3D PET Image Research Articles

Kinetic model-informed deep learning for multiplexed PET image separation

BackgroundMultiplexed positron emission tomography (mPET) imaging can measure physiological and pathological information from different tracers simultaneously in a single scan. Separation of the multiplexed PET signals within a single PET scan is challenging due to the fact that each tracer gives rise to indistinguishable 511 keV photon pairs, and thus no unique energy information for differentiating the source of each photon pair.MethodsRecently, many applications of deep learning for mPET image separation have been concentrated on pure data-driven methods, e.g., training a neural network to separate mPET images into single-tracer dynamic/static images. These methods use over-parameterized networks with only a very weak inductive prior. In this work, we improve the inductive prior of the deep network by incorporating a general kinetic model based on spectral analysis. The model is incorporated, along with deep networks, into an unrolled image-space version of an iterative fully 4D PET reconstruction algorithm.ResultsThe performance of the proposed method was evaluated on a simulated brain image dataset for dual-tracer [18\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$^{18}$$\\end{document}F]FDG+[11\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$^{11}$$\\end{document}C]MET PET image separation. The results demonstrate that the proposed method can achieve separation performance comparable to that obtained with single-tracer imaging. In addition, the proposed method outperformed the model-based separation methods (the conventional voxel-wise multi-tracer compartment modeling method (v-MTCM) and the image-space dual-tracer version of the fully 4D PET image reconstruction algorithm (IS-F4D)), as well as a pure data-driven separation [using a convolutional encoder-decoder (CED)], with fewer training examples.ConclusionsThis work proposes a kinetic model-informed unrolled deep learning method for mPET image separation. In simulation studies, the method proved able to outperform both the conventional v-MTCM method and a pure data-driven CED with less training data.

A quantitative assessment of Geant4 for predicting the yield and distribution of positron-emitting fragments in ion beam therapy

Objective. To compare the accuracy with which different hadronic inelastic physics models across ten Geant4 Monte Carlo simulation toolkit versions can predict positron-emitting fragments produced along the beam path during carbon and oxygen ion therapy. Approach. Phantoms of polyethylene, gelatin, or poly(methyl methacrylate) were irradiated with monoenergetic carbon and oxygen ion beams. Post-irradiation, 4D PET images were acquired and parent 11C, 10C and 15O radionuclides contributions in each voxel were determined from the extracted time activity curves. Next, the experimental configurations were simulated in Geant4 Monte Carlo versions 10.0 to 11.1, with three different fragmentation models—binary ion cascade (BIC), quantum molecular dynamics (QMD) and the Liege intranuclear cascade (INCL++) - 30 model-version combinations. Total positron annihilation and parent isotope production yields predicted by each simulation were compared between simulations and experiments using normalised mean squared error and Pearson cross-correlation coefficient. Finally, we compared the depth of the maximum positron annihilation yield and the distal point at which the positron yield decreases to 50% of peak between each model and the experimental results. Main results. Performance varied considerably across versions and models, with no one version/model combination providing the best prediction of all positron-emitting fragments in all evaluated target materials and irradiation conditions. BIC in Geant4 10.2 provided the best overall agreement with experimental results in the largest number of test cases. QMD consistently provided the best estimates of both the depth of peak positron yield (10.4 and 10.6) and the distal 50%-of-peak point (10.2), while BIC also performed well and INCL generally performed the worst across most Geant4 versions. Significance. The best predictions of the spatial distribution of positron annihilations and positron-emitting fragment production along the beam path during carbon and oxygen ion therapy was obtained using Geant4 10.2.p03 with BIC or QMD. These version/model combinations are recommended for future heavy ion therapy research.

Machine learning techniques based on 18F-FDG PET radiomics features of temporal regions for the classification of temporal lobe epilepsy patients from healthy controls.

