Click on a content entry to get a more detailed information.
Deep Learning Models for PET Scatter Estimation (#4237)
H. Qian1, X. Rui1, B. De Man1
1 GE Global Research, Niskayuna, New York, United States of America
We developed two convolutional neural networks for PET scatter estimation. The first model estimates the multiple scatter distribution from the single scatter distribution in an iterative estimation process. The second model directly predicts the scatter of an object from the emission and attenuation sinograms of the object.
Keywords: Positron Emission Tomography, Scatter Estimation, Deep Learning, Convolutional Neural Network, System Simulations
An Exponential Model For Multiple Scatter in 3D PET (#3390)
I. Hong1, C. Michel1
1 Siemens Medical Solutions USA, Inc, Molecular Imaging, Knoxville, Tennessee, United States of America
Precise 3D scatter correction is essential in 3D PET to achieve quantitative imaging in clinical studies. This is especially true for large patients where multiple scatter constitutes a large fraction of the total scatter. The Single Scatter Simulation algorithm (SSS) was initially developed for 2D PET and single bed position imaging. Later, the method was extended to 3D PET: a relative estimation of the 3D scatter was scaled to the True projection tails, using a mask defined by all LOR’s outside the patient (where the attenuation is below a threshold). Neglecting multiple scatter underestimates the total scatter, especially in 3D. At first, this underestimation was roughly compensated by the use of an activity distribution uncorrected for scatter. Later, an iterative scheme was introduced and the total scatter was estimated from the scaling of the SSS, assuming that the radial scatter shape was identical at all orders. The tail fitting procedure is very sensitive to the projection noise for short frames with low statistics. So avoid tail fitting is highly desirable. This work exploits a new heuristic relationship found between the amplitude of the scatter contributions at all scattering orders. We derived this relationship from Monte-Carlo simulations (GATE) on several scanner models (mCT-4R, Horizon-3R) with multiple phantoms. The relationship holds as long as the lower discriminator of the energy qualification window is larger than 350 keV. It allows us to estimate the amplitude of single scatter by using LOR’s inside the patient (where ACF is larger than a threshold) and avoids tail fitting. This new scatter correction scheme relies only on the knowledge of SSS inside the patient.
Keywords: scatter, tail fitting, PSMA
Accounting for bone during PET/MRI attenuation correction of the pelvis: Preliminary results (#3006)
C. N. Ladefoged1, R. Taylor2, 3, U. Anazodo2, 3, J. D. Thiessen2, 3, M. Fenchel4, L. Højgaard1, F. L. Andersen1
1 Rigshospitalet, Department of Clinical Physiology, Nuclear Medicine and PET, London, Denmark
AIM:/strong> Bones are currently not compensated for during standard Dixon whole-body MR attenuation correction, which results in underestimation of the PET values in and near bone. We here introduce two new machine-learning-based methods that compensate for the missing bone, and apply it to patients imaged over the pelvic area.
METHODS: We included five patients injected with [18F]DCFPyL for PSMA targeted PET imaging on a Siemens PET/MRI scanner. In all patients, we acquired Dixon in- and opposed-phase images and a work-in-progress dual UTE sequence for whole-body imaging. We trained two random forest regressors, one with Dixon images as input and the other with both Dixon and dual-UTE, using a leave-one-out-cross-validation scheme. For feature selection, we extracted local mean and standard deviation information at varying neighborhoods around each voxel for each input image, as well as the subsequently calculated gradient, R1 and R2* maps. Furthermore, we calculated the spatial information, local binary patterns and auto-context features. Four chained random forests were trained using 500 trees in each loop, and refined using the final bone probability map. The two methods (rRFDixon and rRFDixon+dUTE) were evaluated against a co-registered CT reference.
RESULTS: Across all patients, both the Jaccard index and F1 bone similarity measures were slightly higher in rRFDixon+dUTE (0.47 and 0.63) compared to rRFDixon (0.44 and 0.60). The relative %-error in PET uptake was improved over Dixon-AC using both proposed methods in the tumor, bladder, and full pelvis region.
