Discovery ST. An Oncology System Designed For PET/CT. Revision: B Date: 30 Jan Page 1 of 47

Transcription

1 Discovery ST An Oncology System Designed For PET/CT Revision: B Date: 30 Jan 2003 Page 1 of 47

2 TABLE OF CONTENTS 1 Introduction Design Requirements The Design Objective Design Philosophy PET and CT Designs The Primary PET Detector - Which Crystal? New Crystals and Old Crystals Why BGO? Scanner Performance PET Detector Construction The Block The Detector Ring Axial Sampling Performance Implications of the Detector System Transaxial Resolution Sensitivity Scatter Fraction Countrate Performance PET Signal Processing Front End Electronics Signal Processing Performance of Discovery ST as a PET Imager Sensitivity Discovery ST compared to Advance and Discovery LS Recovering Sensitivity on Discovery ST Increasing 2D Axial Acceptance Scatter Fraction Systems Compared Collimation (Septa) Energy Thresholding System Shielding Spatial Resolution Spatial Resolution: Effect of Statistics Statistics & Resolution: What is Required? Resolution Non-uniformity Spatial Resolution: GE vs. Siemens Counting Rate Performance The Objective is Maximizing Return on Injection Summary of Counting Rate Parameters...29 Appendix A: 2D and 3D PET...34 Appendix B: Axial Sampling & Resolution...43 References...45 Page 2 of 47

3 1 Introduction The Discovery ST is an integrated PET/CT system designed for oncology imaging, both for diagnosis and for treatment planning. It was designed as a second-generation PET/CT system after evaluating PET/CT applications in thousands of patients using Discovery LS, the first commercially available PET/CT system. Discovery LS was created by combining an existing CT system with an existing PET scanner, so the first objective was to abolish any limitations resulting from this approach by designing an integrated system from the beginning. This meant that the PET and CT detector systems could use any practical materials, electronics and design. The second objective was to design a system that was the best possible for the oncology application. Designers of previous PET and CT systems have tried to develop general-purpose machines to cover a wide range of applications (neurology, cardiology, pharmaceutical development, etc in PET; vascular studies, biopsy guidance, etc in CT) Combining these objectives with the experience gathered over the last two years with the Discovery LS provided a prioritized list of design objectives. Highest SENSITIVITY to positron events PET SPATIAL RESOLUTION to detect sub-centimeter lesions (6-8 mm) Low SCATTER FRACTION in PET imaging Fast, MULTI-slice CT High-resolution, THIN-slice CT 2D and 3D PET imaging Large patient APERTURE to accommodate Radiation Therapy patients & restraints Simple and reliable operation (easy to use). Based on these priorities a completely new PET detector and electronics system was designed for the Discovery ST. For a number of reasons it was decided not to design a CT especially for this product, but to use an existing design. The special requirements of the CT system were very minor, and by using a widely available CT system increases the flexibility of the system for stand-alone CT applications. In the case of the PET detector there were several good reasons to embarking on a new design, the principal ones being: - The sensitivity of PET could be increased by a new design - Although the patient aperture could be enlarged, it could only be achieved with existing designs by seriously degrading scatter fraction. - The footprint and size of the machine could be significantly reduced if the PET detector system were designed around the CT gantry, instead of placing a second gantry beside it. - By initiating a completely new design it would be possible to take advantage of recent advances in detector materials, light detectors and electronics in order to improve many performance parameters. Page 3 of 47 The result of the Discovery ST program is an entirely new PET/CT system which is a real advance in PET/CT capability, and provides the very best performance in diagnostic oncology, and in oncology therapy planning. When the Discovery ST is evaluated against standard performance tests (NEMA

4 1994; NEMA 2001; IEC) it demonstrates clear superiority over all other PET or PET/CT systems currently available. 2 Design Requirements 2.1 The Design Objective The Discovery ST is an integrated PET/CT system designed for oncology. All other PET/CT systems have been created by combining separate PET and CT systems, which were designed for singlemodality operation [Townsend et al 1998]. Discovery ST is the first system designed from the beginning as an integrated PET/CT system. The PET detector system has never been produced as a stand-alone PET system, but was created expressly for the Discovery ST. The CT system is available separately, but the design brief was for CT that could be used alone or as a PET/CT combined unit. The stand-alone feature is retained here to permit the CT to be installed first, and later upgraded to a PET/CT system. In creating the Discovery ST the objective was to produce the best possible system for use in Oncology. The applications of PET imaging in the diagnosis, staging, and treatment planning of cancer have grown rapidly in both number and variety over the last five years. Although the principal applications of PET imaging were once neurology, cardiology, oncology and pharmacology research, today about 90% of PET scans are performed on cancer patients. At the same time, CT is rapidly becoming established as a fundamental requirement of high-quality PET. Combined PET/CT imaging is becoming the minimum standard for FDG imaging in oncology, while only two years ago it was considered an esoteric luxury. The explosion of PET oncology procedures provides a different focus on the priorities in PET imaging. For oncology the tracer of interest is Fluoro-DeoxyGlucose (FDG). New and more specific tracers are on the way and will greatly improve the value of PET/CT imaging, but it is clear that the current FDG distribution model will apply to these new tracers. They will almost certainly be compounds of Fluorine-18 or longer-lived isotopes (e.g. Copper-64). So we know that for clinical Oncology studies we won t be using large quantities of short-lived tracers (Carbon-11, Nitrogen-13, Oxygen-15), and we can place an effective upper limit on the injected tracer of about 20 millicuries [ACR Standard, 2000; Wahl, 2001] Similarly, we know that the in Oncology the prime task will be locating regions of higher concentration of tracer in a background of lower tracer concentration ( hot-spot imaging ). We know that in the great majority of cases a large volume of the patient will be imaged, usually the whole of the neck, trunk and upper thighs. And finally we know that a concurrent, registered CT image will be required. The result is a very different emphasis on design than was used for all other PET scanners to date. In those designs the it was necessary to compromise the oncology requirements with those of supporting very high tracer concentrations of ultra-short-lived isotopes, detecting subtle reductions in tracer in a region of an organ, performing fast dynamic tracer kinetic studies on the brain, imaging tracers with coincident high-energy gamma emissions (e.g. Iodine-124) and anticipating new PET research applications. GE s design team were able to focus on the parameters of a PET/CT system which are most important for Oncology: detecting malignant tissue, assessing its metabolic activity, identifying its anatomical Page 4 of 47

5 location and that of the surrounding structures, and gathering the information required to deliver the optimal therapy. The essential performance parameters were: a) Highest SENSITIVITY to positron events b) PET SPATIAL RESOLUTION adequate to detect sub-centimeter lesions c) Low SCATTER FRACTION in PET imaging d) Fast multi-slice CT e) High resolution, thin-slice CT f) Capacity to image in 2D PET and 3D PET modes g) A large patient aperture to accommodate patient restraints 2.2 Design Philosophy From almost ten years of clinical experience with the Advance PET Scanner [Degrado et al 1994, Lewellen et al, 1999] and the early experience with the first commercial PET/CT system, Discovery LS, together with careful study of the performance of non-ge PET scanners, the design team set out to understand everything that was good, and every weakness, of existing systems. At the same time they searched the industrial, academic and independent research laboratories to identify new technology and opportunities for innovation in all aspects of the system. With this basic groundwork, the team then set out to question and reassess every element of the system, and to identify the best possible route to perfection. Six Sigma processes were employed at every step. Testing to Six Sigma tolerances was carried out at every phase, and each element of the design selected was required to meet the exacting standards of this rigorous system of quality and reliability. Nothing was taken for granted. No element of concept of the Discovery LS was adapted for the new system unless if proved the best of all possible solutions. In order to perform the numerous comparisons to select the best materials and designs, extensive use was made of computer simulations, but with rigorous practical experiments at every stage to validate the simulation results. Many hundreds of different detector configurations were simulated, and equally many options in the signal processing pathways were simulated to identify the optimal design choices in the Discovery ST. As a result of this process the scintillation crystal block detector size, shape, and construction are entirely new. The photomultiplier tubes that measure the light emitted from the detector blocks are an entirely new design, especially for the Discovery LS. Every element of the acquisition, coincidence processing and signal processing electronics has been completely redesigned. In a few areas, for example the reconstruction processor, the latest designs developed for the Discovery LS have been re-used, but only because they provided the best quality available. 2.3 PET and CT Designs Page 5 of 47 The PET and CT portions of the new system were treated differently in the original outline. With clinical PET moving inexorably toward PET/CT, the focus was on designing a fully integrated system. The capability of operating the PET detector without CT present was given a low priority. On the other hand, CT is in a phase of very rapid technological development, and the great majority of CT systems are CT only. PET/CT represents only a small part of the CT market. For these reasons it was felt best to build the new PET/CT

