System and method for power-efficient multiplexing of high-resolution time-of-flight positron emission tomography modules with inter-crystal light sharing

Abstract

A multiplexing scheme for both energy and time information is provided for a particle detection system having an optical sensor array with multiple optical sensors. Each optical sensor is associated with multiple scintillator modules. The system has a segmented quasi-cylindrical light guide comprising multiple quasi-cylindrical segments. Each segment is associated with multiple optical sensors, where the optical sensors are adjacent. One end of each scintillator module is in contact with its associated optical sensor and the other end is in contact with its associated segment. The multiple optical sensors can be connected to energy readout channels, respectively, such that optical sensors associated with the same segment are not connected to the same energy readout channel. Each energy readout channel has at least two timestamps associated therewith.

Description

[0001] Cross-references to related applications

[0002] This application claims the benefit and priority of U.S. Provisional Application Serial No. 63 / 088,718, filed on October 7, 2020, the entire contents of which are incorporated herein by reference. Technical Field

[0003] This disclosure relates generally to the field of radiation imaging, and more particularly to positron emission tomography (PET). Background Technology

[0004] PET imaging is a powerful technique primarily used for the diagnosis, treatment selection, treatment monitoring, and research of cancer and neuropsychiatric disorders. Despite its high molecular specificity, quantitative properties, and clinical usability, PET has not yet fully realized its potential as a reliable molecular imaging modality, mainly due to its relatively poor spatial resolution. Several attempts have been made to achieve high-resolution PET, including using n-to-one coupling (where n>1) from the scintillator module to the readout pixel (optical sensor). This allows the spatial resolution to be equal to the size of the scintillator module without increasing the cost of the readout side (e.g., optical sensor, connector, readout ASIC). While other attempts include using a monolithic scintillator module and employing nearest-neighbor localization algorithms, n-to-one coupled light sharing is the most commercially viable option because it offers both depth of interaction (DOI) and time-of-flight (TOF) readout capabilities without compromise in sensitivity and / or energy resolution.

[0005] However, with the increase in spatial resolution, the amount of data per PET scan increases dramatically due to the increased number of voxels. Depth encoding is necessary to mitigate parallax and fully realize the benefits of high-resolution PET, but it also exacerbates the data size problem because the number of lines of response (LORs) grows exponentially as a function of the number of DOI bins. Combining high resolution with TOF readout also contributes to the larger data size in PET, as one timestamp is read per pixel per channel, which still makes the process computationally inefficient, even though multiple timestamps are not typically used per event.

[0006] As data increases, the number of connections between optical sensors and readout ASICs also increases, which in practice will increase the heat generated by the device.

[0007] Readout systems typically utilize a one-to-one coupling between readout pixels and channels. However, this readout method is inefficient because not all pixels need to be read out for each event.

[0008] To reduce the computational cost of PET, signal multiplexing techniques have been proposed to reduce data size and complexity. Multiplexing combines signals read from multiple optical sensors (pixels) for each event. However, even with multiplexed signals, the solution must still be able to determine the interactions between the main optical sensors (pixels), the main scintillator module, DOI, and TOF.

[0009] In one or more known systems employing multiplexing, the detector modules used lack deep coding capabilities (therefore, multiplexing readout schemes have not yet demonstrated compatibility with DOI readout), while deep coding capabilities are crucial for achieving system-level spatial resolution uniformity or high temporal resolution capabilities for Time of Flight (TOF). Multiplexing schemes may also affect temporal resolution. Summary of the Invention

[0010] Therefore, a particle detection system is disclosed, which may include an optical sensor array, a scintillator array, and a segmented light guide. The optical sensor array may include a first plurality of optical sensors. Each optical sensor may correspond to one pixel. The scintillator array may include a second plurality of scintillator modules. The number of scintillator modules may be greater than the number of optical sensors. The multiple scintillator modules may contact a corresponding optical sensor at a first end of their respective scintillator modules. The segmented light guide may include a multiple prismatoid segments. The segmented light guide may contact a second end of the second plurality of scintillator modules. Each prismatoid segment may contact a scintillator module that is in contact with at least two different optical sensors. The at least two different optical sensors may be adjacent optical sensors. Each prismatoid segment may be configured to redirect particles between scintillator modules in contact with the corresponding prismatoid segment.

[0011] The system may also include a third or more energy readout channels. Multiple optical sensors can be connected to energy readout channels separately, such that optical sensors associated with the same pseudo-cylindrical segment may not be connected to the same energy readout channel. Each energy readout channel may have at least two timestamps associated with it.

[0012] In one aspect of this disclosure, the optical sensors can be arranged in rows and columns. Adjacent optical sensors in a row can be connected to different energy readout channels, and adjacent optical sensors in a column can be connected to different energy readout channels.

[0013] In one aspect of this disclosure, the system may also include at least two comparators for each energy readout channel, said at least two comparators being connected to multiple optical sensors of the same energy readout channel. Each comparator (of the same energy readout channel) may have a different threshold. The comparators may be connected to either an anode or a cathode.

[0014] In one aspect of this disclosure, the energy readout channel may be connected to the same or different terminals as the time information.

[0015] In one aspect of this disclosure, four optical sensors can be connected to the same energy readout channel.

[0016] In one aspect of this disclosure, there may be different couplings from the scintillator module to the optical sensor, such as four-to-one coupling or nine-to-one coupling.

[0017] In one aspect of this disclosure, the system may further include a first processor configured to bias a first plurality of optical sensors during readout and to receive outputs and at least two timestamps associated with each energy readout channel via a third plurality of energy readout channels.

[0018] In one aspect of this disclosure, the system may further include a second processor in communication with the first processor. The second processor may be configured to determine a timing parameter of an event based on at least two received timestamps.

[0019] In one aspect of this disclosure, the time parameter may be based on a combination of the at least two timestamps. In some aspects, the time parameter may be based at least on the fastest timestamp. In some aspects, the time parameter may be based on a linear regression analysis of the at least two received timestamps.

[0020] In one aspect of this disclosure, the second processor can also be configured to determine the time of flight (TOF) between the coincidence detection modules based on time parameters.

[0021] In one aspect of this disclosure, the second processor may also be configured to determine at least one of a primary interacting pixel, a primary interacting scintillator module, or an interaction depth for an event. In another aspect of this disclosure, the second processor may select at least two timestamps associated with the determined primary interacting pixel to determine timing parameters.

[0022] In one aspect of this disclosure, the second processor can be configured to use a machine learning model to determine the Time of Frame (TOF), the machine learning model being input to at least two timestamps received from the conformity detection module. Attached Figure Description

[0023] Figure 1A A multiplexing scheme according to an aspect of the present disclosure is shown, in which the cathode of an optical sensor is multiplexed to provide energy information, and the anode of an optical sensor is multiplexed to provide multiple timestamps;

[0024] Figure 1B A multiplexing scheme and associated timestamps for an energy channel are shown, wherein the anode of the optical sensor is multiplexed to provide energy information, and the cathode of the optical sensor is multiplexed to provide timestamps;

[0025] Figure 1C A multiplexing scheme and associated timestamp for an energy channel are shown, wherein the anode of an optical sensor is multiplexed to provide energy information and a timestamp;

[0026] Figure 2A A particle detection device having a four-to-one coupling from a scintillator module to an optical sensor is shown according to aspects of this disclosure.

[0027] Figure 2B A particle detection system according to an aspect of this disclosure is shown, wherein there is a four-to-one coupling from a scintillator module to an optical sensor;

[0028] Figure 3A A top view of the segmented light guide and optical sensor for a four-to-one coupling from the scintillator module to the optical sensor is shown, wherein the segments of the segmented light guide have three different designs.

[0029] Figure 3B An example of a 3D view of a segment of a segmented optical guide according to aspects of this disclosure is shown;

[0030] Figure 4 A particle detection system according to an aspect of this disclosure is shown, wherein there is a nine-to-one coupling from a scintillator module to an optical sensor;

[0031] Figure 5 A top view of the segmented light guide and optical sensor for a nine-to-one coupling from the scintillator module to the optical sensor is shown, wherein the segments of the segmented light guide have three different designs.

[0032] Figure 6 A flowchart of a method according to an aspect of this disclosure is shown;

[0033] Figure 7 A flowchart illustrating an example of training and testing a machine learning model for demultiplexing multiplexed ASIC power channels according to aspects of this disclosure is shown.

[0034] Figure 8 An example of a machine learning model for demultiplexing multiplexed ASIC power channels according to aspects of this disclosure is shown;

[0035] Figure 9A and Figure 9BA comparison is shown between the ground truth for four-to-one coupling from the scintillator module to the optical sensor and demultiplexing the multiplexed signal using a machine learning model according to aspects of this disclosure;

[0036] Figure 9C and Figure 9D A comparison is shown between a synthetically reused dataset and an actual reused dataset reused according to aspects of this disclosure;

[0037] Figure 10A and Figure 10B A comparison is shown between the DOI resolution of a correlated particle detection system with a four-to-one coupling from a scintillator module to an optical sensor and the DOI resolution of a particle detection system according to aspects of this disclosure.

[0038] Figure 11A and Figure 11B A comparison is shown between the ground truth for nine-to-one coupling from the scintillator module to the optical sensor and demultiplexing the multiplexed signal using a machine learning model according to aspects of this disclosure;

[0039] Figure 12A and Figure 12B A comparison is shown between the DOI resolution of a correlated particle detection system with a nine-to-one coupling from a scintillator module to an optical sensor and the DOI resolution of a particle detection system according to aspects of this disclosure.

