A method for generating a PET detector signal simulation based on artificial intelligence

Abstract

The application discloses a PET detector signal simulation generation method based on artificial intelligence, comprising the following steps: generating an initial two-photon signal pair meeting initial energy and time difference physical constraints through a two-photon signal generator fusing a Transformer and a U-Net; decoupling mixed noise into multiple independent physical noise components according to a PET noise physical prior library by using a multi-branch decoupling network based on an attention mechanism, and outputting a pure signal and a noise component spectrum; performing linkage adjustment on the pure signal and the noise component according to target parameters set by a user through a space-time-noise collaborative regulator; and finally, outputting a customized signal data format according to a downstream task type. The application realizes high-fidelity, interpretable and controllable coincidence event level signal simulation, effectively solves the core problems of complex modeling, noise distortion, lack of correlation and poor scene adaptability of traditional methods, and significantly improves the efficiency and accuracy of PET detector research and testing.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of PET detector signal simulation technology, and in particular to a PET detector signal simulation generation method based on artificial intelligence. BACKGROUND

[0002] As an important technology of nuclear medical imaging, the simulation generation of PET detector signals plays a key role in device research and development, system calibration, and algorithm optimization. The goal of signal simulation technology is to achieve reproducibility of the real detection process through modeling and generation of double-photon coincidence events, in order to support device performance evaluation and algorithm verification. However, existing signal simulation methods generally have problems such as complex physical modeling, low computational efficiency, and insufficient noise authenticity, making it difficult to meet the application requirements of modern PET systems in multiple scenarios and multiple precision requirements.

[0003] In traditional methods, signal generation usually relies on accurate physical modeling, which jointly simulates multiple links such as detector crystal characteristics, photoelectric conversion, and electronic response. Although this method has certain interpretability, it consumes a large amount of computing resources and is difficult to maintain authenticity and stability in large-scale sample generation and multi-noise environments. In addition, such models can only model single-photon events and lack effective characterization of key coincidence events (energy and time correlation between double photons) in PET imaging, resulting in deviations in time and space consistency between simulated signals and real detection data.

[0004] With the development of artificial intelligence, especially deep learning generation models, researchers have begun to try data-driven signal simulation to reduce the dependence on physical modeling and improve generation efficiency. However, existing AI simulation methods often ignore the physical constraints and interpretability of noise characteristics of PET signals. In particular, in the generation of double-photon coincidence events, there are still problems such as insufficient learning of time and space correlation, difficulty in separating noise mixing, and lack of samples in extreme scenarios (such as high-energy photons and scattering events). These deficiencies make the generated signals unable to fully support detector performance evaluation or system-level optimization of coincidence circuits.

[0005] In addition, current simulation methods generally use fixed output modes, lacking flexibility in adapting to different application scenarios. For example, in tasks such as detector structure optimization, noise suppression algorithm development, and extreme event analysis, different signal fidelity and noise characteristics are often required, but traditional models cannot achieve on-demand regulation, resulting in a balance between accuracy and efficiency in the generated results.

[0006] In summary, the existing PET detector signal simulation method still has limitations in space-time consistency, noise decoupling and task adaptability, and a new signal generation method that can integrate physical constraints and intelligent generation capabilities, with controllable noise and event-level precision, is needed to support the intelligent application of PET systems in research and development, testing and evaluation. SUMMARY

[0007] To overcome the shortcomings of the prior art, the present application provides a PET detector signal simulation generation method based on artificial intelligence, which realizes high-fidelity generation and noise controllable adjustment of two-photon signals at the coincidence event level, thereby improving the physical authenticity and task adaptability of signal simulation and supporting the research and development, optimization and performance evaluation of PET systems.

[0008] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a PET detector signal simulation generation method based on artificial intelligence, comprising:

[0009] Signal generation step: obtain the real PET coincidence event dataset and the detector hardware parameters, input them into the trained two-photon signal generator, and generate the initial two-photon signal pair; the two-photon signal generator adopts a neural network architecture that integrates a Transformer encoder and a U-Net decoder, learns the space-time correlation features between the two photons through the Transformer encoder, and generates the initial signal pair containing mixed noise based on the U-Net decoder, and the training process ensures that the output signal meets the preset energy and time difference physical range through the initial coincidence constraint;

[0010] Noise decoupling step: input the initial two-photon signal pair into the coincidence event noise decoupler, decouple the mixed noise into multiple independent physical noise components based on the pre-constructed PET noise physical prior library, and output the two-photon pure signal and the noise component spectrum; the decoupler adopts a multi-branch decoupling network, each branch corresponds to a type of physical noise, and the cross-decoupling attention mechanism is used to constrain the distribution of the same type of noise between the two photons to meet the annihilation energy conservation;

[0011] Coordinated regulation step: receive the user-set target space-time parameters and target noise parameters, and perform linked adjustment on the two-photon pure signal and the physical noise component through the space-time-noise coordinated regulator; the adjustment includes adjusting the time stamp and spatial coordinates according to the target space-time parameters, and scaling the amplitude of the specified noise component according to the intensity adjustment coefficient in the target noise parameters, and synchronously adjusting the noise distribution of the two photons after scaling to maintain energy conservation;

[0012] Signal output step: according to the specified downstream task type, adapt the customized two-photon signal pair after regulation to the data format matched with the task.