6 system on a leading, high-end CT system and in such a way that new CT developments can be brought to the Discovery ST as rapidly as possible. The result of this philosophy is a system that uses the LightSpeed CT family as the core structure, and the PET detector is designed around this core. The CT system can be operated as a stand-alone CT system by a technician without knowledge or experience of PET. Since the core is a LightSpeed CT, the process of evolving the PET/CT to include new members of the LightSpeed family is simplified. The Discovery ST can easily be upgraded in the future. At the same time the PET system is designed, not as a stand-alone unit, but to be integrated fully with the CT systems, producing a remarkably compact unit, only slightly larger than the stand-alone CT. Page 6 of 47

7 3 The Primary PET Detector - Which Crystal? The primary PET detector in Discovery ST is a BGO block. BGO (Bismuth Germinate) was selected as the primary detector (crystal) material over several alternative materials including LSO (lutetium orthosilicate) and GSO (germanium orthosilicate) for one simply reason only: performance. Although there is a widespread perception that BGO is an old material and LSO and GSO are new and therefore somehow better, the design team analyzed each of these materials (and many others) before selecting BGO. It is worth noting that GSO, far from being new, is the material which BGO replaced in 1986 in the GE/Scanditronix scanner family. 3.1 New Crystals and Old Crystals The reasons for the choice of BGO can be seen from tables 3.1 and 3.2 below. The physical properties of various scintillators are show in these tables[sorensen, 1987; Wahl, 2001] BGO NaI(Tl) GSO LSO Density Effective Atomic No Atten Coeff µ(cm-1) Photofraction ~40% ~15% ~25% ~30% Light Output 20-25% 100% 35% 75% Decay Constant Availability High High Limited Limited Growth Crystal Crystal Crystal Crystal Melting Point Table Physical Properties of Some PET Scintillators It is widely recognized that both GSO and LSO have significantly higher light output, and a faster decay times than BGO. Both of these properties are attractive in PET scanners. They aren t the only parameters, however. In designing an efficient PET detector we must consider the total performance. To understand the impact we need to examine each of the crystal parameters in detail. Density - LSO is slightly denser than BGO, and both are substantially more dense than GSO. However density is not important by itself. Density is one of many factors that contribute to sensitivity or efficiency the ability of the crystal to stop gamma rays of the required energy 511keV. Effective Atomic Number - The effective atomic number is more important in determining the efficiency of a material at stopping energetic gamma rays. A higher atomic number results in more electrons per atom, and the more electrons, the more likely a gamma ray is to interact with an electron. BGO is the best of all the materials in this category. Page 7 of 47 Attenuation Coefficient - The attenuation coefficient, mu (µ) is a measure of the stopping power of the material. The attenuation coefficient depends on the density, but more importantly on the electron density of the material. Now we see that although LSO is more dense than BGO, its attenuation coefficient is lower, meaning that it is less efficient at stopping gamma rays. GSO, as expected is much less efficient.

8 Photo-fraction Stopping gamma rays isn t enough. When a 511 kev gamma ray interacts with any material, it may be absorbed immediately (a photo-electric event) or it may be scattered (Compton event). If the ray is scattered it losses some, but not all, of its energy, but then continues on. A 511 kev gamma ray may be scattered once, twice, or more before eventually stopping in a photo-electric event. In a PET scanner we want as many of the 511 kev gamma rays to be stopped with photo-electric events the first time, with no scattering. If the gamma ray is scattered, then it emits some light from the place where it was first scattered, but the remainder of the light will be deposited at places of later interactions. These piecemeal sources of light usually result in the gamma ray being lost (because the computed energy is very low), or being detected in the wrong position. Photofraction is the fraction (of 511 kev) gamma rays which interact with a single photo-electric event. These are the events which will be detected correctly. The effective sensitivity of a detector material is determined by the Attenuation Coefficient and the Photofraction. Light Output The total light produced when a 511 kev gamma ray is absorbed. More light allows the designer to connect more crystal to each photomultiplier tube, or it (usually) allows the designer to measure the energy of each gamma ray more precisely. Connecting more crystals to each photomultiplier tube doesn t improve performance of the PET scanner; it may save some cost by reducing the number of tubes, but since both GSO and LSO are more expensive than BGO, the net cost usually goes up. If the energy resolution can be improved, that does improve scanner performance by making it possible to reject more scattered radiation. In almost all scintillation materials the energy resolution is directly related to the amount of light produced. The very significant exception is LSO. LSO suffers from what is referred to as nonproportionality of light output. [Mengesha et al, 1992; Balcerzyk et al, 1999; Saodi et al, 1999] What that means is that the amount of light produced varies from one event to the next and therefore the energy resolution is much worse than would be expected from the light output. This effect can be seen in real scanners. The manufacturer s quoted energy resolution for the Accel (CPS Inc), the only LSO scanner, is 25%. The energy resolution of the GE Advance NXi scanner is 20%; and, as we shall see, the energy resolution of the GE Discovery ST system is 17%. So the BGO systems from GE have better energy resolution than the LSO system. In the case of LSO the increase in light output contributes nothing to the performance of the PET scanner. In fact, because of the non-proportionality effect, the performance degrades GSO does not suffer from the same defect as LSO, so the GSO scanner does gain in energy resolution. The only GSO scanner currently available, Allegro (Philips Medical Systems), has a quoted energy resolution on 16% [MEEN, 2002] (but compare that with 17% in the Discovery ST). Decay Constant The Decay Constant is a measure of the time it takes for the light from a scintillation event to appear. When a gamma ray enters a scintillator light begins to appear very quickly and reaches a peak within a few nanoseconds of the gamma ray interaction. Light continues to be released, but with decreasing intensity over the a period of time. The intensity falls with an approximately exponential decay. The Decay Constant is the time it takes to fall to half the peak intensity, and so after about 3 times the decay constant, 95% of the light will have been emitted. The decay constant is important in PET imaging for two reasons. The accuracy of timing is influenced by the decay constant. The faster the light emerges, the better our ability to calculate precisely the time at which the light started to emerge. This is important in determining whether two gamma rays were Page 8 of 47

9 detected at the same time, and hence whether they were the result of a single positron annihilation. A coincidence window is used to determine if two gamma rays were created from a single positron. If the time between the two gamma rays being detected is less than some time, 2τ (two times tau; the reason for two will be explained in section 4.2 below) we consider that they were detected at the same time, and we assume they originated from one positron annihilation. Clearly we need to be able to measure the time that a gamma ray enters the detector to a higher precision than the time of the coincidence window. For example, if we could only measure the detection time to the nearest 20 nanoseconds we could not set the coincidence window to 10 nanoseconds without rejecting a lot of valid events. But if the light from the scintillation event is emitted very quickly (as in GSO or LSO) then we can measure the time with better precision, and so we can reduce the coincidence window and still detect all the true coincidences. As well as true coincidences, there are random coincidences (or randoms ). Randoms are pairs of gamma rays that enter the detectors at the same time purely by chance, even though they didn t come from the same positron annihilation. When millions of gamma rays are being detected, two will occaissionally happen at about the same time. But the longer the coincidence window, the more likely it is that two gamma rays will be detected within window by chance. So by having faster light output the time at which gamma rays occur can be measured more accurately. If the time can be measured more accurately, the coincidence window can be shorter, and therefore randoms are less likely. A crystal with a faster decay time produces a lower randoms rate. This is the principal advantage of a faster decay time, although there is another. The decay time also determines how long an event lasts. With a 300 nsec decay time, it takes about 1000 nsecs for one event to clear and allow the crystal to detect another event. If the decay time is only 50 nsecs, then another event can be detected after only 150 or 200 nsecs. This can be important in determining the total event rate, or counting rate performance of a PET scanner. Both of these parameters (randoms and counting rate) influence the performance of the PET scanner when there is a lot of radioactive tracer in the patient. The rate of random coincidences is proportional to the duration of the coincidence window, but it is also proportional to the square of the singles rate the rate of all gamma rays (good and bad) being detected. At low counting rates the randoms rate is not important (as in most 2D imaging), but at higher rates it becomes much more important. Similarly the time to process an event doesn t affect performance at low event rates, but at high rates one event may occur before the previous event was finished (pile-up) resulting in the loss of both events [Erriksson et al, 1994]. Availability The remaining parameters in the table, availability, growth, and melting point are important for the designer because they affect the cost and the difficulty of obtaining high quality material, but they do not directly impact performance. 3.2 Why BGO? Scanner Performance So having examined the various crystals available we have to choose one. For the Discovery ST the choice was made by taking the basic physical parameters above and converting them to realistic performance parameters (Table 1.2 [MEEN, 2002] ), and then comparing these with the design requirements listed in section 1.1. Page 9 of 47

