DNN-based multiplexing PET detector event positioning method and system, and medium

Abstract

The invention provides a DNN-based multiplexing PET detector event positioning method and system and a medium, and the method comprises the following steps: simulating and obtaining the spatial characteristics and statistical characteristics of inter-crystal scattering events in a PET detector, and dividing CP events into adjacent CP events and long-distance CP events; generating a training data set in combination with an optical simulation and analysis method, wherein the training data set comprises optical distribution data of CP events and optical distribution data of P events; constructing a multiplexing DNN model, training the multiplexing DNN model by using the data set, inputting optical distribution data to be processed into the trained multiplexing DNN model, and completing event classification; according to different event types, a first interaction crystal is selected, and high-precision event positioning is realized through crystal scattering identification.

Description

Technical Field

[0001] This invention relates to the field of medical imaging equipment, and in particular to a method, system and medium for event localization of a DNN-based multiplexed PET detector. Background Technology

[0002] Positron emission tomography (PET) is a key technology in molecular imaging, widely used in fields such as tumor diagnosis, cardiac disease assessment, and nervous system research. A core prerequisite for achieving high-resolution PET imaging is the accurate determination of the three-dimensional interaction positions of annihilated photons within a scintillation crystal array. This includes not only localization in the lateral plane but also the acquisition of depth-direction interaction (DOI) information. In PET detector design, commonly used crystal configurations include monolithic crystals and pixel arrays. Monolithic crystal positioning relies on the propagation characteristics of light on the photodetector and extensive calibration of the light field, offering high cost-effectiveness. However, its spatial resolution has inherent limitations—high levels can only be achieved in relatively thin (≤10mm) scintillators, thus requiring a trade-off between resolution and sensitivity. In contrast, the spatial resolution of pixel arrays is primarily determined by the lateral dimensions of the crystal, while sensitivity is closely related to crystal thickness. Positioning accuracy is also affected by the size ratio of the crystal to the photodetector. In high-resolution applications (such as small animal PET), 1:1 coupling (i.e., each crystal corresponds to a separate photodetector) is difficult to implement due to space and cost constraints, hindering the achievement of fine pixelation and tightly packed crystal arrays. Therefore, a light-sharing scheme is often used in practice, where a group of crystals is coupled to the same photodetector via an optical guide, and event positioning is achieved based on the centroid calculation of the light intensity distribution.

[0003] However, the light-sharing scheme still faces two common challenges: First, cells located at the edge of the crystal array often exhibit performance degradation due to insufficient light sharing between adjacent photodetectors. Specifically, in light-sharing PET detectors, the Anger logic-based localization algorithm relies on centroid calculation, exhibiting significant edge effects. When annihilated photons scatter within the crystal array, traditional methods cannot accurately identify the first interacting crystal, causing localization errors to propagate along the reconstruction chain and ultimately reducing the system's spatial resolution. Second, this scheme struggles to effectively identify and correct inter-crystal scattering (ICS) events occurring within the crystal array. Existing ICS event handling methods (such as maximum likelihood estimation and convex constraint optimization) offer only limited improvements and often come at the cost of sacrificing system sensitivity (e.g., directly eliminating scattering events). Furthermore, the scattering direction assumption based on the Klein-Nishina formula fails in small-sized crystals, with the measured forward to backscatter ratio approaching 1, making it impossible to simply determine the scattering order through energy deposition. Therefore, in such events, annihilated photons interact multiple times between multiple crystals, easily causing localization errors and thus reducing the overall spatial resolution of the system.

[0004] With the rapid development of deep learning technology, deep neural networks have demonstrated outstanding feature extraction and pattern recognition capabilities in various fields such as image recognition and signal processing. Introducing such methods into PET detector localization and intercrystalline scattering (ICS) identification tasks is expected to significantly improve the accuracy of event localization and scattering identification. However, research in this area is still in its early stages, and the technology is not yet mature. There is an urgent need to build an efficient and reliable deep learning system to realize its practical applications. Summary of the Invention

[0005] To overcome the aforementioned technical deficiencies, the present invention aims to provide a DNN-based multiplexed PET detector event localization method, which achieves accuracy and efficiency in localizing different events through deep neural networks.

