Showing papers by "Tianyu Ma published in 2019"

An investigation of quantitative accuracy for deep learning based denoising in oncological PET

Abstract: Reducing radiation dose is important for PET imaging. However, reducing injection doses causes increased image noise and low signal-to-noise ratio (SNR), subsequently affecting diagnostic and quantitative accuracies. Deep learning methods have shown a great potential to reduce the noise and improve the SNR in low dose PET data. In this work, we comprehensively investigated the quantitative accuracy of small lung nodules, in addition to visual image quality, using deep learning based denoising methods for oncological PET imaging. We applied and optimized an advanced deep learning method based on the U-net architecture to predict the standard dose PET image from 10% low-dose PET data. We also investigated the effect of different network architectures, image dimensions, labels and inputs for deep learning methods with respect to both noise reduction performance and quantitative accuracy. Normalized mean square error (NMSE), SNR, and standard uptake value (SUV) bias of different nodule regions of interest (ROIs) were used for evaluation. Our results showed that U-net and GAN are superior to CAE with smaller SUVmean and SUVmax bias at the expense of inferior SNR. A fully 3D U-net has optimal quantitative performance compared to 2D and 2.5D U-net with less than 15% SUVmean bias for all the ten patients. U-net outperforms Residual U-net (r-U-net) in general with smaller NMSE, higher SNR and lower SUVmax bias. Fully 3D U-net is superior to several existing denoising methods, including Gaussian filter, anatomical-guided non-local mean (NLM) filter, and MAP reconstruction with Quadratic prior and relative difference prior, in terms of superior image quality and trade-off between noise and bias. Furthermore, incorporating aligned CT images has the potential to further improve the quantitative accuracy in multi-channel U-net. We found the optimal architectures and parameters of deep learning based methods are different for absolute quantitative accuracy and visual image quality. Our quantitative results demonstrated that fully 3D U-net can both effectively reduce image noise and control bias even for sub-centimeter small lung nodules when generating standard dose PET using 10% low count down-sampled data.

A high sensitivity 4π view gamma imager with a monolithic 3D position-sensitive detector

Abstract: Imaging is an effective way to survey gamma radiation hotspots in nuclear safety, public health and homeland security applications. The existing gamma imagers either equip a collimator , which limits the imaging field of view (FOV) and sensitivity, or work without a collimator but require a high performance detector with sophisticated data processing and image reconstruction methods, e.g. the Compton camera. In this work, we demonstrate the feasibility of a high sensitivity 4 π view gamma imager design concept. Based on the photon event density distribution inside a monolithic 3D position-sensitive detector, the imager reconstructs the 4 π view of the surrounding gamma hotspots that the detector is exposed to. Monte Carlo simulations and experiments were conducted for several configurations of surrounding sparse gamma point sources in the energy range from 0.1 to 1.5 MeV. The results show that the reconstructed 4 π view image can identify multiple gamma hotspots, with the hotspots’ positioning accuracy and resolution dependent on the detector’s intrinsic resolution and the statistics of the acquired data. We conclude that the proposed design can be used for surveying gamma hotspots in nuclear security applications.

Recombination coefficients for Cl on plasma-conditioned yttrium oxide chamber wall surfaces

Abstract: Studies of power-modulated chlorine inductively coupled plasmas (ICPs) bounded by yttria-coated chamber walls are presented. Time-resolved optical emissions from Cl and Xe actinometry trace gas were recorded over the 740–920 nm region as power at 13.56 MHz was modulated between high power and no power. The intensity ratio of Cl-to-Xe emission, proportional to Cl number density, nCl, followed the modulation in power, allowing Cl heterogeneous loss coefficients, γCl, to be obtained from a simple time-resolved, 0-dimensional model of the afterglow period that best matched computed relative changes in nCl at the beginning and end of the powered period, with γCl as the only adjustable parameter. This approach only requires a treatment of diffusion and avoids complications introduced by attempting simulations of the full modulation period. Cl recombination coefficients were determined on the mostly yttria surfaces for Cl2 ICPs (a) immediately after NF3 plasma cleaning (γCl = 0.20), (b) during long exposure to the Cl2 plasma with no substrate bias (γCl = 0.11), and (c) during Si etching with substrate bias (γCl = 0.055-0.070). For Cl2/5% O2 ICPs, these values are 0.28, 0.17, and 0.030, respectively. These results compare favorably to qualitative behavior reported previously for continuous Cl2 and Cl2/O2 ICPs in this yttria-coated chamber.Studies of power-modulated chlorine inductively coupled plasmas (ICPs) bounded by yttria-coated chamber walls are presented. Time-resolved optical emissions from Cl and Xe actinometry trace gas were recorded over the 740–920 nm region as power at 13.56 MHz was modulated between high power and no power. The intensity ratio of Cl-to-Xe emission, proportional to Cl number density, nCl, followed the modulation in power, allowing Cl heterogeneous loss coefficients, γCl, to be obtained from a simple time-resolved, 0-dimensional model of the afterglow period that best matched computed relative changes in nCl at the beginning and end of the powered period, with γCl as the only adjustable parameter. This approach only requires a treatment of diffusion and avoids complications introduced by attempting simulations of the full modulation period. Cl recombination coefficients were determined on the mostly yttria surfaces for Cl2 ICPs (a) immediately after NF3 plasma cleaning (γCl = 0.20), (b) during long exposure to t...