[0013] Further, the PET noise physical prior library comprises at least three of SiPM dark current noise, ADC quantization noise, Compton scattering noise, crystal afterglow noise, power supply noise and crosstalk noise.

[0014] Further, the coincidence event noise decoupler, when performing noise decoupling, first decomposes the initial two-photon signal pair into different frequency bands through wavelet transform, and then causes each branch network to extract and decouple noise components in the corresponding frequency band according to the power spectrum characteristics of the corresponding noise in the PET noise physical prior library.

[0015] Further, the noise decoupling step is implemented through a unified, end-to-end deep learning model.

[0016] The deep learning model is trained through a total loss function composed of coincidence consistency loss and noise rationality loss.

[0017] Wherein, the coincidence consistency loss adopts the average absolute error to calculate the spatial and temporal deviation between the generated signal and the real signal, and the formula is:

[0018] ;

[0019] Wherein, and respectively represent the generated first photon energy value and the second photon energy value, represents the time difference of the generated two-photon signal, represents the average value of the two-photon time difference in the real coincidence event data set, is a time weight coefficient;

[0020] The noise rationality loss adopts KL divergence to calculate the distribution difference between the generated noise and the real noise, and the formula is:

[0021] ;

[0022] Wherein, represents the generated i-th noise distribution, represents the real noise distribution from the noise prior library, represents the noise rationality loss.

[0023] Further, the total loss function is composed of the coincidence consistency loss, the noise rationality loss and a newly added noise distribution stability loss, and the expression is configured as:

[0024] ;

[0025] Wherein, wherein is the noise distribution stability loss, , , are the learnable weight coefficients of each loss, which are dynamically updated by gradient descent method at the early stage of training and fixed at the later stage of training;

[0026] is the hyperbolic tangent function, is the Gaussian error function, is the natural logarithm function;

[0027] is the scale factor of the noise distribution stability loss, which is used to balance the input dynamic range of the tanh function;

[0028] and are the target expected value and tolerance standard deviation of the noise reasonability loss, respectively, which are used to construct the expected constraint for the noise loss;

[0029] is the shape adjustment factor of the stability loss;

[0030] The noise distribution stability loss is defined as the negative log-likelihood of the variance of each noise component in the generated noise component map, and its expression is:

[0031] wherein represents the variance of the i-th noise component, is a very small positive number to prevent numerical overflow.

[0032] Further, in the step of cooperative regulation, when adjusting the time stamp and the spatial coordinate according to the target spatiotemporal parameter, an interpolation and coordinate mapping algorithm under physical constraints is adopted, and the process specifically includes:

[0033] time interpolation adjustment: based on the target time difference and the original time difference, the first photon time stamp and the second photon time stamp are synchronously adjusted by a linear interpolation function, so that the adjusted time difference satisfies: wherein represents the first photon time stamp, represents the second photon time stamp, represents the adjusted time difference, represents the target time difference, which maintains the causality of the photon time sequence in the time interpolation process;

[0034] The spatial coordinate mapping: based on the target spatial distance and the original spatial distance, the first photon detection position and the second photon detection position are synchronously adjusted by a conformal transformation algorithm to ensure that the distance of the two adjusted points is equal to the target spatial distance, and the mapped coordinates are still located within the effective crystal array region of the PET detector.

[0035] Further, the intensity adjustment coefficient in the target noise parameter is defined as , and the value range is , wherein is 0.5, is 2.0;

[0036] When scaling the amplitude of the specified noise component, the synergistic regulation step ensures that the amplitudes of the same type of noise between the two photons change proportionally and synchronously through the cross-decoupling attention mechanism.

[0037] Further, in the synergistic regulation step, when scaling the amplitude of the specified noise component according to the intensity adjustment coefficient, a noise-energy balance constraint mechanism is introduced;

[0038] The noise-energy balance constraint mechanism is configured to: after scaling the amplitude of the i-th type of noise component by times, calculate the single-photon energy deviation caused by the change of this type of noise, and accordingly compensate the pure signal energy of the two photons, so that the absolute value of the deviation between the total energy of the two photons after compensation and the total energy before compensation is not greater than a preset threshold;

[0039] wherein the total energy of the two photons after compensation is:

[0040] , wherein and respectively represent the single-photon energy deviation of the generated first photon and second photon;

[0041] The total energy before compensation is: .