[0006] To achieve the above objectives, this invention discloses a DNN-based multiplexed PET detector event localization method, comprising the following steps: The spatial and statistical characteristics of intercrystalline scattering events in a PET detector were simulated and obtained, and CP events were divided into adjacent CP events and distant CP events. A training dataset is generated by combining optical simulation and analytical methods. The training dataset includes optical distribution data of CP events and optical distribution data of P events. Construct a multiplexed DNN model and train the multiplexed DNN model using the dataset. Input the optical distribution data to be processed into the trained multiplexed DNN model to complete event classification. Based on different event types, the first interacting crystal is selected, and high-precision event localization is achieved through crystal scattering recognition.

[0007] Preferably, the simulation acquires the spatial and statistical characteristics of intercrystalline scattering events in the PET detector, classifying CP events into adjacent CP events and distant CP events, including the following steps: Using GATE to simulate the LYSO PET detector system, multiple events were generated, and the number of scattering interactions experienced by each event before it occurred in the same detector module, as well as the crystal ID of all interactions, were recorded. Based on the crystal index difference between scattering interactions and photoelectric interactions, CP events are divided into adjacent CP events and distant CP events.

[0008] Preferably, the generation of the training dataset by combining optical simulation and analytical methods includes the following steps: Optical GATE simulation generates the mean and variance of row and column light distribution for each crystal, which serve as the basic distribution characteristics of each crystal. Based on the fundamental distribution characteristics and the crystal IDs involved in the CP event, optical distribution data of the CP event is generated using an analytical method. This analytical method is used to represent the optical distribution of the CP event as an energy-weighted combination of the fundamental distributions of the two crystals involved in the event.

[0009] Preferably, the method of generating light distribution data for CP events through analysis includes the following steps: Generate CP event light distribution data according to formulas (1) and (2); (1) (2) in, For the line light distribution involving CP events of crystals a and b, The beam distribution for the CP events involving crystals a and b; The row-based distribution of crystal a For the column-based distribution of crystal b, For the row-based distribution of crystal b, The column-based distribution of crystal b; For energy deposition in crystal a, For energy deposition in crystal b, and keV.

[0010] Preferably, the multiplexed DNN model includes an input layer, multiple hidden layers, and an output layer; wherein, the input layer receives input data compressed into a row and column distribution by optical distribution data, the hidden layer has multiple fully connected hidden layers and one partially connected hidden layer, and uses the ReLU function as the activation function, the output layer uses the softmax function and the sigmoid function as activation functions, and uses the softmax function to perform class probability normalization processing, and the output layer outputs crystal index and energy deposition information.

[0011] Preferably, training the multiplexed DNN model using the dataset includes the following steps: The dataset is input into the multiplexed DNN model, and a composite loss function containing energy regression and crystal index classification is used. The loss function is shown in Equation (3). The multiplexed DNN model is trained by a strategy of warm-up training and alternating loss optimization. (3) in, For energy regression error, The categorical cross-entropy of the crystal index of crystal row a. The classification cross-entropy of the crystal index of row b of crystal. The categorical cross-entropy of the column crystal index of crystal a. The classification cross-entropy of the column crystal index for crystal b; During the warm-up training phase, only the classification cross-entropy term is optimized; during the alternating loss optimization phase, the energy regression term and the classification cross term are used alternately for training according to a specific period until the total number of training rounds reaches the set period.

[0012] Preferably, the event classification includes dividing the events into three categories: P events, adjacent CP events, and distant CP events. The crystal difference of the P event is 0, and the adjacent CP event and the distant CP event are distinguished by a preset crystal window threshold.

[0013] Preferably, the selection of the first interacting crystal based on different event types includes the following steps: In response to the classification of event types based on the spatial and statistical characteristics of inter-crystal scattering events, different first-interacting crystal selection strategies are adopted according to different event types. Specifically, for P events, any crystal index is randomly selected as the location result; for adjacent CP events, the crystal with the highest energy deposition is selected as the first interacting crystal; and for distant CP events, the crystal with the lowest energy deposition is selected as the first interacting crystal.