[0040] Figure 13 An example arrangement of an optical sensor array and segmented light guides according to aspects of this disclosure is shown;

[0041] Figure 14 It also shows, as Figure 1A A table showing the multiplexed energy channels according to aspects of this disclosure;

[0042] Figure 15 An example of a compliance detection module according to aspects of this disclosure is shown;

[0043] Figure 16 A flowchart illustrating an example of training and testing a machine learning model for use in TOF prediction according to aspects of this disclosure is shown.

[0044] Figure 17 Examples of machine learning models for use in TOF predictions according to aspects of this disclosure are shown; and

[0045] Figure 18 A table showing simulation results based on one, two, or three timestamps according to aspects of this disclosure is presented, where DOI correction shows improved conformance temporal resolution using multiple timestamps. Detailed Implementation

[0046] A multiplexing scheme utilizing deterministic light sharing is disclosed, which is achieved using segmented light guides such as those disclosed in U.S. Patent Publication No. 2020 / 0326434, which is incorporated herein by reference. The particle detection system (and apparatus) described herein has a single-ended readable (depth-coded) display having a specific segment pattern of segmented quasi-cylindrical light guides. The light guides have quasi-cylindrical segments, which will be referenced at least to… Figure 3A This is described in detail. According to aspects of this disclosure, the segmented pseudo-pillar light guide 200 has at least three different pseudo-pillar designs, such as a central pseudo-pillar 162, a corner pseudo-pillar 166, and an edge pseudo-pillar 168. The pseudo-pillars are designed to mitigate edge and corner artifacts, thereby achieving uniform crystal recognition performance, even when using the multiplexing scheme described herein.

[0047] Light sharing between scintillator modules 205 is limited to scintillator modules 205 that belong to adjacent or nearby optical sensors 10 (e.g., the nearest neighbor) to generate deterministic and anisotropic light sharing modes between scintillator modules and maximize the signal-to-background ratio on the optical sensors 10 to improve energy resolution and DOI resolution, while preserving high temporal resolution for time of flight (TOF).

[0048] Due to the deterministic light-sharing mode, only a subset of optical sensors 10 (pixels) from the nearest neighbor optical sensors (pixels) is needed to accurately perform the main optical sensor interactions and DOI (and estimate the main scintillator module). This is because the relevant signals will be contained within the optically isolated quasi-pillar segments.

[0049] Figure 1A An example of a reuse scheme according to aspects of this disclosure is shown. For example... Figure 1A As shown, for optical sensors 101 to 10 64 (collectively referred to as 10) (e.g., optical sensor array 210) are multiplexed. In one aspect of this disclosure, multiplexing generates multiple energy readout channels 1001 to 100. 16 Multiplexed outputs Y01 to Y16 (501 to 50) 16 (Collectively referred to as "50"). Multiplexed outputs Y01 to Y16 are inputs to the read ASIC 405. For example... Figure 1A As shown, each energy readout channel has four optical sensors 10. The number of optical sensors, the number of multiplexed optical sensors, and the number of energy readout channels are not limited to 64, 4, and 16, respectively. Other combinations can be used. Figure 1A and 1B As shown, when the cathode is also connected to bias 15, capacitor C can be connected to the cathode between the multiplexed output and the readout ASIC 405.

[0050] Each optical sensor 10 has an anode and a cathode. Figure 1A In this design, the cathode is displayed at the top of the pixel, and the anode is displayed at the bottom of each pixel. In one aspect of this disclosure, a bias 15 can be provided to the cathode via a bias circuit. The bias circuit is not in... Figure 1A As shown in the diagram. The bias circuit may include one or more capacitors and one or more resistors.

[0051] In one aspect of this disclosure, optical sensors 101 to 10 64 The optical sensor array 210 can be arranged in rows and columns. For example, the optical sensor array 210 can be an 8 x 8 readout array. However, the readout array is not limited to 8 x 8 and can be other dimensions, such as 4 x 4 or 16 x 16. In some aspects, the readout array can be an integer multiple of 2. The two-dimensional array can be formed in a plane perpendicular to the longitudinal axis of the scintillator module. In one aspect of this disclosure, the optical sensor 10 can be a silicon photomultiplier (SiPM). In other aspects of this disclosure, the optical sensor 10 can be an avalanche photodiode (APD), a single-photon avalanche photodiode (SPAD), a photomultiplier tube (PMT), or a silicon avalanche photodiode (SiAPD). These are non-limiting examples of solid-state detectors that can be used. The number of optical sensors 10 (pixels) in the device can be based on the application and size of the PET system. Figure 1A In this designation, the optical sensor 10 is labeled "SiPM Pixel". The two digits at the bottom right of each pixel indicate its pixel number. For example, "01" represents the first pixel, and "64" represents the last pixel. These numbers are for descriptive purposes only.

[0052] Figure 13 An example of an optical sensor array 210 with an 8 x 8 configuration (8 rows and 8 columns) is shown. Figure 13 In this context, not all optical sensors are numbered as SiPM pixels "XX", where XX represents a number.

[0053] Optical sensors (SiPM 01 to 08) are in the first row, optical sensors (SiPM 09 to 16) are in the second row... optical sensors (SiPM 57 to 64) are in the eighth row (last row). Optical sensors (SiPM 01, 09, 17, 25, 33, 41, 49, and 57) are in the first column, optical sensors (SiPM 02, 10, 18, 26, 34, 42, 50, and 58) are in the second column... optical sensors (SiPM 08, 16, 24, 32, 40, 48, 56, and 64) are in the eighth column (last column). Figure 13An example arrangement of the pseudo-pillar segments of the segmented pseudo-pillar light guide 200 superimposed on the optical sensor 10 is also shown. The optical sensor 10 is shown as being separated by lines, and the scintillator module (crystal) is indicated by dashed lines.

[0054] like Figure 1A As shown, the cathode of the optical sensor 10 is multiplexed to generate an energy readout channel (via integrator 30). The signal is integrated via integrator 30 to provide energy for one or more events.

[0055] For a given energy channel multiplexing, a specific optical sensor 10 is selected such that optical sensors 10 connected to the same segment of the segmented pseudo-cylindrical light guide 200 are not multiplexed. For example, segment 1 1301 (such as...) Figure 13 The optical sensors 01, 02, 09, and 10 are associated with optical sensors (SiPM). Therefore, light can be shared among optical sensors (SiPM) 01, 02, 09, and 10. According to aspects of this disclosure, these optical sensors can be de-multiplexed. Similarly, segment 2 1302 is associated with optical sensors 56, 63, and 64. Therefore, light can be shared among them. According to aspects of this disclosure, these optical sensors can be de-multiplexed. Similarly, segment 3 1303 is associated with optical sensors (SiPM) 61 and 62. Therefore, light can be shared among them. According to aspects of this disclosure, these optical sensors can be de-multiplexed. Figure 13 The arrangement shown is similar to Figure 3A The arrangement shown.

[0056] Figure 14 An example of a multiplexing scheme for optical sensor 10 is shown, wherein the multiplexed optical sensor in each energy channel is not associated with the same pseudo-cylindrical segment of the segmented pseudo-cylindrical light guide (with... Figure 1A (The scheme is the same as in the example). In this example, at least one optical sensor 10 (pixel) is located between optical sensors connected to the same energy channel.

[0057] For example, in energy channel (ASIC_Energy_01) 1001, optical sensors 101, 103, 105, and 107 are connected to this channel (not all pixels / optical sensors are explicitly labeled with reference numeral 10 for illustrative purposes). Optical sensors 102, 104, 106, and 108 are not connected to energy channel (ASIC_Energy_01). In other aspects of this disclosure, optical sensors 102, 104, 106, and 108 may be connected to energy channel (ASIC_Energy_01) 1001, while optical sensors 101, 103, 105, and 107 may not be connected to energy channel (ASIC_Energy_01) 1001.

[0058] (ASIC_Energy_01)1001 to (ASIC_Energy_08)1008 may also be referred to as row channels in this paper, since the optical sensors in a row are connected to the same channel (also referred to as horizontal channels in this paper).

[0059] (ASIC_Energy_09)1009 to (ASIC_Energy_16)100 16 This can also be referred to as a column channel in this document, because the optical sensors in a column are each connected to the same channel (also known as a vertical channel). For example, in energy channel (ASIC_Energy_09) 1009, optical sensors 109, 10... 25 10 41 10 57 Connected to the same energy channel. Optical sensors 101, 10 17 10 33 10 49 There is no connection to the energy channel (ASIC_Energy_09) 1009. In other aspects of this disclosure, optical sensors 101, 10... 17 10 33 10 49 It can be connected to the energy channel (ASIC_Energy_09) 1009, while the optical sensors 109 and 1009... 25 10 41 10 57 It is not necessary to connect it to the energy channel (ASIC_Energy_09) 1009.

[0060] As mentioned above, the channels are connected so that adjacent pixels in any direction are not connected to the same energy channel.

