Deep learning-based interleaving multiplexing method and PET detector

Abstract

The invention provides an interleaving multiplexing method based on deep learning and a PET detector. The interleaving multiplexing method comprises the steps that S1, all signals in an SiPM array are obtained and evenly distributed to corresponding row signal lines and column signal lines in the SiPM array, the SiPM array is an A * B array, and A and B are integers larger than or equal to 2; s2, marking A * B paths of signals without channel multiplexing as real labels, and calculating according to the A * B paths of signals to obtain synthesized A + B paths of signals as training samples of a deep learning model; s3, on each row signal line, adding a plurality of signals which are arranged in a staggered manner to form a row signal; on each column signal line, adding the remaining multiple signals which are arranged in a staggered manner to form a path of column signal; adding the A-path row signals and the B-path column signals to form A + B-path signals; s4, decoding the A + B paths of signals by using the trained deep learning model to obtain predicted A * B paths of signals; and S5, calculating the predicted A * B paths of signals to obtain a decoded image.

Description

Technical Field

[0001] The present invention relates to an interleaved multiplexing method and a PET detector based on deep learning, belonging to the technical field of nuclear medicine. Background Art

[0002] PET (Positron Emission Tomography) systems with Time of Flight (TOF) and Depth of Interaction (DOI) functions provide good position and time performance, but they require a large number of readout channels, which increases the complexity, power consumption, and cost of the entire PET system.

[0003] For PET systems that require high DOI depth resolution and high TOF time resolution, such as dedicated human brain PET systems, a better electronics readout system is particularly important. As Figure 1 shown, the dedicated human brain DOI-TOF-PET detector uses two layers of crystals. The top layer of crystals is smaller to provide high spatial resolution. The bottom layer of crystals is larger, and the crystal bars correspond one-to-one with the silicon photomultiplier (SiPM) channels to improve the TOF time resolution. The light guide between the two layers introduces a light distribution rich in depth information, enabling the bottom layer of crystals to have the ability to identify DOI. Each SiPM channel uses a dedicated PET chip PETsys TOFPET2ASIC as the electronics for readout. In a high-resolution TOF-DOI PET system, there are thousands of signal channels that need to be read out by electronics, which makes the PET system engineering highly complex, consumes a lot of power, and also requires the design of a complex cooling system, which further increases the already high cost of the PET system.

[0004] Although the number of channels can be reduced through a multiplexing circuit, the existing multiplexing circuit has too many resistors in the signal path, and the signal-to-noise ratio of the multiplexed signal is too poor, which will seriously affect the quality of the crystal resolution map, energy resolution, DOI resolution, and TOF time resolution after multiplexing.

[0005] There are many electronics channels in the detectors of nuclear medicine imaging instruments such as Single Photon Emission Computed Tomography (SPECT), Compton Camera, and Autoradiography, and the complexity, power consumption, and cost of the system are all relatively high.

[0006] Therefore, a better multiplexing method is needed, which not only has a high signal-to-noise ratio but also can reduce complexity, power consumption, and cost while maintaining the performance of the instrument system. Summary of the Invention

[0007] The object of the present invention is to provide an interleaved multiplexing method based on deep learning. After using this interleaved multiplexing method, the TOF, DOI, and spatial resolution performance of the PET detector will basically not decline.

[0008] To achieve the above object, the present invention provides an interleaved multiplexing method based on deep learning, which is applied to the SiPM array of the PET detector and includes the following steps:

[0009] S1. Obtain each signal in the SiPM array of the PET detector, and evenly distribute each signal to the corresponding row signal line and column signal line in the SiPM array. The SiPM array is an A×B array, where both A and B are integers greater than or equal to 2;

[0010] S2. Mark the A×B channels of signals without channel multiplexing as the true labels, and at the same time calculate the synthesized A + B channels of signals according to the A×B channels of signals as the training samples of the deep learning model;

[0011] S3. On each row signal line, add multiple mutually interleaved signals to form one row signal; on each column signal line, add the remaining interleaved multiple signals to form one column signal; add the A row signals and the B column signals to form A + B channels of signals;

[0012] S4. Use the trained deep learning model to decode the A + B channels of signals obtained in step S3 to obtain the predicted A×B channels of signals;

[0013] S5. Calculate the predicted A×B channels of signals obtained in step S4 to obtain a decoded map.