[0014] This invention also discloses a DNN-based multiplexed PET detector event localization system, comprising: The ICS event characterization module is used to simulate and acquire the spatial and statistical characteristics of intercrystalline scattering events in PET detectors, and to classify CP events into adjacent CP events and distant CP events. The training dataset generation module is used to generate a training dataset by combining optical simulation and analytical methods. This training dataset includes optical distribution data of CP events and optical distribution data of P events. The DNN model building and training module is used to build a multiplexed DNN model and train the multiplexed DNN model using the dataset. The optical distribution data to be processed is input into the trained multiplexed DNN model to complete the event classification. The event location and recognition module is used to select the first interacting crystal based on different event types, and achieve high-precision event location through crystal scattering recognition.

[0015] The present invention also discloses a computer-readable storage medium for a DNN-based multiplexed PET detector event localization method, wherein a computer program is stored thereon, and the computer executes any of the described DNN-based multiplexed PET detector event localization methods.

[0016] Compared with existing technologies, the above technical solution has the following advantages: 1. DNN architecture design for multiplexed signals: PET detectors adapted to signal multiplexing compress the light distribution into row and column distributions as input, improving positioning performance in multiplexed scenarios; 2. Energy-guided localization algorithm: Based on the statistical characteristics and energy deposition of ICS events, it achieves accurate selection of the first interacting crystal in a CP event; 3. Hybrid Dataset Training Strategy: Combining optical simulation and analytical methods to generate large-scale and diverse training data improves the network's generalization ability and classification and localization accuracy; 4. High-precision classification and positioning of multiple event types: It can accurately classify different event types such as P, nCP, and dCP, and use corresponding strategies for positioning for different event types, thereby improving the overall positioning performance. Attached Figure Description

[0017] Figure 1 This is a flowchart of the steps of the DNN-based multiplexed PET detector event localization method disclosed in this invention; Figure 2 This is a framework diagram of the DNN-based multiplexed PET detector event localization system disclosed in this invention. Figure 3 This is a schematic diagram of the structure of a multiplexed DNN model disclosed in a preferred embodiment of the present invention; Figure 4This is a flowchart of a localization and scattering identification method disclosed in a preferred embodiment of the present invention.

[0018] Figure labels: 200 - ICS event representation module; 300 - training dataset generation module; 400 - DNN model construction and training module; 500 - event localization and recognition module. Detailed Implementation

[0019] The advantages of the present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments.

[0020] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0021] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a DNN-based multiplexed PET detector event localization method,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0022] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0023] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0024] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0025] In the following description, suffixes such as "module," "part," or "unit" used to denote elements are used only for the convenience of the description of the invention and have no specific meaning in themselves. Therefore, "module" and "part" can be used interchangeably.

[0026] like Figure 1 As shown, to achieve the above objectives, this invention discloses a DNN-based multiplexed PET detector event localization method, comprising the following steps: Step S101: Simulate and obtain the spatial and statistical characteristics of intercrystalline scattering events in the PET detector, and divide the CP events into adjacent CP events and distant CP events; Step S102: Combine optical simulation and analytical methods to generate a training dataset, which includes optical distribution data of CP events and optical distribution data of P events; Step S103: Construct a multiplexed DNN model and train the multiplexed DNN model using the dataset. Input the optical distribution data to be processed into the trained multiplexed DNN model to complete the event classification. Step S104: Select the first interacting crystal according to different event types, and achieve high-precision event localization through crystal scattering recognition.

