Aim/Introduction: [18F]Florbetapir (AV45) positron emission tomography (PET) studies allow for in vivo assessment of amyloid deposition. Presence of amyloid deposition is usually assessed visually, which is highly dependent on observer training and experience. These visual reads can be difficult in case of low amyloid depositions, e.g. in cognitively normal individuals. The aim of this work was to develop, train and validate a Convolutional Neural Networks (CNN) able to discriminate between amyloid negative and positive PET scans. Materials and Methods: 133 AV45-PET images (101 negative and 32 positive) were acquired at 50-70 min post-injection on a Philips Gemini TF-64 PET/CT scanner and visually assessed by an experienced nuclear medicine physician in the context of the SCIENCe project. AV45-PET SUVR images using cerebellum as reference region were first spatially aligned (MNI). A 2D-CNN was developed consisting of 4 convolution layers, two max-pooling layers, batch-normalisation and one sigmoid output layer. Next, 106 (80%) scans were used to train a 2D-CNN. During training data balancing was achieved by oversampling the amyloid positive scans and data augmentation was achieved by rotating, flipping, zooming and translating the images. Overfitting was avoided by applying dropouts. To study the impact of spatial resolution on the CNN model performance, PET images were additionally smoothed with 5 and 10 mm FWHM Gaussian kernels . The 2D CNN model performance was tested using the 27 (20%) unused scans, yielding accuracy, sensitivity and specificity. Results: The model classified amyloid negative and positive scans with 100% accuracy. Applying 5 and 10 mm Gaussian smoothing to the 20% test scans resulted in an accuracy of 96.3% and 81.5%, respectively, and a specificity of 94.4% and 72.2%, respectively. Retraining the 2D CNN model using the lower but matched resolution images recovered the accuracy to 100% and 92.6% and the specificity to 100% and 77.8% for the 5 and 10 mm smoothed data, respectively. Conclusion: A 2D CNN was developed, trained and tested and showed very promising classification performance for amyloid PET scans in a SCD patient cohort. However, decreased image quality/image resolution impacts the performance of the 2D-CNN, which can partly be resolved by retraining the model using the lower and matched resolution images. References: none
Aim/Introduction: Dopamine transporter SPECT imaging with I-123-FP-CIT allows for visualisation of dysfunction of the dopaminergic system, which is characteristic of Parkinson’s Disease (PD). Interpretation of scans based on visual assessment and semi-quantitative analysis imposes limitations as the latter requires a site-specific reference database that is often not available. Our aim was to develop a machine learning (ML)-based approach for interpretation of I-123-FP-CIT scans and determine its added value in clinical practice. Materials and Methods: We retrospectively included a consecutive cohort of 130 patients that underwent I-123-FP-CIT SPECT imaging (Discovery D670, GE Healthcare) and had a clinically confirmed diagnosis. Patients were labelled as either having PD or a diagnosis other than PD (non-PD) and divided into a training set (58 PD, 32 non-PD) and validation set (25 PD, 15 non-PD) using stratified random sampling. The training set was used to build a linear support vector machine (SVM) classifier to discriminate PD from non-PD using I-123-FP-CIT striatal uptake ratios, age and gender as input features. Ratios were obtained by means of semi-quantitative analysis (Xeleris 4.0, GE Healthcare) and comprised specific binding in striatum, caudate nucleus and putamen as well as a putamen/caudate index for both left and right hemisphere. A stratified, 10-times repeated 10-fold cross-validation was conducted to perform model optimization using mean accuracy and F1-score as evaluation measures. Subsequently, the derived SVM model was tested on the validation set. I-123-FP-CIT scans and corresponding ratios of the validation set were scored as either PD or non-PD by two expert nuclear medicine physicians following European guidelines. Overread from a third expert was performed in case of disagreement. Next, their prediction performance was compared to that of the SVM model. Results: The highest mean prediction accuracy and F1-score as found by cross-validation were 94.3% and 0.956, respectively. Testing the derived SVM model on the validation set, an accuracy of 95.0%, sensitivity of 96.0% and specificity of 93.3% were obtained. Prediction performance did not differ from visual assessment of PD, obtaining an equivalent accuracy, sensitivity and specificity of 95.0%, 96.0% and 93.3% (p > 0.99), respectively. Conclusion: ML-based interpretation of I-123-FP-CIT scans results in accurate discrimination of PD from non-PD identical to standard visual assessment, thereby encouraging implementation of this SVM model as diagnostic aid in clinical practice. References: none
