Lightburst Digital Detector

Transcription

1 GE Healthcare Lightburst Digital Detector

2 INTRODUCTION In clinical practice, PET/CT imaging helps clinicians visualize disease at an early stage, before it metastasizes and involves other organs, tissues or vessels. Quantitative PET based on standard uptake values (SUVs) also delivers crucial information regarding a patient s treatment response. It has been documented in literature that patients treated for localized, early stage cancer and those that respond positively to therapy have better outcomes in terms of disease-free survival. Diagnosing cancer at an early stage is largely dependent upon the sensitivity and image quality provided by the PET/CT system. At GE Healthcare, sensitivity and clinical performance have been a core foundation in our PET/CT system design. Sensitivity enables a gain in signal-to-noise ratio (SNR) and a gain in SNR enables a higher quality image (higher resolution image) and the ability to reduce scan time and/or patient dose. In product development, designing and engineering a medical imaging system is not dependent on one specification or technology choice. Rather, it is the combined operations of all the components in the imaging chain that enable the technology to deliver the high-quality imaging and clinical capabilities to help clinicians make confident patient care decisions that positively impact patient outcomes. At GE Healthcare, we ve evaluated and optimized each component of the molecular imaging workflow, from acquisition to reconstruction to report generation and communication with referring physicians, to deliver outcomes-based and comprehensive solutions. In the imaging chain, the detectors are of paramount importance. In this report we describe the Lightburst Digital Detector. We include evaluation of the detector components that can impact sensitivity, energy and timing resolution, scalability, scatter recovery, spatial resolution, counting rate statistics, correction accuracy, image quality and perhaps most important, small lesion detectability in order to provide the ability for clinicians to detect disease at an early stage. used in other diagnostic imaging systems: (1) the annihilation photons have an energy of 511 kev, which is higher than those used in most other diagnostic imaging applications; (2) the need to find coincidence between two photons on opposite sides of the patient, meaning that the detector must be able to resolve the arrival time of a photon to within a few nanoseconds (or less than a nanosecond for a time-of-flight [TOF] PET detector). These requirements have resulted in a two-step detection process for PET detectors, consisting of: A scintillator crystal that converts the high-energy photon into visible or ultraviolet light photons. A light sensor that converts these lower-energy photons into an electrical signal. Scintillators In many ways, the scintillator is the prime determinant of the performance of the detector. If the incident photon does not interact with the scintillator, then no event is recorded no matter how high-performing the downstream sensors and electronics may be. Stopping a large fraction of the 511 kev photons coming from the patient requires: A large detection area so that more of the photons are captured by the detector. A highly sensitive detector made from a high-stopping power material of sufficient depth to stop a high percentage of the photons that enter the detector (see Figure 1). It is important to note that since the PET system must detect two coincident photons, the sensitivity of the detector has a quadratic effect on system sensitivity. For example, a detector that captures 50% of the incoming 511 kev photons only records 25% of the available coincidences. PET Detector design A PET detector converts annihilation photons into an electrical signal that is processed by the front-end electronics into a signal suited for digital signal processing and ultimately into a clinical image. Two key factors make a PET detector unique from those Figure 1. The importance of crystal depth in determining detector sensitivity is demonstrated.