This study aimed to investigate the clinical application of 18F-FDG PET radiomics features for temporal lobe epilepsy and to create PET radiomics-based machine learning models for differentiating temporal lobe epilepsy (TLE) patients from healthy controls. A total of 347 subjects who underwent 18F-FDG PET scans from March 2014 to January 2020 (234 TLE patients: 25.50 ± 8.89 years, 141 male patients and 93 female patients; and 113 controls: 27.59 ± 6.94 years, 48 male individuals and 65 female individuals) were allocated to the training (n = 248) and test (n = 99) sets. All 3D PET images were registered to the Montreal Neurological Institute template. PyRadiomics was used to extract radiomics features from the temporal regions segmented according to the Automated Anatomical Labeling (AAL) atlas. The least absolute shrinkage and selection operator (LASSO) and Boruta algorithms were applied to select the radiomics features significantly associated with TLE. Eleven machine-learning algorithms were used to establish models and to select the best model in the training set. The final radiomics features (n = 7) used for model training were selected through the combinations of the LASSO and the Boruta algorithms with cross-validation. All data were randomly divided into a training set (n = 248) and a testing set (n = 99). Among 11 machine-learning algorithms, the logistic regression (AUC 0.984, F1-Score 0.959) model performed the best in the training set. Then, we deployed the corresponding online website version (https://wane199.shinyapps.io/TLE_Classification/), showing the details of the LR model for convenience. The AUCs of the tuned logistic regression model in the training and test sets were 0.981 and 0.957, respectively. Furthermore, the calibration curves demonstrated satisfactory alignment (visually assessed) for identifying the TLE patients. The radiomics model from temporal regions can be a potential method for distinguishing TLE. Machine learning-based diagnosis of TLE from preoperative FDG PET images could serve as a useful preoperative diagnostic tool.

Non-rigid motion-compensated 3D whole-heart T2 mapping in a hybrid 3T PET-MR system.

Simultaneous PET-MRI improves inflammatory cardiac disease diagnosis. However, challenges persist in respiratory motion and mis-registration between free-breathing 3D PET and 2D breath-held MR images. We propose a free-breathing non-rigid motion-compensated 3D T2 -mapping sequence enabling whole-heart myocardial tissue characterization in a hybrid 3T PET-MR system and provides non-rigid respiratory motion fields to correct also simultaneously acquired PET data. Free-breathing 3D whole-heart T2 -mapping was implemented on a hybrid 3T PET-MRI system. Three datasets were acquired with different T2 -preparation modules (0, 28, 55 ms) using 3-fold undersampled variable-density Cartesian trajectory. Respiratory motion was estimated via virtual 3D image navigators, enabling multi-contrast non-rigid motion-corrected MR reconstruction. T2 -maps were computed using dictionary-matching. Approach was tested in phantom, 8 healthy subjects, 14 MR only and 2 PET-MR patients with suspected cardiac disease and compared with spin echo reference (phantom) and clinical 2D T2 -mapping (in-vivo). Phantom results show a high correlation (R2 = 0.996) between proposed approach and gold standard 2D T2 mapping. In-vivo 3D T2 -mapping average values in healthy subjects (39.0 ± 1.4 ms) and patients (healthy tissue) (39.1 ± 1.4 ms) agree with conventional 2D T2 -mapping (healthy = 38.6 ± 1.2 ms, patients = 40.3 ± 1.7 ms). Bland-Altman analysis reveals bias of 1.8 ms and 95% limits of agreement (LOA) of -2.4-6 ms for healthy subjects, and bias of 1.3 ms and 95% LOA of -1.9 to 4.6 ms for patients. Validated efficient 3D whole-heart T2 -mapping at hybrid 3T PET-MRI provides myocardial inflammation characterization and non-rigid respiratory motion fields for simultaneous PET data correction. Comparable T2 values were achieved with both 3D and 2D methods. Improved image quality was observed in the PET images after MR-based motion correction.

Improved diagnostic accuracy for coronary artery disease detection with quantitative 3D 82Rb PET myocardial perfusion imaging.

To establish requirements for normal databases for quantitative rubidium-82 (82Rb) PET MPI analysis with contemporary 3D PET/CT technology and reconstruction methods for maximizing diagnostic accuracy of total perfusion deficit (TPD), a combined metric of defect extent and severity, versus invasive coronary angiography. In total, 1571 patients with 82Rb PET/CT MPI on a 3D scanner and stress static images reconstructed with and without time-of-flight (TOF) modeling were identified. An additional eighty low pre-test probability of disease (PTP) patients reported as normal were used to form separate sex-stratified and sex-independent iterative and TOF normal databases. 3D normal databases were applied to matched patient reconstructions to quantify TPD. Per-patient and per-vessel performance of 3D versus 2D PET normal databases was assessed with receiver operator characteristic curve analysis. Diagnostic accuracy was evaluated at optimal thresholds established from PTP patients. Results were compared against logistic regression modeling of TPD adjusted for clinical variables, and standard clinical interpretation. TPD diagnostic accuracy was significantly higher using 3D PET normal databases (per-patient: 80.1% for 3D databases, versus 74.9% and 77.7% for 2D database applied to iterative and TOF images respectively, p < 0.05). Differences in male and female normal distributions for 3D attenuation-corrected reconstructions were not clinically meaningful; therefore, sex-independent databases were used. Logistic regression modeling including TPD demonstrated improved performance over clinical reads. Normal databases tailored to 3D PET images provide significantly improved diagnostic accuracy for PET MPI evaluation with automated quantitative TPD. Clinical application of these techniques should be considered to support accurate image interpretation.

