Christian Thibaudeau

LabPET II, an APD-based Detector Module with PET and Counting CT Imaging Capabilities

Computed tomography (CT) is currently the standard modality to provide anatomical reference for positron emission tomography (PET) in molecular imaging applications. Since both PET and CT rely on detecting radiation to generate images, using the same detection system for data acquisition is a compelling idea even though merging PET and CT hardware imposes stringent requirements on detectors. These requirements include large signal dynamic range with high signal-to-noise ratio for good energy resolution in PET and energy-resolved photon-counting CT, high pixelization for suitable spatial resolution in CT, and high count rate capability for reasonable CT acquisition time. To meet these criteria, the avalanche photodiode (APD)-based LabPET II module is proposed as the building block for a truly combined PET/CT scanner. The module is made of two monolithic 4×8 APD pixel arrays mounted side-by-side on a custom ceramic holder. Individual APD pixels have an active area of 1.1×1.1 mm2 at a 1.2 mm pitch. The APD arrays are coupled to a 12-mm high, 8 ×8 LYSO scintillator array made of 1.12 ×1.12 mm2 pixels also at a pitch of 1.2 mm to ensure direct one-to-one coupling to individual APD pixels. The scintillator array was designed with unbound specular reflective material between pixels to maximize the difference between refractive indices and enhance total internal reflection at the crystal side surfaces for better light collection, and the APD quantum efficiency was improved to ~ 60% at 420 nm to optimize intrinsic detector performance. Mean energy resolution was 20 ±1% at 511 keV and 41±4% at 60 keV. The measured intrinsic spatial and time resolution for PET were respectively 0.81 ±0.04 mm FWHM/1.57 ±0.04 mm FWTM and 3.6±0.3 ns FWHM with an energy threshold of 400 keV. Initial phantom images obtained using a CT test bench demonstrated excellent contrast linearity as a function of material density. With a magnification factor of 2, a CT spatial resolution of 0.66 mm FWHM/1.2 mm FWTM, corresponding to 1.18 lp/mm at MTF10%/0.67 lp/mm at MTF50%, was measured, allowing 0.75 mm air holes in an Ultra-Micro Hot Spot resolution phantom to be clearly distinguished.

Fast, accurate and versatile Monte Carlo method for computing system matrix

A new methodology is proposed to mitigate the high computation cost required to derive accurate Monte Carlo (MC) based system matrix for tomographic image reconstruction. The strategy consists of taking advantage of the symmetries between the lines of response to increase the statistics of data collected for the determination of the system matrix coefficients. By using the rotation and axial symmetries of a cylindrical camera, the number of MC generated events can be reduced substantially without affecting the matrix coefficient accuracy. Moreover, using the GATE simulator list-mode saving capabilities for storing coincidence events, single events and/or single hits with all their relevant information, the MC simulation can be performed only once and system matrices for different system configurations be derived from the same simulation. Using for example Positron emission tomography (PET), the processing of the collected MC data can be fine tuned to the imaging system characteristics by setting accordingly the time and energy blurring, the detector efficiencies, the coincidence time window width and the energy thresholds. The system matrix can also include or exclude PET events like scatters and randoms. Furthermore, the system matrix can be computed for different image grids and basis functions without requiring a new MC simulation to be performed.