3 Since most 511 kev photons have multiple interactions in a detector, it is impossible to derive a simple expression for the detector sensitivity. However, an upper bound can be computed using the linear attenuation coefficient of the material; if there is no interaction between the photon and the detector, then no signal can be recorded. Estimates of detector sensitivity must be made by Monte Carlo simulation. However, these require making assumptions regarding the specific geometry of the detector and the detected energy required to qualify as a recorded event. Results from a sample calculation are shown in Figure 2. Here the coincidence sensitivity is plotted against detector depth for a 50 mm x 50 mm lutetium-based scintillator (LBS) detector. The importance of detector depth is underscored by this figure. In a detector with a shorter depth, the probability of the 511 kev photons having any interaction drops substantially by about one-third when the crystal depth is reduced from 25 mm to 15 mm but the probability of capturing sufficient energy to qualify as a valid event drops even faster, at over 40%. becomes one million electrons out) with minimal added noise and excellent time resolution. This high level of performance made PMTs a preferred component, even after they had been replaced by other light sensors (primarily photodiodes) in many other medical and non-medical applications. The most common form of photodiode is the p-i-n photodiode. In a p-i-n photodiode, an incident photon ionizes an atom in the intrinsic region of the structure and that electron drifts through an electric field to be collected at the cathode. While these devices can be very effective in many applications, they provide no signal gain (one photon yields one electron) and, therefore, are not well suited for PET systems and applications. Avalanche photodiodes (APDs) are a form of photodiode used in some PET scanners, both in clinical and research applications. By increasing the bias voltage on the diode, the electron produced by the photon is accelerated to the point where it can ionize other atoms in the lattice, and those electrons can ionize other atoms, and so on (hence, the avalanche ). This gives the device some gain, so that it can effectively detect the scintillation light from an incident 511 kev photon; yet, it is not enough to provide a sufficient signal to achieve sub-nanosecond timing resolution for TOF imaging. Figure 2. Coincidence sensitivity as a function of crystal depth for LBS detectors, as determined by Monte Carlo simulation. The Lightburst Digital Detector uses LBS, which provides very good stopping power (the mean free path of a 511 kev photon in these materials is just over 1 cm), together with the timing resolution needed to perform TOF imaging. Importantly, it also uses crystals that are 25 mm deep to provide high sensitivity. Light sensors: photomultipliers to SiPMs Once the scintillator has converted the annihilation photon into light photons, that light must be converted into an electrical signal. Since the advent of positron imaging in the 1950s, photomultiplier tubes (PMTs) have been the most widely used light sensors in PET detectors. The advantage of the PMT is that it provides a signal gain on the order of one million (1 photon in Silicon photomultipliers (SiPMs) represent a further evolution in the photodiode process. The SiPM is comprised of many microcells on the order of 25 µm µm in each dimension. The bias voltage on the SiPM is increased beyond the breakdown voltage of the device, so that the single ionization caused by an incident photon leads to complete discharge of one microcell, similar to a Geiger-Mueller tube. The number of microcells discharging is proportional to the number of incident light photons, providing the energy signal for discriminating true 511 photons from scattered radiation. SiPMs provide sufficient gain to enable TOF imaging. Because they have a higher photon detection efficiency (PDE) than conventional PMTs (>36% compared to 25% for a vacuum-tube PMT), SiPMs provide the opportunity to realize timing resolution better than PMT-based system designs. GE recognized the value of SiPMs for PET applications as the technology began to develop. SIPMs form the foundation of the SIGNA PET/MR detector because conventional PMTs would not operate inside an MR magnet. The confined, harsh and demanding environment of the PET/MR scanner forced the

4 detector design to be compact, robust in the presence of challenging mechanical, thermal and electrical conditions, and both immune to the powerful RF pulses inherent in MR imaging as well as not interfere with the detection of the low-power resonance signals returned from the patient. The Lightburst Digital Detector As with any technology, incorporating SiPMs into a PET/MR or PET/CT system design involves a number of development decisions to balance competing aspects of the detector design as well as to accommodate any non-ideal properties of the new technology. This section describes some of the key design elements that have been incorporated into the Lightburst Digital Detector. SiPM device Microcell size: As stated above, the SiPM device consists of hundreds or thousands of microcells that can range between 25 µm x 25 µm 2 to 100 µm x 100 µm 2. A design with a larger number of small microcells provides a wider dynamic range for the output signal and a lower dark current (signal pulses that occur when no input photon is present) per microcell. Larger microcells have fractionally less dead area (the non-sensing part of the device devoted to signal traces and the quenching resistor required for the microcell to operate), and therefore have a higher PDE. In our testing a 50 µm x 50 µm microcell provided the best performance, and was able to maintain that performance over a wide operating voltage range. SiPM pixel size: As with the microcells that form the SiPM device, there are competing factors to consider in choosing the size of the individual SiPM. Smaller SiPM pixels minimize device capacitance, which improves timing resolution and can minimize electronic deadtime in processing to boost high count rate performance. On the other hand, a design with fewer, larger SiPMs requires less associated electronics, reducing power consumption and system complexity. The Lightburst digital detector achieves a good balance with a SiPM pixel size of 4 mm 2 x 6 mm 2 and a total of 10,375 microcells. SiPM device packaging: Each individual SiPM device has approximately one millimeter perimeter of dead area. To minimize this dead Figure 3. The SiPM device consists of hundreds or thousands of microcells. area, a SIPM device with a 2x3 array of 4 mm 2 x 6 mm 2 pixels was developed in cooperation with the SIPM supplier. In this design, the dead area between pixels in a device was limited to 250 µm and a large photosensitive area of 12.6 mm 2 x 12.6 mm 2 was achieved. The dead area around the edge of a monolithic device is covered by a beveled light guide that reflects light which would otherwise arrive in the dead area into the photosensitive area of the device, maximizing light collection from the scintillation event. Light sharing crystal block design A 4x3 array of 3.95 mm 3 x 5 mm 3 x 25 mm 3 LBS crystals are coupled to the active area of the 12.6 mm x 12.6 mm SiPM array through a beveled light guide. These crystals are not optically isolated from each other so the scintillation light will spread through the crystals and on to multiple elements of the SiPM array. The scintillating crystal is identified from the relative signal levels arising from the elements of the SiPM array in the same way that a PMT-based block detector operates. The surfaces of the crystals are engineered to control the light propagation among the crystals in order to minimize misidentification of the scintillating crystal. This design was chosen, as opposed to a one-to-one coupling of scintillator crystals to SiPM elements, because: First, and most importantly, the optically isolated crystals in a one-to-one coupling design exhibit poorer light collection than a block detector design. This is likely due to the high number of reflections that the scintillation photons must take down the long, narrow crystals to get to the SiPM, meaning that even slight non-idealities in the reflectors around each crystal cause a substantial loss of signal. Also, one-to-one coupling of crystals to SiPM elements means that they must be the same size and the design cannot optimize each of those parameters independently. High PET spatial resolution requires small crystals, but a small SiPM Figure 4. Crystal block design: (from top) scintillator crystals with light guide; circuit board with SiPMs on top and application-specific integrated circuits (ASICs) underneath; nonconducting spacer; and bottom of block housing. The light-tight enclosure over the block is not shown.