Aim/Introduction: Amyotrophic lateral sclerosis (ALS) is a lethal degeneration of upper and lower motor neurons. Median survival, from onset to death, is extremely variable and, to date, no reliable prognostic factors have been validated. Recent literature suggests a primary role of skeletal muscle and its derived signals in ALS progression. To verify this hypothesis, in the present study we evaluated whether ALS is associated with alteration in FDG uptake and total volume of both psoas as a sample of skeletal muscles already used for prognostic stratification in different disorders. Materials and Methods: We analyzed 54 ALS patients with spinal onset consecutively submitted to PET/CT imaging. These data were compared with the corresponding findings in an age-and sex-matched control population. A computational 3D method was used to extract psoas muscle’s volumes from CT images and to evaluate total muscle volume and attenuation coefficient. In co-registered PET images, FDG accumulation was defined by average standardized uptake value (a-SUV) and the heterogeneity of its distribution expressed by SUV standard deviation (SUV-SD). Psoas volume was normalized for the ideal body weight. Results: Average psoas attenuation coefficient was similar in patients and controls (39±9 vs 38±11 Housfield units, respectively, ns). By contrast, ALS was associated with a significant reduction in Psoas volume (221±74 mL vs 262±85 mL; p<0.01). This difference persisted when normalization for ideal body weight was considered (3.54±1.02 vs 4.12±1.33 mL/Kg; p<0.05). Similarly, at PET imaging, a-SUV was significantly lower in patients than in controls (0.77±0.21 vs 0.90±0.18; p<0.01). Finally, heterogeneity of psoas a-SUV, expressed by SUV-SD, predicted overall survival rate at Kaplan-Meyer analysis (p<0.05) with a predictive power that was confirmed by univariate as well as by multivariate Cox analysis (p<0.02). Conclusion: ALS is associated with a reduction in volume and FDG uptake of psoas muscles. The heterogeneity of glucose metabolism within this muscular district is related to disease aggressiveness. References: none
Aim/Introduction: 123I-meta-iodobenzylguanidine (mIBG) has been used in neurology, and most of the studies have used a heart-to-mediastinal average count ratio (HMR) for diagnosis of Lewy-body diseases. This study aimed to determine abnormal mIBG scan with low cardiac uptake, which is typically observed in Lewy-body diseases, without specifying regions of interest (ROI), using machine learning algorithm. Materials and Methods: Anterior 123I-mIBG images with normal uptake (n=72, 148 scans) and low uptake (n=55, 110 scans) were used as the training database of machine learning including neural network (Mathematica, Wolfram Research). Anterior planar mIBG images (20 minutes and 3 hours) were directly input without any preparations, and the probability of abnormality was judged from 0 (abnormal uptake or absent uptake) to 1 (normal uptake). In addition to visual interpretation of normal/abnormal, conventional quantitation method of HMR was calculated using heart and mediastinum ROI. Accuracy of determining normal or abnormal uptake was examined using receiver operating characteristic (ROC) analysis. The accuracy of the model was also tested with additional consecutive patients with Lewy-body diseases, amyloidosis, and other diseases (n=25, 50 scans; 15 scans with HMR<2.0). Results: The best classifier among those we studied was constructed by an artificial neural network consisting of 10 layers (256x256-matrix grayscale images for encoder, two classes of normal and abnormal as output). Probability of abnormality was 0.997±0.013 in patients with