Study of Silicon Photomultipliers for Positron Emission Tomography (PET) Application Eric Oberla 5 June 29 Abstract A relatively new photodetector, the silicon photomultiplier (SiPM), is well suited for PET applications. It has similar sensitivity and gain to the industry standard photomultiplier tube (PMT), but has advantages such as smaller size and insensitivity to magnetic field. These properties make this detector an active area of research in the PET field. I will study a simplified setup, comprised of two antiparallel SiPM/LSO coupled detectors using a Na-22 positron source. The coincidence timing and energy resolution is determined using two methods, a CAMAC system and a fast oscilloscope. The best coincidence time resolution, 66 ps FWHM, was obtained using the digital oscilloscope. At best, the energy resolution was found to be 16.4% FWHM. Results using two types of SiPM, 16 and 4 pixel, are presented. Data from a GEANT4 simulation of the described setup are also shown. Introduction Positron Emission Tomography (PET) is a medical imaging technique that is used to observe functional processes in vivo. The functional process of interest is observed by introducing a chemical tracer that is metabolized by certain tissues in the body. The tracer is doped with a radioisotope that undergoes positive beta decay (positron emission). A large majority of current PET scans use fluorodeoxyglucose (FDG), a glucose molecule with a hydroxyl group replaced by radioactive 18 F. FDG enters the same metabolic pathways as glucose and is utilized for oncology and brain imaging [1]. Upon injection in the body, the tracer is concentrated in specific tissues, such as a tumor. Positrons emitted from beta decay annihilate with nearby electrons, generating back-to-back 511 kev photons that can be detected by a scintillator crystal coupled to 1
a photodetector. The location of the annihilation event can by narrowed to a line of response (LOR) from two coincident detections [1]. In PET systems with high timing resolution, the annihilation location can be further constrained to a segment on the LOR. This is known as time-of-flight (TOF) PET, in which high timing resolution results in the reduction statistical noise. Advantages of TOF PET include faster computer processing times and better image quality [1-2]. Photodetectors and PET The standard photodetector in clinical PET machines is the photomultiplier tube (PMT). The high gain, fast response and high sensitivity of PMTs have made them a viable detector for PET, but there exist several drawbacks. One, the bulky size of PMTs puts a limit on the spatial resolution of a detector [1]. PMTs are also highly sensitive to magnetic fields. This makes it impossible to implement PET with magnetic resonance imaging (MRI), considered the future of biomedical imaging because of the promise of simultaneous metabolic and anatomical information [3-4]. An area of active research is the study and application of silicon photomultipliers (SiPM). SiPMs are relatively compact and insensitive to magnetic fields while achieving roughly the same gain and sensitivity as PMTs [4-5]. The SiPM is a semiconductor device made up of avalanche photodioide (APD) pixels connected in parallel. Each APD is an individual photon counter and the sum of all the APD pixels is the output of the SiPM. The APDs are operated in Geiger Mode, where the bias voltage applied is greater than the reverse breakdown voltage resulting in a large internal electric field. An incident photon causes a carrier to be injected into this electric field creating a large pulse that can be put into electronics. The specifications of the SiPM vary widely, but for PET applications the greater number of pixels (lower fill factor) is desired