[0061] In one aspect of this disclosure, a subset of optical sensors in a row connected to an energy channel is offset in columns from a subset of optical sensors in adjacent rows connected to its energy channel. For example, optical sensors 101, 103, 105, and 107 connected to (ASIC_Energy_01) 1001 are in columns C1, C3, C5, and C7, respectively. Therefore, optical sensors 109 and 100 are also in columns C1, C3, C5, and C7. 11 10 13 10 15 Instead of connecting to (ASIC_Energy_02) 1002, the optical sensor 10 can be located in columns C2, C4, C6, and C8. 10 10 12 10 14 10 16Connect to (ASIC_Energy_02)1002.

[0062] In one aspect of this disclosure, a subset of optical sensors connected to a column of an energy channel is offset in row from a subset of optical sensors connected to a row of its energy channel. For example, optical sensors 109, 1009 are connected to (ASIC_Energy_09) 1009. 25 10 41 10 57 They are located in rows R2, R4, R6, and R8, respectively. Therefore, optical sensors 10 are also located in rows R2, R4, R6, and R8. 10 10 26 10 42 10 58 (In column C2) it is not necessary to connect to (ASIC_Energy_10)100 10 Instead, it is the optical sensors 102 and 103 in rows R1, R3, R5, and R7. 18 10 34 10 50 Connected to (ASIC_Energy_10) 100 10 .

[0063] According to aspects of this disclosure, the same optical sensor for which energy is reused is also reused to generate at least two timestamps, such as time information. Figure 1A As shown, the anode of the optical sensor is time-multiplexed, and the multiplexed time output is... Figure 1A The numbers shown are X01 to X16 (551 to 55). 16 (collectively referred to as "55"). X01 to X16 can be input to the read ASIC 405. In one aspect of this disclosure, anodes can be used because they are generally faster than cathodes. The anodes are connected to at least two comparators 20 (two timestamps) within the read ASIC 405. Figure 1A As shown, there are three comparators 20 associated with each energy channel. Each comparator 20 is associated with a different voltage threshold V_th1, V_th2, and V_th3. The voltage thresholds can correspond to different photon counts. The same three voltage thresholds can be used for comparators associated with different energy channels ASIC_Energy_01 to ASIC_Energy_16 (collectively referred to as "100"). When the multiplexed voltage exceeds the corresponding threshold, the output of the corresponding comparator 20 will change (e.g., for ASIC_Energy_01, X01_T1, X01_T2, and X01_T3... for ASIC_Energy_16, X16_T1, X16_T2, and X16_T3). The change time of each comparator can be used as a timestamp.

[0064] In some aspects of this disclosure, timestamps can be combined to determine the time parameters of the detection device (also referred to herein as a detection module) for an event. These time parameters can then be used to determine the Time of Response (TOF) between coincidence detection devices. The TOF can be determined by taking the difference between the time parameters of two relative detection devices (coincidence). Figure 15 Two detection devices (e.g., detection module 1 1501 and detection module 2 1502) and a radiation source 1500 between them are shown. The radiation source 1500 can be aligned with the center of the two detection devices. Coincidence time resolution (CTR) is a measure of the accuracy (jitter) of repeated TOF measurements at the same location of the radiation source 1500. CTR is determined by taking the half-maximum full width (FWHM) of the TOF distribution at a given fixed location.

[0065] CTR can be improved by using multiple timestamps. In some respects, using multiple timestamps can improve CTR through leading-edge slope estimation or waveform shape estimation. Leading-edge slope estimation or waveform shape estimation can be performed using machine learning. For example, convolutional neural networks (CNNs) can be used, as will be described later.

[0066] In other respects, the connection to the readout ASIC 405 can be reversed, and the multiplexed output 55' of the connected anode can be used as... Figure 1B The energy channels shown are, for example, ASIC_Energy_01'. The multiplexed output 50' of the cathode can be used for timestamps, for example, X01_T1', X01_T2', and X01_T3'. For discussion purposes, Figure 1B Only one multiplexed energy channel (and associated timestamp) is shown; however, other channels may have similar configurations.

[0067] In other respects, the same terminal (e.g., anode or cathode) can be used for both energy and time information. For example, as Figure 1C As shown, the anode of the optical sensor 10 can be multiplexed, such that the same multiplexed output 55'' is connected to the integrator 30 and the comparator 20 to generate one or more energy channels (e.g., ASIC_Energy_01) and timestamps (e.g., X01_T1'', X01_T2'', and X01_I3''). Similar to... Figure 1B For the purpose of discussion, Figure 1C Only one reused energy channel (and associated timestamp) is shown; however, other energy channels may have similar configurations.

[0068] Multiple outputs Y01 to Y16 and multiplexed outputs X01 to X16 can be connected to a readout ASIC 405 (also referred to herein as a first processor). The readout ASIC 405 may include a comparator 20 and an integrator 30. Time is recorded as the output changes. The readout ASIC 405 may also include an analog-to-digital converter for digitizing signals from the optical sensor array 210 and circuitry for controlling bias. The readout ASIC 405 may also include a communication interface for transmitting the digitized signals to a remote computer 400 (also referred to herein as a second processor) via a synchronization board 410. The synchronization board 410 synchronizes readings from different detection devices / readout ASICs within the PET system. Figure 2B The system shown depicts only one detection device; however, in reality, multiple detection devices are connected to the synchronization board 410. These multiple detection devices may include… Figure 15 The relative detection devices (detection modules 1 and 2) 1501 and 1502 are shown. Each detection device may have the four-to-one readout multiplexing described herein. Figure 2B The reflector 215 is omitted from the original text. However, each detection device will have a reflector 215.

[0069] Figure 2A A particle detection device with a four-to-one coupling 202 from scintillator modules to an optical sensor is shown according to aspects of this disclosure. Each scintillator module 205 can be made of lutetium yttrium silicate (LYSO) crystal. The scintillator module 205 is not limited to LYSO, and other types of crystals that emit photons in the presence of incident gamma radiation, such as lutetium silicate (LSO), can be used. Figure 2A In this design, the optical sensor array is represented as a SiPM array 210. However, as mentioned above, this array is not limited to SiPMs. A scintillator module 205 contacts the surface of the SiPM array 210 at its first end. Although Figure 2A The gap between the scintillator module 205 and the SiPM array 210 is shown, but in reality, the scintillator module 205 is attached to the SiPM array 210 by an optical adhesive or epoxy resin. The optical adhesive or epoxy resin does not alter the path of the particles or light or attenuate it (if any change occurs, it is minimal). This gap is shown to illustrate the travel of particles from the first end of the scintillator module to the SiPM array (pixel). The scintillator module 205 contacts the surface of the segmented pseudo-pillar light guide (PLGA 200) at its second end. A reflector 215 is located above the PLGA 200. In one aspect of this disclosure, the reflector 215 may comprise barium sulfate (BaSO4). In other aspects, the reflector 215 may comprise other reflective materials. In one aspect of this disclosure, the reflector 215 may be used between each scintillator module 205. The reflector 215 may also fill any gaps between the segments of the segmented pseudo-pillar light guide 200.

[0070] Figure 3A This diagram shows a view of a segmented quasi-cylindrical light guide and optical sensors with four-to-one coupling from the scintillator modules to the optical sensors, where the segments of the light guide have three different designs. The lower left corner of the diagram is a plan view showing the relative arrangement of the scintillator modules (2 x 2) for each optical sensor. Figure 3A This is also referred to as a "crystal". For illustrative purposes, only a subset of the array is shown. Three different designs of the pseudo-prism segments are shown using different hashing notations, such as central pseudo-prism 162, corner pseudo-prism 166, and edge pseudo-prism 168. Central pseudo-prism 162 and edge pseudo-prism 168 are shown with hashing in opposite directions, while corner pseudo-prism 166 is shown with cross-hashing notations. Figure 3A The upper right corner shows examples of three different designs (sectional and perspective views). The corner pseudo-prism 166 can contact the scintillator module 205, which contacts three different optical sensors (three pixels). The edge pseudo-prism 168 can contact the scintillator module 205, which contacts two different optical sensors (two pixels). The center pseudo-prism 162 can contact the scintillator module 205, which contacts four different optical sensors (four pixels).

[0071] Figure 3A 142 and 144 are used to identify two adjacent optical sensors. For example... Figure 3A As shown, the outline of the pseudo-prism is generally triangular. However, in another aspect of this disclosure, the shape of the pseudo-prism may be approximately at least one of the following: at least one prism, at least one antiprism, at least one truncated pyramid, at least one dome, at least one parallelepiped, at least one wedge, at least one pyramid, at least one truncated pyramid, at least a portion of a sphere, or at least one cuboid. Figure 3B Examples of certain 3D shapes are shown (five different shapes of the segment). For example, the shape can be 1) a cuboid, 2) a pyramid, 3) a combination of a cuboid and a pyramid, 4) a triangular prism, and 5) a combination of a cuboid and a triangular prism. The combination of a cuboid and a triangular prism is shown in... Figure 3A As shown in the figure, the cuboid forms the base of the triangular prism.

[0072] In one aspect of this disclosure, each pseudo-cylindrical segment of the segmented pseudo-cylindrical light guide 200 is offset from the optical sensor. In some aspects, it is offset according to the scintillator module. In this aspect of the disclosure (and in the case of a four-to-one coupling from the module to the sensor), each scintillator module can share light with other scintillator modules from different optical sensors (pixels). For example, when a photon enters the pseudo-cylindrical segment (segment of the light guide) after interacting with gamma rays from scintillator module 205, the photon (i.e., particle 300) is effectively redirected to adjacent scintillator modules (of different pixels) due to the geometry, thereby improving the light sharing rate between optical sensors (pixels).