[0042] Further, in the signal generation step, the Transformer encoder extracts the spatio-temporal correlation features between the two photons through a multi-head self-attention mechanism, specifically including:

[0043] Spatio-temporal feature embedding: the first photon energy value, the first photon timestamp, and the first photon detection position are encoded into a first feature vector, and the second photon energy value, the second photon timestamp, and the second photon detection position are encoded into a second feature vector. The position encoding information representing the photon serial number and the detector crystal ring position is fused into the first feature vector and the second feature vector;

[0044] Correlation feature extraction: the first feature vector and the second feature vector are spliced and input to the multi-head self-attention layer of the Transformer encoder; the self-attention layer is configured to calculate the interaction attention weight between the first feature vector and the second feature vector to explicitly model the spatiotemporal correlation of the pair of two-photon signals, and output a correlation feature vector; the correlation feature vector at least contains the implicit representation of the adjusted time difference and the implicit representation of the spatial probe point connection vector;

[0045] Signal decoding: input the correlation feature vector into the U-Net decoder, and based on the correlation feature vector and the bottom layer information transmitted by the skip connection, the U-Net decoder reconstructs an initial pair of two-photon signals that meet the initial coincidence constraint.

[0046] Further, the signal output step dynamically reconstructs the customized pair of two-photon signals and enhances the information according to the type of downstream task, specifically including:

[0047] When the downstream task is component-level optimization, the output module disassembles the customized pair of two-photon signals to generate a noise-signal pair contrast dataset indexed by a single detector component; the dataset separates each type of physical noise component after decoupling from the corresponding pure signal, and is associated with the detector hardware parameters to form a two-dimensional mapping table for component performance analysis;

[0048] When the downstream task is system-level testing, the output module sorts 100,000 to 1,000,000 of the customized pair of two-photon signals by timestamp, synthesizes a continuous coincidence event stream, and generates a corresponding system timing atlas; the system timing atlas at least contains a time difference distribution histogram of each coincidence event and a curve of coincidence count rate change over time, which is used for calibration of system dead time and coincidence resolution time;

[0049] When the downstream task is extreme event evaluation, the output module filters out signals with energy values greater than a preset upper energy threshold or less than a lower energy threshold, or with a spatial scattering angle greater than an upper scattering angle from the customized pair of two-photon signals, constructs an extreme event signal library, and labels each signal in the library with its corresponding physical parameters, including energy deposition, scattering angle, and noise category.

[0050] The beneficial effects of the present application are:

[0051] The PET detector signal simulation generation method based on artificial intelligence proposed in the present application adopts a neural network architecture that fuses a Transformer encoder and a U-Net decoder, and introduces coincidence event constraints, noise decoupling, and spatiotemporal collaborative regulation mechanisms, which has the following beneficial effects compared to the prior art:

[0052] 1. Event-level high-fidelity generation:

[0053] The spatiotemporal correlation features between two photons are learned by a Transformer encoder, and coincidence constraints are introduced in the U-Net decoder, so that the generated signal pairs have high physical consistency in the energy conservation and time difference range, and can be directly used for coincidence circuit simulation and system-level testing.

[0054] 2. Fine decoupling of physical noise:

[0055] The mixed noise is separated into multiple independent components with physical meaning by using a multi-branch noise decoupling network and a cross-decoupling attention mechanism, and the SiPM dark current, ADC quantization error, Compton scattering and other noises are accurately modeled, the noise decoupling precision is high, and the detector component-level optimization is supported.

[0056] 3. Cooperative and controllable adjustment of time and space and noise:

[0057] Through the time-space-noise cooperative regulator, the timestamps, spatial coordinates and noise intensities can be adjusted according to user-set parameters, the personalized generation and multi-scene adaptation of signals are realized, and the specific needs of different testing tasks are met.

[0058] 4. Energy conservation and noise consistency guarantee:

[0059] The cooperative regulation mechanism automatically maintains the energy conservation of the two photons when adjusting the noise components, avoids signal distortion, and ensures that the generated data is consistent with the real detection data in terms of physical characteristics.

[0060] 5. Generation efficiency and task adaptability improvement:

[0061] The automatic simulation framework based on the deep generation model avoids the high computational complexity of traditional physical modeling, realizes the rapid generation of millions of samples, and supports customized output of different task types (system-level testing, component-level optimization, extreme event evaluation), improves the practicality and efficiency of the model in industrial research and development. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 is the step flow chart of the PET detector signal simulation generation method in the present application. DETAILED DESCRIPTION

[0063] The present application will be further described in detail below in combination with the drawings and examples. It should be noted that the same parts are denoted by the same reference numerals in the following description. It should be noted that the words "front", "back", "left", "right", "up" and "down" used in the following description refer to the directions in the drawings, and the words "bottom surface" and "top surface", "inner" and "outer" refer to the directions towards or away from the geometric center of a particular part.

[0064] Embodiment one, refer to Figure 1 For the first embodiment of the application, the embodiment provides a PET probe signal simulation generation method based on artificial intelligence. The method runs in a computer software and hardware integrated environment, and can realize high-fidelity generation of two-photon signals at coincidence event level, noise decoupling and task output. The system of the embodiment includes the following functional modules:

[0065] Two-photon signal generator; coincidence event noise decoupler; space-time-noise collaborative controller; coincidence consistency-noise rationality double-loss module; downstream task adaptive output module.