because of the high light input to the detector [3,5]. Studies of SiPMs have yielded promising results. Measurements of the inherent coincidence timing resolution using a splitted laser gave timing on the order of 1 ps, suggesting the viability of SiPM for TOF PET [6]. The coincidence timing between two LYSO/SiPM coupled detectors has been measured by Kim et al to be at best 24 ps [7]. However, there are challenges to implementing SiPMs into a full scale PET machine. There is a moderate dependence of gain with bias voltage, so each SiPM would require its own voltage control. Temperature dependence is also an issue. Other challenges exist and are discussed in reference [7], but the potential benefits of SiPM photodetectors for PET make their implentation a promising next step in detector design. 2
Experiment Characterization of SiPM We first wanted to test for ourselves some basic properties of SiPMs. We focus on the 16 pixel, 25 µm 2 pixel size SiPM (Hamamatsu S1362-25C) because of the large number of photons incident on the photocathode in PET. This sacrifices active area, but it ensures that the photocathode will not be saturated. We will also analyze the properties of the 4 pixel, 5 µm 2 pixel area SiPM (Hamamatsu S1362-5C). Gain and dark count rate were measured using a single SiPM without an LSO crystal or radioactive source. Detector was kept in black box for both experiments. For gain measurement, SiPM output was amplified using 3 db preamplifier and signal was read out using a fast oscilloscope (TDC6154, 2 GS/s). Oscilloscope was triggered at 5 mv increments from 5 mv to 5 mv and 15 events were collected at each trigger level. The result is the energy of the photoelectron (p.e) peaks after gain from both the SiPM and preamp, shown Figure 1. By a linear fit of the amplified energy (charge) of the 1,2,3 and 4 p.e. peaks vs. the unamplified photon energy,the gain of the 16 pixel SiPM was determined to be 2. 1 5. Similarly, the gain of the 4 pixel SiPM was found to be 6.2 1 5. The dark count rate was also measured. The SiPM is a solid-state device with inherent noise due to thermal excitations of the APD pixels, causing uncorrelated photon counts. To make any reliable measurements, it is important to understand and quantify this noise. In order measure the dark count, signal was counted at different levels of noise. The SiPM output was sent to a 1x preamplifier, split and sent to two discriminator inputs. Trigger levels of the discriminator were set to.5 p.e. and 1.5 p.e. respectively. Figure 1: Photoelectron spectrum of SiPM (left peak corresponds to 1 p.e., etc.) Figure 2: Dark count rate vs. bias voltage, 16 pixel 3
These levels varied with applied bias voltage and were set by observing the preamplified pulse in oscilloscope. The discriminated signal was sent to a Lecroy 2551 1 MHz scaler to give a counts per acquisition time for each threshold level. Bias voltage was varied from 7.1 to 71.9 V and the result if shown in Figure 2. For an operating voltage of 71. V, it is clear that the dominatant noise is from single photoelectron counts. Although the noise rate is high (~11 khz), the energy is much smaller than a useful PET signal so it is easily filtered. However, if our goal was to count individual photons, the SiPM would not be a good photodetector. Methods Two 1x1x1 mm 3 LSO crystals were attached to a pair 16 or 4 pixel SiPMs using a small amount of fast-drying, transparent epoxy. The crystals were wrapped in teflon tape to keep optical photons internally reflected. Both SiPMs were operated at a bias voltage of 71. V. A 22 Na source was used as a positron emitter and placed between the two detectors, as shown in Figure 3. CAMAC System A CAMAC data acquisition system with NIM electronics was used to measure coincidence timing and energy resolution. The SiPM output was put directly in a LeCroy model 612 fast amplifier (1x) and then split using a linear fan-in/fan-out module (LeCroy model 428F). One branch was put in a LeCroy 623B discriminator with a minimum threshold of 3 mv. The discriminated signal was sent to a coincidence unit set to AND logic so as to fire for a coincidence event. Figure 3: Experimental setup 4