[0073] Figure 4 Another example of a particle detection system according to aspects of this disclosure is shown. Figure 4 In this configuration, there is a nine-to-one coupling between the scintillator module and the optical sensor. The optical sensor 10 is multiplexed with the aforementioned four-to-one readout (e.g., ...). Figure 1A and 2B (As shown) Connect to the read ASIC 405 in the same way. Similar to... Figure 2B The readout ASIC 405 is connected to the computer 400 via a synchronization board 410. The synchronization board synchronizes readings from different detection devices / readout ASICs within the PET system. Figure 4 The system shown depicts only one detection device; however, multiple detection devices are actually connected to the synchronization board 410. These multiple detection devices may include, for example... Figure 15 The relative detection devices (detection modules 1 and 2) 1501 and 1502 are shown. Each detection device has the four-to-one readout multiplexing described herein. Figure 4 The reflector 215 is omitted. However, each detection device will have a reflector 215. The computer 400 may include at least one processor, memory, and a user interface, such as a keyboard and / or display. The operator can use the user interface to specify the readout interval or period.

[0074] In one aspect of this disclosure, each pixel (except for the four corner pixels) may have nine scintillator modules 205. The corner pixels may have four scintillator modules. Figure 5 A segment of the optical guide is shown. Similar to... Figure 3A The lower left corner displays segments with different mixed labels, indicating different designs. Figure 5The lower left portion shows only a representative portion of array 220. The solid lines around a group of scintillator modules or crystals in the lower left refer to pixels (SiPM pixels), while the dashed lines refer to modules or crystals. Three different designs of the pseudo-pillar segments are shown using different mixed notations, such as the central pseudo-pillar 162, the corner pseudo-pillar 166, and the edge pseudo-pillar 168. The central pseudo-pillar 162 and the edge pseudo-pillar 168 are shown with mixed notations in opposite directions, while the corner pseudo-pillar 166 is shown with intersecting mixed notations. The contour of the corner pseudo-pillar 166 in a 9 x 1 configuration may differ from that in a 4 x 1 configuration because only corner pixels in a 9 x 1 configuration may have 4 x 1 coupling. Figure 5 The right side shows several different central pseudo-cylindrical positions relative to the pixels (and scintillator modules). Figure 5 Not all SiPM pixels (optical sensors) are shown on the right side. Figure 5 The diagram shows nine central pseudo-pillars to illustrate nine different primary interacting scintillator modules (primary interactions). For example, when the primary interacting scintillator module is module 139 (the central scintillator module in this segment), this segment directs particles to four adjacent optical sensors / pixels 142, 144, 146, and 148. Figure 5 The "X" in the diagram refers to the main interacting scintillator module. Segments 132 and 134 may not be adjacent to each other, but they appear to be adjacent in the diagram.

[0075] In this configuration, the corner pseudo-pillar 166 can redirect particles between the ends (ends in contact with the segment) of a set of five scintillator modules (three different optical sensors / pixels). The edge pseudo-pillar in this configuration can also redirect particles between the ends (ends in contact with the segment) of the five scintillator modules (two different optical sensors / pixels).

[0076] In other configurations, even the corner optical sensor / pixel 10 can contact nine scintillator modules 205.

[0077] In one aspect of this disclosure, the scintillator module 205 may have a tapered end, as described in PCT application serial number US21 / 48880, filed September 2, 2021, entitled “Tapered Scintillator Crystal Module and Method of Use Thereof,” the contents of which are incorporated herein by reference. The tapered end is a first end, such as a scintillator module / optical sensor interface.

[0078] As described above, the deterministic light-sharing scheme caused by the segmented light guide 200 ensures that light sharing between scintillator modules occurs only between scintillator modules coupled to the same optically isolated quasi-cylindrical light guide, and this allows the reuse described herein to retain high centroid, TOF and DOI, as well as energy resolution.

[0079] Figure 6 A flowchart of a method according to an aspect of this disclosure is shown. For purposes of description, the functions described below are performed by a processor of computer 400. In S600, the processor (via synchronization board 410) instructs a readout ASIC 405 to read signals from the optical sensor array. This may be in the form of a frame synchronization command. When the readout ASIC 405 receives the instruction, it causes power to be supplied to the optical sensor array 210. In some aspects of this disclosure, there are switches controlled to close to provide bias. The readout ASIC 405 receives multiplexed signals Y01 to Y16 50 for energy (via connection) and (via connection) multiplexed signals X01 to X16 55 for time (or vice versa), or receives one multiplexed signal for both. The multiplexed signals Y01 to Y16 50 can be integrated to obtain the corresponding energy channels, digitized and synchronized (via synchronization board 410), and transmitted to computer 400. The multiplexed signals X01 to X16 can be sent to comparator 20. Each comparator outputs the values ​​T1, T2, and T3 associated with each energy channel. The time changes can be recorded, digitized, and transmitted to a computer 400. Although Figure 1A Three comparators are shown, but the number of comparators is not limited to three. In some aspects of this disclosure, at least two comparators can be used.

[0080] In other aspects of this disclosure, the outputs of the comparators can be combined via one or more logic gates before being transmitted to a computer. For example, the output from a first comparator can be sent to a logic gate, the output from a second comparator can be sent to a different logic gate, and so on.

[0081] In one aspect of this disclosure, computer 400 includes a communication interface. In some aspects, the communication interface may be a wired interface.

[0082] In S605, the processor receives digitized signals and time signals from each energy channel ASIC_Energy 01 to ASIC_Energy 16 100, such as the digitized outputs from comparator 20 (as associated with each energy channel). In other aspects, the processor may receive digitized signals from outputs combined via logic gates. In some aspects of this disclosure, the digitized signals from each energy channel ASIC_Energy 01 to ASIC_Energy 16 100 are associated with channel identifiers, enabling the processor to identify which digitized signals correspond to which channel. The digitized signals may be stored in memory. In one aspect of this disclosure, computer 400 has a preset mapping that identifies which pixels are connected to the corresponding (multiplexed) channels. This mapping may be stored in memory.

[0083] In 610, the processor can identify a subset of energy channels ASIC_Energy_01 to ASIC_Energy_16 with the highest digitized signal (e.g., highest X energy) for each event. Each event is determined relative to a time window. The window for an event begins when the SiPM initially senses one or more particles. The window "opens" for a set time period. The set time period can be a few nanoseconds. Particles detected within this window (from any SiPM) are grouped together and considered to belong to the same event. In one aspect of this disclosure, the number of associated energy channels can be based on the location of the event. For example, if the main interaction is located at the center of the array (associated with the central pseudo-pillar 162), the number of associated energy channels can be four. The processor can identify four energy channels with the four highest digitized signals for that event. When the main interaction is located at the corner pseudo-pillar 166, the processor may only need to identify the three energy channels associated with the three highest digitized outputs. When the main interaction is located at the edge pseudo-pillar 168, the processor may only need to identify the two energy channels associated with the two highest digitized outputs.

[0084] Assuming the light sharing is optically isolated by the segments, the interacting master optical sensors (pixels) can be determined based on the relationship between the energy channels and certain highest digitized signals. This relationship allows for the unique identification of adjacent optical sensors based on the patterns of the energy channels with certain highest digitized signals. In S615, the processor can determine the master interacting optical sensors (pixels). For example, if the master interacting optical sensor is the center, the processor can use a stored mapping to determine the relative positions of the four identified energy channels associated with these four highest signals. This reduces the master optical sensors (from 16 possible sensors / pixels connected to the identified channels) to four neighboring optical sensors / pixels. For example, when these four highest channels are energy channels ASIC_Energy_02, ASIC_Energy_03, ASIC_Energy_10, and ASIC_Energy_11, the processor can identify SiPM pixels 10, 11, 18, and 19 as neighboring optical sensors, such as adjacent pixels. The processor can then determine which of the four energy channels has the highest signal. The optical sensor associated with the energy channel with the highest sensor value (among the four neighboring optical sensors after reduction) is identified as the primary optical sensor / pixel (primary interaction). For example, when the maximum signal of the four energy channels is ASIC_Energy_03, the processor can determine that the primary interaction optical sensor (pixel) is 19 (which is reduced from 17, 19, 21, and 23 connected to ASIC_Energy_03).

[0085] When the primary interacting optical sensor is at a corner, the processor can use a stored mapping to determine the relative positions of the three identified energy channels associated with the three highest signals. Otherwise, the processor can still use four energy channels with four highest signals. This reduces the primary interacting optical sensor to three neighboring optical sensors / pixels. The processor can then determine which of the three energy channels has the highest signal. The optical sensor associated with the energy channel having the highest signal (among the reduced three neighboring optical sensors) is identified as the primary optical sensor / pixel (primary interacting).

[0086] When the primary interacting optical sensor is an edge optical sensor (associated with an edge quasi-pillar), the processor can use a stored mapping to determine the relative positions of the two identified energy channels associated with the two highest signals. Alternatively, the processor can still use four energy channels with four highest signals. This reduces the primary interacting optical sensor to two adjacent optical sensors / pixels. The processor can then determine which of the two energy channels has the highest signal. The optical sensor associated with the energy channel having the highest signal (among the two adjacent optical sensors in the reduced set) is identified as the primary interacting optical sensor / pixel.