10 BGO NaI(Tl) GSO LSO GE Advance UGM CPET+ Philips Allegro CTI Accel Rel. Coincidence Sensitivity 100% 15% 50% 80% Coincidence Window 12 nsec 12 nsec 8 nscec 6 nsec Peak NECR 3D 38 kcps?? 48kcps 58kcps Energy Resolution 20% 15% 16% 25% Scatter Fraction 30% 30% 30% 35% Table Performance Properties of PET Scintillators in Scanners BGO values are from the GE Advance Nxi, GSO from the Philips Allegro, and LSO values from the CTI Accel. All values from When the physical properties are applied to performance parameters in actual PET scanners the crystal material affects Sensitivity, Counting Rate parameters, and Scatter. The relatively small differences in attenuation coefficient and photo-fraction between BGO and LSO result in an advantage of just over 10% for BGO in detecting single gamma rays. But in PET we need to detect both gamma rays to detect a single event, so the effective sensitivity of LSO is actually 20% lower than BGO for coincidence events. GSO is substantially less efficient, so the effective sensitivity is actually about half that of BGO. For high counting rates, both LSO and GSO offer significant advantages due to lower randoms rates and a shorter event duration. In practical scanners the randoms rate of GSO and LSO is about half of that in BGO scanners. Energy resolution is of value only in that it helps separate scattered from unscattered gamma rays. It is important, therefore to look at the Scatter Fraction, rather than to examine the energy resolution in isolation. What this shows is that, although the energy resolution of GSO is slightly better than BGO, in practice the scatter fraction of BGO can match GSO and beat LSO. So we find that the only remaining advantage of both GSO and LSO is at high counting rates. And in each case we must accept a reduction in sensitivity, 20% with LSO and 50% with GSO to obtain it. When we look at the design objectives for the Discovery ST we find that the prime goal is maximizing sensitivity while high tracer activities (and hence high counting rates) is a very low priority. As a result BGO clearly offers the best choice. The sensitivity and scatter characteristics are significantly better than the alternatives, while the main disadvantage (high counting rates) isn t a requirement for oncololgy imaging. Page 10 of 47

11 4 PET Detector Construction 4.1 The Block The primary detector for the PET unit is a BGO block detector of 36 crystals with a single, four-anode square photomultiplier in each block. This can be compared to competitive systems using single-anode, circular photomultipliers, and the Discovery LS with two dualanode rectangular photomultiplier tubes per block. The advantage of an integrated (multianode) tube is that there is less loss of light between the tubes. Square (or rectangular) tubes also provide full coverage of the crystals, so that light cannot emerge between the tubes where there is no photocathode surface, as occurs with round tubes. Non-GE Design Light Detected Discovery LS uses two Dual- PMTs per block Light Lost GE Designs Light Detected No Light Lost Discovery ST uses one Quad- PMT per block Figure 4.1 Detector Structures: light loss between PMTs degrades performance Discovery ST uses specially designed four anode tubes (Quad PMTs: Discovery LS and Advance use dual-anode tubes) and new electronics. The new electronics is able to correct event timing individually for each crystal rather than on a block basis, so the larger timing variation seen across larger PMT s can be accommodated. The result is a detector block that is extremely efficient at collecting all the light available from a scintillation, and which has even better energy resolution than the Advance Nxi (17% compared to 20% for the Advance). The crystal dimension on Discovery ST is 6.3 mm in both the axial and transaxial directions, and 30 mm deep. This can be compared to 4mm x 8mm x 30 mm crystals in the Advance NXi, and 6.8 mm x 6.8 mm x 24mm in the LSO scanner. As other parameters, the crystal dimensions were selected to give the best overall performance compatible with the design objectives. Page 11 of 47

12 Discovery ST Detector Figure Blocks are assembled to make the detector ring. The assembled detector contains 24 rings of 420 crystals. 4.2 The Detector Ring 4.3 Axial Sampling The detector ring diameter is chosen to permit optimized image quality with a bore size of 70 cm, in contrast to the 60 cm bore of the Advance system. The prime reason for this increase is to allow patients about to undergo Radiation Therapy to be imaged in the positions in which they will be treated. The Discovery ST bore diameter of 70 cm means that there is no reduction in the tunnel diameter as the patient is moved from the CT scan plane to the PET scan volume reducing the likelihood of interference or claustrophobia issues, and increasing the proportion of patients who can be scanned and providing adequate space for Radiation Therapy Planning applications. The detector architecture of Discovery ST uses four blocks of six crystals in the axial direction, and therefore has 24 rings of crystals in contrast to 18 rings on Advance over the same 15 cm axial field of view. This results in effective axial sampling (slice spacing) of 3.3 mm for Discovery ST, an improvement of 26% over the Discovery LS axial sampling of 4.25 mm Page 12 of 47

13 4.4 Performance Implications of the Detector System Page 13 of Transaxial Resolution The crystal dimension on Discovery ST is 6.3 mm in both the axial and transaxial directions. The crystral dimensions are the primary determinant of spatial resolution. The Discovery ST provides better resolution than the Discovery LS in axial direction, while sacrificing some resolution within the slice. The result is nearly isotropic resolution (the same actual resolution in all directions). While the transaxial resolution is not as good as the Discovery LS, it matches the best that can usually be expected for clinical imaging in Oncology, where the limiting resolution is determined by counting statistics. In certain neurological studies with very large tracer concentrations, sufficient information may be collected to justify resolutions below 6.5.mm, but this rarely happens in FDG cancer studies Sensitivity Sensitivity usually refers to overall sensitivity, expressed as the total trues events detected per second for a given amount of tracer in the imaged volume (counts per second per kilobequerel). In assessing image quality, it is also relevant to examine slice sensitivity, which is the number of events per image (slice). Improving the axial sampling necessarily reduces slice thickness, which in turn reduces per slice sensitivity. This is a particular problem for 2D imaging, where the reduction goes as slice thickness squared, so a 26% reduction in slice thickness reduces sensitivity by 45%. In the Discovery ST this loss of sensitivity has been bought back by increasing the axial acceptance angle by reducing the detector diameter. These design actions restore the per-slice sensitivity to a value equivalent to that of Discovery LS. However, Discovery ST has 24 slices compared to Discovery LS with 18, so the overall 2D system sensitivity for Discovery ST is 50% higher than that of Discovery LS. A significant element of this increased 2D sensitivity is related to the septa. In order to increase the detector bore (to accept larger patients), the septa must be shorter. This in turn produces a larger acceptance angle and an increase in sensitivity. At the same time it allows more scattered radiation to reach the detector, so the 2D scatter fraction is increased. Improvements in 3D acquisition, and the smaller detector diameter, mean that the 3D sensitivity for Discovery ST is also higher than that of Discovery LS, by about 20% Scatter Fraction The increased axial acceptance angle improves system sensitivity, and also increases scatter fraction. On Discovery ST, the estimated scatter fraction is 16% in contrast to 10% for Discovery LS. Note that 16% is equivalent to the scatter fraction reported for the best CPS systems in 2D. The selected geometry and resulting scatter fraction value were chosen to optimize the noise equivalent count rate (NECR) of the system. The result is a system which has an NECR that exceeds that of Discovery LS in head imaging, and comes within 10% of Discovery LS in large bodies, the worst case for scatter and randoms. The net result is that clinical images will look as good or better on Discovery ST than Discovery LS Countrate Performance The improvements made in Discovery ST electronics result in a dramatic improvement in system countrate capability. Discovery ST shows greater than a factor of two improvement over Discovery LS in countrate, and greater than a factor of four improvement over BGO systems produced by CPS. These improvements in electronics also translate into shorter calibration times reduced by a factor of 50 or 100. This permits the use of a single, weaker source and eliminates the need for using two transmission sources for calibrations or quality assurance.