Figure 4: Electronic schematic for coincidence timing measurement using CAMAC system This output was used for the ADC gate and the TDC start. The other branch of the fan-out module was delayed and sent to the ADC channel inputs for energy information. To extract timing information, the analog signal was amplified 1x and then put into a discriminator. The discriminator level was set to the lowest possible level above the noise threshold that was observed in oscilloscope. This was necessary to get the best possible timing resolution from our system. Both discriminator outputs were delayed with a 4 ns cable and sent to the stop channels in the TDC for timing information. See Figure 4 for a schematic of the timing electronics. TDC Calibration TDC needed to be calibrated to interpret data. Discriminated pulse from a function generator was sent to TDC with, 2, 4 ns delays. By measuring TDC values to the delays, conversion factor of 41 ps/bit was obtained. Fast Oscilloscope In addition, the coincidence timing and energy resolution were analyzed using a Tektronix DPO7 series 4 GS/s digital oscilloscope. With the same set-up as in Figure 3, the unamplified SiPM signals were sent to the scope and coincidence events were captured using a logic AND trigger. Each channel was sampled at 2 GS/s and an event was 5 ns, providing 1 points per saved waveform with a time interval of.5 ns between samples. 5
Results/Analysis From CAMAC Setup Results from 16 pixel SiPM shown in Figure 5. Fourty-thousand coincident events in total were collected at a rate of about 1 Hz. The right peak in the charge (energy) 2 spectrum is due to the gamma photon depositing all 511 kev in the scintillator. The left peak is of less enery, in which the gamma has Compton scattered, only depositing a fraction of its energy in the crystal. These events degrade the time resolution and are filtered out for timing measurements. The disparity in relative peak size was most likely due to coupling differences between the SiPM surface and the LSO crystal of the two detectors. The coincidence timing is presented before (Delta T) and after (Delta T: cut) filtering unwanted events. The timing resolution considering only those events in the 511 kev peak is shown to be 82 ps FWHM. The best energy resolution (from SiPM 1) is 23.9 % FWHM. The discriminator threshold level for timing was set at 5 mv for this timing result. See Figure 6 for timing resolution as a function of varying this threshold. Temperature sensitivity of the SiPM became apparent during this experiment as ADC spectrum clearly shifted as time progressed. There has been shown to be a considerable gain dependence on temperature in SiPMs [4] and this might be the cause of the spectrum shift. This makes getting good energy resolution at low coincident rates a difficult task! SiPM Entries 4 14 Mean 93.28 RMS 58.88 12 Integral 3.967e+4 χ 2 / ndf 528.7 / 438 1 Constant 68.84 ±.82 Mean 162.2 ±.2 8 Sigma 19.38 ±.19 6 4 2 5 1 15 2 25 Charge (pc) Entries 4 Delta T: T1 T2 7 6 5 4 3 2 1 Mean 4.255 RMS 1.854 Integral 3.83e+4 2 2 4 6 8 1 12 14 Time (ns) SiPM 1 Entries 4 2 Mean 12.3 18 RMS 64.93 Integral 3.968e+4 16 χ 2 / ndf 342.8 / 317 14 Constant 69.88 ±.83 Mean 18.9 ±.2 12 Sigma 18.36 ±.18 1 8 6 4 2 2 18 16 14 12 1 8 6 4 2 5 1 15 2 25 Charge (pc) Entries 426 Delta T: T1 T2 cut Mean 4.189 RMS.4298 Integral 414 χ 2 / ndf 38.93 / 44 Constant 182.3 ± 3.8 Mean 4.184 ±.6 Sigma.353 ±.5 2 2 4 6 8 1 12 14 Time (ns) Figure 5: Results: 16 pixel SiPM, top: Energy Spectrum, bottom: Coincidence Timing 6