abnormal or low uptake, and 0.003±0.001 in patients with normal uptake (t ratio=663, p<0.0001). A dataset of 196 scans were randomly selected for the training, and 60 scans (abnormal/normal = 30/30) were used as the validation set. This processing was repeated 10 times. Average ROC area under the curve was 0.97 (range 0.93-1.00), accuracy 0.91 (0.87-0.97), true positive rate 0.88 (0.80-0.97), and true negative rate 0.93 (0.90-0.97). Calculated HMR was 2.71±0.70 in normal cases, and 1.78±0.71 in abnormal cases (t ratio=9, p<0.001), and some overlap of ranges was observed for HMR. In additional patients for the validation, when thresholds of HMR=2.0 and probability of 0.5 were used for contingency analysis, 86% of the data showed complete agreement (Pearson X2=23, p<0.0001). Conclusion: Machine learning with appropriately trained neural network successfully classified mIBG images into normal and abnormal scans. Since anterior image was the only input without predefined ROI on the heart and mediastinum, machine learning could be a novel approach to classify patients with Lewy-body diseases. References: None
Aim/Introduction: Comparison, longitudinal assessment, and quantification of multi-center PET data are challenging mainly due to differences in scanner resolutions. Harmonization methods reduce inter-center variability and facilitate comparisons of such PET data, typically by smoothing the data to match a lower target resolution across used scanners. In this work, we investigated the impact of harmonizing ADNI-PET data on the classification of FDG and amyloid images into Alzheimer’s (AD) and normal-controls (NC) using a convolutional-neural-network (CNN). Materials and Methods: Longitudinal FDG and AV45 datasets corresponding to 479 subjects (222-AD, 257-NC) were obtained from the ADNI database for a total of 801 FDG (343-AD, 458-NC) and 654 AV45 (168-AD, 486-NC) datasets. For each dataset, both non-harmonized (summed image of co-registered dynamic frames in native space) and harmonized (processed image that included smoothing with scanner-specific filter to achieve 8mm FWHM), registered to MNI space, were separately used to train two identical CNNs designed to classify AD versus NC subjects. A 10-fold cross-validation scheme was used to evaluate classification performances. For each dataset, seven coronal slices (7x91x91) sparsely covering the brain were selected as CNN inputs, and went through four layers, each composed of a 2D convolution and a max-pooling operation. Outputs of these layers were aggregated and used for class prediction (AD or NC). Each CNN was trained to minimize the cross-entropy error using Adam optimizer with leaky ReLu activation for each layer, using a learning rate of 0.05 and dividing by 10 every 50 epochs until convergence. The classification performance using harmonized and non-harmonized data from the same subjects were compared and McNemar test was used to assess the effect of harmonization on classification performance. Results: Classification sensitivity, specificity, accuracy, and area-under-the-curve (AUC) were as follows: i) FDG (88.3%, 91.7%, 90.2%, and 0.956) for non-harmonized versus (87.1%, 90.8%, 89.3%, and 0.955) for harmonized data; and ii) AV45 (76.7%, 93.5%, 89.2%, and 0.921) for non-harmonized versus (78.2%, 89.9%, 87.0%, and 0.919) for harmonized data, respectively. No significant difference in class assignment consistency was found for FDG (McNemar p=0.90) but this difference was significant for AV45 (p=0.027). Conclusion: We demonstrated that deep learning potentially overcomes the variability in multi-center PET FDG and amyloid data for image-based classification of AD versus NC subjects using CNN. For both tracers accuracy was higher for non-harmonized data. We carefully suggest that harmonization of multi-center brain PET data is not required for accurate disease classification with deep learning. References: none