14 Page 14 of 47

15 5 PET Signal Processing 5.1 Front End Electronics All elements of the PET system have been re-examined, and the majority of the signal processing path of the Discovery ST is a new design. The front-end refers to the initial signal processing that occurs immediately after the photomultiplier tubes. The objectives of the front end electronics in any PET scanner are: - to detect when a gamma ray is absorbed in a crystal - to identify the time at which the event entered - to compute the energy of the event - to compute the position of the event (which crystal in the block) - to perform all of the above as accurately as possible The Discovery ST converts the analog signals from the photomultiplier tubes to digital signals immediately. There are four signals available from each of the four-anode tubes, and these are constantly digitized at a rate of 48 MHz. At the same time event detection circuitry detects the beginning of a gamma ray event in the block. Integration of the signals is a digital process involving summing each of the digital values measured during the integration time. Because BGO is a relatively slow scintillator, integration proceeds for 700 nanoseconds. In more primitive PET scanner designs a long integration time can result in problems due to pile-up. Integrated Signal is Digitized (Once) adc adc Analog Processing and Integration Add Noise. Digital Processing Reduces Noise, and Improves Energy Resolution Original Signal is Digitized many times Figure 5.1 The Signals from the 4-anode PMTs are digitized every 21 nanoseconds and integrated digitally Pile-up occurs when a new gamma ray event occurs before all the light from the previous event has been emitted. When this happens the first event is likely to be contaminated with light from the second event, and the second with some residual light from the first. As a result both events should be rejected. The likelihood of pile-up is directly related to the total event rate, and the time it takes for all the light to be emitted from the crystal. Complex electronics are frequently used to clip the signals so the duration of the electronic signal is shorter than the period of light emission in the crystal. The Discovery ST uses a much simpler solution. Page 15 of 47

16 Pile-up is directly related to the event rate. The front end electronics is required to recognize and process every gamma ray entering the detector blocks, whether or not it turns out to be a good (coincidence) event. The total event rate (in the whole scanner) must be very high, but the event rate in a single crystal is quite small. With over ten thousand crystals in the detector the event rate in each crystal is quite small. With a total event rate of 100 million gamma rays per second, the event rate in a single crystal is only 10,000 per second. The probability of pile-up occurring within the crystal is very low indeed, and is never a problem. Pile-up becomes a problem when the signals from many crystals are combined in the electronics components. Each photomultiplier group (four anodes) views 36 crystals, so the likelihood of pile-up in the photomultiplier is 36 times more likely than pile-up in a crystal. For an overall event rate of 100 million events per second, the event rate in each block will be about 350,000 per second. At this rate the probability of pile-up with a 700 nanosecond integration time is less than 3%. The solution then, is to ensure that very small units of the detector are processed individually, at moderate event rates, rather than combining signals from large volumes of crystal and exacerbating the problem. 5.2 Signal Processing The major advantage of digital electronics is that digital electronics doesn t add noise. Every analog operation, whether amplification, conversion, attenuation, integration or transmission, adds some noise, however small. Digital processes do not. By converting the signals emerging from the photomultiplier tubes to digital signals as rapidly as possible, without passing them through long cables or analog circuits, the signal noise is kept to a minimum. This benefit appears in many places, but especially in improved energy resolution and in system stability. Once the front-end electronics has identified an event (a single gamma ray interaction), the values associated with that event must be routed to a central clearing-house that can search for a matching gamma ray from the same positron event. That matching is based on timing. For each single gamma ray event that was detected at time, t, we need to identify any other event, in any other detector block, that occurred within τ (half the duration of the coincidence window) before or after that time. For example, if a gamma ray is detected in block 22 at time seconds (measured to the nearest nanosecond), and the coincidence window is 12 nanoseconds (τ=6 nanoseconds) then we need to see if any events occurred elsewhere in the time from to nanoseconds. Detectors Processing Channels Detectors Processing a) Each processor connected to N detectors b) Each detector connected to N processors Figure 5.2 Smart Multiplexing ensures that multiple events aren t lost Page 16 of 47

17 The processing module which performs this process must be very fast and very efficient to deal with such a complex searching task in a very short time. It turns out that one of the most difficult problems is getting the data from the many detector units to the coincidence processing unit. Events are occurring very frequently and at totally random times, so the design must allow for the events to be transferred to the coincidence processor at very high rates but without running into each other. The problem becomes very similar to the pile-up problem of the front-end electronics, except that it isn t possible to isolate small units with low count rates. The solution is an intelligent multiplexing system. The conventional approach is to provide a number of portals into the processor, each of which can function independantly, and to connect a number of detector blocks to each portal. For example, 8 portals may each have 32 detector blocks connected. This can be very efficient, but the problem is, what happens if 2, or 3 or the 32 blocks connected to one portal generate events at the same time. Only one event can be transferred, and the portal is blocked to the others. The Discovery ST connects multiple portals to multiple detector blocks, so each detector is connected to several portals. When a detector block wishes to transfer an event to the coincidence processor it simply uses the next free portal. The advantage of this intelligent multiplexing is seen in the count rate performance of the system. When choosing BGO over LSO for the primary detectors it was accepted that count rate performance was being sacrificed for sensitivity. Despite this choice, the sophistication of the electronics has produced a system whose count rate performance (peak trues and peak NECR) are even better than current LSO systems. Page 17 of 47

18 6 Performance of Discovery ST as a PET Imager 6.1 Sensitivity Performance of PET scanners is best assessed using standard test procedures. The two widely accepted standards are the NEMA standard, NU-2, and the IEC standard. There are two revisions of the NEMA standard that are widely used, NEMA NU-2, 1994 is the original standard, and NEMA NU-2, 2001 is the revision published in The IEC standard is very similar to the NEMA 1994 standard upon which it is based. In section 1.1 it was determined that the most important single parameter impacting the performance in oncology imaging was sensitivity. Sensitivity is a direct measurement of the information collected for each unit of tracer in the patient. Sensitivity is measured in units of true counts collected for each microcurie (or kilobecquerel) of tracer present in a defined phantom. (In NEMA 2001 sensitivity is expressed as counts per second per kilobecquerel. In NEMA 1994 sensitivity is expressed as counts per second per microcurie per milliliter in a defined phantom.) The sensitivity of the Discovery ST, measured by each of the NEMA standards, in 2D and 3D modes are given below: 2D Sensitivity 3D Sensitivity NEMA kcps/µci/ml 1280 kcps/µci/ml NEMA cps/kbq 7.5 cps/kbq Discovery ST compared to Advance and Discovery LS Discovery ST has 50% more rings than Advance and Discovery LS but has the same sensitivity per-slice, therefore Discovery ST has 50% higher system sensitivity than Discovery LS. The Discovery ST detector diameter is 5% smaller than Discovery LS, which increases 3D sensitivity proportionally. In addition, Discovery LS acquires 3D data assuming that reprojection processing will be used, which in turn works best by limiting the 3D axial acceptance angle. Discovery ST assumes the availability of Fourier rebinning, which permits opening up the 3D acquisition to the full acceptance volume. This permits a 15% increase in sensitivity. In combination these two factors provide a 20% increase in 3D sensitivity for Discovery ST. 2D/3D Sensitivity (kcps/µci/cc) (Nema NU2 1994) Discovery ST Discovery LS Accel (LSO) Allegro(GSO) HR+(BGO) 2D TRUES D TRUES (per Slice) D TRUES Figure 6.1 System Sensitivity of Various PET Scanners Page 18 of 47

19 6.1.2 Recovering Sensitivity on Discovery ST In 3D imaging the sensitivity is determined almost entirely by geometry and the stopping power of the detectors. The geometry determines the probability that both gamma rays emerging from a positron annihilation will enter detector crystals, and the stopping power determines the probability that gamma rays entering the crystal will be stopped. The solid angle is larger if the axial field-of-view is longer, and the detector ring diameter is smaller. The stopping power is determined by the thickness, and the properties, of the crystal material. Until the arrival of Discovery ST, the Discovery LS (and the Advance NXi) was the system with the highest sensitivity. The Discovery LS collects about 18% more trues than the next most sensitive system, the Exact HR+ (1060 kcps/uci/ml versus 900 kcps/uci/ml). The principal reason for this difference is the geometry. Although the HR+ has a similar size, the detector construction leaves numerous gaps between the crystals. Gamma rays entering these gaps are less likely to be detected. Other competitive systems have lower sensitivity than the HR+, and much lower sensitivity than either of the Discovery models, primarily because of crystal thickness. Discovery LS and ST each have 30mm thick crystals, while most competing systems have only 20mm. The LSO system uses 24mm thick crystals, but this is equivalent to only 20mm of BGO. The GSO system uses only 20mm of GSO, so has a much reduced sensitivity. The GSO scanner recovers some of the lost sensitivity by increasing the axial field-of-view, which provides a significant increase in sensitivity in 3D mode. Discovery ST achieves an increase in sensitivity even over the Discovery LS by improved detector packing, and by a slight reduction in the diameter of the detector ring. When operating in 2D mode the same factors (geometry and stopping power) determine sensitivity, but now it is the slice thickness, not the entire axial field-of-view that determines the counts in each slice. The sensitivity of the Discovery ST is much greater than the Discovery LS in 2D mode, due principally to a shorter collimator. Hi-Resolution Direct Plane Hi-Resolution Cross Plane Hi-Sensitivity Direct Plane Hi-Sensitivity Cross Plane Hi-Sensitivity includes photons that would be lost Hi-Sensitivity Works only near central axis Figure 6.2 How Slices are combined in 2D Imaging Increasing 2D Axial Acceptance Rather than losing the events that miss a narrow detector ring, the detectors in the neighboring rings are used to detect these events. On Discovery LS this scheme is referred to as high-sensitivity mode. Page 19 of 47