[0027] Specifically, firstly, GATE simulation is preferably used to obtain the spatial and statistical characteristics of intercrystalline scattering (ICS) events in a subdivided LYSO PET detector, dividing Compton + photoelectric (CP) events into adjacent CP events (nCP) and distant CP events (dCP). Here, the Compton + photoelectric (CP) event is a Compton effect and photoelectric effect event, and the P event is a photoelectric effect event; all of these events are scattering events. In one embodiment of the invention, gamma rays are emitted to simulate and generate events, recording the number of scattering interactions experienced by each event before it undergoes photoelectric interaction in the same detector, as well as the crystal IDs of all interactions. Based on the crystal index difference between scattering and photoelectric interaction, CP events are divided into adjacent CP events and distant CP events. Secondly, a training dataset is generated by combining optical simulation and analytical methods, including both optical simulation and generating CP event data based on analytical methods. This step is used to generate large-scale and diverse training data to improve the generalization ability and classification and localization accuracy of subsequent deep neural networks. On the one hand, a pencil beam gamma source is simulated, and photoelectric events are collected for adjacent crystals. Based on the number of photons arriving at each SiPM, the light distribution of multiple SiPM detector arrays is simulated, and the light distribution of the entire photodetector array is deduced. The light distribution of the multiple SiPM detector arrays is combined and compressed into row and column distributions. The mean of the row and column light distributions and the variance of the row and column light distributions for each crystal are calculated as the basic distribution characteristics of each crystal, thus realizing optical simulation. On the other hand, by utilizing the basic light distribution characteristics of each crystal and combining the crystal IDs involved in the CP event, row and column light distribution data for different CP events are generated through energy weighting. These row and column light distribution data are then randomly combined with the light distribution data of photoelectric events. All light distribution data are normalized to generate light distribution data for CP events through analytical methods, including nCP light distribution data and dCP event light distribution. In other words, this analytical method is used to represent the optical distribution of CP events as an energy weighted combination of the basic distributions of the two crystals involved in the event. Furthermore, CP event light distribution data is generated according to formulas (1) and (2). (1) (2) in, For the line light distribution involving CP events of crystals a and b, The beam distribution for the CP events involving crystals a and b; The row-based distribution of crystal a For the column-based distribution of crystal b, For the row-based distribution of crystal b, The column-based distribution of crystal b; For energy deposition in crystal a, For energy deposition in crystal b, and keV.

[0028] Secondly, for the design of DNN architecture for multiplexed signals, a multiplexed DNN model is constructed to adapt to PET detectors with multiplexed signals and improve the positioning performance in multiplexed scenarios. Finally, based on the Compton kinematics principle, the first interacting crystal is selected according to different event types (P, nCP, dCP) to achieve high-precision event positioning and inter-crystal scattering recognition.

[0029] Preferably, in step S103, the multiplexed DNN model includes an input layer, multiple hidden layers, and an output layer; wherein, the input layer receives input data compressed into row and column distributions by optical distribution data; the hidden layer has multiple fully connected hidden layers and one partially connected hidden layer, and uses the ReLU function as the activation function; the output layer uses the softmax function and the sigmoid function as activation functions; wherein, the softmax function is used for class probability normalization, and normalizing the optical distribution data can eliminate energy differences and SiPM overvoltage fluctuations. Due to the influence of the current environment, the output layer outputs crystal index and energy deposition information. Furthermore, the input layer of this multiplexed DNN model contains 12 neurons, corresponding to 6-dimensional row light distribution and 6-dimensional column light distribution, respectively. The network structure includes three fully connected hidden layers, each with 512 neurons, and one partially connected hidden layer containing 936 neurons. The output layer has 93 neurons, with the first four groups of 23 output neurons each using the softmax activation function, and the last output neuron using the sigmoid activation function. All hidden layers use ReLU as the activation function. Figure 3 This is a schematic diagram of the structure of a multiplexed DNN model disclosed in a preferred embodiment of the present invention, as shown below. Figure 3 As shown, normalized row (red) and column (green) light distribution data are input into a neural network, and the row and column crystal indices with the highest probability are selected as interacting crystals.

[0030] Preferably, in step S103, the multiplexed DNN model is trained using the dataset, including: inputting the dataset into the multiplexed DNN model, using a composite loss function that includes an energy regression term and a crystal index classification term, the loss function being shown in formula (3), and training the multiplexed DNN model through a strategy of warm-up training and alternating loss optimization. (3) The loss function consists of five terms: the first term is the energy regression error, and the remaining four terms are the classification cross-entropy of the row and column crystal indices, respectively. For energy regression error, The categorical cross-entropy of the crystal index of crystal row a. The classification cross-entropy of the crystal index of row b of crystal. The categorical cross-entropy of the column crystal index of crystal a. The classification cross-entropy of the column crystal index for crystal b; During the warm-up training phase, only the classification cross-entropy term is optimized; during the alternating loss optimization phase, the energy regression term and the classification cross-entropy term are used alternately for training according to a specific period until the total number of training rounds reaches the set period; in a preferred embodiment of the present invention, the training process first performs a 20-epoch "warm-up" phase, during which only the classification cross-entropy term is optimized; then, the energy regression term and the classification cross-entropy term are used alternately for training at a period of 5 epochs until the total number of training rounds reaches 600 epochs.