[0066] The system operating environment includes:

[0067] Computing device: Intel Xeon Gold 6330 CPU, NVIDIA A100 GPU (video memory 80GB), memory 128GB;

[0068] Operating system and framework: Ubuntu 22.04 LTS operating system, Python 3.10 environment, and deep learning framework PyTorch 2.0;

[0069] Input data: contains real PET coincidence event data set and detector hardware parameters (crystal array density, SiPM gain curve, energy response characteristics, etc.).

[0070] Working principle of embodiment one:

[0071] (1) Signal generation step

[0072] This step is completed by the two-photon signal generator.

[0073] The real PET coincidence event data set and the detector hardware parameters are input into the neural network with a fusion Transformer encoder and a U-Net decoder.

[0074] The Transformer encoder part is used to learn the space-time correlation features between two photons, including time difference distribution, spatial coordinate angle, energy complementarity relationship, etc. Through the multi-head self-attention mechanism, the model can capture the dynamic correlation of two photons in the energy and time dimensions.

[0075] The U-Net decoder part is responsible for generating two-photon initial signal pairs under the guidance of this correlation information, including energy, timestamp and spatial coordinate information. The U-Net structure maintains high-resolution features through skip connection, realizing high-fidelity reconstruction of signal details.

[0076] In the model training phase, the system introduces an initial coincidence constraint mechanism to ensure that the output signal meets the physical rationality range of energy and time difference , effectively avoiding non-physical deviation of the signal.

[0077] The training process of the two-photon signal generator and coincidence event noise decoupler adopts a phased strategy:

[0078] In the first phase, the two-photon signal generator is first trained, and an initial coincidence constraint is applied;

[0079] In the second phase, the generator parameters are frozen, and the coincidence event noise decoupler is trained;

[0080] In the third phase, the overall model is jointly fine-tuned with a lower learning rate.

[0081] Technical effects:

[0082] The generated signal in this phase is superior to traditional methods in terms of energy conservation and time matching accuracy, with a coincidence event consistency error of less than 5%, providing high-precision input for subsequent noise decoupling.

[0083] (2) Noise decoupling step

[0084] This phase is executed by the coincidence event noise decoupler, and its core is a multi-branch decoupling network.

[0085] First, the initial two-photon signal pair is input into the decoupler. The system has a built-in PET noise physical prior library, which contains statistical characteristic parameters of various noise types, including:

[0086] SiPM dark current noise, ADC quantization noise, Compton scattering noise, crystal afterglow noise, power supply interference noise, and signal crosstalk noise.

[0087] The decoupler works through the following process:

[0088] Wavelet transform preprocessing layer: decomposes the signal into different frequency bands to match the power characteristics of different noises in the frequency domain.

[0089] Multi-branch neural network structure: each branch is responsible for identifying and reconstructing a specific noise component.

[0090] Cross-decoupling attention mechanism: introduces an energy conservation constraint when processing two-photon signals, i.e., the superposition of any similar noise in E1 and E2 does not destroy the annihilation energy balance.

[0091] The output includes two parts:

[0092] Two-photon pure signal (ideal signal component after noise removal);

[0093] Noise component atlas (identifying the distribution proportion of each type of noise in the energy and time dimension).

[0094] Technical effects:

[0095] This module realizes the fine separation of physical noise, and the decoupling accuracy can reach more than 92%. By independently obtaining the intensity information of each type of noise, the detector electronics components (such as TDC, SiPM module) can be optimized.

[0096] (3) Synergistic regulation step

[0097] The joint regulation of signal and noise is realized by the space-time-noise synergistic regulator.

[0098] The input is the decoupled pure signal and noise atlas, as well as the target parameters set by the user:

[0099] Target space-time parameters: including the expected two-photon time difference and detection spatial distance;

[0100] Target noise parameters: including specific noise intensity adjustment coefficient .

[0101] The control logic includes:

[0102] Space-time regulation: using time interpolation and spatial coordinate mapping algorithm to make the generated signal meet the two-photon time difference and detection spatial distance requirements;

[0103] Noise regulation: according to Value scaling the amplitude of the specified noise, and through the energy conservation coupling algorithm, keep balance between And , to ensure physical consistency.

[0104] Technical effects:

[0105] It can flexibly generate two-photon signals in different scenarios, support detector resolution testing, noise resistance algorithm optimization, and extreme event evaluation tasks.

[0106] (4) Signal constraint and optimization step

[0107] This stage realizes model optimization by a consistency-noise rationality double-loss module.

[0108] The module includes:

[0109] Consistency loss: by calculating the energy and time difference deviation between the generated signal and the real signal, to maintain physical matching;

[0110] Noise rationality loss: by KL divergence to measure the difference between the generated noise and the real noise distribution, to ensure the realism of the noise.

[0111] This stage is trained by Adam optimizer (learning rate 1e-4) until the model converges.