[0087] In the S620, the processor can determine the DOI. The DOI can be determined using the following formula:

[0088]

[0089] Pmax is the digitized value associated with the energy channel having the highest signal (highest energy) for that event, and P is the sum of the digitized signals associated with the identified subset of energy channels for that event, which can also be calculated after subtracting Pmax as needed. Since segment optics isolate adjacent optical sensors associated with the segment, this summation is actually the ratio of the energy associated with the primary interacting optical sensor to the sum of the energies of adjacent sensors. Once the processor identifies the primary interacting optical sensor, it knows how many energy channels (up to M energy channels) to sum, for example, 4 for a central quasi-prism optical sensor, 3 for a corner quasi-prism optical sensor, and 2 for an edge quasi-prism optical sensor.

[0090] The ratio can then be converted into depth using the following formula.

[0091] DOI = m * w + q (2)

[0092] Where m is the slope between DOI and w according to the best-fit linear regression model, and q is the intercept to ensure that the DOI estimate begins at DOI = 0 mm. Parameters m and q can be determined in advance for the scintillator module 205.

[0093] Therefore, according to aspects of this disclosure, the multiplexed energy signal can be used to determine the DOI and the master interacting optical sensor without requiring demultiplexing techniques such as machine learning or lookup tables described herein to demultiplex the energy signal. In other aspects of this disclosure, the DOI can be calculated after demultiplexing the multiplexed energy signal according to aspects of this disclosure and subsequently calculated from the demultiplexed energy signal, where Pmax is the digitized value associated with the optical sensor / pixel having the highest demultiplexed value, and p is the sum of all demultiplexed values ​​for each optical sensor / pixel.

[0094] In one aspect of this disclosure, the primary interacting scintillator module can be estimated using multiplexed energy signals based on the relative amplitudes of the four highest energy channels. Using the example identified above, when the four highest energy channels ASIC_Energy_02, ASIC_Energy_03, ASIC_Energy_10, and ASIC_Energy_11 are determined, assuming a light-sharing scheme for the central segment (e.g., a quasi-pillar), the upper-left scintillator module associated with SiPM 19 can be estimated as the primary interacting scintillator module. Using relative amplitudes, the processor can identify the primary optical sensor (pixel), vertical / horizontal neighbors, and diagonal neighbors. Diagonal neighbors may have the lowest energy of the identified subset of energy channels. Horizontal / vertical neighbors may have close energies; for example, the energy channel outputs may be nearly equal. Adjacent optical sensors identified using subsets of energy channels may be associated with the same segment (due to light sharing).

[0095] Although the primary interaction optical sensor and primary interaction scintillator module can be estimated as described above, they can be determined after the energy signal in the demultiplexed energy channel 100 as described herein, due to scattering and noise.

[0096] In the S625, the processor can demultiplex the multiplexed energy signal from energy channel 100 to achieve full optical sensor resolution. For example, the processor acquires multiplexed energy signals from energy channels ASIC_Energy 01 to ASIC_Energy 16 100 and generates M x M information energy channels (the number of optical sensors in the system), where M is the number of rows and columns. For example, for an 8 x 8 readout array, there are 64 demultiplexed energy channels.

[0097] In one aspect of this disclosure, the transformation is based on a pre-stored machine learning model. See later for reference. Figure 7and Figure 8 The process of generating a machine learning model is described in detail. Specifically, the processor can retrieve a stored machine learning model and use the multiplexed energy signal as input to output a demultiplexed energy signal corresponding to the 64 energy channels of an 8 x 8 array.

[0098] In other respects, the processor can use a stored lookup table that correlates the multiplexed energy signals with demultiplexed energy signals at full energy channel resolution. This lookup table can be created using experimental data obtained from unmultiplexed energy channels. For an 8 x 8 array, this lookup table can be created from experimental data taken from 64 energy channels across multiple events. For example, data can be obtained from 64 energy channels of a single event. The multiplexed data can be generated by the processor (based on software multiplexing), which will then... Figure 1A The same energy channels are added together to generate data for 16 energy channels (adding 4 energy channels together). This data for 16 energy channels is then associated with data for 64 energy channels for later use. This process can be repeated for multiple events to create multiple correspondences, such as from 64 energy channels to 16 energy channels. Subsequently, when multiplexed data is obtained from the readout ASIC 405, the processor looks up the 64 energy channel data. The processor can select the 64 energy channel data corresponding to the 16 energy channel data that is closest to the actually detected energy channel data. "Closest" can be defined as the smallest root mean square error or mean square error. However, other parameters can be used to determine the closest stored 16 energy channel data in the lookup table. In other aspects of this disclosure, the processor can interpolate the 64 energy channel data based on the difference between the closest stored 16 energy channel datasets (e.g., two closest datasets).

[0099] In the S630, the processor uses the demultiplexed energy signal (e.g., a signal representing the energy from each optical sensor) to calculate the energy-weighted average. The energy-weighted average can be calculated using the following formula:

[0100]

[0101] Where x i and y i p represents the x and y positions of the i-th readout optical sensor (pixel). i is the digital signal read from the i-th optical sensor (pixel), N is the total number of optical sensors (pixels) in the optical sensor array, and P is the sum of the digital signals from all optical sensors (pixels) for a single gamma-ray interaction event.

[0102] In S635, the processor can determine the primary interacting scintillator module based on the energy-weighted average calculated for each scintillator module 205. The scintillator module 205 with the highest calculated energy-weighted average can be determined as the primary interacting scintillator module. The optical sensor (pixel) associated with the scintillator module 205 with the highest calculated energy-weighted average can be determined as the primary interacting optical sensor (pixel).

[0103] In other aspects of this disclosure, instead of determining all three features, such as the primary interaction optical sensor (pixel), the primary interaction scintillator module, and the DOI, the processor may determine only one of the three features or any combination of these features, such as at least one of the three features.

[0104] In S640, the processor can determine the timing parameters for the event (for the detection device). These timing parameters can then be used to determine the Time of Flight (TOF) between the detection devices (e.g., detection module 1 1501 and detection module 2 1502). The timing parameters can be determined based on one or more timestamps received from the readout ASIC 405. In one aspect of this disclosure, since the primary interaction optical sensor (pixel) may have already been identified, the processor can use one or more timestamps associated with that energy channel to determine the timing parameters. The processor can retrieve one or more timestamps associated with that energy channel from memory. For example, when SiPM 19 is identified as the primary interaction optical sensor (pixel), the processor can retrieve X03_T1, X03_T2, and X03_T3 from memory. These timestamps are obtained from comparator 20 (the leading edge detector). In some aspects, the processor may retrieve only X03_T1, as the primary interaction optical sensor can typically have the fastest timestamp. In one aspect of this disclosure, the processor may perform linear regression to determine the temporal parameters of an event using retrieved timestamps (e.g., X03_T1, X03_T2, and X03_T3). In other aspects of this disclosure, the processor may retrieve machine learning models to predict the time-of-flight (TOF) between conformance detector devices (e.g., detection module 1 1501 and detection module 2 1502).

[0105] Machine learning models can be based on neural networks. However, machine learning models are not limited to NNs. Other machine learning techniques, such as state vector regression, can be used. In some aspects of this disclosure, the neural network can be a convolutional neural network (CNN), which will be described later.

[0106] Using multiple timestamps can improve the resolution of CTR because it can eliminate jitter.

[0107] In other aspects, the processor can use the first few determined timestamps to determine the timing parameters. In some aspects of the invention, the foremost timestamp can be determined by combining the timestamp output from the comparator with a minimum voltage threshold via logic gates. Furthermore, timestamps can be determined in the same manner.

[0108] In other respects, time parameters can be determined before the primary interaction optical sensors (pixels) are determined.

[0109] Figure 7 A flowchart illustrating an example of training and testing a machine learning model for conversion or demultiplexing of energy channels according to aspects of this disclosure is shown. The generation of one or more machine learning models for conversion or demultiplexing of energy channels can be performed on computer 400. In other aspects, different devices can perform the generation of models for conversion or demultiplexing of energy channels, and the models are subsequently transferred to computer 400.

[0110] Different machine learning models (for demultiplexing) can be configured for different scintillator modules / optical sensor arrays. For example, a first machine learning model (for demultiplexing) can be used for a four-to-one coupling from the scintillator module to the optical sensor array, and a second machine learning model (for demultiplexing) can be used for a nine-to-one coupling from the scintillator module to the optical sensor array (and a third machine learning model for a sixteen-to-one coupling).

[0111] Different machine learning models (for demultiplexing) can be used for different scintillator modules (dimensions). For example, with the same coupling (e.g., a four-to-one coupling from the scintillator module to the optical sensor array), different ML models (for demultiplexing) can be used for scintillator modules with dimensions of 1.5 mm x 1.5 mm x 20 mm and 1.4 mm x 1.4 mm x 20 mm. To obtain a dataset for training / testing, a particle detection device comprising an array of scintillator modules, segmented light guides, and an optical sensor array (connected to a readout ASIC) can be exposed to a known particle source. Instead of multiplexing via connections to the readout ASIC as in aspects of this disclosure, the optical sensor array is connected to the readout ASIC via N connections, where N is the number of optical sensors 10 in the optical sensor array. The device can be exposed to different depths and experience multiple events. In S700, each event records a digitized signal from each channel (e.g., 64 channels). This full-channel resolution is taken as the ground truth for evaluating the model (during testing).