CONCLUSIONS: The added value of the full model versus the simple model using only Dixon-images as input seems limited, but could be due to limited variation in the small dataset. The average PET-bias in the tumor did not exceed 5% using any of the AC-methods for these five patients. Both proposed methods show improvements over standard MR-AC across all metrics.
Keywords: PET/MRI, Attenuation correction, whole-body, bone, PSMA, prostate
Potential benefits of incorporating energy information when estimating attenuation from PET data (#3592)
L. Brusaferri1, A. Bousse1, N. Efthimiou4, E. Emond1, D. Atkinson2, S. Ourselin3, B. Hutton1, S. Arridge3, K. Thielemans1
1 University College London, Institute of Nuclear Medicine, London, United Kingdom of Great Britain and Northern Ireland
In current combined PET/MRI systems, accurate attenuation correction is one of the biggest challenges. This is due to the fact that MR signal is not directly correlated to tissue density and, therefore, to attenuation. A way to overcome this limitation is to incorporate additional knowledge about the photon attenuation derived from the PET emission data. However, attempts to estimate attenuation coefficients from emission data only have shown limited success unless PET time-of-flight (TOF) is available. Recently, it has been proposed to take advantage of scattered coincidences to deduce further information both on the annihilation position and density distribution. In this work, we show that to fully exploit scatter information, measuring the energy of each annihilation photon in the same energy window is inadequate. We show results with simulations of single scatter events for simple geometric phantoms and more realistic XCAT phantoms. If the same energy window is used for both detectors, very little scattered counts are detected. By contrast, multiple energy window measurements led to a large increase in the number of detected scatter counts. To evaluate the dependence of the scattered counts on the density image, we also show results for the Jacobian of the forward scatter model in terms of the density image. Similar to the detected counts, the Jacobian indicates that using different energy windows gives more information on attenuation. Our implementation uses caching and other techniques to be able to significantly reduce the computational time both in scatter simulation and in Jacobian calculation; the finite energy resolution of the scanner has also been taken into account and included in the model. Our results show that incorporating data from multiple energy windows into the attenuation estimation should be feasible.
Keywords: Scatter, PET, Attenuation Estimation, Energy Resolution, Energy Window, Jacobian
Accounting for Breathing Pattern Variability and Baseline Shift in Event-by-Event Respiratory Motion Correction in PET Using Dynamic Internal-External Motion Correlation (#3421)
Y. Lu1, K. Fontaine1, J. - D. Gallezot1, S. Ren1, T. Mulnix1, C. Liu1, R. E. Carson1
1 Yale University, New Haven, Connecticut, United States of America
Background: For respiratory motion correction (MC), we previously proposed a listmode-based non-rigid event-by-event (EBE) approach based on internal-external motion correlation (INTEX). This approach assumes the INTEX relationship is valid for the entire PET data acquisition. However, irregular respiratory patterns (RP) with varying amplitudes may change the INTEX correlation during the scan. In addition, baseline shift may change the reference location of MC, thus introducing additional errors.
Methods: In this work, we first investigated the impact of motion amplitude change and baseline shift on the INTEX correlation. We proposed a new model of Dynamic-INTEX (D-INTEX), which dynamically modulates the INTEX correlation in response to varying motion amplitudes and baseline shifts. D-INTEX first analyzes a set of predefined frames, each with minimal RP variation and baseline shift, and then uses this frame-dependent INTEX information to predict the varying INTEX correlation for the whole scan followed by EBE MC reconstruction throughout the scan. We evaluated D-INTEX using two human datasets with 18F-FPDTBZ (pancreas tracer) and compared it with conventional INTEX, which utilizes a single correlation throughout the scan.
Results: The initial results showed that the INTEX correlation varies when RP changes and baseline shifts. D-INTEX outperformed conventional INTEX in terms of organ resolution and contrast recovery. For shorter frames without sufficient counts to derive reliable INTEX, the motion model predicted by D-INTEX was very robust to noise and provided reliable motion estimation for even 1-min frames.