[0112] Technical effects:

[0113] This double constraint mechanism ensures the physical reasonableness of the signal and the authenticity of the noise, so that the generated signal reaches more than 95% similarity in the energy and time difference distribution of the real sample.

[0114] (5) Signal output step

[0115] This stage is completed by the downstream task adaptation output module.

[0116] According to different application scenarios, the system formats the adjusted signal for output:

[0117] For component-level optimization tasks, output the single-photon noise data and hardware parameter correspondence table for device anti-interference performance verification.

[0118] For system-level testing tasks, output the coincidence event sequence and timing atlas for coincidence circuit delay and synchronization precision adjustment.

[0119] For extreme event evaluation tasks, output the high-energy photon and large-angle scattering signal library for detector limit response performance evaluation.

[0120] Technical effects:

[0121] Realize multi-task output mode, adapt to 12 types of PET system development needs, and signal generation efficiency is improved by more than 200 times compared with traditional Monte Carlo simulation.

[0122] Technical effects of embodiment one:

[0123] Through the cooperative operation of the above modules, the embodiment realizes the following comprehensive technical advantages:

[0124] The generated signal highly conforms to the real physical law in energy, time and space characteristics;

[0125] Realize independent decoupling and adjustable control of noise components, support directional noise research;

[0126] Can generate two-photon signals with different space-time and noise configurations on demand, significantly improving system development efficiency;

[0127] The overall generation efficiency is improved by about 200 times, meeting the needs of industrialized rapid verification and algorithm testing.

[0128] Embodiment two, as a second embodiment of the present application, provides a PET detector signal simulation generation method based on artificial intelligence, which is consistent with the system architecture of embodiment one.

[0129] Working principle of embodiment two:

[0130] (1) Construction of PET noise physical prior library:

[0131] In the noise decoupling step of this embodiment, a PET noise physical prior library is introduced, which contains at least three representative types of physical noise: SiPM dark current noise, ADC quantization noise, Compton scattering noise, crystal afterglow noise, power supply noise, and signal crosstalk noise.

[0132] The data of various types of noise are obtained by statistical analysis of experimental signals collected by an actual PET system, including their power spectral density, mean value, variance, and time correlation characteristics.

[0133] The existence of this prior library provides physical guidance and constraints for subsequent noise extraction and identification, avoiding the "black box" problem of purely data-driven models.

[0134] Technical effects:

[0135] By introducing the noise prior library, the model can obtain physical reference in noise type identification and energy distribution fitting, significantly improving the decoupling accuracy and noise interpretation.

[0136] (2) Wavelet multi-branch noise decoupling based on frequency domain characteristics:

[0137] When the coincidence event noise decoupler performs noise decoupling, it first performs wavelet transform on the initial two-photon signal pair, decomposing the original time-domain signal into sub-signals in different frequency bands.

[0138] Each frequency band corresponds to the main energy distribution interval of a specific noise, for example:

[0139] The low-frequency band corresponds to power supply noise (50 / 60 Hz);

[0140] The medium-frequency band corresponds to crystal afterglow noise and dark current noise;

[0141] The high-frequency band corresponds to ADC quantization noise and crosstalk noise.

[0142] After frequency domain decomposition, multiple branch sub-networks within the decoupler perform noise extraction and separation in the corresponding frequency band according to the power spectrum characteristics in the prior library. The noise components output by each branch network are fused by the convergence layer to form a complete noise component spectrum, and a pure two-photon signal after noise removal is simultaneously generated.

[0143] Technical effects:

[0144] The use of wavelet frequency domain decomposition enables the model to accurately identify the spectral characteristics of various types of noise, avoiding the aliasing problem between different noises, improving the noise decoupling accuracy by about 15%, and significantly improving the time-domain stability of signal restoration.

[0145] (3) End-to-end noise decoupling model and joint loss optimization mechanism:

[0146] The noise decoupling process of the embodiment is implemented through an end-to-end deep learning model, and the optimization of the model is driven by a total loss function.

[0147] The total loss function includes the following three parts:

[0148] 1. Coincidence loss , used to constrain the spatiotemporal matching degree of the generated signal and the real signal, and its calculation formula is:

[0149] ;

[0150] Wherein, and represent the generated first photon energy value and the second photon energy value, represent the time difference of the generated two-photon signal, represent the average value of the time difference of the two photons in the real coincidence event data set, is a time weight coefficient, represent the characteristic energy of the two photons generated when a positron and an electron annihilate, and the time weight coefficient in the embodiment can be 100;

[0151] 2. Noise rationality loss, which calculates the distribution difference between the generated noise and the real noise using KL divergence, and its formula is:

[0152] ;

[0153] Wherein, represents the generated i-th noise distribution, represents the real noise distribution from the noise prior library, represents the noise rationality loss.

[0154] 3. Noise distribution stability loss, defined as the negative log-likelihood of the generated noise variance, to prevent the noise distribution from fluctuating sharply during training:

[0155] , wherein represents the variance of the i-th noise component, is a very small normal number to prevent numerical overflow.