Contrastive Learning for Prediction of Alzheimer's Disease Using Brain 18F-FDG PET.

Brain 18F-FDG PET images are commonly-known materials for effectively predicting Alzheimer's disease (AD). How-ever, the data volume of PET is usually insufficient, which is unfavorable to train an accurate AD prediction networks. Fur-thermore, the PET image is noisy with low signal-to-noise ratio, and simultaneously the feature (metabolic abnormality) used for predicting AD in PET image is not always obvious. Such charac-teristics of 18F-FDG PET images hinder the existing deep learning networks to learn the feature of lesion (i.e., glucose metabolism abnormality) effectively, which leads to unsatisfactory classifica-tion performance and poor robustness. In this paper, a contrastive-based learning method is proposed to address the challenges of PET image inherently possessed. Firstly, the slices of 3D PET image are amplified by cropping the image of anchors (i.e., an augmented version of the same image) to generate extended train-ing data. Meanwhile, contrastive loss is adopted to enlarge inter-class feature distances and reduce intra-class feature differences using subject fuzzy labels as supervised information. Secondly, we construct a double convolutional hybrid attention module to enhance the network to learn different perceptual domains where two convolutional layers with different convolutional kernels (7 × 7 and 5 × 5) are constructed. Moreover, we recommend a diagnosis mechanism by analyzing the consistency of predicted result for PET slices alone with clinical neuropsychological assessment to achieve a better AD diagnosis. The experimental results show that the proposed method outperforms the state-of-the-arts for brain 18F-FDG PET images while remaining satisfactory computational performance, and hence demonstrate the advantage of the method in effectively predicting AD.

An Investigation of Stochastic Variance Reduction Algorithms for Relative Difference Penalized 3D PET Image Reconstruction.

Penalised PET image reconstruction algorithms are often accelerated during early iterations with the use of subsets. However, these methods may exhibit limit cycle behaviour at later iterations due to variations between subsets. Desirable converged images can be achieved for a subclass of these algorithms via the implementation of a relaxed step size sequence, but the heuristic selection of parameters will impact the quality of the image sequence and algorithm convergence rates. In this work, we demonstrate the adaption and application of a class of stochastic variance reduction gradient algorithms for PET image reconstruction using the relative difference penalty and numerically compare convergence performance to BSREM. The two investigated algorithms are: SAGA and SVRG. These algorithms require the retention in memory of recently computed subset gradients, which are utilised in subsequent updates. We present several numerical studies based on Monte Carlo simulated data and a patient data set for fully 3D PET acquisitions. The impact of the number of subsets, different preconditioners and step size methods on the convergence of regions of interest values within the reconstructed images is explored. We observe that when using constant preconditioning, SAGA and SVRG demonstrate reduced variations in voxel values between subsequent updates and are less reliant on step size hyper-parameter selection than BSREM reconstructions. Furthermore, SAGA and SVRG can converge significantly faster to the penalised maximum likelihood solution than BSREM, particularly in low count data.

Multi-task multi-scale learning for outcome prediction in 3D PET images

Background and Objectives:Predicting patient response to treatment and survival in oncology is a prominent way towards precision medicine. To this end, radiomics has been proposed as a field of study where images are used instead of invasive methods. The first step in radiomic analysis in oncology is lesion segmentation. However, this task is time consuming and can be physician subjective. Automated tools based on supervised deep learning have made great progress in helping physicians. However, they are data hungry, and annotated data remains a major issue in the medical field where only a small subset of annotated images are available. Methods:In this work, we propose a multi-task, multi-scale learning framework to predict patient’s survival and response. We show that the encoder can leverage multiple tasks to extract meaningful and powerful features that improve radiomic performance. We also show that subsidiary tasks serve as an inductive bias so that the model can better generalize. Results:Our model was tested and validated for treatment response and survival in esophageal and lung cancers, with an area under the ROC curve of 77% and 71% respectively, outperforming single-task learning methods. Conclusions:Multi-task multi-scale learning enables higher performance of radiomic analysis by extracting rich information from intratumoral and peritumoral regions.