1.1 Timing Resolution - varying Threshold Level 1.5 1 Time [ns].95.9.85.8 5 1 15 2 25 3 Threshold Level [mv] Figure 6: Coincident timing as a function of threshold level See figures 19 and 2 in the appendix for raw data. Results from 4 pixel SiPM shown in figure 7. Coincidence timing was 85 ps FWHM and the best energy resolution was 16.4% FWHM. SiPM Entries 4 16 Mean 97.2 RMS 52.33 14 Integral 3.988e+4 12 χ 2 / ndf 56.7 / 4 Constant 126.5 ± 1.4 1 Mean 154.8 ±.1 Sigma 1.82 ±.9 8 SiPM 1 Entries 4 Mean 111.7 RMS 6.22 1 Integral 3.975e+4 χ 2 / ndf 248.8 / 237 8 Constant 82.73 ±.99 Mean 177.9 ±.2 6 Sigma 16.94 ±.2 6 4 2 4 2 5 4 3 2 5 1 15 2 25 Charge (pc) Entries 4 Delta T: T1 T2 Mean 5.997 RMS 2.411 Integral 3.753e+4 5 1 15 2 25 Charge (pc) Entries 4913 Delta T: T1 T2 cut 22 2 18 16 14 12 1 8 Mean 5.648 RMS.4961 Integral 4911 χ 2 / ndf 89.94 / 61 Constant 215.9 ± 4. Mean 5.64 ±.1 Sigma.3634 ±.42 6 1 4 2 2 2 4 6 8 1 12 14 Time (ns) 2 2 4 6 8 1 12 14 Time (ns) Figure 7: Results: 4 pixel SiPM, top: Energy Spectrum, bottom: Coincidence Timing 7
Amplitude(mV) Event# 1 : Disc1 2 2 Event χ / ndf Entries 2479 / 18 Mean x p Mean y 189.6 ± 3.736 8 p1 RMS x 4.9 ±.7756 RMS y 6 Integral Amplitude(mV) 4 4 6 2 Amplitude(mV) 8 1 5 1 15 2 25 3 Event# 1 : Disc2 2 Time(ns) Amplitude(mV) 4 45 5 55 6 65 Time(ns) Event χ 2 / ndf Entries 1252 / 18 Mean x p Mean y 231.9 ± 3.31 8 p1 RMS x 5.66 ±.7756 RMS y 6 Integral 4 4 6 2 8 1 5 1 15 2 25 3 Time(ns) 4 45 5 55 6 65 Time(ns) Figure 8: Sample coincident event acquired using digital oscilloscope. Figure 9: Zoom into rising edge of event in fig 8. Red line is linear fit over n sample points at a specified voltage threshold. Coincidence timing is extracted by examining difference in x-intercept of fits between channels. From Oscilloscope Using the 4 pixel SiPMs, 4 coincidence events in total were collected. Example waveforms from a coincidence event are shown in Figure 8. A linear fit was performed on the rising edge of each waveform at a specified threshold as presented in Figure 9. Coincidence time was found by examining the difference between x-intercept values of the fit between the two channels. The best timing was extracted by fitting to 1 points, corresponding to 5 ps, below threshold level. Again, by selecting only those events in which all 511 kev of gamma photon energy was deposited, the coincidence timing was found to be 66 ps FWHM. This is shown in Figure 11. The threshold was set to 3 mv for this result, and coincidence timing as a function of fitting threshold is provided in Figure 1. No improvement in time resolution was found for a threshold lower than 3 mv since that was just above the noise level of the signal. The energy spectrum was determined by integrating discretely under each waveform. The integrated charge was found by implementing the algorithm (from Ohm s Law): Q = 1 R Nsample i=1 V [i] t i = t R 8 Nsample i=1 V [i] (1)
11 Timing Resolution vs. Threshold FWHM Coincidence Timing Resolution [ps] 1 9 8 7 6 3 4 5 6 7 Threshold for linear fit [mv] Figure 1: Timing resolution as a function of the linear fit threshold to digitized waveform Delta T: T1 T2 Entries 4 18 16 14 12 1 8 6 4 2 8 Mean 4.766 RMS 2.451 Integral 3664 15 1 5 5 1 Time (ns) SiPM Entries 4 22 Mean 42.69 2 RMS 17.62 18 Integral 4 16 χ 2 / ndf 1.68 / 14 14 Constant 193.6 ± 5.9 12 Mean 55.8 ±.17 1 Sigma 5.917 ±.15 Delta T: T1 T2: Cut Entries 748 Mean 4.76 5 4 3 2 1 4 RMS 2.457 Integral 681 χ 2 / ndf 91.73 / 51 Constant 45.54 ± 2.65 Mean 5.155 ±.13 Sigma.2893 ±.112 15 1 5 5 1 Time (ns) SiPM 1 Entries 4 Mean 88.41 