LabPET II, an APD-based PET detector module with counting CT imaging capability

CT imaging is currently the standard modality to provide anatomical reference in PET molecular imaging. Since both PET and CT rely on detecting radiation to generate images, it would make sense to use the same detection system for data acquisition. Merging PET and CT hardware imposes stringent requirements on detectors, including wide dynamic range with high signal-to-noise ratio for good energy resolution in both modalities, high pixellisation for high spatial resolution, and very high count rate capabilities. The APD-based LabPET II module is proposed as the building block for a truly combined PET/CT scanner. The module is made of two 4 × 8 APD pixel monolithic arrays mounted side by side unto a custom ceramic holder, with each element having an active area of 1.1 × 1.1 mm2 at a 1.2 mm pitch, coupled to a 12-mm high LYSO scintillator block array. While a previous version of the module was made of pyramidal shaped crystals (1.35 × 1.35 / 1.2 × 1.2 mm2, top/bottom), a recent version was designed with a simpler rectangular geometry (1.2 × 1.2 mm2), better reflective material optimizing the shift of refractive index at crystal interface, and enhanced APD quantum efficiency to improve intrinsic detector performance. Mean energy resolution was improved to 20 ± 1% (formerly 24 ± 1%) at 511 keV and to 41 ± 4% (formerly 48 ± 3%) at 60 keV. These intrinsic detector performance characteristics make the LabPET II module suitable for counting CT imaging with efficient energy discrimination. Initial phantom images obtained from a CT test bench demonstrated excellent contrast linearity as a function of material density and spatial resolution of 0.61 mm FWHM / 1.1 mm FWTM, corresponding to 1.3 lp/mm at MTF10% / 0.73 lp/mm at MTF50%, which allowed 0.75 mm air holes in an Ultra Micro resolution phantom to be clearly distinguished.

Real-time discreet SPAD array readout architecture for time of flight PET

Single photon avalanche diode (SPAD) arrays have proven themselves as serious candidates for time of flight positron emission tomography (PET). Discrete SPAD readout schemes mitigate the low-noise requirements of analog schemes and offer very fine control over threshold levels and timing pickup strategies. On the other hand, a high optical fill factor is paramount to timing performance in such detectors, and consequently space is limited for closely integrated electronics. Nonetheless, a production, daily used PET scanner must minimize bandwidth usage, data volume, data analysis time and power consumption and therefore requires a real-time readout and data processing architecture as close to the detector as possible. We propose a fully digital, embedded real-time readout architecture for SPAD-based detector. The readout circuit is located directly under the SPAD array instead of within or beside it to overcome the fill factor versus circuit capabilities tradeoff. Since the overall real-time engine provides all the required data processing, the system needs only to send the data required by the PET coincidence engine, significantly reducing the bandwidth requirement. A 3D prototype device was implemented in 2 tiers of 130 nm CMOS from Global Foundry / Tezzaron featuring individual readout for 6 scintillator channels. The timing readout is provided by a first photon discriminator and a 31 ps resolution time to digital converter, while energy readout and event packaging is done in real-time using synchronous logic from a CMOS standard cell library, all fully embedded in the ASIC. The dedicated serial output line supports a sustained rate 2.2 Mcps in PET acquisition mode, or 170 kcps in an oscilloscope mode for offline validation and development.Single photon avalanche diode (SPAD) arrays have proven themselves as serious candidates for time of flight positron emission tomography (PET). Discreet readout schemes mitigate the low-noise requirements of analog schemes and offer very fine control on threshold levels and timing pickup strategies. A high optical fill factor is paramount to timing performance in such detectors, and therefore space is limited for closely integrated electronics. On the other hand, a production, daily used PET scanner must minimize bandwidth usage, data volume, data analysis time and power consumption, and therefore requires a real-time readout and data processing architecture as close to the detector as possible. We propose a vertically integrated architecture that allows placement of the real-time data extraction directly under the detector while maintaining high optical fill factor for optimal timing performance. This paper focuses on the real-time data acquisition engine. It reduces transmitted data by a factor of 8 in standard operational mode. Combined with small local memory buffers, this significantly reduces acquisition dead time. Finally, auxiliary circuits accelerate channel self-diagnosis, essential for large PET systems. A prototype device featuring individual readout for 6 scintillator channels was fabricated. Timing readout is provided by a first photon discriminator driving a TDC, while energy reading and event packaging is done using standard logic. The dedicated serial output line supports a sustained rate of 170k counts per second (cps) in waveform mode, while the standard operational mode supports 2.2M cps. In normal PET acquisition conditions (up to 1250 cps/mm2), this provides

