.325MBq/mL and 2.45MBq/mL, respectively) were scanned using GE-D690 PET/CT camera. Images were reconstructed with OSEM (standard) and the Q.Clear regularized reconstruction method (β-parameter ranging 0-8000), both with the time-of-flight option enabled. The reconstruction parameters were optimized for image quality by evaluating the background noise and contrast-to-noise ratios; while, the hot and cold contrast recovery coefficients were used for evaluating image quantification accuracy. Then, the optimized reconstruction parameters were applied to PET/CT scans obtained from five subjects following 90Y-radioembolization studies. Dose maps of healthy liver parenchyma and tumours were estimated using the MIM software. The variabilities (coefficients of variation (COV)) of dose volume histograms, Dmax, Dmean, D70, and V100Gy for the entire range of investigated reconstructions were assessed. Results: Generally, superior image quality was obtained from Q.Clear compared to OSEM reconstructions. The β-parameter optimized for best image quality was found to be in the range of 2000 - 2500 and equal to 300 for best quantification accuracy. In order to estimate how different reconstructions influence dose distributions, β-parameter ranged from 300-8000 were applied to patient studies. For Dmean, D70, and V100Gy, the COVs were <7% for all five patients; whereas, the COVs for Dmax were >15%. Conclusion: There was a clear advantage of using regularized (Q.Clear) over conventional OSEM reconstruction to maximize image quality and quantitative accuracy for PET imaging studies performed after 90Y-radioembolization studies. However, the β-parameter used in Q.Clear reconstructions needs to be carefully selected based on the study objectives, since its value will largely influence the visual quality of reconstructed images, as well as quantification accuracy and dose distribution in tumours and healthy liver. References: Non
Aim/Introduction: Scandium-44 is a promising PET nuclide with a half-life of 3.97 h and a β+ fraction of 94.27 %. Several preclinical and clinical trials have been performed using Sc-44 labeled tracers e.g. Sc-44-PSMA-617 or Sc-44-DOTATOC. In contrast to the F-18, Sc-44 emits additional γ quanta (E= 1157 keV), which may influence the PET data acquisition. In this study we investigated the differences in coincidence energy spectra of Sc-44 compared to F-18 and analyzed its influence on quantification.
Materials and Methods: A series of phantom measurements with Sc-44 and F-18 at different activity concentrations were performed using Mediso nanoScan PET/MR. Raw data were analyzed with respect to different energy windows. After iterative 3D reconstruction, a quantitative image VOI analysis was performed by using PMOD. Results: In comparison to F-18, Sc-44 showed a different energy spectrum of coincidences with a higher Compton peak, lower 511 keV peak and higher background due to its additional γ quanta. For Sc-44, 16.9-20.1 % of acquired coincidences were detected in an energy window of 400-600 keV in contrast to 19.3-23.7 % for F-18. On average, the reconstructed activity of Sc-44 was 19.5% lower than that of F-18. Conclusion: The co-emission of γ quanta of Sc-44 leads to a different energy spectrum in the raw data in comparison to F-18. As a consequence, lower activity was detected for Sc-44 in the reconstructed images. Therefore, an additional correction factor should be applied in further studies, in particular for quantification and dose calculations. References: none