[0014] As a further improvement of the present invention, in step S3, before each signal is added to other mutually interleaved signals, it is first filtered through a filtering structure.

[0015] As a further improvement of the present invention, the filtering structure is a diode or a resistor or a capacitor.

[0016] As a further improvement of the present invention, in step S2, the synthesized A + B channels of signals are calculated using the mathematical model m = M×s + n, where m represents the vector of the A + B channels of signals, s represents the vector of the complete SiPM pixel values of all signals in the SiPM array, n represents the vector of minute noises in each signal, and the matrix M represents the multiplexing network that maps all signals to A + B encoded readout channels.

[0017] As a further improvement of the present invention, the multiplexing ratio of the SiPM array is

[0018] As a further improvement of the present invention, in step S5, the centroid algorithm is used to calculate the predicted A×B path signals obtained in step S4.

[0019] As a further improvement of the present invention, the calculation formula of the centroid algorithm is: where i1, i2,... i A ... i B respectively represent the corresponding row signals and column signals.

[0020] As a further improvement of the present invention, the deep learning model in step S2 includes, but is not limited to, a convolutional neural network model. After the input charge integration, the convolutional neural network model undergoes two downsamplings and two upsamplings in sequence, and finally obtains the charge integration of A×B paths.

[0021] The present invention also aims to provide a PET detector applying the above interleaved multiplexing method, and the TOF, DOI, and spatial resolution performance of the PET detector will basically not decrease.

[0022] To achieve the above object, the present invention provides a PET detector applying the aforementioned interleaved multiplexing method based on deep learning.

[0023] As a further improvement of the present invention, the PET detector is formed by coupling a scintillation crystal array one-to-one to an SiPM array, and a layer of light guide sheet is coupled to the incident surface of the PET detector.

[0024] The beneficial effects of the present invention are as follows: By adding the multiple signals arranged alternately on each row signal line to form a row signal, and adding the multiple signals arranged alternately on each column signal line to form a column signal, the interleaved multiplexing method of the present invention not only significantly reduces the number of channels that need to be processed by the electronics, but also the spatial resolution, depth resolution, and time resolution performance of the PET or SPECT detector and instrument will basically not decrease, greatly reducing the number of electronics, and reducing the complexity, power consumption, and heat generation of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 is a schematic structural diagram of a conventional DOI-TOF-PET detector dedicated for the human brain.

[0026] Figure 2 is a schematic structural diagram of the interleaved multiplexing circuit of the present invention.

[0027] Figure 3 is Figure 2 a schematic structural diagram after adding a Schottky diode on the basis of

[0028] Figure 4 is the decoding graph calculated by the traditional multiplexing circuit using the centroid algorithm.

[0029] Figure 5 is the decoding graph calculated by the interleaved multiplexing circuit of the present invention using the centroid algorithm.

[0030] Figure 6 is the average depth DOI resolution before and after using the interleaved multiplexing circuit.

[0031] Figure 7 is the average time resolution without ICS events before and after using the interleaved multiplexing circuit.

[0032] Figure 8 is the structural diagram of the convolutional neural network model of the present invention. Detailed implementation manners

[0033] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] The present invention discloses a PET detector and an interleaved multiplexing method based on deep learning. The interleaved multiplexing method based on deep learning is mainly applied to a novel interleaved multiplexing circuit of a PET detector. After using the interleaved multiplexing circuit and the interleaved multiplexing method, the TOF time resolution, DOI depth resolution and spatial resolution performance of the PET detector basically do not decrease.

[0035] In this embodiment, the PET detector is preferably formed by coupling a scintillator crystal array of 3.0×3.0×15 mm 3 one-to-one to an SiPM array, and a layer of light guide sheet is coupled to the incident surface of the PET detector. The scintillator crystal is an inorganic scintillator crystal, and its material can be BGO (bismuth germanate), LSO (lutetium silicate), LaBr3 (lanthanum bromide), LYSO (lutetium yttrium silicate), YSO (yttrium silicate), GAGG (gadolinium aluminum gallate), BaF2 (barium fluoride), NaI (sodium iodide), CeI (cesium iodide). It should be noted that: the specific size of the scintillator crystal array here is only exemplary and can be adjusted according to actual situations without limitation.