Respiratory 4D-Gating F-18 FDG PET/CT Scan for Liver Malignancies: Feasibility in Liver Cancer Patient and Tumor Quantitative Analysis.

PurposeTo evaluate the potential clinical role and effectiveness of respiratory 4D-gating F-18 FDG PET/CT scan for liver malignancies, relative to routine (3D) F-18 FDG PET/CT scan.Materials and MethodsThis study presented a prospective clinical study of 16 patients who received F-18 FDG PET/CT scan for known or suspected malignant liver lesions. Ethics approvals were obtained from the ethics committees of the Hong Kong Baptist Hospital and The Hong Kong Polytechnic University. Liver lesions were compared between the gated and ungated image sets, in terms of 1) volume measurement of PET image, 2) accuracy of maximum standardized uptake value (SUVmax), mean standardized uptake value (SUVmean), and 3) accuracy of total lesion glycoses (TLG). Statistical analysis was performed by using a two-tailed paired Student t-test and Pearson correlation test.ResultsThe study population consisted of 16 patients (9 males and 7 females; mean age of 65) with a total number of 89 lesions. The SUVmax and SUVmean measurement of the gated PET images was more accurate than that of the ungated PET images, compared to the static reference images. An average of 21.48% (p < 0.001) reduction of the tumor volume was also observed. The SUVmax and SUVmean of the gated PET images were improved by 19.81% (p < 0.001) and 25.53% (p < 0.001), compared to the ungated PET images.ConclusionsWe have demonstrated the feasibility of implementing 4D PET/CT scan for liver malignancies in a prospective clinical study. The 4D PET/CT scan for liver malignancies could improve the quality of PET image by improving the SUV accuracy of the lesions and reducing image blurring. The improved accuracy in the classification and identification of liver tumors with 4D PET image would potentially lead to its increased utilization in target delineation of GTV, ITV, and PTV for liver radiotherapy treatment planning in the future.

Evaluation and optimization of radiomics features stability to respiratory motion in 18F‐FDG 3D PET imaging

To evaluate the impact of respiratory motion on radiomics features in 18 F-fluoro-2-deoxy-D-glucose three dimensional positron emission tomography (18 F-FDG 3D PET) imaging and optimize feature stability by combining preprocessing configurations and aggregation strategies. An in-house developed respiratory motion phantom was imaged in 3D PET scanner under nine respiratory patterns including one reference pattern. In total, 487radiomics features were extracted for each respiratory pattern. Feature stability to respiratory motion was first evaluated by metrics of coefficient of variation (COV) and relative difference (RD) in a fixed preprocessing configuration. Further, one-way ANOVA and trend analysis were performed to evaluate the impact of preprocessing configuration (voxel size, discretization scheme) and aggregation strategy on feature stability. Finally, an optimization framework was proposed by selected feature-specific configurations with minimum COV value, and the diagnostic performance was validated in stable versus unstable features and fixed versus optimal features by a set of 46 patients with lung disease. PET radiomics features were sensitive to respiratory motion, only 79/487 (16%) features were identified to be very stable in the fixed configuration. Preprocessing configuration and aggregation strategy had an impact on feature stability. For different voxel size, bin number, bin size and aggregation strategy, 188/487 (39%), 70/487 (15%), 148/487 (30%), and 38/95 (29%) features appeared significant changes in feature stability. The optimized configuration had the potential to improve feature stability compared to fixed configuration, with the COV decreased from 22% ±24% to 16% ±20%. Regarding the diagnostic performance, the stable and optimal configuration features outperformed the unstable and fixed configuration features, respectively (AUC 0.88, 0.87 vs. 0.83, 0.85). Respiratory motion shows considerable impact on feature stability in 3D PET imaging, while optimizing preprocessing configuration may improve feature stability and diagnostic performance.

Innovative procedure for measuring left ventricular ejection fraction from 18F-FDG first-pass ultra-sensitive digital PET/CT images: evaluation with an anthropomorphic heart phantom