[0031] Preferably, in step S103, the event classification includes dividing the event into three categories: P event, adjacent CP event, and distant CP event. The crystal difference of the P event is 0, and the adjacent CP event and the distant CP event are distinguished by a preset crystal window threshold.

[0032] Preferably, in step S104, selecting the first interacting crystal according to different event types includes: In response to classifying event types based on the spatial and statistical characteristics of intercrystalline scattering events, different first-interacting crystal selection strategies are adopted according to different event types. Specifically, for P events, any crystal index is randomly selected as the location result; for nCP events, the crystal with the highest energy deposition is selected as the first interacting crystal; and for dCP events, the crystal with the lowest energy deposition is selected as the first interacting crystal. In particular, the event location process includes classifying event types based on the statistical characteristics of intercrystalline scattering events and accurately determining the position of the first interacting crystal based on the classification result to achieve precise location.

[0033] Figure 4 This is a flowchart of a localization and scattering identification method disclosed in a preferred embodiment of the present invention. In one embodiment of the present invention, as shown... Figure 4As shown, the light distribution data to be processed is acquired and input into a trained deep neural network. Energy regression yields crystal index classification and energy deposition information. First, it is determined whether the event is a P event. If it is a P event, a crystal index is randomly selected and the crystal coordinates are directly output, which is the position of the crystal hit by the photon. Next, it is determined whether it is an nCP event. If it is an nCP event, the crystal with the highest energy deposition is selected as the first interacting crystal output. Specifically, the crystal with the largest energy deposition is selected, and its position is used as the output. The reason is that only when the first crystal hit has the highest energy will it not pass through many crystals. If it is not an nCP event, i.e., a dCP event, the crystal with the lowest energy deposition is selected as the first interacting crystal output. Specifically, the crystal with the smallest energy deposition is selected, and its position is used as the output. The reason is that only when the photon hits the crystal with the lowest energy on the first time will subsequent crystals at greater distances be hit. Based on the above positioning results, the ICS event is labeled, the scattering event is identified, and the scattering event is accurately located.

[0034] Example 1: (1) Experimental setup A high-resolution PET scanner system based on GATE v8.2 simulation was employed. This system consists of four detector panels, each integrating 4×4 detector modules. Each module contains 23×23 LYSO crystals, with individual crystal dimensions of 0.785×0.785×20 mm. 3 (Crystal spacing is 0.85 mm), the overall module size is 19.6 × 19.6 × 20 mm. 3 To estimate deep interactions (DOI), each detector module is connected to a photodetector array at both ends. Each array consists of 6×6 silicon photomultiplier tubes (SiPMs), with each SiPM unit measuring 3.16×3.16 mm. 2 The center-to-center spacing is 3.3 mm, which is used to read the flashing light signal from both ends.

[0035] (2) ICS event representation A point source was placed at the center of the field of view, emitting 511 keV back-to-back gamma rays to simulate approximately 60 million events. The number of scattering interactions experienced by each event before the photoelectric interaction occurred within the same detector block, as well as the crystal IDs of all interactions, were recorded. Based on the crystal index difference between scattering and photoelectric interaction (i.e., the crystal window), CP events were divided into nCP events and dCP events: considering the index difference of 0 to 3 crystals, the forward / backward scattering ratio under different crystal window conditions was analyzed.

[0036] (3) Generation of training dataset ① Optical Simulation: A 511 keV pencil-beam gamma source was simulated by moving it perpendicular to the front face of the crystal. For each of the 66 adjacent crystals (1 / 8 of the entire array), 40,000 photoelectric events were collected. Based on the number of photons reaching each SiPM, the light distribution of the two 6×6 SiPM arrays (front and rear) was simulated separately, and the light distribution of the entire array was deduced using symmetry. After combining the front and rear light distributions, the original 6×6 light distribution was compressed into a 1×6 row distribution and a 1×6 column distribution. Finally, the mean and variance of the row and column light distributions for each crystal were calculated as their basic distribution characteristics.

[0037] ② Generate CP event data based on analytical methods: Utilizing the basic light distribution and variance of each crystal, combined with the crystal ID involved in the CP event, approximately 127 million different row-column light distributions for CP events are generated through energy weighting. This data is then randomly combined with the light distributions of approximately 19 million photoelectric events, and all light distributions are normalized to eliminate energy dependence.