5 size limits the number of microcells per SiPM element, which further limits the dynamic range of the devices and requires a non-linearity correction. A light sharing design eliminates this issue by allowing independent optimization of the crystal and SIPM designs with the opportunity to better maximize total system performance. One-to-one coupling has a detrimental impact on sensitivity due to the large fraction of 511 kev photons that scatter from crystal to crystal within the detector block. A light sharing design like the Lightburst Digital Detector efficiently collects all the energy from multiple interactions in the block. When coupled with the Compton Scatter Recovery feature described later in this paper, this design enhances the sensitivity of the detector by reconstructing many multiple-interaction events which otherwise would be lost. A one-to-one coupling design requires the addition of extensive circuitry or computational power to accomplish what the light sharing design does more easily. Three of these optically isolated arrays are placed in a row to construct a block detector; four of those blocks are then packaged together with their associated electronics into a single light-tight housing measuring approximately 64x50 mm. Readout architecture with low impedance ASIC The SiPM signal readout electronics, implemented in a custom ASIC, was designed to preserve the integrity of the SiPM signal shape, capitalize on the benefits of increased PDE and minimize power consumption. A low input impedance current buffer amplifier was used on the front end to minimize the impact of the SiPM s capacitance which would otherwise lengthen the signal, degrading count rate capability and timing performance. To Figure 5. ASIC power consumption. achieve high timing resolution, the SiPM signals are transmitted directly into the low noise and high bandwidth buffer amplifier. The amplifier outputs are summed and shaped to generate a timing signal and an energy validation signal with two adjustable comparator thresholds. The timing signal is digitized by an external TDC with a 13 ps time bin size and the energy, X and Z signals are used for integration (see Figure 5). Any impact of electronic noise on timing performance is minimized by placing the ASIC as close as possible to SiPMs on the back side of the circuit board housing the SIPMs. Another important design consideration is that ASIC power consumption is minimized by utilizing low bandwidth amplifiers for the energy and positional signals. This is important because high power consumption would be a source of local heating that would affect the performance of the SiPM. Compact detector unit with liquid cooling Like all other solid-state photosensors, SiPMs have a temperature dependence in their breakdown dark voltage (gain) and dark current. Complicating this is the fact that due to high gain in the device (>10 6 ), any signal generated in a SiPM and all electronics (including the ASICs) placed near the SiPMs are a heat source that would change temperature of the SiPMs. This forces the system design to include provisions for liquid cooling of the detector modules to keep SiPM temperature constant within 0.10ºC. In addition, a temperature compensation algorithm is applied to keep the SiPM signal level constant in the event that the module temperature does undergo transient variations. Scalable detector design in axial direction The Lightburst Digital Detector module has been designed to be scalable with different system configurations constructed from differing numbers of mounted detector units. The modularity of the detector design makes it possible to upgrade from one configuration to another as a field service procedure (Table 1).