[0112] In S705, a multiplexed energy signal can be generated by adding a preset number of energy channels for each event. In one aspect of this disclosure, the processor, according to... Figure 1A The multiplexing scheme shown adds signals from the same optical sensor to obtain a multiplexed signal. This is to simulate the hardware multiplexing described herein. For example, the processor can add signals from four optical sensors together to reduce the number of energy channels to 16. The computer-based multiplexed signal can be stored in memory. In the S710, the processor divides the computer-based multiplexed energy signal generated for each event into a dataset for training and a dataset for testing. In some respects, 80% of the computer-based multiplexed energy signal can be used for training, and 20% can be used for testing and validation. Other partitions, such as 75% / 25% or 90% / 10%, can be used. In some respects, the partition can be randomized.

[0113] The machine learning model (used for demultiplexing) can be based on a neural network. However, the machine learning model is not limited to NNs. Other machine learning techniques, such as state vector regression, can be used. In some aspects of this disclosure, the neural network can be a convolutional neural network (CNN). Additionally, in some aspects of this disclosure, the CNN can be a shallow CNN with a U-NET architecture. Hyperparameters, including the number of convolutional layers, filters, and optimizers, can be iteratively optimized.

[0114] Figure 8 An example of a CNN with a U-NET architecture is shown.

[0115] U-Net consists of an input layer 800 with reused data (16 x 1, which can be reshaped into a 4 x 4 x 1 matrix before being fed into the CNN). Following the input layer 800 can be a series of 2D convolutions, such as... Figure 8 The 807 / 809 convolutional layers. Convolutional layers 807 and 809 can have 32 different 4 x 4 matrices (also called "filters").

[0116] Following convolutional layers 807 / 809 are max-pooling layers 811 to reduce the 2D dimension to 2 x 2, additional convolutional layers 813 / 815 each with 64 filters, and another max-pooling layer 817 to reduce the 2D dimension to 1 x 1. After being reduced to a 1 x 1 dimension, the matrix passes through several convolutional layers 819 / 821, each with 128 filters, and then undergoes a dilation path to restore it to its original 4 x 4 dimension, completing the "U" shape.

[0117] The dilation path comprises a series of upsampled convolutional layers 823 / 829, whose features are fused with corresponding layers 825 / 831 of equal dimension and convolutional layers 827 / 833 with 64 / 32 filters, respectively. The output layer 837 can be a convolutional layer with 4 filters to provide a 4 x 4 x 4 matrix, which can then be shaped to be associated with an 8 x 8 readout array. All convolutional layers in the U-Net can have 2 x 2 filters with a span of 1 and can be followed by a rectified linear unit (ReLU) activation function. Conceptually, the U-Net can be formulated to demultiplex a single 4 x 4 matrix (based on a computer-multiplexed signal) fed into the input layer into an 8 x 8 matrix (equal to the number of optical sensors in the array). Note that the shape of the input layer (the dimension of the matrix) and the number of filters in the output layer can be varied based on the readout array used. For example, the input matrix can be 16 x 1. Alternatively, a multiplexed input matrix with a smaller dimension can be used.

[0118] In S715, the model described above can be trained using the training dataset, with the training dataset input at position 800. In S720, the model can be tested using the test dataset, with the test dataset input at position 800. The optimizer can be a modified version of the Adam optimizer. The initial learning rate can be 1.0. In S725, evaluation parameters can be used to evaluate the model's performance. For example, the evaluation parameter can be the mean squared error (MSE). However, the evaluation parameter is not limited to MSE.

[0119] Once the model is validated using the evaluation parameters, it can be stored in the memory (in the computer 400) in the S730 or transferred to the computer 400 for later use.

[0120] Figure 16A flowchart illustrating an example of training and testing a machine learning model for use by a Time-of-Flight (TOF) predicting the conformation detection device (e.g., detection module 1 1501 and detection module 2 1502). At S1600, a dataset for training / testing is obtained. In some aspects of this disclosure, the dataset can be obtained experimentally using a known radiation source 1500 at a location between detection module 1 1501 and detection module 2 1502. The position of the radiation source 1500 can be controlled by a precision motor stage. In one aspect of this disclosure, the increment between positions is controlled to be less than the expected CTR. In some aspects, the increment can be less than producing a CTR of 100 ps. The range of motion of the radiation source 1500 can be 0 to 50 cm. Multiple events can be detected at each position. For example, the number of events can be 1000. In other aspects, the number of events can be 5000. In other aspects, the number of events can be 10000. In one aspect of this disclosure, the radiation source 1500 can be used for 511 keV gamma-ray absorption. At least two thresholds can be used for front-end detection.

[0121] In other respects, the dataset used for training / testing can be acquired by simulating events using parameters of the actual detection modules, including the length, width, and height of the scintillator modules, the optical response of silicon photomultipliers with various single-photon time resolutions, coupling (e.g., 4 to 1), reflector filling between scintillator modules, optical sharing segments (shapes), the known response of the scintillator modules to 511 keV gamma-ray interactions, the size of the SiPM, and the efficiency of the scintillator modules and the SiPM.

[0122] In S1605, the dataset can be divided into a training set and a testing set. In some cases, 80% of the collected dataset can be used for training, and 20% for testing and validation. Other partitions can be used, such as 75% / 25% or 90% / 10%. In some cases, the partition can be randomized. In others, a certain proportion of the dataset can be reserved for both training and validation to ensure that overfitting does not occur.

[0123] Figure 17 An example of a CNN that can be used to predict Time-of-Flight (TOF) (output TOF) according to aspects of this disclosure is shown. The input may have one input layer. The input layer may include at least two timestamps for each event from each conformation detection device (e.g., detection module 1 1501 and detection module 2 1502). Figure 17As shown, three timestamps (three thresholds) are used for each detection device. Therefore, the input layer is 1 @ 3 x 2 (three timestamps for two detection devices). The input layer is fed into convolutional layer 1702. Convolutional layer 1702 has 64 filters. The second convolutional layer 1704 has 128 filters. The filters can be 1 x 1 filters with a span of 1 and have a rectified linear unit (ReLU) activation function. The CNN can also include two fully connected (dense) layers 1706 and 1708 with ReLU activation. Dense layer 1706 has 256 filters (weights), and dense layer 1708 has 64 filters (weights). The output Toff (e.g., TOF) is output through a fully connected layer (dense layer 1710). This layer 1710 has one filter with linear activation.

[0124] The model described above can be trained using the training dataset in S1610, with the training dataset input at 1700. The model can be tested using the test dataset in S1615, with the test dataset input at 1700. Stochastic gradient descent (SGD) with momentum can be used for training optimization with an initial learning rate of 0.01. In S1620, evaluation parameters can be used to evaluate the model's performance. For example, the evaluation parameter could be the mean squared error (MSE). However, the evaluation parameter is not limited to MSE.

[0125] Once the model for predicting TOF has been validated using the evaluation parameters, it can be stored in memory (in computer 400) or transferred to computer 400 for later use in S1625.

[0126] Tests and simulations

[0127] The above multiplexing schemes and demultiplexing using one or more machine learning models for demultiplexing multiplexed energy channels were tested for four-to-one coupling of the scintillator module and the optical sensor array, as well as for nine-to-one coupling of the scintillator module and the optical sensor array.

[0128] The scintillator modules are fabricated using LYSO and coupled at one end to an 8 x 8 SiPM array (optical sensor array) and at the other end to a segmented pseudo-cylindrical light guide as described above. The four-pair-to-one coupled scintillator module array for the scintillator modules and optical sensor array consists of a 16 x 16 array of 1.4 mm x 1.4 mm x 20 mm, while the nine-pair-to-one coupled scintillator module array for the scintillator modules and optical sensor array consists of a 24 x 24 array of 0.9 mm x 0.9 mm x 20 mm.

[0129] Standard floodlight data acquisition was achieved by uniformly exposing two scintillator module arrays (and sensors) to a 3MBq Na-22 sodium point source (1mm effective diameter) placed at a distance of 5 cm (at different depths). Depth collimation data was obtained at five different depths (2, 6, 10, 14, and 18 mm) along a 20 mm scintillator module length to evaluate DOI performance using lead collimation (1mm pinhole). Data readout was accelerated using an ASIC (TOFPET2) and a FEB / D_v2 readout board (PETsys Electronics SA). Computer-based multiplexing was performed as described above to achieve 16 x 1 scintillator module-to-energy channel multiplexing for four-to-one coupling from the scintillator module to the optical sensor, and 36 x 1 scintillator module-to-energy channel multiplexing for nine-to-one coupling from the scintillator module to the optical sensor.

[0130] Computer-based multiplexing peak filtering is performed using a ±15% energy window for each scintillator module. Only events where the highest signal is more than twice the second highest signal are accepted, in order to exclude Compton scattering events with peaks.

[0131] The demultiplexing of the energy signal generated by computer-based multiplexing was performed using the aforementioned method via machine learning (CNN with U-Net architecture). 80% of the total dataset was used for U-Net training. 10% of the training dataset was reserved for training validation to ensure no overfitting occurred. A modified version of the Adam optimizer, Adadelta, was used for training optimization.

[0132] Batch sizes of 500 and 1000 epochs were used for training. The training loss was calculated by taking the average difference between the model estimates and the ground truth for all events in each epoch. The model was trained to reduce the loss between consecutive epochs until a global minimum was found. Model convergence was observed by plotting the training and validation loss curves as a function of epochs and ensuring that they reached asymptotic behavior with approximately equal minimums.