BackgroundLeft ventricular ejection fraction (LVEF) is usually measured by cine-cardiac magnetic resonance imaging (MRI), planar and single-photon emission-computerized tomography (SPECT) equilibrium radionuclide angiocardiography (ERNA), and echocardiography. It would be clinically useful to measure LVEF from first-pass positron-emission tomography/computed tomography (PET/CT) radionuclide angiography, but this approach has been limited by fast radiotracer diffusion. Ultra-sensitive digital PET systems can produce high-quality images within 3-s acquisition times. This study determined whether digital PET/CT accurately measured LVEF in an anthropomorphic heart phantom under conditions mimicking radiotracer first-pass into the cardiac cavities.MethodsHeart phantoms in end-diastole and end-systole were 3D-printed from a patient’s MRI dataset. Reference left ventricle end-diastole volume (EDV), end-systole volume (ESV), and LVEF were determined by phantom weights before/after water filling. PET/CT (3-s acquisitions), MRI, and planar and SPECT ERNA were performed. EDV, ESV, and/or LVEF were measured by manual and automated cardiac cavity delineation, using clinical segmentation softwares. LVEF was also measured from PET images converted to 2D “pseudo-planar” images along the short axis and horizontal long axis. LVEF was also calculated for planar ERNA images. All LVEF, ESV and EDV values were compared to the reference values assessed by weighing.ResultsManually calculated 3D-PET-CT-based EDV, ESV, and LVEF were close to MRI and reference values. Automated calculations on the 3D-PET-CT dataset were unreliable, suggesting that the SPECT-based tool used for this calculation is not well adapted for PET acquisitions. Manual and automated LVEF estimations from “pseudo-planar” PET images were very close/identical to MRI and reference values.ConclusionsFirst-pass “pseudo-planar” PET may be a promising method for estimating LVEF, easy to use in clinical practice. Processing 3D PET images is also a valid method but to date suffers from a lack of well-suited software for automated LV segmentation.

Computer-Aided Diagnosis System for Alzheimer's Disease Using Positron Emission Tomography Images.

Alzheimer's disease (AD) is a kind of neurological brain disease. It is an irretrievable, neurodegenerative brain disorder. There are no pills or drugs to cure AD. Therefore, an early diagnosis may help the physician to make accurate analysis and to provide better treatment. With the advent of computational intelligence techniques, machine learning models have made tremendous progress in brain images analysis using MRI, SPECT and PEI. However, accurate analysis of brain scans is an extremely challenging task. The main focus of this paper is to design a Computer Aided Diagnosis (CAD) system using Long-Term Short Memory (LSTM) to improve classification rate and determine suitable attributes that can differentiate AD from Healthy Control (HC) subjects. First, 3D PET images are preprocessed, converted into many groups of 2D images and then grouped into many subsets at certain interval. Subsequently, different features including first order statistical, Gray Level Co-occurrence Matrix and wavelet energy of all sub-bands are extracted from each group, combined and taken as feature vectors. LSTM is designed and employed for classifying PET brain images into HC and AD subjects based on the feature vectors. Finally, the developed system is validated on 18FDG-PET images collected from 188 subjects including 105 HC and 83 AD subjects from ADNI database. Efficacy of the developed CAD system is analyzed using different features. Numerical results revealed that the developed CAD system yields classification accuracy of 98.9% when using combined features, showing outstanding performance.

Automatic Diagnosis of Alzheimer's Disease and Mild Cognitive Impairment Based on CNN + SVM Networks with End-to-End Training.

Alzheimer's disease (AD) is an irreversible neurodegenerative disease, and, at present, once it has been diagnosed, there is no effective curative treatment. Accurate and early diagnosis of Alzheimer's disease is crucial for improving the condition of patients since effective preventive measures can be taken in advance to delay the onset time of the disease. 18F-Fluorodeoxyglucose positron emission tomography (18F-FDG PET : PET) is an effective biomarker of the symptom of AD and has been used as medical imaging data for diagnosing AD. Mild cognitive impairment (MCI) is regarded as an early symptom of AD, and it has been shown that MCI also has a certain biomedical correlation with PET. In this paper, we explore how to use 3D PET images to realize the effective recognition of MCI and thus achieve the early prediction of AD. This problem is then taken as the classification of three categories of PET images, including MCI, AD, and NC (normal controls). In order to get better classification performance, a novel network model is proposed in the paper based on 3D convolution neural networks (CNN) and support vector machines (SVM) by utilizing both the excellent abilities of CNN in feature extraction and SVM in classification. In order to make full use of the optimal property of SVM in solving binary classification problems, the three-category classification problem is divided into three binary classifications, and each binary classification is being realized with a CNN + SVM network. Then, the outputs of the three CNN + SVM networks are fused into a final three-category classification result. An end-to-end learning algorithm is developed to train the CNN + SVM networks, and a decision fusion algorithm is exploited to realize the fusion of the outputs of three CNN + SVM networks. Experimental results obtained in the work with comparative analyses confirm the effectiveness of the proposed method.