6 Table 1. Scalable detector design makes it possible to upgrade from one configuration to another. Not all configurations are available in all regions. Reliability from PET/MR rugged design The GE Lightburst Digital Detector was designed for extreme robustness. For its initial deployment in the GE SIGNA PET/MR, the GE Lightburst Digital Detector has provided high performance and stability even when embedded in an MR bore during simultaneous MR imaging. MRI sequences impose high magnetic fields, transmit kilowatts of pulsed RF energy and listen for femtowatt RF signals in response. In addition, MRI uses spatially modulated magnetic fields that induce eddy currents and as well as heat and vibration on nearby electronics. Heat induced on body coil shielding near the PET detector varies spatially and dynamically by up to 30ºC. Given the much more benign environment of PET/CT, the GE Lightburst Digital Detector can be expected to be even more stable, reliable and robust in its Discovery MI configuration. Compton scatter recovery As mentioned above, one important capability of the Lightburst Digital Detector design is its ability to record events that are Compton scattered in the detector. Approximately 30% of the 511 kev photons that interact in the LBS scintillator have a photoelectric event and deposit all of their energy in a single interaction. Most interactions are Compton scatter, which creates a secondary photon that will either escape the detector or have another interaction within the detector. Since a single Compton interaction by a 511 kev photon can deposit no more than 340 kev at the interaction site, if the secondary photon escapes the detector the signal collected from only the first interaction cannot pass the lower energy threshold and, therefore, will not be recorded. Figure 6. The scattered event on the left is automatically recorded by the block detector. The event on the right requires Compton Scatter Recovery in order to be properly recorded. In a large block detector design such as those in previous generations of GE PET/CT scanners, many of the secondary photons interact within the same detector block, and the signals from the two interactions are read together as if a single interaction took place. However, the smaller dimensions of the Lightburst Digital block mean that more of the secondary photons escape the block and interact in another block. Without additional circuitry designed to find and reconstitute those events, the detector will lose them and sensitivity of the system will be reduced. In the Lightburst Digital system design, adjacent blocks communicate with each other to recover these scattered events. 1 When this communication was first enabled in SIGNA PET/MR, it recovered an additional 20% of the true event sensitivity of the scanner. The Discovery MI System The Discovery MI system design provides high sensitivity and options for long axial coverage. This capability enables faster patient scans and the use of lower tracer injections. In addition, the acquisition system provides the data throughput and inline processing technology needed to use fast-decaying and/or fast-clearing tracers at higher doses, in applications such as cardiac imaging. With a range of configuration options, users can tailor the Discovery MI system to their current practice and to allow extensibility for the future. Systems can be installed with 15 cm or 20 cm axial coverage and later be upgraded to 20 cm or 25 cm of axial coverage. Scalability in the reconstruction engine also allows the user to select the performance to meet their needs and to increase system capacity in the future if needed. In addition to the high-performing PET components, the Discovery MI system incorporates Revolution EVO to deliver highperformance diagnostic CT scans with several dose-lowering options, such as ASiR-V iterative reconstruction. PET performance The NEMA NU methods 2 are used to evaluate the performance of the Discovery MI 4-ring configuration. The following performance results were performed at two clinical sites: Stanford University (Stanford, CA) and Uppsala University

7 Table 2. Spatial resolution. (Uppsala, Sweden). 3 Spatial resolution FDG point sources (approximately 1 mm long and 1 mm in diameter) were created using capillary tubes. The small sources were suspended at radial offsets of 1 cm, 10 cm and 20 cm and two axial positions: center and 1/8 from the edge of the AFOV. Data were collected for 1 min (>500 kcps) at each position. The full width at half maximum (FWHM) and full width at tenth maximum of the point sources were quantified at all locations, using the NEMA-specified filtered back projection algorithm without anodization, as well as non-tof ordered-subset expectation maximization without pointspread function (non-tof OSEM). These are the results measured at two sites. doms subtraction. Note that the results presented here (and in the count rate test results that follow) are reported as described in the NEMA standard, without increasing the measured count rate values to account for the effective benefits of time-of-flight imaging. Calculations of effective sensitivity and effective NECR are generally based on SNR analysis, such as that presented by Budinger TF 4, appropriate for images reconstructed with filtered backprojection, and does not apply to images reconstructed with iterative algorithms. Counting rate statistics The patient table was moved to its lowest setting and the NEMA scatter phantom was mounted from its ends above the patient table so that the 20 cm diameter phantom was centered in the transaxial field of view of the scanner. A line source (70 cm in length, 3.2 mm in inner diameter) was filled with a calibrated activity of roughly 817 MBq of 18F-FDG and inserted into the NEMA scatter phantom. Twenty-four frames Sensitivity Plastic tubing (70 cm in length, 1 mm in inner diameter) was filled with an average calibrated activity of approximately 20 MBq of 18F-FDG, allowed to decay for 250 min to reach an activity that generates less than 5% randoms and fixed at both the center of the FOV and at a vertical radial offset of 10 cm using positioning scans and a positioning apparatus. One-minute scans were taken with 5 aluminum sleeves with increasing thicknesses (NEMA-NU sensitivity prescribed phantom). NEMA sensitivity calculations were performed after ran- Figure 7. NEMA sensitivity measurements. (A) Sensitivity of different axial slices. (B) Sensitivity as a function of the number of attenuating aluminum sleeves. Figure 8. NEMA counting rate measurements. (A) Counting rate vs. activity. (B) Scatter fraction vs. activity. NEC=noise-equivalent counts.