Toward truly combined PET/CT imaging using PET detectors and photon counting CT with iterative reconstruction implementing physical detector response

PURPOSE This paper intends to demonstrate the feasibility of truly combined PET/CT imaging and addresses some of the major challenges raised by this dual modality approach. A method is proposed to retrieve maximum accuracy out of limited resolution computed tomography (CT) scans acquired with positron emission tomography (PET) detectors. METHODS A PET/CT simulator was built using the LabPET™ detectors and front-end electronics. Acquisitions of energy-binned data sets were made using this low spatial resolution CT system in photon counting mode. To overcome the limitations of the filtered back-projection technique, an iterative reconstruction library was developed and tested for the counting mode CT. Construction of the system matrix is based on a preregistered raster scan from which the experimental detector response is obtained. PET data were obtained sequentially with CT in a conventional manner. RESULTS A meticulous description of the system geometry and misalignment corrections is imperative and was incorporated into the matrix definition to achieve good image quality. Using this method, no sinogram precorrection or interpolation is necessary and measured projections can be used as raw input data for the iterative reconstruction algorithm. Genuine dual modality PET/CT images of phantoms and animals were obtained for the first time using the same detection platform. CONCLUSIONS CT and fused PET/CT images show that LabPET™ detectors can be successfully used as individual X-ray photon counting devices for low-dose CT imaging of the anatomy in a molecular PET imaging context.

Digital Identification of Fast Scintillators in Phoswich APD-Based Detectors

Significant progress has been made in the last 15 years to improve the spatial resolution of small animal PET scanners, mainly by reducing the size of detector pixels. Spatial resolution can be further improved by using phoswich scintillator assemblies, either to increase pixellization or to measure the depth-of-interaction. A number of high-density, fast and high light output Ce-activated Lu-based scintillating materials with a range of decay times are now available, which has widened the list of potential candidates for applications in medical imaging. Most of them have suitable emission wavelength (above 400 nm) for readout with APDs, but their decay times are often too similar for accurate identification in phoswich detectors using standard analog pulse shape discrimination techniques. This study investigates the spectroscopic characteristics of these fast Lu-based scintillators and their diverse combinations into phoswich assemblies for high resolution PET detectors. Crystal identification was assessed using advanced numerical methods derived from signal recognition theory to fit a mathematical model to the digitized APD output signals and then to discriminate the crystal of interaction based on model parameters. Identification errors were evaluated as the overlapping peak area in the model parameter spectra. Identification is virtually error-free for decay time differences (Δτ) larger than 20 ns, while the measured error is generally less than 5% for Δτ > 10 ns. Whereas the Δτ between crystals is the major factor influencing identification performance, others factors such as the initial photon emission rate and the decay time also affect the identification accuracy. The phoswich pair consisting of LSO:Ce, Ca (τ = 32 ns) and LGSO (10%Gd-0.75%Ce) (τ = 45 ns) achieves the best overall performance for the PET application.

Iterative CT reconstruction using LabPET™ detector modules

A computed tomography (CT) simulator has been built using the LabPET™ detectors and front-end electronics. Acquisitions of energy-binned data sets were made using this low spatial resolution CT system in photon counting mode. To retrieve maximum accuracy out of these acquisitions, an iterative reconstruction library was developed and tested. Construction of the system matrix is based on a preliminary raster scan from which the experimental detector response is obtained. A meticulous description of the system geometry and misalignment corrections is imperative and must be incorporated into the matrix definition to achieve good image quality. Using this method, no sinogram pre-correction or interpolation is necessary and measured projections can be used as raw input data for the iterative reconstruction algorithm. CT and fused PET/CT images show that LabPET™ detectors can be successfully used as individual X-ray photon counting devices for low-dose CT imaging of the anatomy in a molecular PET imaging context.