Unsupervised 3D PET-CT Image Registration Method Using a Metabolic Constraint Function and a Multi-Domain Similarity Measure

High-resolution CT images can clearly display anatomical structures but does not display functional information, while blurred PET images can display molecular and functional information of lesions but cannot clearly display morphological structures. Therefore, accurate PET-CT image registration, which is used for anatomical structure and functional information fusion, is a prerequisite for early stage cancer diagnosis. However, some hypermetabolic anatomical structures, such as brain and bladder, have low registration accuracy. To solve this problem, a 3D unsupervised network based on a metabolic constraint function and a multi-domain similarity measure (3D MC-MDS Net) is proposed for 3D PET-CT image registration. Specifically, a metabolic constraint model is established based on the standard uptake value (SUV) distribution of hypermetabolic regions such as brain, bladder, liver and heart, which reduces the excessive distortion on displacement vector field (DVF) caused by hypermetabolic anatomical structures in PET images. A DVF estimator is built based on 3D unsupervised convolutional neural networks and a spatial transformer is used for warping 3D PET images to 3D CT images. The generated registration results (PET image patches) and the original 3D CT image patches are used for calculating the spatial domain similarity (SD similarity) and frequency domain similarity (FD similarity). Finally, the loss function of the entire registration network is constructed by a weighted sum of SD similarity, FD similarity and a smoothness of DVF. A dataset consisted of 170 whole-body PET-CT images is used for registration accuracy evaluation. The proposed unsupervised registration network, 3D MC-MDS Net, can accurately learn the 3D registration model by using the training dataset with the metabolic constraint model, which significantly improves the registration accuracy.

Reducing Respiratory Motion-Induced Errors in PET and CT Using an Iterative Image-to-Image Reconstruction Algorithm

Respiratory motion may change SUVmax and SUVpeak in PET images and compromise target definition and treatment response assessment in radiation therapy. While 4D imaging techniques may help address these issues, the time required for 4D data acquisition, image reconstruction and administrative costs keep 3D PET imaging staying as the mainstream in clinical practice. The purpose of this study is to explore the potential of using 4DCT information to remove motion-induced errors in 3D PET/CT images. A novel iterative image-to-image (III) reconstruction algorithm was developed. Specifically, a motion-blurred PET image was considered as a weighted average of PET images acquired at different respiration phases. PET images at each phase were assumed to be deformed from a virtual reference image. Image-intensity-based B-Spline registrations were performed on 4DCT from the end-inhale to the other phases. Resultant deformation maps were used to establish a set of composition equations between the virtual reference image and the blurred PET image. A maximum likelihood estimation maximization (MLEM) algorithm was used to solve these equations. The solution of these equations was smoothed by minimizing total gradient variations before generating the virtual reference image. A 10-phase 4D dataset with the resolution of 3.64×3.64×3.24 for PET and1.37×1.37×2.15 for CT was acquired from a lung cancer patient and used to benchmark the computed results. Different from previous studies, the III-reconstruction was performed in the image space, so there is no any tracking device or sorting operation required during the PET image acquisition. For the blurred PET, III-reconstructed PET and PET0 (which is the 4D-PET at the end-inhale phase), their relative SUVpeak are 10.2, 12.7 and 12.8, and SUVmax are 9.2, 12.5 and 13.1. With PET0 taken as the baseline, the III-reconstruction reduced the motion-induced error from 20.3% to 0.8% for SUVpeak, and from 29.8% to 4.6% for SUVmax. The mean SUV difference between the III-reconstructed PET and PET0 was 1.8% in the tumor region, and 13.2% over the whole image domain. Large differences between the two PET images were located mainly in the chest wall where ribs were mis-registered about 6 mm due to discontinuous motion between the ribs and lungs. The III-reconstruction algorithm was also applied to the 4DCT image. Image contrast ratios for the blurred 3DCT, III-reconstructed CT, and CT0 are 1.8, 3.7 and 4.0, respectively. The III-reconstruction method can be used to correct motion-induced errors in SUVmax and SUVpeak and improve image contrasts in free-breathing PET and CT images. This method does not require sorting projection data, thereby reducing financing cost and clinical burden. Further investigation of this reconstruction method is warranted.

Detection and segmentation of lymphomas in 3D PET images via clustering with entropy-based optimization strategy.