[0133] Figure 9A and 9BA qualitative comparison is shown of the actual energy signals (without multiplexing) from the outputs of each of the multiple optical sensors in a four-to-one coupling from the scintillator module to the optical sensors, and the predictions obtained from a training / testing machine learning model of the computer-based multiplexed signals using the multiplexing scheme described herein (demultiplexed). The results appear similar. For example, the comparison shows that perfect scintillator module separation is achieved in all center, edge, and corner scintillator modules with and without (per-pixel channel) computer-based multiplexing. U is on the x-axis, and V is on the y-axis.

[0134] Figure 9C An example of a synthetic dataset (computer-based multiplexed energy data) is shown, which is generated (multiplexed) by adding the outputs of four sensors in a similar manner to that described above, where full-resolution (e.g., 64) sensor outputs are read. Figure 9D An example of a multiplexed dataset is shown, generated from readings of a multiplexed energy signal from a readout ASIC 405, which is connected to a sensor array 210 via the multiplexing scheme described above. Figure 9C and Figure 9D The comparison shows that the datasets are very similar, but differ slightly due to imperfect model convergence. Figure 9C and Figure 9D The mapping in U' and V' spaces is shown, which is performed to show the channels in a square.

[0135] Figure 10A and Figure 10B A comparison is shown between the DOI resolution of a correlated particle detection system with four-to-one coupling of scintillator modules to optical sensors at five different depths (2, 6, 10, 14, and 18 mm) and the DOI resolution of a particle detection system according to aspects of this disclosure. This comparison is made for a central optical sensor in an optical sensor array and another central optical sensor in the same array. In Figure 10A, a "conventional" calculation method is used. In the conventional method, Equation 1 is calculated using the highest energy signal (Pmax based on the optical sensor or pixel), and P is calculated from the sum of each energy channel (unmultiplexed, therefore all 64 energy channel values ​​are added). Figure 10B In this method, the DOI is calculated directly using computer-based multiplexed energy signals. For example, Pmax is determined as the highest signal among 16 computer-based multiplexed energy signals, and P is determined based on the sum of the four highest signals among the 16 computer-based multiplexed energy signals.

[0136] For unreused data (Figure 10A) and reused data ( Figure 10BThe DOI estimation distributions are similar. The average DOI resolution for all measured depths is 2.32 mm full width at half maximum (FWHM) for the unreused data (Fig. 10A), and for the reused data ( Figure 10B It is 2.73 mm FWHM.

[0137] Figure 11A and 11B A qualitative comparison is shown of the actual energy signals (unmultiplexed) from the outputs of each of the multiple optical sensors in a nine-to-one coupling from the scintillator module to the optical sensors, and the predictions obtained from a computer-based multiplexed energy signal trained / tested using the multiplexing scheme described herein (demultiplexed). Excellent scintillator module separation is achieved in the center and edge scintillator modules, in the unmultiplexed data ( Figure 11A ) and reused data ( Figure 11B They have comparable performance.

[0138] Figure 12A and Figure 12B A comparison is shown between the DOI resolution of a correlated particle detection system with nine-to-one coupling of scintillator modules to optical sensors at five different depths (2, 6, 10, 14, and 18 mm) and the DOI resolution of a particle detection system according to aspects of this disclosure. This comparison is made for a central optical sensor in an optical sensor array and another central optical sensor in the same array. In Figure 12A, a "conventional" calculation method is used. In the conventional method, Equation 1 is calculated using the highest energy signal (Pmax based on the optical sensor or pixel), and P is calculated from the sum of each energy channel (unmultiplexed, therefore all 64 energy channel values ​​are added). Figure 12B In this method, the DOI is calculated directly using computer-based multiplexed energy signals. For example, Pmax is determined as the highest signal among 16 computer-based multiplexed energy signals, and P is determined based on the sum of the four highest signals among the 16 computer-based multiplexed energy signals.

[0139] For unreused data (Figure 12A) and reused data ( Figure 12B The DOI estimation distributions are similar. The average DOI resolution for all measured depths is 3.8 mm full width at half maximum (FWHM) for the unreused data (Fig. 12A), and for the reused data ( Figure 12B The thickness is 3.64 mm FWHM.

[0140] For a four-to-one coupling from the scintillator module to the optical sensor, the percentage errors of CNN predictions using the energy-weighted averaging method for the x and y coordinates are 2.05% and 2.15%, respectively, and for a nine-to-one coupling from the scintillator module to the optical sensor, the percentage errors are 2.41% and 1.97%, respectively. For a four-to-one coupling from the scintillator module to the optical sensor, the percentage error of the total detection energy per event after multiplexing data following CNN prediction is 1.53%, and for a nine-to-one coupling from the scintillator module to the optical sensor, it is 1.69%.

[0141] The tests described above demonstrate that, due to the deterministic light sharing derived from the segmented quasi-cylindrical light guide, any differences in system performance are minimized by using the multiplexing scheme described herein. Note that observed differences may stem from experimental conditions, such as the use of a 3MBq Na-22 sodium point source (1 mm effective diameter). Multiplexing results in data output from the optical sensor array to the readout ASIC and the connection. Minimizing the data file size is particularly critical because the field displacement to the DOI PET, depending on the readout scheme and DOI resolution (which determines the number of DOI bins), can increase the effective number of response lines (LORs) by more than two orders of magnitude.

[0142] As described above, using multiple timestamps for each energy channel can improve the CTR of the system. To demonstrate this improvement, events were simulated in software for two coincidence detection modules. In this simulation, the energy channels were not multiplexed. However, due to the presence of light sharing, and because the aforementioned multiplexing does not multiplex optical sensors associated with the same quasi-cylindrical segment, the CTR (and the corresponding time in each module) should not be affected. Each detection module was simulated as a 16 x 16 LYSO array, with each scintillator module being 1.5 mm x 1.5 mm x 20 mm. There was 4-to-1 coupling. The size of each SiPM (pixel) was simulated as 3.2 mm x 3.2 mm. The segmented quasi-cylindrical light guide described herein was used in the simulation. As described herein, the segmented quasi-cylindrical light guide increases the light sharing ratio of all scintillator modules coupled to the same quasi-cylindrical segment, thus introducing a depth-encoded signal. Reflective material between the scintillator modules was also included in the simulation.

[0143] The 511 keV gamma-ray absorption was simulated as a spherical source (0.1 mm in diameter) emitting light equal to the LYSO yield (approximately 27,000 photons / MeV). Events were distributed based on the Beer-Lambert law of lutetium photoelectric absorption depth-dependent in the scintillator. Energy deposition curves as a function of time were generated for each absorption and convolved with the photoresponse of silicon photomultipliers with various single-photon time resolutions (SPTR = 10, 50, and 100 ps). The total peak energy resolution was simulated at 10%.

[0144] Timestamps are generated based on three trigger thresholds (n=5, 10, and 50 photons) corresponding to the number of photons collected on the readout side (examples of voltage thresholds described in this paper). Uniformly distributed time offsets corresponding to position offsets from (0 to 50 cm) are used. t off = 0-1667 ps (also referred to as TOF in this paper) is added to the timestamp from one of the two crystals for each coincidence pair. This is to simulate the actual movement of the radiation source between the coincidence detection modules.

[0145] Although the ground truth DOI was known in the simulations, the DOI parameters used for CNN training and testing were calculated separately using an energy-weighted averaging method. Three separate simulations were performed: two coincidence center crystals, two coincidence edge crystals, and two coincidence corner crystals, to independently characterize the temporal performance in these three regions. 60,000 events were simulated in each of the three cases, with a total of 30,000 coincidence pairs per simulation.

[0146] Batch sizes of 20 and 100 epochs were used for training. The ground truth for each conformation pair in the test dataset was calculated against the CNN output. t off The difference in values ​​is used to calculate the standard deviation of the error distribution characterizing the CTR of the CNN. The CNN performance accuracy is characterized as follows: each training case is run 10 times, with data regrouped and redistributed between the training and test datasets, and the mean and standard deviation of the CTR values ​​are calculated for each case.

[0147] Using the above CNN ( Figure 17 The Time-of-Flight (TOF) prediction was performed. Six different input layers were used: 1 timestamp, 2 timestamps, 1 timestamp with DOI correction, 2 timestamps with DOI correction, 3 timestamps, and 3 timestamps with DOI correction. Figure 18 A table showing the results is provided. The table is sorted by SPTR 100, 50, and 10. For each SPTR, six different input layers are shown. It can be seen that the CTR is improved when two or three timestamps are used instead of one. For example, for an SPTR of 100, when one timestamp is used for the SiPM associated with the central quasi-pillar 162, the CTR is 195 (SD 2.3), however, when two timestamps are used, the CTR is 136 (SD 1.1), and when three timestamps are used, the CTR is even lower at 124 (SD 1.5). Other quasi-pillar designs (corners and edges) show similar improvements. This improvement is even more pronounced when DOI correction (depth encoding is "yes") is used. Figure 18As you can see, the best result (highlighted) is three timestamps with depth encoding as "yes". Figure 18 A comparison with the "classic CTR" is shown. The classic CTR does not use any machine learning; instead, it uses the detected timestamp difference to calculate the Time of Rendering (TOF) (without deep encoding). With deep encoding, a linear regression between TOF and DOI is performed to determine the contribution of the DOI to the TOF, and this contribution is subtracted to obtain an accurate TOF estimate.