8 of data were taken, with the first 17 frames taken as 15-min acquisitions and the last 7 frames taken as 25 min acquisitions followed by 25 min delays (i.e., at 50 min intervals). NEMA- NU specified methods were used to derive the trues, randoms, scatter and noise-equivalent counting rate (NECR) Table 3. Counting rate measurements. from the prompts dataset in each frame. Randoms were estimated using singles rates. Correction accuracy The system s counting rate accuracy, which compares the measured activity with the expected activity and is dependent on the system corrections used, was found from a linear fit of the activity concentrations measured below peak NECR. In addition to attenuation and scatter corrections, randoms and dead-time corrections were performed using singles-based randoms subtraction and pileup correction, respectively. The dimensions of the reconstructed image matrix were 128x128, with a pixel size of 1.41 mm x 1.41 mm. The maximum deviation from expected activity at Uppsala was 3.86% at an activity of 1.19 kbq/ml, whereas at Stanford it was 2.43% at an activity of kbq/ml. Image quality The background region of the NEMA image-quality phantom and the 10 mm, 13 mm, 17 mm and 22 mm diameter spheres were filled with 18F-FDG activity concentrations of, respectively, 4.7 kbq/ml and 18.8 kbq/ml at Stanford and 5.1 kbq/ml and 20.4 kbq/ml at Uppsala, yielding a 4:1 sphere-to-background concentration ratio. The 28 mm and 37 mm diameter spheres were filled with nonradioactive water. The scatter line source used to measure NEMA counting rate statistics was filled with roughly 118 MBq of 18F-FDG and threaded through the body phantom. Three separate acquisitions of the image-quality phantom were taken (as prescribed in the NEMA-NU standard) with decay-adjusted acquisition times of 271 sec, 279 sec and 282 sec, consistent with a 151 mm axial step for each bed position, and reconstructed with the standard GE Healthcare clinical algorithm (TOF OSEM - PSF) and block-sequential regularized expectation maximization algorithm with PSF (Q.Clear) 5 and beta-value of 50, VPFX reconstruction Q.Clear reconstruction Uppsala, coronal view Stanford, coronal view Uppsala, transverse view Stanford, transverse view Table 4. Contrast recovery and background variability. Figure 9. Phantom images reconstructed with TOF OSEM PSF (left column) and Q.Clear (right column). Top two rows show transverse slices through center of all spheres. Bottom two rows show coronal slices through 10/13-mm spheres.

9 yielding noise levels similar to TOF OSEM. The average and SD of the contrast recovery and background variability were quantified over the 3 sets of data replicates. Corrections for randoms, scatter, CT-based attenuation, dead time and normalization were applied. The dimensions of the reconstructed image matrix were 384 mm x 384 mm x 71 mm, with a pixel size of mm x mm and a slice thickness of mm. Figure 9 shows reconstructed images from the NEMA-IQ phantom. Energy and timing resolutions The average system photopeak energy resolution was 9.44% +/- 0.07% FWHM at Stanford and /- 0.05% FWHM at Uppsala. The average system coincidence time resolution was /- 2.6 ps FWHM at Stanford and /- 2.7 ps FWHM at Uppsala. Research The Discovery MI scanner is a research-friendly versatile system thanks to its high sensitivity and spatial resolution, large axial coverage and excellent randoms reduction and scatter rejection for TOF imaging. The system is capable of handling up to 12.8 mcps, yet it can acquire at higher rates through unbiased rejection of events. Just as important for research, Discovery MI can image a wide range of tracers F18, Ga68, C11, Y90, N13H3, 015, and Rb I124, to name a few. Image reconstruction capabilities combine TOF, regularized OSEM and PSF with accurate quantitation (Q.Clear). Several current areas of research enabled by Discovery MI include: Low-dose studies with Y90 Theraspheres and Ga68. Y90 Theraspheres could be used in post-treatment exams to validate correct positioning of the implant and delivered dose, as well as in radiation treatment planning by using accurate quantitation from Y90 PET images. Gallium s challenging chemistry requires a high sensitivity scanner to take maximum benefit of the limited activity available from the generator. Discovery MI may enable optimum use of available dose for multiple patients. Shorter acquisitions. Discovery MI s high sensitivity and wide axial FOV coverage allow shorter acquisitions without compromising diagnostic quality. As a result, it can free up time for research-oriented activities without impacting revenue. High image quality, low-dose cardiac studies with Rb. In order to achieve high count rate imaging during the perfusion phase of the exam, the scanner must also handle very high count rates during the flow phase of the exam. Discovery MI s high count rate capabilities enables optimum image quality output from both phases of the cardiac exam. Further, with a high sensitivity Discovery MI enables the opportunity for high image quality at lower-dose cardiac studies. Small lesion detection. High resolution and sensitivity may allow for small lesion detection studies, such as early detection of prostate cancer (PSMA). Scalability and large AFOV. Brain imaging can benefit through an activity input function obtained directly from imaging of the carotid artery on the neck. Neuro inflammation. The Discovery MI could benefit these types of studies with its high sensitivity to address low uptake and high resolution to detect small features. Cardiac vulnerable plaque imaging. Similar to neuro inflammation, detection of vulnerable plaque is challenged by low uptake and a small lesion size; these studies could benefit from Discovery MI s high sensitivity and resolution. Faster processing of images with multi-queuing (PARC 4) also enables research studies with a larger sampling of the controlled variables of the study, for example, examining the benefits of TOF or regularized recon in cardiac imaging. Access to the raw data for research purposes has been enhanced with the use of HDF5 data format. HDF5 is a well-known data structure for which many off-the-shelf utilities are available allowing limited access to the raw data in sinogram mode and list mode. Discovery MI PET/CT and small lesion detectability To understand the interactions between all the components of the Discovery MI system and resulting PET image quality, several measures are available such as those in the NEMA NU-2 standard. 2 The NEMA measurements are defined to test intrinsic system performance in a manner that compares all systems to the same measure. Small lesion detectability is not currently part of the standard and, therefore, requires a different approach. PET has historically excelled as a quantitative imaging modality, providing the clinical capability to characterize small differences in tracer uptake throughout the body. The clinical task in FDG-PET oncology is usually driven by the need to see and treat cancer, including detection, localization and characteriza-