Conclusions: Respiratory pattern change and baseline shift could cause INTEX correlation variation. The proposed D-INTEX method dynamically adjusts INTEX to account for such variabilities. Although the initial results are very promising, substantial additional evaluation is required in order to validate the assumptions and application of the D-INTEX method.
Keywords: Dynamic INTEX, D-INTEX, INTEX, Respiratory Motion Correction, Event-by-event, Pattern Variability, Baseline Shift, PET
A systematic study on factors influencing the accuracy of MLAA (#2762)
W. Zhu1, T. Feng1, Y. Dong2, J. Bao2, H. Li1
1 UIH America Inc, houston, Texas, United States of America
Maximum likelihood of attenuation and activity (MLAA) for time-of-flight (TOF) PET works well in theory. However in real systems it is subject to various types of errors. This paper provides a systematic study on the accuracy of MLAA due to errors of different factors, and further ranks the relative significance of these factors.
The influence of various factors on the accuracy of MLAA was investigated as follows. For random correction, scatter correction, TOF timing kernel and detector timing offsets, a gradual increasing bias was applied upon the simulated ground truth as estimation error. The accuracy of MLAA-generated umap versus error of each factor was then computed and plotted. Furthermore, the significance of these factors was compared. A torso XCAT phantom was used in simulation. The bias for random and scatter estimation was set to vary between 0~10%. The timing kernel in reconstruction was set to vary from 860~1260ps (true timing resolution 1.06ns). A timing offset table was scaled and applied on the zero-offset TOF data to simulate detector timing drifts.
In the simulated study, the bias of random estimation incurred smallest error for MLAA-generated umap, while the timing offset incurred the largest. The normalized root mean square error (NRMSE) for umap was 8.2% when bias of random estimation was 10%, and was 9% with 10% scatter estimation bias. An overestimation (250ps) or underestimation (-150ps) of timing resolution resulted in close NRMSE (9.3%) for the umap, indicating underestimation of timing resolution is more severe for MLAA. On the other hand, average 100ps drift (<10% timing resolution, more likely to happen than other errors) made MLAA results visually unacceptable.
For practical system status, timing offset and timing kernel are more important than random or scatter estimation to the overall accuracy of MLAA. The conclusion may contribute towards a better system development.
Keywords: MLAA, RANDOM, SCATTER, TIMING RESOLUTION, OFFSET
Cerenkov Radiation Energy Transfer Imaging Combined with probe-based confocal laser endomicroscopy for precise image-guide tumor resection (#2851)
S. Zheng1, 2, Y. Qu1, X. Zhang3, Z. Hu2, J. Tian2, H. Liu1
1 General Hospital of Chinese People’s Armed Police Forces, Department of Gastroenterology, Beijing, Beijing, China
Cerenkov luminescence imaging (CLI) has been extensively studied in preclinical intraoperative navigation. However, the weak luminescence intensity limits the application of CLI. In this study, we used the clinical contrast agent -- fluorescein sodium to enhance CLI signals by Cerenkov Radiation Energy Transfer (CRET), and applied probe-based confocal laser endomicroscopy (pCLE) imaging to detect the tumor margins at cellular level to achieve surgical guidance. Phantom studies were firstly conducted to investigate the signal intensity and penetration ability of CRET imaging. Then in vivo verifications of CRET imaging were done on subcutaneous 4T1 mice models, and pCLE was applied to show the tumor margins. The experimental results showed that the emission wavelength of CRET imaging was shifted to 540nm and signal peak of CRET is over 5 times higher than CLI. The combined imaging of CRET and pCLE exhibited high performance of detecting small size tumors and identify tumor margins. In conclusion, we proposed a novel dual-modality image-guided tumor resection method. This method was capable to preoperatively detect tumors with strengthened optical signals and intraoperatively reveal the tumor margins. All the experimental goals in this study were achieved by using clinical proved pharmaceuticals. Thus, we believe this proposed dual-modality image-guided tumor resection method owns considerable potential in clinical translation.
Keywords: Cerenkov Radiation Energy Transfer, probe-based confocal laser endomicroscopy, image-guide tumor resection