[0036] The interleaved multiplexing circuit includes a circuit board and an SiPM array integrated on the circuit board. The SiPM array is an A×B array, where both A and B are integers greater than or equal to 2.

[0037] The interleaved multiplexing method based on deep learning of the present invention includes the following steps:

[0038] S1. Obtain each signal in the SiPM array of the PET detector, and evenly distribute each signal to the corresponding row signal line and column signal line in the SiPM array;

[0039] S2. Mark the A×B channel signals without channel multiplexing as real labels, and at the same time calculate the synthesized A + B channel signals based on the A×B channel signals as the training samples of the deep learning model;

[0040] S3. On each row signal line, add multiple signals arranged alternately to form one row signal; on each column signal line, add the remaining alternately arranged multiple signals to form one column signal; add the A row signals and the B column signals to form A + B channel signals;

[0041] S4. Use the trained deep learning model to decode the A + B channel signals obtained in step S3 to obtain the predicted A×B channel signals;

[0042] S5. Calculate the predicted A×B channel signals obtained in step S4 to obtain a flood histogram.

[0043] Optionally, in step S3, before each signal is added to other alternately arranged signals, it can also be filtered through a filtering structure. In this embodiment, the filtering structure is a Schottky diode, as Figure 3 shown. Of course, in other embodiments, the filtering structure can also be other types of diodes; or, the filtering structure can also be a resistor or a capacitor; or, the filtering structure is a diode and a resistor; or, the filtering structure is a diode and a capacitor; as long as the same filtering effect can be achieved, there is no limitation here. In addition, the number of diodes / resistors / capacitors connected to each signal can be 1, 2, or more than 2, and there is no limitation here.

[0044] In step S2, the synthesized A + B channel signals are calculated through software using the mathematical model m = M×s + n, where m represents the vector of the A + B channel signals, s represents the vector of the complete SiPM pixel values of all signals in the SiPM array, n represents the vector of minute noises in each signal, and the matrix M represents the multiplexing network that maps all signals to A + B encoded readout channels. The multiplexing ratio of the SiPM array is

[0045] In step S2, the deep learning model is preferably a convolutional neural network model. After the charge integral (QDC) is input into the convolutional neural network model, it undergoes two downsamplings and two upsamplings in sequence, and finally obtains A×B channels of charge integral. Of course, in other embodiments, other deep learning models can also be selected, such as Generative Adversarial Networks (GAN) and Diffusion Models, as long as they can achieve the same function, and there is no limitation here.

[0046] In step S5, the centroid algorithm is used to calculate the predicted A×B channel signals obtained in step S4 to obtain a decoded map. Optionally, the calculation formula of the centroid algorithm is: where i1, i2,... i A ... i B respectively represent the corresponding row signals and column signals. Of course, in other embodiments, other parsing algorithms can also be selected for calculation, and there is no limitation here.

[0047] In this embodiment, the decoded map is directly translated from flood histogram; of course, it can also be translated into position spectrum, crystal resolution map, crystal decoded map, position decoded map, panoramic field image, etc., and there is no limitation here.

[0048] Next, specific embodiments and appendices will be combined Figure 2-8 to specifically describe the technical solution of the present invention.

[0049] As Figure 2 shown, in this embodiment, the SiPM array is an 8×8 array, that is, both A and B are 8. At this time, step S1 is specifically: obtaining 64 signals in the SiPM array of the PET detector, and equally distributing the 64 signals to the corresponding row signal lines and column signal lines in the SiPM array.

[0050] Step S2 is specifically: collecting 64 ordinary charge integral (QDC) signals through a non-channel multiplexing circuit board, and marking the 64 QDC signals as real labels. At the same time, using the mathematical model m = M×s + n to perform software calculation on the 64 QDC signals to obtain 16 synthesized QDC signals as the training samples of the Convolutional Neural Network (CNN) model. Where m represents the 16×1 vector of the encoded readout signal, s represents the 64×1 vector of the complete SiPM pixel values of all QDC signals, n represents the 16×1 vector of the tiny noise in each QDC signal, and the 16×64 matrix M represents the multiplexing network that maps all 64 pixels to 16 encoded readout channels:

[0051]

[0052] In each row, four positions are set to 1, and the other positions are set to 0. Specifically, the column numbers of 1 in the first 8 rows correspond to Figure 2 The channel number of the Mul A-Mul H connection in the last 8 rows corresponds to the column number 1 Figure 2 The channel number of the Mul 1-Mul 8 connection.