[0156] The three together constitute the total loss function:

[0157] The total loss function is composed of the coincidence loss, the noise rationality loss, and a newly added noise distribution stability loss, and its expression is configured as:

[0158] ;

[0159] wherein, is the total loss function, is the noise distribution stability loss, , , are the learnable weight coefficients of each loss, which are dynamically updated by gradient descent method at the initial stage of training and fixed at the later stage of training;

[0160] is the hyperbolic tangent function, is the Gaussian error function, is the natural logarithm function;

[0161] is the scale factor of the noise distribution stability loss, which is used to balance the input dynamic range of the tanh function;

[0162] and are the target expected value and tolerance standard deviation of the noise rationality loss, respectively, which are used to construct the expected constraint for the noise loss;

[0163] is the shape adjustment factor of the stability loss.

[0164] Working principle:

[0165] During the training process, the network automatically adjusts the weights of each item to dynamically balance the physical consistency of the signal and the rationality of the noise distribution. As the iteration proceeds, the tanh function suppresses the gradient influence of large error samples, the erf function controls the stable convergence of the noise loss, and the log function maintains the smooth evolution of the noise variance, thereby achieving convergence and physical constraint.

[0166] Technical effects:

[0167] The introduction of the joint loss makes the model training more stable, the physical deviation of the generated signal is less than 3%, the KL divergence of the noise distribution is less than 0.15, and the noise fluctuation standard deviation is reduced by about 20%.

[0168] (4) Noise synergistic regulation and energy balance compensation mechanism

[0169] In the space-time-noise synergistic regulation step, the embodiment further guarantees the physical rationality of the signal by introducing a noise-energy balance constraint mechanism.

[0170] When the user sets the noise intensity adjustment coefficient The system scales the amplitude of the specified noise component when it varies in the interval [0.5, 2.0]. At this time, the collaborative controller ensures that the amplitudes of the same type of noise between the two photons change proportionally and synchronously through the cross-decoupling attention mechanism, that is, the noise enhancement or weakening process maintains the energy conservation relationship.

[0171] Further, an energy compensation mechanism is introduced:

[0172] When the amplitude of the i-th type of noise is amplified or reduced, the system calculates the single-photon energy deviation caused by this and , and respectively represent the single-photon energy deviation of the generated first photon and second photon;

[0173] Through the energy reverse compensation algorithm, the energy of the double-photon pure signal is adjusted so that the deviation between the total energy of the compensated double photons and the total energy of the double photons before compensation is less than a preset threshold (such as 2keV).

[0174] wherein the total energy of the compensated double photons is:

[0175] wherein and respectively represent the single-photon energy deviation of the generated first photon and second photon;

[0176] The total energy before compensation is: .

[0177] Technical effects:

[0178] This mechanism ensures that the overall signal energy is not imbalanced during noise adjustment, and the generated signal maintains physical reality under the constraint of double-photon energy conservation, with simultaneous optimization of time deviation and energy deviation, and is suitable for coincidence circuit delay testing and detector energy spectrum calibration scenarios.

[0179] Technical effects of embodiment two:

[0180] Compared with embodiment one, the following further technical improvements are realized in this embodiment:

[0181] Enhanced noise recognition capability: The multi-branch wavelet decoupling model improves the frequency domain separation accuracy by 15%, effectively avoiding noise aliasing.

[0182] Improved training stability: After introducing the noise distribution stability loss, the model convergence speed is improved by about 25%, and the generated signal distribution is smoother.

[0183] More accurate energy conservation: The noise-energy balance constraint controls the double-photon energy deviation within 2keV.

[0184] Application scalability enhancement: through adjustment It can switch freely between different tasks and supports component-level noise optimization and system-level dynamic testing.

[0185] Example 3, the third embodiment of the present invention, further optimizes the following two core aspects based on the overall framework of Example 1:

[0186] In the signal generation step, a Transformer encoder with a multi-head self-attention mechanism is introduced to extract the spatiotemporal correlation features between two photons, and combined with a U-Net decoder to achieve signal reconstruction that conforms to physical constraints.

[0187] In the coordinated control step, a time interpolation and spatial conformal mapping algorithm based on physical constraints is adopted to achieve precise spatiotemporal regulation that conforms to event-level signals.

[0188] This embodiment also runs on a GPU-accelerated environment (NVIDIA A100), with inputs including real PET coincidence event data and detector hardware parameters, and outputs a two-photon signal pair that satisfies the spatiotemporal parameter constraints of the target.

[0189] Working principle of Example 3:

[0190] (1) Signal generation stage

[0191] The core of this stage is a dual-path generation architecture that integrates the Transformer encoder and the U-Net decoder. Its working principle is as follows:

[0192] a. Spatiotemporal feature embedding

[0193] First, the input matching event data is feature-encoded:

[0194] The energy value of the first photon The timestamp of the first photon Detection position of the first photon Vectorize to form the first eigenvector. ;

[0195] Energy value of the second photon The second photon's timestamp , and the detection position of the second photon Vectorize to form the second feature vector. .