20 This differs from the case of the cross-plane, where coincidence is allowed between detectors of two adjacent rings. The cross-plane is acquired to detect sources that lie between physical rings. In the case of high-sensitivity mode, coincidence is allowed between two rings that are not directly adjacent, in order to enhance the sensitivity of the ring that lies between them. The concept of high sensitivity is also applied to cross-planes themselves to enhance their sensitivity. In this case, coincidence is allowed between detectors in the two rings that lie outside the rings used for the cross plane. Discovery LS users acquire in high-sensitivity mode almost exclusively because counts are critical in PET imaging. This mode increases sensitivity by using ring offset of 0 and offset of ±2 for direct slices, and using offset of ±1 and offset of ±3 for cross slices. Discovery ST increases sensitivity by increasing the accepted ring difference adding offsets of ±4 and ± Scatter Fraction Scatter fraction is the most direct measurement of the bad data included in a scan. Scatter occurs as gamma rays emerging from the patient are scattered in the patient, table-top, or other material in or near the imaging field-of-view. In clinical studies the amount of scatter depends on the size and shape of the patient, and on the distribution of tracer within the patient. The NEMA specification of scatter fraction is intended to indicate the relative effectiveness of the scanner at preventing scattered events from being counted. When patients are imaged the scatter fraction will almost always be larger than the specified scatter fraction. The actual scatter fraction in clinical imaging depends principally on the shielding and the energy resolution, and on the size of the patient. In 2D imaging the scatter fraction in a large patient may be double the NEMA scatter fraction, and in 3D imaging of large patients the scatter fraction can easily be two or three times the NEMA value. Despite the high dependence on patient size (weight), the NEMA value is still a very important guide to how effective the scanner is at rejecting scattered events. This is not the same as scatter correction. Scatter correction attempts to correct for scattered events, which have been collected. Scatter correction is a poor substitute for scatter rejection, because although the distribution of included scatter can be calculated fairly accurately, when it is removed there is a large increase in statistical noise Systems Compared The Discovery LS has a lower Scatter Fraction in 2D than any competitor, including Discovery ST. This is due to the longer collimator septa: 11.7 cm vs. 6 cm of tungsten. Scanners using thin crystals, or crystals with lower stopping power (e.g. LSO or GSO) are less efficient at detecting unscattered radiation (good events) but are as efficient at detecting the lower energy scattered radiation. GSO systems suffer from this problem, but the relatively good energy resolution of GSO allows more scatter to be rejected, giving a net scatter fraction that is similar to the Discovery LS and Discovery ST. SCATTER FRACTION (Nema NU2 1994) Page 20 of 47

21 Discovery ST Discovery LS Accel (LSO) Allegro(GSO) HR+(BGO) 2D 16% 10% 23% - 16% 3D 29% 30% 48% 30% 35% Figure 6.3 Scatter Fractions of Various PET Scanners For 2D imaging, even the worst-performing system (Accel, the LSO system) performs adequately, because the scatter content is still within a range where a good scatter correction algorithm can correct for it without introducing an overwhelming noise content. In 3D imaging, however, all systems have a much higher scatter content. The values in the table above are those measured using the NEMA method, and are appropriate for a 20cm diameter object (patient). For realistic patients the values are much higher, easily exceeding 70 or 80%. When the scatter content is that high it overwhelms any reasonable correction algorithm. It isn t that the scatter distribution can t be calculated: it can, and with considerable accuracy. The problem is noise. When more than 80% of the collected events are subtracted (because they are scatter and randoms), even if the quantity to be subtracted is known accurately, the statistical noise is extremely high. This is the fundamental problem with scatter. In large patients the total fraction of gamma rays scattered is very high indeed. If the detector collects these scattered events, then, even though the distribution and quantity of scattered events is known, the process of correcting for scatter results in a large increase in statistical noise. If at the same time as increasing scatter, we also increase randoms, as we do in 3D imaging, the problem becomes worse. The scanner design should therefore be carefully tailored to minimize the amount of scatter accepted. The biggest reduction in scatter can be achieved by the use of septa (collimation) to reduce cross-plane events. This results in what is termed 2D imaging. It can be seen from the table above that 2D imaging reduces the scatter fraction by at least 50% in all scanners. For 3D imaging the principal tools for reducing scatter are energy discrimination and shielding. Page 21 of Collimation (Septa) Collimation reduces scatter by severely restricting the directions of gamma rays entering the detectors. Collimation consists of thin discs of tungsten (or lead in some systems) placed between the detectors and the patient, and between the rings of individual crystals. The effect is that only gamma rays traveling nearly perpendicular to the patient axis are able to enter the detectors without striking the collimator discs. Because of the selection of thinner slices for Discovery ST, the optimum collimator length for NECR leads to the selection of a shorter collimator. This has the fortuitous consequence of optimizing close to the desired patient port diameter of 70 cm. The collimator length for Discovery ST is therefore shorter than the Discovery LS collimator 6.4 cm vs cm. Since scatter fraction goes roughly as the inverse of the collimator length, one would expect the shorter collimator length to nearly double the scatter fraction in Discovery ST. However, the scatter fraction is also proportional to slice thickness, which is 25% smaller on Discovery ST. So rough calculation would place the scatter fraction on Discovery ST at 10% ( ) ( ) 14%.

22 The actual scatter fraction in 2D is 16% (Nema NU2 2001). Patient port (as large as possible) Collimator (as long as possible) Detector ring (as small as possible) Figure 6.4 Competing interests of patient aperture Energy Thresholding Another means of reducing scatter is energy discrimination. This method relies on the fact that a gamma-ray photon that has undergone Compton scattering lost some energy in the process. Therefore detected photons that have an energy much less than 511 kev are probably scattered and should be rejected. The detectors are not able to measure the energy of the incident gamma-ray accurately. In fact, the energy measurement is only accurate to around 17-20%, depending on the scanner design. An unscattered gamma ray of 511 kev could easily be measured at 410 kev purely due to the uncertainty in the detector. The threshold at which we reject events must be fairly low if we are to avoid discarding unscattered gamma rays. The Discovery LS uses a threshold of 300 kev (or 40% below) for 2D imaging, and 375 kev for 3D imaging (where scatter is more prevalent). Increasing the threshold reduces scatter but can also impact sensitivity to true events. Recent measurements have shown that for Discovery ST (with an energy resolution of 17%), increasing the threshold to 380 kev reduces scatter without a significant impact on system sensitivity. Number of Photons Unscattered Positron Emission Photons Scattered Photons Pileup 375 kev 511 kev 650 kev Energy Figure 6.5 Energy Thresholding to reduce Scatter Page 22 of System Shielding The detectors are very sensitive to radiation, yet the designer wishes them to detect only radiation emerging from that part of the patient that is in the field-of-view inside the detector ring. When imaging the chest there is substantial radiation originating in the

23 patient s head, legs, and probably bladder, all of which are outside the field-of-view. In order to reduce the amount of such radiation reaching the detectors, end shields are used on either side of the detector ring. Note that as the ring radius gets smaller, or as the patient port gets bigger, the region to which the detector is exposed increases. The performance of both the Discovery LS and the Discovery ST comply with the original design specifications, and, in the case of the Discovery LS, provide the same performance in a PET/CT system as in the PET-only version of the same design. This is not the case with PET/CT systems produced by other vendors. Competing products (Biograph/Reveal) are produced with end shields closing to 56cm in the PET-only versions (Accel and HR+), but in the PET/CT models the patient aperture is increased to 70 cm, so the shielding is very significantly reduced. This has a dramatic effect on the performance of these systems (Brasse,D et al, SNM 2002), increasing the scatter fraction and randoms, and reducing the Noise-Equivalent counting rate performance. Fixed 93cm ring 60cm aperture 82cm ring 70cm aperture Collimator Detector crystal Figure End Shielding on the Advance and Discovery ST The Discovery ST was designed from the start to have a large (70cm) patient aperture, and so a great emphasis in the design was placed on energy resolution, as well as making the detector ring larger and end shielding thicker to minimize scatter. The results of this careful work can be seen in the measured scatter fraction. Page 23 of 47