[0148] The terms "segment" and "quasi-cylindrical segment" are used interchangeably in this document. The terms "segmented optical guide," "quasi-cylindrical optical guide," and "segmented quasi-cylindrical optical guide" are also used interchangeably in this document.

[0149] As used herein, terms such as “a,” “an,” and “the” are not intended to refer to a single entity, but rather to include its general category, which can be illustrated by specific examples.

[0150] As used herein, terms qualified in the singular are intended to include those qualified in the plural, and vice versa.

[0151] References to “one aspect,” “some aspects,” “a few aspects,” or “one aspect” in the specification indicate that the described aspects(s) may include a particular feature or characteristic, but each aspect may not necessarily include a particular feature, structure, or characteristic. Furthermore, such phrases do not necessarily refer to the same aspect. Additionally, when a particular feature, structure, or characteristic is described in conjunction with one aspect, it is considered that the combination of other aspects affects such feature, structure, or characteristic, whether or not explicitly described, and this is within the knowledge of those skilled in the art. For the purposes described below, the terms “upper,” “lower,” “right,” “left,” “vertical,” “horizontal,” “top,” “bottom,” and their derivatives should refer to means relative to a floor and / or oriented as it is in the figures.

[0152] Any range of values ​​mentioned in this document explicitly includes every value contained within that range (including fractions and integers). For illustrative purposes, the range of “at least 50” or “at least approximately 50” mentioned in this document includes integers 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, etc., and fractions 50.1, 50.2, 50.3, 50.4, 50.5, 50.6, 50.7, 50.8, 50.9, etc. In further illustrative terms, the range of “less than 50” or “less than approximately 50” mentioned in this document includes integers 49, 48, 47, 46, 45, 44, 43, 42, 41, 40, etc., and fractions 49.9, 49.8, 49.7, 49.6, 49.5, 49.4, 49.3, 49.2, 49.1, 49.0, etc.

[0153] As used herein, the term "processor" can include a single-core processor, a multi-core processor, multiple processors located in a single device, or multiple processors that are wired or wirelessly connected to each other and distributed across a network of devices, the Internet, or the cloud. Therefore, as used herein, a function, feature, or instruction performed or configured to be performed by a "processor" can include a function, feature, or instruction performed by a single-core processor, a function, feature, or instruction performed jointly or collaboratively by multiple cores of a multi-core processor, or a function, feature, or instruction performed jointly or collaboratively by multiple processors, wherein each processor or core does not need to perform each function, feature, or instruction individually. For example, a single FPGA or multiple FPGAs can be used to implement the functions, features, or instructions described herein. For example, multiple processors can allow for load balancing. In another example, a server (also referred to as a remote or cloud) processor can perform some or all of the functionality on behalf of a client processor. The term "processor" also includes one or more ASICs as described herein.

[0154] As used herein, the term “processor” may be replaced by the term “circuit”. The term “processor” may refer to processor hardware (shared, dedicated, or grouped) that executes code and memory hardware (shared, dedicated, or grouped) that stores the code executed by the processor, or a portion thereof.

[0155] Furthermore, in some aspects of this disclosure, a non-transitory computer-readable storage medium includes electronically readable control information stored thereon, which is configured to enable the functional aspects described herein to be implemented when the storage medium is used in a processor.

[0156] Furthermore, any of the foregoing methods can be embodied in the form of a program. This program can be stored on a non-transitory computer-readable medium and, when run on a computer device (including a processor), is adapted to perform any of the foregoing methods. Therefore, a non-transitory tangible computer-readable medium is suitable for storing information and for interacting with a data processing facility or computer device to perform the program of any of the foregoing embodiments and / or the method of any of the foregoing embodiments.

[0157] A computer-readable medium or storage medium can be an internal medium housed within the main body of a computer device, or a removable medium arranged such that it can be separated from the main body of the computer device. As used herein, the term computer-readable medium does not include transient electrical or electromagnetic signals propagated through a medium (such as on a carrier wave); therefore, the term computer-readable medium is considered tangible and non-transitory. Non-limiting examples of non-transitory computer-readable media include, but are not limited to, rewritable non-volatile memory devices (including, for example, flash memory devices, erasable programmable read-only memory devices, or mask read-only memory devices); volatile memory devices (including, for example, static random access memory devices or dynamic random access memory devices); magnetic storage media (including, for example, analog or digital magnetic tape or hard disk drives); and optical storage media (including, for example, CDs, DVDs, or Blu-ray discs). Examples of media having built-in rewritable non-volatile memory include, but are not limited to, memory cards; and media having built-in ROM, including, but not limited to, ROM cartridges; and so on. Furthermore, various information about the stored image (e.g., attribute information) may be stored in any other form or may be provided in other ways.

[0158] The term memory hardware is a subset of the term computer-readable media.

[0159] The aspects and examples described in this disclosure are intended to be illustrative and not limiting, and are not intended to represent every aspect or example of this disclosure. While the essential novel features of this disclosure as applied to various specific aspects have been shown, described, and pointed out, it will also be understood that various omissions, substitutions, and changes can be made to the form and details of the illustrated devices and their operation without departing from the spirit of this disclosure. For example, all combinations of those elements and / or method steps that are explicitly intended to perform substantially the same function in substantially the same manner to achieve the same result are within the scope of this disclosure. Furthermore, it should be recognized that structures and / or elements and / or method steps shown and / or described in conjunction with any disclosed form or aspect of this disclosure can be incorporated as a general issue of design selection into any other disclosed or described or suggested form or aspect. In addition, various modifications and variations can be made without departing from the spirit or scope of this disclosure as set forth literally in the appended claims and in their legally recognized equivalents.

Claims

1. A particle detection system, comprising: An optical sensor array comprising a plurality of optical sensors, each optical sensor in the array corresponding to a pixel; A scintillator array comprising a second plurality of scintillator modules, the second plurality of scintillator modules being larger than a first plurality of optical sensors, wherein a plurality of scintillator modules are in contact with a corresponding optical sensor at a first end of the respective scintillator module; and The segmented light guide includes multiple pseudo-cylindrical segments. The segmented light guide contacts the second end of a plurality of second scintillator modules. Each pseudo-cylindrical segment contacts a scintillator module that is in contact with at least two different optical sensors, said at least two different optical sensors being adjacent optical sensors. Each pseudo-cylindrical segment is configured to redirect particles between scintillator modules in contact with the corresponding pseudo-cylindrical segment. There are a third or more energy readout channels, in which multiple optical sensors are connected to energy readout channels respectively, such that optical sensors associated with the same quasi-cylindrical segment are not connected to the same energy readout channel, and each energy readout channel has at least two timestamps associated with it. The particle detection system also includes at least two comparators for each energy readout channel, the at least two comparators being connected to multiple optical sensors of the same energy readout channel, each of the at least two comparators having a different threshold, and wherein a timestamp is the time at which the corresponding comparator outputs a change based on a comparison with the corresponding threshold.

2. The particle detection system of claim 1, wherein, The at least two timestamps are three timestamps.

3. The particle detection system according to claim 1, wherein, The at least two comparators are connected to each corresponding anode of the plurality of optical sensors.

4. The particle detection system according to claim 1, wherein, The at least two comparators are connected to each corresponding cathode of the plurality of optical sensors.

5. The particle detection system according to claim 1, wherein, The third plurality of energy readout channels and the at least two comparators are connected to different terminals of the optical sensor.

6. The particle detection system according to any one of claims 1 to 5, wherein, The number of optical sensors connected to the same energy readout channel is four.

7. The particle detection system according to any one of claims 1 to 5, wherein, There is a four-to-one coupling from the scintillator module to the optical sensor.

8. The particle detection system according to any one of claims 1 to 5, wherein, There is a nine-to-one coupling from the scintillator module to the optical sensor.

9. The particle detection system according to any one of claims 1 to 5, further comprising a first processor configured to bias a first plurality of optical sensors during readout and to receive outputs via a third plurality of energy readout channels and at least two timestamps associated with each energy readout channel.

10. The particle detection system according to claim 9, further comprising a second processor communicating with the first processor, wherein, The second processor is configured to determine the timing parameters of the event based on at least two received timestamps.

11. The particle detection system according to claim 10, wherein, The time parameter is based on a combination of the at least two timestamps.

12. The particle detection system according to claim 10, wherein, The second processor is configured to determine the time of flight (TOF) between coincidence detection modules based on time parameters.

13. The particle detection system according to any one of claims 10 to 12, wherein, The second processor is also configured to determine at least one of the following: the primary interacting pixel of the event, the primary interacting scintillator module, or the interaction depth.

14. The particle detection system according to claim 13, wherein, The second processor is configured to select at least two timestamps associated with the determined primary interacting pixels to determine the timing parameters.

15. The particle detection system according to claim 12, wherein, The second processor is configured to use a machine learning model to determine the Time of Frame (TOF), the machine learning model being input to at least two timestamps received from the conformity detection module.

16. The particle detection system according to claim 10 or claim 11, wherein, The time parameter is based at least on the fastest timestamp.

17. The particle detection system according to claim 10 or claim 11, wherein, The time parameter is based on linear regression analysis using at least two received timestamps.

18. The particle detection system according to any one of claims 1 to 5, 10 to 12, and 14 to 15, wherein, The first plurality of optical sensors are arranged in rows and columns, wherein adjacent optical sensors in a row are connected to different energy readout channels, and adjacent optical sensors in a column are connected to different energy readout channels.

Images