Aim/Introduction: 18F-fluoromethylcholine (FCho) has gained interest as a PET radiotracer in the assessment of gliomas, but full kinetic analysis has not yet been reported in this setting. Following its transport across the blood-brain-barrier (BBB), FCho is intracellularly phosphorylated by choline-kinase. Our aim is to characterize the uptake kinetics of FCho in high-grade-gliomas (HGG), as a required step for its rational use as a clinical diagnostic tool. Materials and Methods: Eighteen patients with suspected initial diagnosis of HGG underwent a 45min dynamic brain PET scan immediately after a FCho bolus injection, from which a VOI-based tumoral time-activity-curve was extracted. Plasma input functions were obtained using manual arterial blood sampling from a radial catheter, and parent fractions for metabolite correction were determined by thin-layer-chromatography. These data were fitted to several 1- and 2-tissue-compartment models, the best of which was selected through the Akaike-Information-Criteria (AIC). Parametric images of the main kinetic rate constants were also generated for each patient. We finally assessed the correlation between the kinetic parameters and the tumor SUV on static images, both for VOI-based and voxel-based approaches. Results: The 2-tissue-compartment model accounting for blood volume fraction (vB) yielded the lowest AIC score in all patients and was selected as the best model. The reversible step parameter k4 was either small compared to the other exchange rate constants (in 12/18 patients) or even fitted to zero (in 6/18 patients): K1 [median(range): 0.14 (0.06-0.30)ml·cm-3·min-1], k2 [0.11 (0.01-0.28)min-1], k3 [0.16 (0.00-0.46)min-1 ], k4 [0.02 (0-0.03)min-1], vB [0.08 (0.04-0.15)]. The perfusion/BBB transport rate constant K1 and the influx macroparameter Ki =K1*k3 /(k2+k3) were clearly the most identifiable parameters and both showed a positive linear correlation (r2=0.66 and 0.69 respectively) with the SUVmax of the tumor on the VOI-based analysis. Parametric images of Ki were obtained through a voxel-based Patlak analysis, and a good correlation was shown between these Ki parametric images and SUV images in each patient (r2=0.95[median]; 0.62-0.99[range]). Following surgery, a glioblastoma was confirmed in 16 patients, while the other two had a diagnosis of metastasis. Conclusion: FCho uptake kinetics in HGG is best explained by an irreversible or near irreversible 2-tissue-compartment model with blood volume fraction vB. The good linear relationship between the SUV and the K1 rate constant suggests a strong influence of the perfusión/BBB transport step on FCho tumor uptake. Clinical routine SUV images may serve as an adequate surrogate for the Ki influx rate parametric images. References: none
Aim/Introduction: Aromatase is an enzyme that converts androgens to estradiol and estrone in the brain and it is considered to play a role in different personality traits such as aggression and sexual behaviour. The aim of this study was to evaluate tracer kinetic models, as well as effect of scan duration on outcome parameters, for quantitative analysis of the novel aromatase PET ligand [11C]cetrozole. Materials and Methods: Data from ten subjects, with three 90 min dynamic [11C]Cetrozole PET scans each at baseline and after two different challenges, were included in this study. Arterial blood was sampled for measurement of blood radioactivity and metabolite analysis, to obtain a plasma-input function, in eight of the scans. VOI-based analysis was performed using single-tissue (1TCM) and two-tissue (2TCM) reversible plasma-input compartment models and Logan graphical analysis, in the eight scans with plasma data. In addition, simplified reference tissue models (SRTM) and reference Logan analysis were performed for all thirty scans with cerebellum as reference region. The optimal reference model was used for evaluation of decreased scan duration of 60 min. Five VOIs were included in the evaluation; thalamus, hypothalamus, putamen and raphae, obtained from a probabilistic VOI template on co-registered T1-MRI images, and amygdala, obtained by manually drawing a 70% isocontour VOI on the uptake images. Correlation and agreement between the plasma-input and reference methods, as well as for the full and decreased scan duration, were assessed by linear regression. Results: 2TCM performed better than 1TCM according to Akaike criteria in all cases except one, but was not able to robustly determine individual parameters. Plasma Logan distribution value ratio (DVR) agreed well with the 2TCM DVR and was used for validation of the reference models. Correlation and agreement between plasma Logan DVR-1 and reference Logan DVR-1 was very high (R2=1.00, slope=1.00) and slightly