24 6.3 Spatial Resolution Spatial Resolution: Effect of Statistics 100k counts 1,000k counts 10,000k counts No filter 4 mm filter mm filter Clinical imaging operating range Page 24 of 47 Figure 6.7 The effect of counts on Spatial Resolution The statistical noise introduced by the small number of counts in a PET image can obscure details of the image. In order to reduce the effect of the noise, the data is filtered during reconstruction. Analysis of noise and resolution in images is not simple, but there are some good general principals to bear in mind: For a typical image with about a million counts, it is not unusual to see a filter cutoff of 7-10 mm used to counter statistical noise. Clearly with other resolution-limiting factors on the order of 1-2 mm, this one factor dominates PET image resolution. Secondly, each time the counts in the image are increased by a factor of 10, the filter cutoff for eliminating noise can be reduced by a factor of 2. This is a practical expression of a general relationship that counts go as the cube of the image bandwidth. The relationship between counts and resolution is heavily dependent on the image itself. This relationship will be very different for a hot spot on a cold background, a hot spot on a warm background, a cold spot on a hot background, multiple hot spots on a cold background, and so forth Statistics & Resolution: What is Required? The Advance NXi scanner, and a number of other scanners currently available, were originally designed as general-purpose PET scanners with the expectation of performing imaging in neurology, cardiology, and physiology research applications. This included, for example, applications of Carbon-11, Oxygen-15, and Nitrogen-13, where very large amounts of tracer might be infused, and where the objective is quite different from that in oncology imaging. The Discovery ST, on the other hand, is designed for the much more restricted oncology applications only. Where 10 to 20 mci of tracer will be used, and where the prime objective is to detect and characterize physically small (less than 10cm

25 diameter) concentrations of tracer in excess of the prevailing background. With this limitation of objectives it is easier to deal with the trade-offs that present themselves to the system designer. With a 10 mci injection of tracer into a 70 kg patient (the WHO Standard Man ) in 2D imaging mode, the average concentration becomes µci/cc. A sensitivity of 6.3 kcps/µci/cc per slice (Advance NXi) leads to 6, = 909 counts per second in a slice. So to get to 100,000 counts takes 100, = 110 seconds, or about 2 minutes, and getting to 1,000,000 counts takes 20 minutes. Based on the experience with the Advance NXi, therefore, we deduce that most current clinical imaging operates at around 100,000 to 1,000,000 counts per slice. Note that at 100,000 counts, we can just resolve the 12.7 mm pins, and at 1,000,000 counts we can clearly see the 6.4 mm pins (not the 4.8). Based on this simplistic analysis, we would expect that 6.5 mm resolution is an appropriate target for oncology imaging. The principal design choice influencing spatial resolution is the size of the entrance face of the individual crystal elements in the detector. By making the crystals smaller, the potential spatial resolution is improved. However, as with almost every other design decision, there are a number of interactions whenever anything is changed. There are two fundamental limits on resolution in a PET scanner, - non-co linearity and positron range. The positron range is the distance a positron may travel in the patient before annihilation. If the range is more than about 1.5 millimeters then it will become a significant factor in resolution of clinical images. Fortunately the range of F-18 positrons is very small. Noncolinearity refers to the fact that the two gamma rays resulting from a position annihilation may not travel in exactly opposite directions. There may be a small angle between them. Non-colinearity becomes more important when the detectors are far apart, and becomes less important when the detectors are close together. In other words, it is more important with larger detector rings. In practice, for F-18 imaging, the combined effect of non-colinearity and positron range is (for a whole body scanner) about 2.5 mm FWHM at the center. With crystals smaller than this it should, in principal, be possible to achieve a resolution under 3mm, but not much better. The Advance NXi scanner and the Discovery LS used crystals with an entrance face of 8mm x 4 mm. 4 mm in the transaxial direction and 8 mm in the axial direction. This produces a central resolution of 4.8 mm within the slice (transaxial direction) and a rather worse resolution in the axial direction. The Discovery ST uses crystals whose entrance face is 6.3 x 6.3 mm. This yields a more isotropic resolution, but the central axis resolution within the slice is inferior to the Discovery LS (see Figure 6.10) Resolution Non-uniformity Spatial resolution of a PET scanner isn t the same everywhere in the field of view. The resolution is best on the central axis (i.e. at the center of each ring of detectors) and degrades for objects away from the central axis. Page 25 of 47

26 Plot of Resolution for a PET Scanner FWHM Radial Tangential Distance from Center of Image Figure 6.8 Resolution degrades away from the center Spatial resolution in the radial direction degrades toward the edge of the scanner field of view. The change in radial resolution is due to the use of a circular ring of crystals to detect annihilation events. The effect is illustrated above. Gamma rays hitting the detector crystals at an angle may travel through the first crystal encountered and stop in a neighboring crystal. If the gamma ray stops in a neighboring crystal and is detected there, the location of the original event will be shifted toward the center. Note that as the ring diameter gets smaller, this effect will be increasingly pronounced. Page 26 of 47

27 Advance/Discovery LS = 92.7 cm Discovery ST = 88.1 cm Accel/HR+ = 82.7 cm Figure 6.9 Comparing the size of detector rings Since the detector diameter for Discovery ST is halfway between that of Discovery LS and that of Siemens scanners (HR+ & Accel), Discovery ST shows slightly more resolution degradation (non-uniformity) toward the edge of the field of view than Discovery LS, and not as much degradation as is seen on HR+, Accel, or Biograph. The net result is reflected in the following comparison for transaxial resolution Spatial Resolution: GE vs. Siemens The NEMA resolution measurements of the competitor and GE PET systems are given in Figure 6.10 below. These resolution differences reflect the basic design decisions of crystal size and ring diameter. The Siemens HR+ and Accel have a smaller ring diameter than Discovery LS or Discovery ST (83 cm compared to 93 cm and 88 cm) The Discovery ST ring diameter represents the midpoint: 88.1 cm. In the transaxial direction the crystal size of Discovery LS and the HR+ are nearly identical (GE = 3.94 mm; Siemens = 4.05 mm) The crystal size of Discovery ST and Accel are similar (6.3 mm and 6.8 mm). The smaller rings on HR+ and Accel improve resolution at the center by reducing the non-colinearity effect. This provides an advantage in neurology imaging. The larger rings on Discovery LS & Discovery ST improve the resolution away from the central axis by reducing the radial mispositioning effect. This is a significant advantage in body imaging, for example axillary lymph nodes are typically at a radius of about 20 cm. Page 27 of 47

28 2D SPATIAL RESOLUTION (Nema NU2 1994) Discovery ST Discovery LS Accel Allegro HR+ 2D SPATIAL RESOLUTION 1cm 6.2 mm 4.8 mm 6.0 mm cm 6.8 mm 5.4 mm 6.7 mm cm 7.2 mm 6.2 mm 8.6 mm mm 3D SPATIAL RESOLUTION 1cm 6.2 mm 4.8 mm 6.0 mm cm 6.8 mm 5.4 mm 6.5 mm cm 7.2 mm 6.2 mm???? 6.8 mm 2D SPATIAL RESOLUTION 1cm 6.2 mm 4.8 mm 4.5 mm cm 6.8 mm 5.4 mm 5.9 mm cm 7.5 mm 6.6 mm?? mm* 3D SPATIAL RESOLUTION 1cm 6.2 mm 6.0 mm 4.6 mm cm 6.8 mm 6.3 mm 6.5 mm cm 7.5 mm 6.6 mm?????? Figure 6.10 Comparison of Scanner Spatial Resolutions [ECRI, 2002] 6.4 Counting Rate Performance The Objective is Maximizing Return on Injection Page 28 of 47 Counting Rate performance is of greatest importance when the scanner is exposed to large quantities of radioactive tracer. In imaging of very short-lived tracers this is a most important parameter, but it is much less important in oncology studies. Since all current oncology studies are performed with Fluorine-18 the total injected tracer is unlikely to exceed 20 mci. Future radiopharmaceuticals for oncology are likely to be based on fluorine-18, or on tracers of longer half-life. This means that the injected activities are unlikely to be greater than those of today. This is the reason that the Discovery ST design focused more on sensitivity than on counting rate performance. The objective of this system is to extract as much information as possible from the current dose levels, rather than provide the capability to perform at very high tracer concentrations which are unlikely to be used with the Discovery ST.