10 tion of tumor uptake. One of the most difficult clinical tasks is distinguishing small, low-contrast tumors from background noise imaging with PET is always limited by injected dose and allowable imaging time. Two general approaches have emerged to expand the PET system characterization measurements beyond NEMA to include quantitative assessment of lesion detectability by using the receiver-operator characteristic (ROC) curve. One approach utilizes inserted lesions into patient data; a second approach is to use a phantom with known features. In a patient-based approach using clinical data with inserted realistic lesions of known size, contrast and location, Wangerin et al., demonstrated that a convergent penalized likelihood image reconstruction resulted in improved lesion detectability. 6 This approach has also been used by others to study other system factors such as the use of TOF imaging. 7-9 To more readily compare systems and algorithms without requiring patient data, a detectability phantom design and measurement approach has been defined and applied to several PET/CT systems The main goal of a detection phantom is to enable measurement of the area under the ROC curve (AUC) to determine the degree of accuracy of a detection task. The AUC, which ranges from 0.5 (guessing) to 1.0 (always accurately detected), is often converted into the detectability index, d, often called the observer signal-to-noise ratio (SNR), 13 using: d = 2erf -1 (2AUC - 1) filled with 4 mm diameter acrylic balls, they create spherical contrast features without cold walls (the nylon is porous). Using a template, the features can be reliably placed into regular patterns as the phantom is built layer-by-layer. The goal of the feature placement is to enable non-overlapping PET sub-images centered on the features in this case, approximately 4 cm x 4 cm sub-images were used. In addition, the background solution, diluted by the imperfectly-packed plastic balls, forms a textured background that is more like clinical imaging than the uniform water bath generally seen in phantoms. Scanning the same phantom or phantoms across multiple systems for comparison requires careful planning so that the data acquired are useful for lesion detection. Typically, this means that the AUC for all systems under all conditions will range between Preliminary ranging studies enabled definition of both the proper feature sizes (at ~3:1 contrast) as well as the clinically realistic count levels needed to support generation of sub-scan data from PET list-mode data. The goal A Detection measurement methodology using a phantom in radionuclide imaging Robust measurement of lesion detection requires large numbers of independent lesion-present (LP) and lesion-absent (LA) images. To accomplish this using a single phantom, the contrast features were spaced in regular patterns such that sub-images centered on the features could be extracted from the image data. With co-planar features, multiple sub-images could be extracted from areas between the planes for feature-absent sub-images. B D C E The main idea behind the detection phantom design is to dilute the activity in the phantom with close-packed, non-absorbing plastic balls, then defining hot spots in the matrix by introducing 3D-printed dodecahedrons (spherical cages) into the matrix of solid plastic balls. The dodecahedrons do not admit the plastic balls into their interior, creating an effective activity concentration about three times the surrounding area (see Figure 10). The nylon contrast features have 0.75 mm diameter side elements and are robust to re-use. When placed into the phantom tank Figure 10. (a) Schematic drawing of dodecahedrons designed to externalize balls; (b) four 3D-printed dodecahedron sizes from 4 mm to 7 mm; (c) 4 mm to 6 mm contrast features in a background of 4 mm acrylic balls; (d) 5 mm to 6 mm features placed in a layer of a body-size PET phantom; (e) assembled body-size PET phantom with 5 layers of contrast features.