[0038] (4) Construction and training of deep neural networks ① Network Architecture: Based on the Keras and TensorFlow-GPU v2.6.2 framework, implemented on an NVIDIA GeForce RTX2060 GPU. The input layer has 12 neurons; the network contains 3 fully connected hidden layers, each with 512 neurons, and 1 partially connected hidden layer with 936 neurons; the output layer has 93 neurons. All hidden layers use the ReLU activation function; in the output layer, the first four groups of 23 neurons each use the softmax activation function, and the last neuron uses the sigmoid activation function.

[0039] ② Training Process: The loss function consists of one energy regression loss term and four crystal index classification cross-entropy loss terms. Training begins with a 20-round "warm-up" period, during which only the classification cross-entropy term is optimized. Afterward, the energy regression term and classification cross-entropy term are alternately optimized every 5 rounds, for a total of 600 training rounds. The learning rate is set to 0.001, and mini-batch training is used. Dropout = 0.5 is applied to all hidden layers during training to prevent overfitting.

[0040] (5) Effect evaluation Three point sources emitting isotropic 511 keV gamma rays were simulated, located 3 cm along the Z-axis on the front of the detector and distributed at different positions in the XY plane. The L1 distance between the DNN prediction results and the simulated real-world positioning data was calculated and compared with the positioning performance of the traditional Anger logic. The results show that this technique exhibits higher positioning accuracy in both P and CP events: the crystal classification accuracy reaches 90% for P events and 82% for CP events, demonstrating a significant improvement in overall positioning performance compared to the existing Anger logic technique.

[0041] like Figure 2 As shown, to achieve the above objectives, the present invention also discloses a DNN-based multiplexed PET detector event localization system, comprising: The ICS event characterization module is used to simulate and acquire the spatial and statistical characteristics of intercrystalline scattering events in PET detectors, and to classify CP events into adjacent CP events and distant CP events. The training dataset generation module is used to generate a training dataset by combining optical simulation and analytical methods. The training dataset includes optical distribution data of CP events and optical distribution data of P events. The DNN model building and training module is used to build a multiplexed DNN model and train the multiplexed DNN model using the dataset. The optical distribution data to be processed is input into the trained multiplexed DNN model to complete the event classification. The event location and recognition module is used to select the first interacting crystal based on different event types, and achieve high-precision event location through crystal scattering recognition.

[0042] It should be noted that this system corresponds to the above-mentioned event localization method for multiplexed PET detectors based on DNN, and the remaining undescribed parts refer to the content of the above method.

[0043] To achieve the above objectives, the present invention also provides a computer-readable storage medium comprising multiple storage media, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, server, app store, etc., wherein a computer program is stored thereon, and the program performs corresponding functions when executed by a processor. The computer-readable storage medium of this embodiment is used to execute any of the above-described DNN-based multiplexed PET detector event localization methods.

[0044] It should be noted that the embodiments of the present invention have better implementability and are not intended to limit the present invention in any way. Any person skilled in the art may use the above-disclosed technical content to change or modify it into equivalent effective embodiments. However, any modifications or equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A method for event localization in a DNN-based multiplexed PET detector, characterized in that, Includes the following steps: The spatial and statistical characteristics of intercrystalline scattering events in a PET detector were simulated and obtained, and CP events were divided into adjacent CP events and distant CP events. A training dataset is generated by combining optical simulation and analytical methods. The training dataset includes optical distribution data of CP events and optical distribution data of P events. Construct a multiplexed DNN model and train the multiplexed DNN model using the dataset. Input the optical distribution data to be processed into the trained multiplexed DNN model to complete event classification. Based on different event types, the first interacting crystal is selected, and high-precision event localization is achieved through crystal scattering recognition.

2. The event localization method for a DNN-based multiplexed PET detector as described in claim 1, characterized in that, The simulation acquires the spatial and statistical characteristics of intercrystalline scattering events in the PET detector, classifying CP events into adjacent CP events and distant CP events, including the following steps: Using GATE to simulate the LYSO PET detector system, multiple events were generated, and the number of scattering interactions experienced by each event before the photoelectric interaction occurred in the same detector module was recorded, along with the crystal ID of all interactions. Based on the crystal index difference between scattering interactions and photoelectric interactions, CP events are divided into adjacent CP events and distant CP events.