[0196] To enhance the model's perception of the spatial topology of the PET system, and The following additional information is embedded in each of them:

[0197] Photon Index Encoding, used to identify the sequential relationship of photons in coincidence events;

[0198] Crystal Ring Positional Encoding, reflecting the relative angle and level information of the detection module in the PET ring array.

[0199] Through this embedding, the model can simultaneously perceive the multi-dimensional correlation of time, space and energy in subsequent attention learning.

[0200] Technical effects:

[0201] The feature embedding layer establishes a high-dimensional representation space for photon events, enabling the network to explicitly distinguish the relationship between double photons under different detection positions and energy states, providing a physical basis for subsequent self-attention learning.

[0202] b. Correlation feature extraction

[0203] The first feature vector and the second feature vector are spliced and input into the Multi-head Self-Attention Layer in the Transformer encoder.

[0204] This layer calculates the interaction attention weight between double photons, specifically including:

[0205] Calculate the query Q, key K and value V for each feature vector.

[0206] Calculate the interaction correlation between the query of the first feature vector and the key of the second feature vector to capture the influence of the first photon on the second photon features.

[0207] At the same time, calculate the weight between the query of the second feature vector and the key of the first feature vector to realize the bidirectional coupling of information.

[0208] The multi-head attention mechanism obtains multiple independent spatio-temporal correlation features through parallel calculation of multiple subspaces, each corresponding to different physical semantics, for example:

[0209] Part of the attention head learns energy coupling relationship ;

[0210] Another part learns time synchronization features Where represents a preset time difference threshold;

[0211] Some heads capture spatial geometric features (the angle and distance distribution of the two photon detection points).

[0212] Finally, the associated feature vector output by the Transformer contains: an implicit representation of the time difference; an implicit representation of the connecting vector of the detection points; an implicit constraint expression of the energy conservation relationship.

[0213] Technical effects:

[0214] Through the multi-head attention mechanism, the model can autonomously learn the spatio-temporal coupling features of the two-photon signal pair without explicitly defining physical formulas, significantly improving the coincidence accuracy of the generated signal in the time and energy dimensions (energy deviation ≤ 2%, time difference deviation ≤ 5%).

[0215] c. Signal decoding and initial coincidence constraint

[0216] The U-Net decoder receives the F_assoc vector output by the Transformer and fuses the detailed features from the bottom layer through a skip connection mechanism to generate the initial two-photon signal pair.

[0217] During the decoding process, an initial coincidence constraint module is introduced to dynamically detect whether the output signal meets the following conditions:

[0218] Energy constraint: ;

[0219] Time constraint: .

[0220] If the output does not meet the conditions, the system automatically triggers gradient-based constraint backpropagation to adjust the decoding weights to ensure physical validity.

[0221] Technical effects:

[0222] Through the constraint-driven decoding mechanism, the generated signal automatically meets the energy and time conservation laws of PET coincidence events, and the generation efficiency is improved by about 30% compared to traditional discriminator-based GAN models.

[0223] (2) Collaborative regulation phase

[0224] In the collaborative regulation step, to further achieve signal fine-tuning under the requirements of the target task, the embodiment uses an interpolation and coordinate mapping algorithm under physical constraints, whose principles are as follows:

[0225] a. Time interpolation adjustment

[0226] The system adjusts the photon timestamps according to the target time difference and the original time difference through a linear interpolation function:

[0227] ;

[0228] and respectively represent the first photon timestamp and the second photon timestamp after difference adjustment.

[0229] The causality of the time sequence is maintained during the interpolation process (i.e. or the original order remains unchanged), ensuring physical rationality.

[0230] Technical effects:

[0231] High-precision control of coincidence event time difference is achieved, with a time adjustment error of not more than 1%, suitable for coincidence circuit time delay testing and time resolution optimization.

[0232] b. Spatial coordinate mapping

[0233] To achieve target adjustment in the spatial dimension, the system maps the photon detection coordinates based on the target detection distance and the original distance using a conformal transformation algorithm:

[0234] Calculate the original detection point connection vector: ;

[0235] Under the premise of maintaining the angle and shape, map to the target distance:

[0236] ; wherein, represents the rotation matrix.

[0237] Re-project the adjusted coordinates , into the PET crystal array coordinate system to ensure that they are still located in the effective detection area.

[0238] Technical effects:

[0239] The spatial mapping process adjusts the target distance while maintaining geometric conformality, ensuring that the signal distribution matches the detector structure, and the spatial fidelity of the generated signal is improved by about 12%, avoiding the problem of signal mapping beyond the effective crystal area.

[0240] Technical effects of embodiment three:

[0241] Compared with embodiments one and two, the present embodiment achieves the following significant technical improvements:

[0242] Enhanced high-dimensional spatiotemporal correlation modeling capability: through the multi-head self-attention mechanism, the model explicitly learns the relationship between double-photon energy, time, and space, making it more general in physical feature capture;

[0243] ​Physical controllability of space-time adjustment is improved: the adjustment mode of interpolation combined with conformal transformation is adopted, and both time and space adjustment are physically reasonable.