29 The design brief for the Discovery ST team was somewhat different. They were asked to produce a system that would provide that maximum return of information for injected doses of Fluorine-18 in the clinical range (about 3 to 20 mci). This meant first, to maximize sensitivity, but also to minimize information losses, from any source, in the clinical imaging range Summary of Counting Rate Parameters The Nema Standard issued in 2001 describes a range of counting rate values to be measured and reported. The three responses to be measured are total counts (or prompts), trues, randoms, scatter and NEC. The total counts is the total number (of coincidence) events registered by the system; trues are genuine coincidences arising from positron annihilations (after randoms and scatter have been subtracted). Randoms and scatter are both invalid events which degrade image quality. NEC is the Noise Equivalent Counts, a value which is intended to represent the effective value of the information. As such, NEC deserves some discussion. If all the information collected by the scanner was trues, meaning that every event was from a positron annihilation within the field-of-view, then the NEC would be the same as total and true ; and randoms and scatter would both be zero. When there are random events and scattered events present, then the trues value is less than total (by the value of the randoms and scatter; total is trues +randoms+scatter. However when randoms and scatter are subtracted from total to get trues the result is not the same as if there had never been randoms and scatter. Because of the statistical nature of the data, all the information contains noise. Suppose we consider a single transaxial slice from a patient that contains one million counts (a lot). If these events are distributed in 5,000 picture elements (pixels) there will be an average of 200 per pixel, but the differences between pixels is what creates the image. Because of the nature of counting statistics, there will be random variation in the number of events per pixel, and the size of the variation is related to the number of events in the pixel. Mathematically, the standard deviation of the number of counts is equal to the square root of the number of counts. So, if the image was collected entirely from trues, the standard deviation of the pixels would be about 14 (for a pixel content of 200), or about 7%. Suppose that instead of collecting only trues, we also collected an equal number of randoms and scattered events (not an unlikely number, when we look at the scatter fractions). Now the matrix has 400 events per pixel, but we know that 200 of them are bad events. Before we subtract these events, the statistical variation in the data is determined by the total number of events collected: 400. The standard deviation is therefore 20 counts (5%). However, if we now subtract the 200 bad counts (scatter and randoms) the standard deviation of the result is still 20 counts (not 14). When randoms or scatter are collected with the data, even if we can correct for them, they leave behind a higher level of statistical noise. Page 29 of 47 The Noise-Equivalent-Count is a concept that tries to take account of this added noise. From the actual numbers of true events, scattered events, and random events, the NEC formula computes the equivalent number of trues that would give the same precision if the data were collected in an ideal situation with no randoms or scatter. It is called noise-equivalent because it computes the number of trues-only events that would need

30 to be collected to achieve the same statistical characteristics (statistical noise) as actually were obtained after subtracting scatter and randoms. The NEMA standard asks that each of these values (totals, trues, random, scatter, and NEC) be plotted as a function of the radioactivity present in the scanner, over a range of activities from very low up to at least the activity that produces the highest NEC rate possible. In addition it asks that the maximum NECR, and the maximum trues rate, and the activity concentrations where those occur, be reported. Activity concentration is based on a distribution in a 22 liter phantom that is 20 cm in diameter and 70 cm long, placed centrally in the scanner. The peak Trues and peak NECR values (where known) are given in Figure 6.11 below. 2D COUNTING RATE PERFORMANCE (NEMA NU2 2001) Discovery ST Discovery LS Accel Allegro HR+ PEAK TRUES (kcps) at µci/ml Activity PEAK NECR (kcps) at µci/ml Activity D COUNTING RATE PERFORMANCE PEAK TRUES(kcps) at µci/ml Activity PEAK NECR (kcps) at µci/ml Activity Figure 6.11 Comparison of Scanner Count Performance [ECRI, 2002] It is difficult to interpret figure 6.11 because it only provides a small amount of information, and the data provided is unlikely to be relevant to oncology imaging. A more useful comparison can be made from the graphs required by the NEMA standard, and most especially by the NECR graph. This graph has two extremely useful characteristics. It shows the noise equivalent counts which is as good a measure as we can get of the useful information collected, and the graph shows how it changes for any reasonable quantity of tracer in the patient. (It just requires a little interpretation.) Page 30 of 47

31 NECR 3D NEMA 2001: GE vs. Others Discovery ST Accel Advance Exact HR Average Tracer Concentration Figure 6.12 Comparison of PET Scanner Counting (NECR) Performance Figure 6.12 shows the NECR curves for GE systems and a number of competitive PET scanners in 3D mode. The height of each curve (Y-axis) shows the effective rate at which good information is collected (the Noise-Equivalent Count Rate) for given average concentrations of tracer in the patient (X-axis). In comparing counting rate performance it is important to consider the average concentration, rather than the specific concentrations in organs or in tumor. Because gamma rays are emitted at all angles, the counting rates in all detector blocks are very similar. Also counting rate performance is determined by the total rate at which events (good and bad) are entering the detectors. In 2D imaging the counting rate performance is determined by to total amount of radioactivity within the field-of-view. In 3D imaging, the performance depends on the total amount of radioactivity within about 40 to 50 cms of the field-of-view. This difference is because in 2D the septa prevent radiation from outside the scanner field-of-view reaching the detectors, but in 3D all radiation originating from about 40 cm on either side of the imaging portion of the scanner can reach the detectors. Although this radiation isn t included in the images, the scanner needs to process it, and so it impacts the overall performance. The average tracer concentration (X-axis) is related to the total injected dose by through the total patient volume. To obtain the average concentration of tracer in the patient divide the total injected dose by the volume of the patient. For example, if 10 mci of FDG is injected into a patient weighing 220 lbs, or 100 kilograms (or with a total volume of 100 liters), then the average concentration in the patient will be (10 mci)/(100 liters) or (10,000 µci)/(100,000 mls) = 0.1 µci/ml. If the same dose of FDG is injected into a patient weighing 110 lbs (50 kg) the concentration will be (10,000 µci)/(50,000 mls) = 0.2 µci/ml. Figure 6.13 shows the same graphs as 6.12, but the X-axis has been converted from concentration to show the injected dose that would give that concentration in a Standard Man as defined by the World Health Organization (70 kg, or 154 pounds). Page 31 of 47

32 NECR 3D NEMA 2001: GE vs. Others Discovery ST Accel Advance Exact HR Total Activity in 70 kg (154 lb) Patient C Figure 6.13 Comparison of Scanner NECR related to patient dose To perform the convertion, each value on the x-axis has been multiplied by 70 liters to give the total activity that would be needed to reach that concentration in a patient whose volume was 70 liters. From this graph it is easy to see that the area of interest for oncology is on the left (shaded) side, where the total activity is less than 20 millicuries. The values on the x-axis are the total activities in the patient at the time of scanning. It is standard practice to wait for an uptake period of 45 minutes to one hour after injection before starting to scan. It is also recommended that patients be asked to empty the bladder prior to starting a scan, so the activity injected would have to be at least 50% larger than the values in Fig 6.13 to achieve the concentrations in Fig Discovery ST Biograph LSO Discovery LS Biograph BGO 0 10 Activity 20 in Patient at Scan time 30 [mci] Figure 6.14 PET/CT Scanners: Counting (NECR) Performance Page 32 of 47

33 Page 33 of 47 It has been pointed out [Badawi et al, 2001] that the values of average concentration obtained using the method defined in NEMA NU-2 (2001) don t always correspond to values obtained in patients, especially when the patients are large.