1 RMS 42.72 Integral 3998 8 χ 2 / ndf 34.71 / 27 Constant 83.28 ± 2.92 6 Mean 13.4 ±.3 Sigma 1.69 ±.31 6 4 2 2 2 4 6 8 1 12 14 Charge(pC) 5 1 15 2 25 Charge(pC) Figure 11: Results from scope: 4 pixel SiPM, top: Coincidence Timing, bottom: Energy Spectrum where R is the scope imput impedence, t, the time interval between sample points, is a constant and V [i] is the signal amplitude at each point. The number of sample points, NSample, was 1 for each event. The energy spectrum is shown in Figure 11. A Gaussian fit to the 511 kev peak gives an energy resolution at best of 19.3% FWHM 1. 1 Note: SiPM channels and 1 switched between Fig 11 and Fig. 7 when comparing energy spectrum. 9
Figure 12: Simulation setup. Green lines indicate gamma path. Simulation Study A GEANT4 simulation was created to reproduce the experimental setup presented in Figure 3. The world set-up is shown in Figure 12. Two 1x1x1 mm 3 LSO crystals were attached to a pair of SiPMs and back-to-back 511 kev gamma photons were created directly between the two detectors, labeled up and down. For each event, the gamma trajectory in the crystal was simulated in a stepwise fashion, generating photons by photoelectric or Compton processes. Once produced, each photon was tracked until hitting the detector surface where it was recorded and terminated. The emission spectrum and light yield (3, photons/mev) of LSO was implemented as well. Five-thousand simulation events were taken in 1 event intervals. The data was Photon Spectrum Entries 261653 2 18 16 14 12 1 8 6 4 2 5 4 3 1 Mean 431.7 RMS 3.42 Integral 2.617e+6 25 3 35 4 45 5 55 6 65 7 75 Wavelength (nm) Photon Spectrum After P.D.E. Entries 641356 Mean 43.1 RMS 29.33 Integral 6.414e+5 PDE of SiPM Entries.35 Mean x Mean y.3.25.2.15.1.5 RMS x RMS y Integral 25 3 35 4 45 5 55 6 65 7 75 Wavelength (nm) Photon Spectrum Entries 261653 2 18 16 14 12 1 8 6 4 2 3 1 Mean 431.7 RMS 3.42 Integral 2.617e+6 25 3 35 4 45 5 55 6 65 7 75 Wavelength (nm) Photon Spectrum After P.D.E. Entries 129274 1 3 1 Mean 428.8 RMS 28.97 PDE of SiPM Entries Mean x.6 Mean y.5.4.3.2.1 RMS x RMS y Integral 25 3 35 4 45 5 55 6 65 7 75 Wavelength (nm) 3 8 Integral 1.293e+6 2 6 1 4 25 3 35 4 45 5 55 6 65 7 75 Wavelength (nm) 2 25 3 35 4 45 5 55 6 65 7 75 Wavelength (nm) Figure 13: 16 pixel, top-right: Incident photon spectrum, top-left: PDE, bottom: Photon spectrum after applying PDE Figure 14: 4 pixel, top-right: Incident photon spectrum, top-left: PDE, bottom: Photon spectrum after applying PDE 1
analyzed for both 4 and 16 pixel SiPMs. The number of incident photons at the detector surface is shown in both Figures 15 and 16. It is identical for both SiPMs since the same generated data was used for both. To find the number of fired pixels, we used the expression given by [5]: Nfired = Npixel ( 1 exp ( )) Nphoton P DE Npixel where Nphoton is the number of incident photons and the photon detection effeciency (2) # of incident photons : up Entries 36 Mean 248 25 RMS 157 Integral 3 2 15 # of incident photons : down Entries 36 22 Mean 2267 2 RMS 158 18 Integral 35 16 14 12 1 1 8 6 5 4 2 1 2 3 4 5 6 1 2 3 4 5 6 # of fired pixels: up Entries 36 4 Mean 412.3 # of fired pixels: down Entries 36 5 Mean 393.2 3.5 RMS 214.3 RMS 215.7 3 Integral 286 4 Integral 288 2.5 3 2 1.5 2 1 1.5 1 2 3 4 5 6 7 8 9 1 # of pixels 1 2 3 4 5 6 7 8 9 1 # of pixels Figure 15: 16 pixel, top: Photons incident on SiPM, bottom: Fired pixels # of incident photons : up Entries 36 Mean 248 25 RMS 157 Integral 3 2 15 # of incident photons : down Entries 36 22 Mean 2267 2 RMS 158 18 Integral 35 16 14 12 1 1 8 6 5 4 2 1 2 3 4 5 6 1 2 3 4 5 6 # of fired pixels: up Entries 36 7 Mean 327.9 # of fired pixels: down Entries 36 6 Mean 319.9 6 5 RMS 91.38 Integral 297 5 RMS 94.7 Integral 32 4 4 3 3 2 2 1 1 5 1 15 2 25 3 35 4 45 5 # of pixels 5 1 15 2 25 3 35 4 45 5 # of pixels Figure 16: 4 pixel, top: Photons incident on SiPM, bottom: Fired pixels 11