3. The event localization method for a DNN-based multiplexed PET detector as described in claim 2, characterized in that, The process of generating the training dataset by combining optical simulation and analytical methods includes the following steps: Optical GATE simulation generates the mean and variance of row and column light distribution for each crystal, which serve as the basic distribution characteristics of each crystal. Based on the fundamental distribution characteristics and the crystal IDs involved in the CP event, optical distribution data of the CP event is generated by an analytical method. The analytical method is used to represent the optical distribution of the CP event as an energy-weighted combination of the fundamental distributions of the two crystals involved in the event.

4. The event localization method for a DNN-based multiplexed PET detector as described in claim 3, characterized in that, The process of generating optical distribution data for CP events using analytical methods includes the following steps: Generate CP event light distribution data according to formulas (1) and (2); （1） （2） in, For the line light distribution involving CP events of crystals a and b, The beam distribution for the CP events involving crystals a and b; The row-based distribution of crystal a For the column-based distribution of crystal b, For the row-based distribution of crystal b, The column-based distribution of crystal b; For energy deposition in crystal a, For energy deposition in crystal b, and keV.

5. The event localization method for a DNN-based multiplexed PET detector as described in claim 1, characterized in that, The multiplexed DNN model includes an input layer, multiple hidden layers, and an output layer. The input layer receives input data compressed from optically distributed data into a row and column distribution. The hidden layer has multiple fully connected hidden layers and one partially connected hidden layer, and uses the ReLU function as the activation function. The output layer uses the softmax function and the sigmoid function as activation functions, and uses the softmax function to perform class probability normalization. The output layer outputs crystal index and energy deposition information.

6. The event localization method for a DNN-based multiplexed PET detector as described in claim 1, characterized in that, Training the multiplexed DNN model using the dataset includes the following steps: The dataset is input into the multiplexed DNN model, and a composite loss function containing an energy regression term and a crystal index classification term is used. The loss function is shown in formula (3). The multiplexed DNN model is trained by a strategy of warm-up training and alternating loss optimization. （3） in, For energy regression error, The categorical cross-entropy of the crystal index of crystal row a. The classification cross-entropy of the crystal index in row b of crystal. The categorical cross-entropy of the column crystal index of crystal a. The classification cross-entropy of the column crystal index for crystal b; During the warm-up training phase, only the classification cross-entropy term is optimized; during the alternating loss optimization phase, the energy regression term and the classification cross term are used alternately for training according to a specific period until the total number of training rounds reaches the set period.

7. The event localization method for a DNN-based multiplexed PET detector as described in claim 1, characterized in that, The event classification includes dividing events into three categories: P events, adjacent CP events, and distant CP events. The crystal difference of the P events is 0, and the adjacent CP events and distant CP events are distinguished by a preset crystal window threshold.

8. The event localization method for a DNN-based multiplexed PET detector as described in claim 7, characterized in that, The selection of the first interacting crystal based on different event types includes the following steps: In response to the classification of event types based on the spatial and statistical characteristics of inter-crystal scattering events, different first-interacting crystal selection strategies are adopted according to different event types. Specifically, for P events, any crystal index is randomly selected as the location result; for adjacent CP events, the crystal with the highest energy deposition is selected as the first interacting crystal; and for distant CP events, the crystal with the lowest energy deposition is selected as the first interacting crystal.

9. A DNN-based multiplexed PET detector event localization system, characterized in that, include: The ICS event characterization module is used to simulate and acquire the spatial and statistical characteristics of intercrystalline scattering events in PET detectors, and to classify CP events into adjacent CP events and distant CP events. The training dataset generation module is used to generate a training dataset by combining optical simulation and analytical methods. The training dataset includes optical distribution data of CP events and optical distribution data of P events. The DNN model building and training module is used to build a multiplexed DNN model and train the multiplexed DNN model using the dataset. The optical distribution data to be processed is input into the trained multiplexed DNN model to complete the event classification. The event location and recognition module is used to select the first interacting crystal based on different event types, and achieve high-precision event location through crystal scattering recognition.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer executes the DNN-based multiplexed PET detector event localization method according to any one of claims 1-8.

Images