34 Appendix A: 2D and 3D PET A.1 Introduction Because pairs of gamma rays are emitted in all directions with equal probability, some of the lines-of-response (LOR) will be perpendicular to the principal axis of the patient (and the scanner) but the majority will not. Another way of looking at it, is to consider the individual crystal elements that detect the two gamma rays. If we consider the detector as a series of detector rings where each ring is one crystal wide in the axial direction, and a complete circle around the patient, then a 2D study is one in which both gamma rays are detected in the same ring. When that happens the source of the two rays must lie in the thin slice of patient that lies inside the ring. In principle, a study using only those events whose Lines-of-Response are perpendicular to the principal axis is a Two-Dimensional study. A study which also uses events whose LORs are not perpendicular to this axis is a Three-Dimensional study. In a 2D study, all the gamma rays originate in the same slice as they are detected. In a 3D study, the paths (LORs) of the gamma rays cross several slices. In PET, 2D and 3D are used to refer to which data is used to form the images, not the resulting images. In all cases, the result is a 3D volume, which can be represented as transaxial, sagittal, coronal, or oblique slices. Fig A.1. 2D and 3D Lines of Response (LORs). In figure A.1, the blue arrow represents an LOR that is included in a 2D acquisition. The red arrow represents an LOR that would be included in a 3D acquisition, but not in a 3D acquisition. For an LOR to be included in a 2D acquisition, both of its end points must be detected within the same slice. Since a 3D acquisition uses all of the detected events (or most of them), and a 2D acquisition uses only some of the events, 3D acquisition has a higher sensitivity than 2D. Page 34 of 47

35 A.2 Two-Dimensional Acquisition A two-dimensional acquisition collects only LORs which are perpendicular (or nearly perpendicular) to the patient s principal axis, and each transaxial slice is reconstructed from LORs which transit that slice, and only that slice. Figure A.2 shows three LORs from three positron annihilations in a patient. All of these events originate in the same transaxial slice of the patient. The blue arrow shows an LOR which is approximately perpendicular to the principal axis of the patient. Both gamma rays from this event are detected in the same ring of detectors. The two red arrows illustrate events which originate in the same slice as the blue arrow, but which do not qualify for inclusion in a true 2D acquisition, because the two gamma rays were not detected in the same ring of detectors. However, in some high sensitivity 2D modes of operation, one of these two events could be included. Because the two gamma rays were detected two rings apart, it is highly likely that the event occurred in the slice mid-way between the two detector rings that detected the event. If events like these are to be included, they will be treated exactly as if they had been detected in the central ring, and information that they were really tilted LORs will be discarded. Fig A.2. LORs contributing to a Transaxial Slice The technique of adding such non-perpendicular LORs (those in which the two gamma rays were detected in different rings, equidistant from the ring in which the event is assumed to originate) to the true LORs (those in which both gamma rays were detected within the same ring) is called spanning. Spanning improves the sensitivity of 2D acquisitions (by using events that would otherwise be rejected), but at the same time it reduces spatial resolution, especially at larger radii. For positron annihilations which occur very close to the center of the detector ring, spanning adds counts and places them in the correct position. For events that occur away from the center of the ring, the true position of the event isn t in the central ring, so placing it in the central ring introduces a positional error, and hence degrades spatial resolution in the axial direction. Page 35 of 47

36 Fig A.3. LORs contributing to a Cross-Plane Transaxial Slice A special case of spanning is in generating cross-plane slices. If the two gamma rays are detected in adjacent rings, then there is a high probability that the origin was between the centers of the two rings. All events of this type are used to reconstruct cross-plane slices, with one such slice between each pair of detector rings. A. 3 Three-Dimensional Acquisition A three-dimensional study collects all events, regardless of which rings are involved. This includes all of the events that might be included in a 2D study with a very large span. The critical difference is that in a 3D study, the events which are detected in different detector rings are recorded at their true angles. They are not combined with data obtained from another ring. For example, if an event is detected in rings 2 and 10, then clearly the LOR traversed the slices within rings 3 to 9, and could have originated in any of these slices. Most 2D acquisitions would discard such an event. In a very high sensitivity 2D acquisition that was accepting events up to a span of 8, this event would be treated as if both gamma rays had been detected by ring 6. In a 3D acquisition the event would be stored as traversing rings 2 to 10, and when reconstructed, the 3D reconstruction would attempt to place the event in its correct position. Page 36 of 47

37 Figure A.4. LORs acquired in 3D The essential feature of 3D acquisition is that all events are collected, even those whose LORs don t lie within a transaxial plane, and the true angles of LORs are recorded and used in reconstruction. The major advantage of 3D acquisition is that many more events are collected in any acquisition time, because all events are acceptable, whereas in 2D acquisition only a fraction of the detected events are useable. A. 4 The Role of Septa The two principal sources of loss of image quality in PET scanning are scatter and random events. Both of these sources of noise are chiefly determined by out-of-field activity. In clinical imaging there is often more radioactivity outside the scanner field-of-view than inside, and this out-of-field activity is the dominant source of both scatter and random events. In the diagram in Figure A.5, the large concentrations of radioactivity normally present in the brain and bladder, are the source of noise events. The arrows marked S1 represent typical paths of scattered radiation. The lower arrow represents one gamma ray which is directed to one of the detectors near the lower portion of the figure. The other gamma ray from this event starts in the opposite direction and would not be detected, but it is scattered in the patient, and so changes direction in such a way as to be detected near the top of the ring. This event generates a valid coincidence which (erroneously) places the event near the center of the field of view. The arrows marked R are rays which contribute to random coincidence events. Gamma rays may enter directly, from source to detector in a straight line, or may be scattered in the patient, or surrounding material, into a detector. The great majority of these are single gamma rays (only one gamma ray of the pair is Page 37 of 47

38 detected). If the rate of such single events is high, then a significant number of random coincidences will be detected. Figure A.5. Scatter and Random Events The final source of noise is illustrated by the blue arrows, S2, and arises from activity within the field-of-view. Here the direction of the gamma rays should result in only one ray (the lower one) being detected, but the second ray is scattered in the patient such that it is also detected. The source of these rays is now incorrectly located on the line joining the positions where the rays were detected. All of these sources of noise, or incorrectly positioned events, can be reduced by well placed septa. The gamma rays from most good events travel almost perpendicular to the scanner axis, and to the entrance face of the detectors, while most of the rays associated with noise events travel at oblique angles to the axis and detector faces. If thin layers of attenuating material (lead or tungsten, for example) are placed in between the rings of detectors and perpendicular to the entrance faces of the detectors, then a large number of the scatter and random events can be stopped. Figure A.6 shows that septa placed between the detector rings can intercept a large fraction of scattered rays. The septa also stop a large fraction of rays from outside the field-of-view which contribute to random coincidences. Because septa are so effective at reducing the main sources of image degradation, (scatter and random events) they are used routinely in 2D acquisitions. Adding septa can have a dramatic impact on image quality, but there are also drawbacks to septa. Septa reduce the sensitivity of the scanner because they reduce the effective area of the detectors. If the septa are too long they may also reduce the sensitivity of the cross-plane slices. Both the length and thickness of the septa must be carefully designed to give the maximum reduction in noise, while minimizing the impact on sensitivity. Page 38 of 47

39 Figure A.6. Septa positioned to Reduce Scatter and Random Events In current PET scanners, septa are designed to optimize one configuration of 2D imaging (for example, with a span of 1, or a span of 3) and are removable. This allows the he septa to be removed completely for 3D imaging. (There are also PET scanners with no septa, which are intended to be used in 3D mode only.) In principle, septa can also be effective in 3D imaging as well. Septa designed for 3D imaging cannot be spaced at the detector spacing. Such septa are primarily aimed at removing radiation that originates outside the field-of-view of the scanner. The objective is to reduce both random and scattered events created by gamma rays traveling at angles significantly different from the transaxial plane. A.5 Scatter in 2D and 3D acquisitions Scatter refers to coincident events in which one or both of the gamma rays changes direction at some point after creation, and before detection. Gamma rays may change direction by scattering, or colliding with an electron. In doing so the gamma ray gives up some energy (to the electron) and changes direction. This process is call Compton scattering. Gamma rays may undergo Compton scattering in the patient, the patient table, septa, or other mechanical structures in or near the scanner field-of-view. The majority of scattering occurs in the patient. When a gamma ray undergoes Compton scattering, as well changing direction, it also loses energy. There is a direct relationship between the change in direction (scattering angle) and the energy lost, with greatest energy loss occurring at large angles, and very little energy loss at small scattering angles. Gamma ray scattering and energy transfer is exactly analogous to the interaction of billiard balls. If a moving ball (the gamma ray) strikes a stationary ball (the electron) then the direction of the moving ball will change, its speed (energy) will be reduced, and the stationary ball will move away from the impact. The greatest energy transfer will be when the moving ball hits the stationary ball squarely and is stopped or even reverses direction. The smallest energy transfer occurs with a glancing blow, when the moving ball is deflected only slightly. Page 39 of 47 Even though one gamma ray of a pair is scattered, it may reach the detector at the same time as its twin. In such a case the two events are taken as a valid coincidence pair. If the scattered ray has given up a very large fraction of its energy it may be rejected for that reason. If the loss of energy isn t large, the event will be accepted. However, since