Fully 3D iterative CT reconstruction using polar coordinates.

PURPOSE This paper demonstrates the feasibility of fully 3D iterative computed tomography reconstruction of highly resolved fields of view using polar coordinates. METHODS System matrix is computed using a ray-tracing approach in cylindrical or spherical coordinates. By using polar symmetries inherent to the acquisition geometry, the system matrix size can be reduced by a factor corresponding to the number of acquired projections. Such an important decrease in size allows the system matrix to be precomputed, and loaded all at once into memory prior to reconstruction. By carefully ordering the field of view voxels and the sinogram data, reconstruction speed is also enhanced by a cache-oblivious computer implementation. The reconstruction algorithm is also compatible with the ordered-subsets acceleration method. A final polar-to-Cartesian transformation is applied to the reconstructed image in order to allow proper visualization. RESULTS The ray-tracing and reconstruction algorithms were implemented in polar representation. Large 3D system matrices were calculated in cylindrical and spherical coordinates, and the performance assessed against Cartesian ray-tracers in terms of speed and memory requirements. Images of analytical phantoms were successfully reconstructed in both cylindrical and spherical coordinates. Fully 3D images of phantoms and small animals were acquired with a Gamma Medica Triumph X-O small animal CT scanner and reconstructed using the manufacturers software and the proposed polar approach to demonstrate the accuracy and robustness of the later. The noise was found to be reduced while preserving the same level of spatial resolution, without noticeable polar artifacts. CONCLUSIONS Under a reasonable set of assumptions, the memory size of the system matrix can be reduced by a factor corresponding to the number of projections. Using this strategy, iterative reconstruction from high resolution clinical and preclinical systems can be more easily performed using general-purpose personal computers.

Cylindrical and spherical ray-tracing for CT iterative reconstruction

Recent events showed that dose administration has become a major concern regarding clinical computed tomography (CT) examinations. A simple solution to lower this dose is the use of more appropriate CT image reconstruction methods. Iterative reconstruction has not yet fully reached clinical applications, mostly due to computer speed and memory limitations. A method is proposed to compute a 3D system matrix using a polar-like voxel representation, allowing the modeling of highly resolved fields of view with a minimal random-access memory usage. As opposed to conventional ray-tracing, which is usually done over Cartesian coordinates, the proposed method uses cylindrical or spherical coordinates. The azimuthal symmetries present in those systems allow to compute only the detector responses of the first projection and use them as-is for all other projections. During reconstruction, only the azimuthal index has to be recomputed, and this is done using a simple addition. The ray-tracing and reconstruction algorithms were implemented and images of analytical phantoms were successfully reconstructed. Fully 3D images of phantoms and small animals acquired with a Gamma Medica Triumph X-O™ CT scanner are also presented to demonstrate the accuracy and robustness of the proposed approach. Under a reasonable set of assumptions, the memory size of the system matrix can be reduced by a factor corresponding to the number of projections. Using this strategy, iterative reconstruction from high resolution clinical and preclinical systems can be performed on general-purpose personal computers.

Comparison of analytical and algebraic 2D tomographic reconstruction approaches for irregularly sampled microCT data

APD-based detectors with individual channel readout were developed for multi-crystal applications and have been implemented for the detection of annihilation radiation in the LabPETTM micro-scanner. The use of these APD-based detectors in X-ray imaging is currently being assessed with a microCT demonstrator in order to later combine PET and CT in one apparatus. This paper is focused on the tomographic reconstruction of the X-ray transmission data acquired with this demonstrator. Two aspects of the acquisition geometry need to be carefully considered: the radius of the detector arc and the irregular sampling of the detector bins. A specific shift- variant filtered backprojection formula derived to account for the detector curvature is applied to equiangularly resampled projection data while the simultaneous algebraic reconstruction technique is applied to both resampled and original projections. Images of physical phantoms reconstructed from measured projections using the different methods are presented and compared.