Design and Development of a 3D Position-Sensitive Detector for 4π View Gamma Imager Based on Dual-Ended Readout Technique

Abstract: Gamma imager capable of monitoring, detection and localization of radioactive sources is of critical importance for nuclear safety and homeland security with the increasing use of radioactive sources in applications of nuclear technology. A 4π view gamma imager based on gamma ray occlusion technique is under development in our lab. The key component of the gamma imager is a 3D position-sensitive detector, which enables 4π field of view of the imager. In this work, we propose a 3D positionsensitive detector design for the 4π view gamma imager based on dual-ended readout (DER) technique. The detector was composed of a GAGG block coupled to an SiPM array at both crystal ends. DOI calibration was conducted with a collimated fan beam 18F source and DOI resolution was evaluated. Energy calibration was performed for each 3D position region of the detector by incorporating 3D position information to correct for the energy response inconsistency at different 3D position regions inside the detector. Various radioisotopes including 99mTc, 131I, 137Cs and 22Na were tested. An exponential factor was applied to compensate for energy response non-linearity due to SiPM saturation effect at higher gamma energy. The measured average DOI resolution was 1.96 ± 0.21 mm for all the four DER detector modules, indicating excellent 3D positioning accuracy of the detector design. With the proposed energy calibration approach, characteristic photopeaks of each radioisotope can be clearly identified at the correct energy position on the measured energy spectra. The achieved energy resolution was 9.76% at 662 keV, demonstrating good energy performance of the detector design, which could benefit radioisotope identification capability of the gamma imager. To conclude, the proposed 3D position-sensitive detector design based on DER technique features good 3D positioning and energy performance and is feasible for the 4π view gamma imager.

Development and Performance Evaluation of a 4π View Radiation Imaging System

Abstract: In nuclear safety, public health and homeland security, detection of radioactive sources is essential. The existing systems have limits in FOV and/or sensitivity. Based on gamma photon event density distribution inside a 3D position-sensitive detector, position of gamma source can be reconstructed with MLEM. The whole system consists of detector, data processing board, panoramic camera and PC. The 3D position-sensitive detector is made of a GAGG crystal array, coupling with two SiPM array on the top and bottom surface. We designed a rotation system to perform the calibration experiment of system matrix. With different radioactive source (99mTc, 18F, 137Cs), dozens of position (θ, φ) are tested to calculate positioning biases. Energy resolution and energy linearity are tested with 137Cs, 60Co, 232Th. The preliminary test of low counts and two point sources imaging has been performed with 99mTc in the experiment. For 99mTc, 18F, 137Cs, mean of positioning bias (PB θ , PB φ ) is (0.81, 0.52), (1.05, 1.07), (2.35, 1.72) and standard deviation of (PB θ , PB φ ) is (0.55, 0.36), (1.02, 1.06), (1.42, 1.54), respectively. Energy resolution is 9.76%(137Cs). The energy spectrums show good energy linearity. Only 5000 counts can give good positioning accuracy. The system can clearly separate two 99mTc point source at angle distance of 20°. In this study, we developed the 4π view radiation imaging system, which can be used in real practical applications. The system has good positioning and energy performance. The preliminary test shows that the system also has good potential in low count and two point sources imaging.

Direct Estimation of Neurotransmitter Activation Parameters in Dynamic PET Using Regression Neural Networks

Abstract: Current pharmacokinetic models, such as the linear parametric neurotransmitter PET (lp-ntPET) model have been developed to detect and quantify transient changes in receptor occupancy caused by variations in the concentration of endogenous neurotransmitters. However, it often performs poorly when applied at the voxel level due to high statistical noise. In this paper, we propose a new method to detect transient changes in neurotransmitter concentration in dynamic PET data using deep learning. Activation onset time and response magnitude of neurotransmitter were directly estimated using a convolution neural network (CNN) and compared to the lpntPET model. Computer simulations, as well as realistic GATE simulations were used to generate dynamic PET data, representing a [11C]raclopride study, with a known range of activation onset times and response magnitudes, across a wide range of noise levels. Results showed that the proposed neural network had better quantitative performance in estimating activation onset time and response magnitude than the conventional lp-ntPET method, especially where noise is high.