[0053] Step S3 is as follows: on each row signal line, four signals arranged in an interlaced manner are selected and added (e.g. Figure 2 As shown), a row signal is formed. Since there are 8 row signals in total, the 8 row signals Mul 1-Mul 8 are added together. On each column signal line, the remaining 4 staggered signals are selected and added together (as shown Figure 2 As shown in FIG. 1 , a column signal is formed. Since there are 8 column signals in total, they are added to form 8 column signals Mul A-Mul H. Thus, a total of 8×8 signals are multiplexed to 8+8 signals. This interleaved multiplexing method greatly reduces the amount of resistors used and reduces the number of channels to 1 / 4 of the original.

[0054] The interleaved multiplexing method has three major advantages: 1. Each signal only needs to be added once, instead of dividing each signal into two equal parts like the traditional row and column addition multiplexing circuit. The signal amplitude does not need to be reduced to half of the original, which improves the signal-to-noise ratio; 2. There is no need to use two resistors or capacitors to divide the signal into row and column signals, which avoids the increase of RC parameters in the signal path, reduces the RC low-pass filtering effect of high-speed pulse signals, and improves the time resolution; 3. The single row (column) signal is changed from the original 8-way signal addition to the current 4-way noise addition. Correspondingly, the original 8-way noise is also changed to the current 4-way noise addition, which increases the signal-to-noise ratio and greatly avoids the loss of TOF time, DOI and spatial information.

[0055] Furthermore, if Figure 3 As shown, the present invention Figure 2 In the interleaved multiplexing circuit, before the interleaved signals are added, they are first filtered through a diode. Since the diode is unidirectional, a fast Schottky diode with the lowest conduction voltage (such as 200mV) can be selected. The Schottky diode can cut off 200mV of the lowest part of each high-speed pulse signal. Since a large part of the cut high-speed pulse signal is noise, the signal-to-noise ratio of the remaining high-speed pulse signal can be further improved.

[0056] Step S4 is specifically as follows: Collect 16 actual QDC signals through plugging in an 8 + 8 channel interleaved multiplexing circuit, and use the trained convolutional neural network model to decode these 16 QDC signals to obtain 64 predicted QDC signals.

[0057] Step S5 is specifically as follows: Use the centroid algorithm to calculate the 64 predicted QDC signals obtained in Step S4 to obtain Figure 5 the decoded graph shown. The calculation formula of the centroid algorithm is:

[0058] i1, i2,... i8 = MuL1, Mul2,... Mul8, i9, i 10 ,

[0059] …i 16 = MuL A , Mul B ,... Mul H .

[0060] To verify the performance of the aforementioned trained convolutional neural network model, after inserting the interleaved multiplexing circuit, without decoding through the convolutional neural network model, directly use the centroid algorithm for calculation, and the obtained decoded graph is as Figure 4 shown. Comparing Figure 4 and Figure 5 , it can be seen that: Figure 4 each crystal bar in Figure 5 is similar to garbled code, and each crystal bar in

[0061] can be clearly distinguished. Figure 6 In addition, the average DOI depth resolution of the PET detector is also calculated through the 64 predicted QDC signals. As shown in Figure 6 b, after using the interleaved multiplexing circuit, the DOI depth resolution is 3.8 mm; in contrast,

[0062] in Figure 7 a, when the interleaved multiplexing circuit is not used at all and 64 electronics channels are directly used to read out the signals of the PET detector, including all inter-crystal scattering (ICS) events, the calculated average time resolution is 250 ps. As shown in Figure 7As shown in Fig. b, when using the interleaved multiplexing circuit of the present invention, only 16 electronic channels are used to read the signals of the PET detector. After removing all ICS events, the calculated average time resolution deteriorates from 237 ps to 268 ps. It can be seen that the time resolution only deteriorates by ~30 ps.

[0063] As Figure 8 shown, the convolutional neural network model structure used in this embodiment is based on the traditional U-net model. The input data is the 4×4 charge integration (QDC) of 16-channel data, which passes through two downsamplings and two upsamplings in sequence to obtain the 8×8 charge integration.