lower for SRTM binding potential (BPND) values (R2=0.96, slope=0.86). Reference Logan was used for analysing the shortened data sets showing a high correlation and agreement for DVR-1 values between the full scan length and the 60 min data sets (R2=0.98, slope=0.94). Conclusion: Reference Logan DVR-1 and SRTM BPND values were highly correlated with plasma Logan DVR-1. Reference Logan showed a higher agreement to the plasma-input model and was chosen as the optimal reference model for VOI-based analysis. Reference Logan generated robust and quantitatively accurate results for a shortened scan, indicating that 60 min scan duration is sufficient for [11C]cetrozole PET. References:
Aim/Introduction: [11C]UCB-J is a novel radioligand that binds to synaptic vesicle glycoprotein 2A (SV2A) which is present in neuronal presynaptic terminals. [11C]UCB-J Positron emission tomography (PET) imaging enables quantification of brain synaptic density in individual regions and may be used to study response to therapies in clinical trials. The main objective of this study is to determine the 28-day test-retest repeatability (TRT) of quantitative [11C]UCB-J brain PET imaging in Alzheimer’s disease (AD) patients and healthy controls. Materials and Methods: In this ongoing multicentre study 9 healthy and 5 AD subjects have been included (age 64.2±5.9, 7 male - 7 female). Subjects underwent two 60 minutes dynamic [11C]UCB-J PET scans (radiotracer dose of 370±58 MBq) with an interval of 28 days. Arterial blood sampling and metabolite analysis were performed to generate a metabolite corrected plasma input function. Various compartmental models +/- blood volume correction (VB) have been evaluated and the optimal model(s) has been assessed using Akaike criterion. Volumes of distribution (VT) and tracer delivery (K1) have been estimated using both 1T2k_VB and 2T4k_VB models. Data were also analyzed using a simplified reference tissue model (SRTM) with centrum semiovale (SO, white matter) (Koole et al., EJNMMI 2019) as reference region, providing estimates of BPND. We also examined the effect of plasma input model choice on model preferences, parameter estimation and TRT. Results: After intravenous injection, [11C]UCB-J showed rapid kinetics. Based on AIC, both 1T2k_VB and 2T4k_VB described the [11C]UCB-J kinetics equally well. VT obtained from both models were similar, suggesting that a 1T2k_VB model can be used to assess the in vivo kinetics. No significant difference in semiovale VT between AD and healthy subjects were observed (p>0.05). Preliminary analysis showed that the mean TRT for VT and SRTM BPND were -5.7 %±9.7 and -3.2%±14.5, respectively, for all subjects across all the different brain regions. Whole brain grey matter TRT for VT and SRTM BPND were -3.4%±4.3 and -2.6%±6.0 respectively. Conclusion: [11C]UCB-J kinetics can be well described by a plasma input single tissue compartment model and reliable fits can be obtained with a 60 min scan duration for both plasma input and reference tissue models, consistent with earlier reports (Finnema et al., JCBFM 2018). Our preliminary analysis shows TRT performance for VT and BPND of <10% and <15% (1 SD), respectively, across brain regions and indicates adequate repeatability of [11C]UCB-J PET for application in longitudinal clinical research. References: None
Aim/Introduction: To introduce ViQuant, a fully-automated MATLAB-based processing pipeline for PET/MRI to non-invasively determine the cerebral metabolic rate of glucose (CMRGlc) images. Materials and Methods: The ViQuant pipeline requires the following inputs: (1) PET list-mode data, (2) attenuation correction (AC) map, (3) MR angiography (MRA) images, (4) T1-w MRI and (5) MR navigators. No other user interaction is required. The processing pipeline includes three main components: (i) a component (IC) generating an image-derived input function (IDIF) , (ii) a component that creates MRGlc images(QC) and (iii) a report-generating component (RC). The IC component calculates an IDIF by first defining a volume-of-interest through automated segmentation of the MRA. This is followed by MR navigator-based motion correction (MC) and an MR-based partial volume correction (PVC). In addition, AC maps are aligned with the PET emission data prior to reconstruction. Once an accurate IDIF is calculated, it is used