Lymphoma detection and segmentation from PET images are critical tasks for cancer staging and treatment monitoring. However, it is still a challenge owing to the complexities of lymphoma PET data themselves, and the huge computational burdens and memory requirements for 3D volume data. In this work, an entropy-based optimization strategy for clustering is proposed to detect and segment lymphomas in 3D PET images. To reduce computational complexity and add more feature information, billions of voxels in 3D volume data are first aggregated into supervoxels. Then, such supervoxels serve as basic data units for further clustering by using DBSCAN algorithm, in which some new feature attributes based on physical spatial information and prior knowledge are proposed. In addition, more importantly, an entropy-based objective function is constructed to search the most appropriate parameters of DBSCAN to obtain the optimal clustering results by using a genetic algorithm. This step allows to automatically adapt the parameters to each patient. Finally, a series of comparison experiments among various feature attributes are performed. 48 patient data are conducted, showing the combination of three features, supervoxel intensity, geographic coordinates and organ distributions, can achieve good performance and the proposed entropy-based optimization scheme has more advantages than the existing methods. The proposed entropy-based optimization strategy for clustering by integrating physical spatial attributes and prior knowledge can achieve better performance than traditional methods.

Scatter Correction Based on GPU-Accelerated Full Monte Carlo Simulation for Brain PET/MRI.

Accurate scatter correction is essential for qualitative and quantitative PET imaging. Until now, scatter correction based on Monte Carlo simulation (MCS) has been recognized as the most accurate method of scatter correction for PET. However, the major disadvantage of MCS is its long computational time, which makes it unfeasible for clinical usage. Meanwhile, single scatter simulation (SSS) is the most widely used method for scatter correction. Nevertheless, SSS has the disadvantage of limited robustness for dynamic measurements and for the measurement of large objects. In this work, a newly developed implementation of MCS using graphics processing unit (GPU) acceleration is employed, allowing full MCS-based scatter correction in clinical 3D brain PET imaging. Starting from the generation of annihilation photons to their detection in the simulated PET scanner, all relevant physical interactions and transport phenomena of the photons were simulated on GPUs. This resulted in an expected distribution of scattered events, which was subsequently used to correct the measured emission data. The accuracy of the approach was validated with simulations using GATE (Geant4 Application for Tomography Emission), and its performance was compared to SSS. The comparison of the computation time between a GPU and a single-threaded CPU showed an acceleration factor of 776 for a voxelized brain phantom study. The speedup of the MCS implemented on the GPU represents a major step toward the application of the more accurate MCS-based scatter correction for PET imaging in clinical routine.

Automatic PET cervical tumor segmentation by combining deep learning and anatomic prior

Cervical tumor segmentation on 3D 18FDG PET images is a challenging task because of the proximity between cervix and bladder, both of which can uptake 18FDG tracers. This problem makes traditional segmentation based on intensity variation methods ineffective and reduces overall accuracy. Based on anatomy knowledge, including ‘roundness’ of the cervical tumor and relative positioning between the bladder and cervix, we propose a supervised machine learning method that integrates convolutional neural network (CNN) with this prior information to segment cervical tumors. First, we constructed a spatial information embedded CNN model (S-CNN) that maps the PET image to its corresponding label map, in which bladder, other normal tissue, and cervical tumor pixels are labeled as −1, 0, and 1, respectively. Then, we obtained the final segmentation from the output of the network by a prior information constrained (PIC) thresholding method. We evaluated the performance of the PIC-S-CNN method on PET images from 50 cervical cancer patients. The PIC-S-CNN method achieved a mean Dice similarity coefficient (DSC) of 0.84 while region-growing, Chan-Vese, graph-cut, fully convolutional neural networks (FCN) based FCN-8 stride, and FCN-2 stride, and U-net achieved 0.55, 0.64, 0.67, 0.71, 0.77, and 0.80 mean DSC, respectively. The proposed PIC-S-CNN provides a more accurate way for segmenting cervical tumors on 3D PET images. Our results suggest that combining deep learning and anatomic prior information may improve segmentation accuracy for cervical tumors.