[0064] In summary, 1. The present invention adopts an interleaved multiplexing circuit different from the conventional multiplexing circuit. Although the multiplexing ratio of both circuits is 4:1, since only 4 out of 8 signals in each row (column) of the interleaved multiplexing circuit are connected, reducing the input of half of the signals, the conventional decoding algorithm will cause the decoding graph to be unable to be decoded or decoded incorrectly. However, with the cooperation of the basic deep learning algorithm, perfect decoding of the decoding graph can be achieved. 2. Through the training and learning of the deep learning algorithm, the present invention can recover the QDC information of the (64) channels before multiplexing. This QDC information can not only decode the decoding graph through the centroid algorithm, but also calculate the accurate depth (DOI) information of each crystal bar. 3. By using the interleaved multiplexing circuit with Schottky diodes added, the pulse signal of each channel is reduced by about 200 mv by the Schottky diode, further reducing the noise in the pulse signal and obtaining a high TOF time resolution, which is only ~30 ps worse than the time resolution of the PET detector without using the multiplexing circuit at all.

[0065] In conclusion, by introducing the interleaved multiplexing circuit and the interleaved multiplexing circuit with diodes into the design of PET or SPECT detectors, the present invention not only greatly reduces the number of channels that the electronics need to process, but also keeps the spatial resolution and depth resolution of the PET or SPECT detector and instrument completely unchanged, and the time resolution only deteriorates slightly, with almost no performance degradation, greatly reducing the number of electronics and reducing the complexity, power consumption and heat generation of the system.

[0066] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep learning-based interleaving multiplexing method applied to a SiPM array of a PET detector, characterized in that: The steps include: S1. Acquire each signal in the SiPM array of the PET detector, and evenly distribute each signal to the corresponding row signal line and column signal line in the SiPM array, wherein the SiPM array is an A×B array, wherein A and B are both integers greater than or equal to 2; S2, marking the A×B signal without channel multiplexing as the true label, and calculating the synthesized A+B signal based on the A×B signal as the training sample of the deep learning model; S3, on each row signal line, add the multiple signals arranged in an interlaced manner to form a row signal; on each column signal line, add the remaining multiple signals arranged in an interlaced manner to form a column signal; add the A row signal and the B column signal to form an A+B signal; S4, using the trained deep learning model to decode the A+B signal obtained in step S3 to obtain a predicted A×B signal; S5. Calculate the predicted A×B path signals obtained in step S4 to obtain a decoding graph.

2. The interleaving multiplexing method according to claim 1, characterized in that: In step S3, each signal is filtered by a filtering structure before being added with other signals arranged in an interlaced manner.

3. The interleaving multiplexing method according to claim 2, characterized in that: The filtering structure is a diode, a resistor or a capacitor.

4. The interleaving multiplexing method according to claim 1, characterized in that: In step S2, the mathematical model m=M×s+n is used to calculate the synthesized A+B signal, where m represents the vector of the A+B signal, s represents the vector of the complete SiPM pixel values ​​of all signals in the SiPM array, n represents the vector of the tiny noise in each signal, and the matrix M represents the multiplexing network that maps all signals to A+B encoding readout channels.

5. The interleaving multiplexing method according to claim 1, characterized in that: The multiplexing ratio of the SiPM array is 6. The interleaving multiplexing method according to claim 1, characterized in that: In step S5, the predicted A×B path signals obtained in step S4 are calculated using a centroid algorithm.

7. The interleaving multiplexing method according to claim 6, characterized in that: The calculation formula of the centroid algorithm is: Among them, i1, i2, ...i A …i B Represent the corresponding row signal and column signal respectively.

8. The interleaving multiplexing method according to claim 1, characterized in that: The deep learning model in step S2 includes a convolutional neural network model. After the charge integral is input, the convolutional neural network model undergoes two downsamplings and two upsamplings in sequence to finally obtain the charge integral of the A×B path.

9. A PET detector, characterized in that: The PET detector applies the deep learning-based interleaving multiplexing method described in any one of claims 1-8.

10. The PET detector according to claim 9, characterized in that: The PET detector is formed by coupling a scintillation crystal array to a SiPM array in a one-to-one ratio, and a light guide sheet is coupled to the incident surface of the PET detector.

Images