11 of the ranging studies was to find the scan durations needed to generate four sub-scans for each scan on each system with each phantom all at near-clinical activity concentrations in the phantom. The use of four sub-scans increases by 4x the number of independent (LP, LA) images available for observer assessments. A channelized Hotelling observer with three difference-of-gaussian channels and internal noise was used. 9 This type of numerical observer has been shown to mimic human observer performance. 13,14 To enable comparisons between multiple PET/CT systems at the same activity concentration and scan duration, the phantom(s) are intended to be scanned at least two times per system. This enables generation of multiple detectability versus scan duration estimates at each given activity concentration (to generate a function of d versus scan duration), which is followed by fitting of the d (fixed duration) results at the two (or more) actual activity concentrations (per system) to enable estimation of the final d (fixed duration, fixed activity concentration) for each system and phantom. For all of the above, comparisons can be made for multiple image reconstruction methods and settings for those methods. Resulting images from a 20 cm head-size phantom (cylinder) with 4 mm contrast features and a 20 cm, 35 cm and 40 cm body-size phantom (extended oval) are shown in Figure 11. Comparison of small lesion detectability between systems and reconstruction algorithms The head-size and body-size phantoms were used to compare multiple configurations of the Discovery MI system to the Discovery MI DR system (the predecessor system). The main Table 5. Comparison of Discovery MI DR and Discovery MI. Lesion-present system characteristics/measurements that contribute to differences in small lesion detectability are summarized in the Table 5. Each measurement utilized six reconstruction settings to enable cross-comparison between algorithms. Two measurements per phantom and system using the head-size and body-size phantoms were acquired at phantom activity concentrations within a factor of 2, at 7.0 kbq/ml and 4.0 kbq/ml, respectively. Comparisons between the model observer SNR at 7.0 kbq/ml and 4.0 kbq/ml and using a scan duration of 2.5 min are shown in Figure 12, with the relative improvements from TOF and system sensitivity clearly evident with the increase in detection SNR. The relative influence of TOF and system sensitivity can be seen in the results. For instance, the influence of TOF is larger on the body-size phantom as compared to the head-size phantom. Further, the change in observer SNR by adding another PET detector ring (15 --> 20 cm or 20 --> 25 cm) is approximately the same or larger than the change from non-tof to TOF, using the same reconstruction method (BSREM). Lesion-absent Figure 11. PET sub-images for lesion-present (left) and lesion-absent (right) with body-size phantom. the phantom contained five layers of seventeen or eighteen (5,6) mm features per layer; only 82 images are shown since one feature had trapped air and was not useful for detection assessment. Clinical impact of the Discovery MI technology The SiPM-based PET detector in the Discovery MI was first implemented in the SIGNA PET/MR system. A clinical assessment of the PET/MR compared to PET/CT were first presented by Iagaru, et al, 16 and confirmed in 36 patients that all lesions seen on PET/CT (imaged first) were also seen on the PET/MR. There are several image comparisons in that work that demonstrate differences in image noise and apparent resolution between the systems. Results from a recent study using Discovery MI PET/CT compared to GE s prior generation PET/CT (Discovery 690 or 600) were presented at the Society of Nuclear Medicine and Molecular

12 Model Observer SNR (d') Head size Phantom CHO d' with 4 sub scans (N = 164), 2.5 min. scan, 7.0 avg. kbq/ml 4.0 mm features in 20 cm dia. phantom DMI DR DMI 15 DMI 20 DMI 25 AUC osem osem+psf bsrem tof osem tof osem+psf tof bsrem Reconstruction Method Model Observer SNR (d') Body-size Phantom CHO d' with 4 sub-scans (N = 328), 2.5 min. scan, 4.0 avg. kbq/ml (5.0, 6.0) mm features in 21x36 cm phantom DMI DR DMI 15 DMI 20 DMI 25 AUC osem osem+psf bsrem tof-osem tof-osem+psf tof-bsrem Reconstruction Method Figure 12. Model observer SNR for head-size (top) and body-size (bottom) phantoms. Using the 4 sub-scan data enabled generation of (164, 328) LP/LA sub-images per phantom, thus providing adequate data for estimation of 95% confidence interval error bars. A second y-axis of AUC is included to aid interpretation.