(PDE) is provided in [5] and plotted in Figures 13 and 14. Since the photon spectrum and PDE peak at the same energy, the PDE effectively reduces the photon spectrum by a constant factor as shown in Figures 13 and 14. So, in equation (2), PDE was taken to be a constant:.2 for 16 pixel,.48 for 4 pixel SiPM. The number of fired pixels is presented in Figures 15 and 16 for the 16 pixel and 4 pixel SiPM, respectively. The 4 pixel SiPM shows clear saturation as most of the events cause all 4 pixels to fire. The energy spectrum is directly related to the number of fired pixels and is shown in Figures 17 and 18. A distinction between the Compton and 511 kev peak is clear when using the 16 pixel SiPM. This is not the case with the 4 pixel SiPM because of saturation. Figure 17: 16 pixel: Energy spectrum from simulation Figure 18: 4 pixel: Energy spectrum from simulation, clearly shows signal saturation Discussion The best coincidence timing resolution was found to be 66 ps FWHM using waveworm sampling. This is good enough for TOF-PET applications. The timing could likely be improved by amplifying the SiPM signal before sending to oscilloscope, allowing for a linear fit to a lower relative threshold level. The best energy resolution, 16.4 % FWHM, was found using the 4 pixel SiPM. The SiPM dark noise was shown to be dominated by low amplitude signals that are easily filtered from a typical PET signal. Contrary to intuition, it appears the the 4 pixel SiPM would be better suited for PET applications because of its better energy resolution. Because of the large number of incident photons (on the order of thousands), a detector with 4 pixels is expected to be saturated as shown in the simulation data (Figure 18). The experimental results suggest otherwise, the detector does not become saturated. There is a distinct 511 kev peak and events in this peak can be clearly analyzed for timing. A possible explanation for this is the incorrect assumption that one pixel can detect only one photon. For a 4 pixel SiPM, the duration of a single photoelectron pulse is roughly 3 ns as shown in Figure 21 in the Appendix. An LSO signal, however, has 12
a width of about 2 ns (see Figure 8). This means that each pixel may fire multiple times during an event, avoiding the saturation predicted by the simulation. In that case, equation (2) can t be applied blindly without considering this effect. References [1] T.K. Lewellen 28 Recent Developments in PET Detector Technology Phys. Med. Biol 53 R287-R317 [2] W.W. Moses 27 Recent Advances and Future Advances in TOF PET Nucl. Instrum. Methods Phys. Res. 58 919-924 [3] W.R. Leo 1994 Techniques for Nuclear and Particle Physics Experiment, 2nd Rev Edition, Springer. [4] S. Espana, et al 28 Performance Evaluation of SiPM Detectors for PET Imaging in the Presence of Magnetic Fields IEEE submitted November 14,28 [5] Hamamatsu 28 Multi-Pixel Photon Counter Specification Sheet [6] Q. Xie et al 27 Performance Evaluation of MPPC for PET Imaging 27 IEEE Nucl. Sci. Symp. Conf. Record 969-974 [7] C.L.Kim et al 28 MPPC for TOF PET Detector and its Challenges 28 IEEE Nucl. Sci. Symp. Conf. Record 3586-359 Acknowledgements Many thanks to Heejong Kim whose help and guidance were greatly appreciated. He provided the CAMAC acquisition programs, the GEANT4 simulation as well as Figures 1 and 21. Thanks to the UChicago Electronics Shop for allowing the use of their expensive oscilloscope. 13
Appendix Figure 19: Ch ADC spectrum, t= Figure 2: ADC spectrum, t = 12 hours later, 511 kev peak shifted off scale Figure 21: Single photoelectron pulse, 4 pixel SiPM 14