to generate MRGlc images based on a voxel-wise Patlak analysis. Finally, a report is generated with a list of regional MRGlc values.To validate the pipeline, 10 healthy volunteers underwent [18F]FDG test-retest PET/MRI examinations in an integrated PET/MR (Siemens Biograph mMR). The imaging protocol consisted of a 60 min list-mode PET acquisition, with parallel MR acquisitions consisting of MRA, MR navigators and T1-w MR. Arterial blood samples (AIF) were collected as reference standard. Pseudo-CT images derived from T1-w MR were used for AC. Quantification accuracy of non-invasive determination of MRGlc was assessed against the reference standard of MRGlc values derived using arterially sampled blood (AIF) based on the absolute percentage difference (APD) in regional MRGlc values determined in the following 6 brain regions: corpus callosum (CC), brainstem (BS), cerebellum (CB), thalamus (TH), anterior cingulate cortex (ACC) and the superior frontal cortex (SFC). Results: The APD between CMRGlc values obtained from AIF and IDIF were: (5.9 ±3.2%) for CC, (5.9 ±3.3%) for BS, (5.8 ±3.4%) for CB, (5.5 ±3.1%) for ACC, (5.8 ±3.14%) for TH and (5.9 ±3.3%) for the SFC. The total processing time was ~6h on a dedicated high-end PC. Conclusion: We have developed a fully-automated open-source processing pipeline which allows non-invasive determination of absolute MRGlc values in a clinical setting. The obtained CMRGlc values were validated to be within 6% of those determined using arterial sampling. At present, our processing pipeline is limited to data obtained from a Siemens PET/MR scanner due to its dependency on the e7 reconstruction tools. References: None
Aim/Introduction: F18-Flutemetamol/Vizamyl PET is a validated surrogate biomarker to reveal the amyloid brain status of patients presenting neurodegenerative disorders, either by visual or semi-quantitative analysis (SUVr). However, alternative tracers and quantification methods complicates the possibility to compare clinical results between centers. In 2015, Klunk proposed a methodological approach based on C11-PIB PET to harmonize and reports semi-quantitative results on the same reference level scale (the so-called centiloid scale). This approach was then endorsed by the Global Alzheimer’s Association Interactive Network (GAAIN) to help centers to calibrate and report their data whatever the type of Amyloid PET tracer and quantification method used. We report here the sequential calibration steps based on this approach to convert Flutemetamol SUVr to the centiloid scale using the PNEURO3.9 software.
Materials and Methods: The PNEURO 3.9 (PMOD,Zurich) maximum_probability_workflow based on 3DT1 MRI segmentation was used to retreat C11-PIB/FLUT PET scans and compute SUVrcentil. values (cortical centil.vois/whole cerebellum-reference). Two historical patients population were downloaded from the GAAIN website to qualify the PNEURO3.9 workflow. Step1:PMOD3.9 workflow validation for C11-PIB PET: n=79 (34 YHC-45 AD;C11-PIB/MRI; GAAIN, Klunk). Step2:FLUT versus C11-PIB SUVrcentil. comparison and convertion of SUVrcentil. to centiloid scale values; n=74 (24 YHC&50 AD; F18-Flutemetamol/C11-PIB/MRI; GAAIN, GE Healthcare).Statistical analysis and data plot fitting were computed using PRISM 8.0. Results: Step1 population analysis performed by PNEURO3.9 demonstrated an excellent correlation between the PMOD3.9 Centiloid scale values derived from the C11-PIB SUVrcentil. compared to the Klunk reported data (R2: 0.9988;Y=1.004X-0.06600), validating the PMOD3.9 workflow based on the author criteria (R2 >0.98; slope/intercept between 0.98;1.02/-2;2). For the step2 population, an excellent correlation was found between the PMOD 3.9 and published SUVrcentil. for the n=74 pairs of F18-FLUT/C11-PIB (R2: 0.9945/0.9959). The correlation between the F18-FLUT and C11-PIB SUVrcentil. was also very similar to those reported (R2:0.95 vs 0.96). The PMOD3.9 derived-equation to convert FLUT SUVrcentil. to centiloid scale values was Centil.val=115.4xFLUTSUVr-113.2 (R2:0.9501) compared to Centil.val=116.5xFLUTSUVr-114.7 (R2:0.9617).
Conclusion: PNEURO 3.9 workflow seems a robust method to reproduced GAAIN data and derived centiloid scale values from Flutemetamol SUVrcentil. values.
References: Klunk et al. Alzheimers Dement, 2015. http://www.gaain.org/centiloid-project