13 Figure 13. Comparison images from L. Baratto, et al, between PET (left pair) and fused PET/CT (right pair). Imaging conference (June 2017, Denver, CO) and demonstrated that in 39 patients, all 83 lesions seen on the prior generation PET/CT scan were seen on the Discovery MI, as well as an additional 29 areas of focal FDG update. 17 Representative images from this study (Figure 13) show a qualitative improvement in image quality in the Discovery MI. Conclusion As described in this whitepaper, development of the Lightburst Digital Detector encompassed an array of design considerations that would deliver the sensitivity, resolution, image quality, reliability and clinical confidence that clinicians need to detect disease at an early stage when treatment may be most effective. Most importantly, the Lightburst Digital Detector is positively impacting clinical excellence by enabling visualization of small lesions, fast scan times and lower dose as well as propelling new research opportunities with novel tracers for detecting and monitoring disease. References 1. Wagadarikar AA, Ivan A, Dolinsky S, et al. Sensitivity Improvement of Time-of-Flight (TOF)-PET Detector Through Recovery of Compton Scattered Annihilation Photons. IEEE NSS/MIC 2012 Conference Record, pp NEMA Standards Publication NU : Performance Measurements of Positron Emission Tomographs National Electrical Manufacturers Association, Rosslyn, VA, Hsu DF, Ilan E, Peterson WT, et al. Studies of a Next-Generation Silicon-Photomultiplier Based Time-of-Flight PET/CT System. J Nucl Med 2017 Apr 27. pii: jnumed doi: / jnumed [Epub ahead of print]. 4. Budinger TF. Time-of-flight positron emission tomography: Status relative to conventional PET. J Nucl Med, 1983; 24: Lantos J, Iagaru A, Levin CS. Scanner dependent noise properties of the Q.Clear PET image reconstruction tool. IEEE Xplore Digital Library website. document/ /. Published Accessed May 16, Wangerin KA, Ahn S, Wollenweber S, et al. Evaluation of lesion detectability in positron emission tomography when using a convergent penalized likelihood image reconstruction method. J Med Imaging, 2017; 4(1): El Fakhri G, Surti S, Trott CM, et al. Improvement in Lesion Detection with Whole-Body Oncologic Time-of-Flight PET. J Nucl Med, 2011; 52(3): Surti S, Scheuermann J, El Fakhri G, et al. Impact of Time-of-Flight PET on Whole-Body Oncologic Studies: A Human Observer Lesion Detection and Localization Study. J Nucl Med, 2011; 52(5): El Fakhri G, Santos PA, Badawi RD, et al. Impact of acquisition geometry, image processing, and patient size on lesion detection in whole-body 18F-FDG PET. J Nucl Med, 2007; 48(12): Wollenweber SD, Kinahan PE, Alessio AM. A Phantom Design for Assessment of Detectability in PET Imaging. Med Phys, 2016; 43(9): Wollenweber SD, Kinahan PE, Alessio AM. A phantom design and assessment of lesion detectability in PET imaging. Proc SPIE, 2017; 10136:101361E E Wollenweber S, Kinahan P, Alessio A. Application of Lesion Detectability Phantoms for Model Observer Assessment in PET Imaging. J Nucl Med, May 2017; 58(1): Abbey CK, Barrett HH, Eckstein MP. Practical issues and methodology in assessment of image quality using model observers. Proc. SPIE - Int Soc Opt Eng, 1997; 3032: Gifford HC, King MA, Pretorius PH, Wells RG. A comparison of human and model observers in multislice LROC studies. IEEE Trans Med Imaging, 2005; 24(2): Wollenweber SD, Tsui BMW, Lalush DS, et al. Comparison of Hotelling observer models and human observers indefect detection from myocardial SPECT imaging. IEEE Trans Nucl Sci, 1999; 46(6): Iagaru A, Mittra E, Minamimoto R. Simultaneous Whole-Body Time-of-Flight 18F-FDG PET/MRI A Pilot Study Comparing SUV max Wtih PET/CT and Assessment of MR Image Quality. Clin Nucl Med, 2015; 40(1): Baratto L, Park S, Davidzon G, et al. SiPM PET/CT vs. Standard PET/CT: A Pilot Study Comparing Semi- Quantitative Measurements in Normal Tissues and Lesions. J Nucl Med, May 2017; 58(1):432.

14 GE Healthcare provides transformational medical technologies and services that are shaping a new age of patient care. Our broad expertise in medical imaging and information technologies, medical diagnostics, patient monitoring systems, drug discovery, biopharmaceutical manufacturing technologies, performance improvement and performance solutions services help our customers to deliver better care to more people around the world at a lower cost. In addition, we partner with healthcare leaders, striving to leverage the global policy change necessary to implement a successful shift to sustainable healthcare systems. Imagination at work 2017 General Electric Company, doing business as GE Healthcare. All rights reserved. The copyright, trademarks, trade names and other intellectual property rights subsisting in or used in connection with and related to this publication are the property of GE Healthcare unless otherwise specified. Reproduction in any form is forbidden without prior written permission from GE Healthcare. JB50594XX