Joint cardiac and respiratory motion estimation for motion-corrected cardiac PET-MR

Respiratory and cardiac motion can strongly impair cardiac PET image quality and tracer uptake quantification. Standard gating techniques can minimize these motion artefacts but suffer from low signal-to-noise ratio because only a small percentage of the total data is utilized. Motion correction approaches have been proposed to overcome this problem but require accurate knowledge of such physiological motion. Here we present a joint PET-MR motion estimation approach which combines complimentary dynamic image information from simultaneously acquired MR and PET to ensure improved cardiac and respiratory motion estimation for motion-corrected image reconstruction (MCIR) of PET images. A 3D triple-echo Dixon MR scan is used both for calculation of MR-based attenuation correction (AC) maps and estimation of physiological motion. PET listmode data is obtained simultaneously to the MR acquisition which is used for a joint motion estimation and reconstruction of the final MCIR PET. In a first step, dynamic cardiac and respiratory motion resolved 4D MR and PET images are reconstructed. These image series are used in a joint image registration to estimate non-rigid cardiac and respiratory motion fields. In a second step, the motion fields are utilized in a MR MCIR to obtain cardiac and respiratory resolved dynamic MR-based AC maps. In the last step, the non-rigid motion fields and the dynamic AC maps are applied in a PET MCIR to obtain the final motion-corrected PET images. PET-MR data has been obtained in six patients without any known heart disease. Motion amplitudes were between 5.6 and 16 mm, with higher values in the basal compared to the mid-ventricular and apical segments. The proposed joint PET-MR motion estimation provided more accurate motion estimation than using either modality separately. The underestimation of PET uptake due to respiratory and cardiac motion artefacts in the AC maps was up to 17%. The average increase in uptake values using MCIR was 23% ± 10% (p < 0.0001), with values of 28% ± 11% (p < 0.0001) for basal, 21% ± 8% (p < 0.0001) for mid-cavity and 17% ± 7% (p < 0.0001) for apical segments. With the proposed scheme we could ensure high PET image quality and improve local PET uptake quantification by up to 30%. Attenuation correction and motion information was obtained from the same PET-MR raw data, which was obtained during free-breathing to minimize scan times and to increase patient comfort.

A new methodology to derive 3D kinetic parametric FDG PET images based on a mathematical approach integrating an error model of measurement

BackgroundIn FDG-PET, SUV images are hampered by large potential biases. Our aim was to develop an alternative method (ParaPET) to generate 3D kinetic parametric FDG-PET images easy to perform in clinical oncology.MethodsThe key points of our method are the use of a new error model of PET measurement extracted from a late dynamic PET acquisition of 15 min, centered over the lesion and an image-derived input function (IDIF). The 15-min acquisition is reconstructed to obtain five images of FDG mean activity concentration and images of its variance to model errors of PET measurement. Our approach is carried out on each voxel to derive 3D kinetic parameter images. ParaPET was evaluated and compared to Patlak analysis as a reference. Hunter and Barbolosi methods (Barbolosi-Bl: with blood samples or Barbolosi-Im: with IDIF) were also investigated and compared to Patlak. Our evaluation was carried on Ki index, the net influx rate and its maximum value in the lesion (Ki,max).ResultsThis parameter was obtained from 41 non-small cell lung cancer lesions associated with 4 to 5 blood samples per patient, required for the Patlak analysis. Compare to Patlak, the median relative difference and associated range (median; [min;max]) in Ki,max estimates were not statistically significant (Wilcoxon test) for ParaPET (− 3.0%; [− 31.9%; 47.3%]; p = 0.08) but statistically significant for Barbolosi-Bl (− 8.0%; [− 30.8%; 53.7%]; p = 0.001), Barbolosi-Im (− 7.9%; [− 38.4%; 30.6%]; p = 0.007) or Hunter (32.8%; [− 14.6%; 132.2%]; p < 10− 5). In the Bland-Altman plots, the ratios between the four methods and Patlak are not dependent of the Ki magnitude, except for Hunter. The 95% limits of agreement are comparable for ParaPET (34.7%), Barbolosi-Bl (30.1%) and Barbolosi-Im (30.8%), lower to Hunter (81.1%). In the 25 lesions imaged before and during the radio-chemotherapy, the decrease in the FDG uptake (ΔSUVmax or ΔKi,max) is statistically more important (p < 0.02, Wilcoxon one-tailed test) when estimated from the Ki images than from the SUV images (additional median variation of − 2.3% [− 52.6%; + 19.1%] for ΔKi,max compared to ΔSUVmax).ConclusionNone of the four methodologies is yet ready to replace the Patlak approach, and further improvements are still required. Nevertheless, ParaPET remains a promising approach, offering a non-invasive alternative to methods based on multiple blood samples and only requiring a late PET acquisition. It allows deriving Ki values, highly correlated and presenting the lowest relative bias with Patlak estimates, in comparison to the other methods we evaluated. Moreover, ParaPET gives access to quantitative information at the pixel level, which needs to be evaluated in the perspective of radiomic and tumour response.Trial registrationNCT 02821936; May 2016.