[0244] Higher physical consistency of generated signals: the energy conservation error is controlled within 2 keV, the time deviation is ≤5%, and the space deviation is ≤3%, meeting the requirements of system-level simulation and detector performance evaluation.

[0245] Improved model stability and application flexibility: it can adaptively run under different PET system structures (ring array, half-ring array), and has high portability and retrainability.

[0246] In embodiment four, the PET detector signal simulation generation method based on artificial intelligence is mainly aimed at different needs of downstream tasks. Through dynamic data structure reconstruction and information enhancement in the signal output step, task-adaptive output of signals is realized.

[0247] The output module automatically selects different data organization methods and information expansion strategies according to the task type, supporting three typical applications:

[0248] Component-level optimization; system-level testing; extreme event evaluation.

[0249] The system runs on a GPU-accelerated signal simulation platform, relying on the customized two-photon signal pairs and their noise components generated by the aforementioned embodiments to realize the conversion of signals to task data.

[0250] Working principle of embodiment four:

[0251] (1) Component-level optimization mode

[0252] When the downstream task type is component-level optimization, the output module disassembles and indexes the customized two-photon signal pairs to form a "noise-signal pair data set" for a single detector component.

[0253] a. Data structure reconstruction process

[0254] First, the module decomposes each pair of photon signals to extract their noise component spectrum (including SiPM dark current, ADC quantization noise, crystal afterglow, etc.) and pure signal part;

[0255] Then, according to the detector hardware parameters (crystal array position, SiPM number, gain curve, etc.), an index relationship is established;

[0256] Finally, a two-dimensional mapping table is generated, with component identification as the index key, and the column items including noise type, noise amplitude, signal energy, timestamp, and hardware response parameters.

[0257] b. Information enhancement strategy

[0258] The output module adds the correlation index of noise and signal (such as the linear correlation coefficient r of noise power spectral density and energy deviation) in the mapping table, and automatically calculates the influence weight of each type of noise on the signal energy and time characteristics.

[0259] c. Technical effects

[0260] This mode realizes the directional correspondence between noise components and hardware characteristics, so that researchers can directly analyze the noise sensitivity of different components (such as different SiPMs or ADCs) through the mapping table, supporting electronic anti-interference design and component selection optimization. Compared with traditional experimental testing, the component performance evaluation period is shortened by more than 50%.

[0261] (2) System-level test mode

[0262] When the downstream task type is a system-level test, the output module performs coincidence event stream reconstruction and timing atlas generation operations to verify the overall performance of the PET system.

[0263] a. Continuous coincidence event stream construction

[0264] The module sorts one hundred thousand to one million pairs of customized two-photon signals in ascending order of timestamp, and synthesizes a continuous coincidence event stream. 、

[0265] Calculate two indicators for each event: and

[0266] Generate an event stream data structure, where each node contains the energy sum, time difference, detection coordinates, and noise proportion.

[0267] b. System timing atlas generation

[0268] The module generates two key atlases based on the event stream:

[0269] ① Time difference distribution histogram: reflects the system time resolution;

[0270] ② Coincidence count rate curve over time: used to analyze the system dead time and trigger efficiency.

[0271] Various statistical features (such as time jitter variance and event pile-up rate) can be added to the atlas to enhance the timing visualization at the system level.

[0272] c. Technical effects

[0273] ​​The system-level output realizes event serialization and time feature reconstruction, so that the signals generated by the model can be directly used for PET coincidence circuit delay calibration, trigger logic verification and time synchronization algorithm optimization. It has been verified that the results of timing test using the analog signal are less than 3% error compared with real experiment, which greatly improves the efficiency of system-level simulation.

[0274] (3) Extreme event evaluation mode

[0275] When the downstream task type is extreme event evaluation, the output module performs high-energy and large-angle event screening in the customized two-photon signal set to construct an extreme event signal library.

[0276] a. Event screening criteria

[0277] The module automatically selects samples according to the following three types of conditions:

[0278] Energy higher than the preset upper threshold (e.g. or );

[0279] Energy lower than the preset lower threshold (e.g. or );

[0280] Spatial scattering angle > upper limit scattering angle (e.g. 70°).

[0281] b. Database structure design

[0282] The screened extreme signals are stored in the "extreme event signal library", and each record contains the following fields:

[0283] Energy value of two photons ; Time difference ; Spatial scattering angle; Energy deposition; Corresponding noise category and its proportion.

[0284] At the same time, the system adds physical parameter labels (such as "high-energy single-photon event", "multiple scattering coincidence event", "low signal-to-noise ratio extreme sample", etc.) to each record.

[0285] c. Technical effects

[0286] This module realizes the structured sampling and enhanced labeling of rare extreme events, making it possible to evaluate the performance of PET detectors under large-angle scattering, high-energy photons and atypical coincidence events. The similarity between the extreme signals generated by the model and the real high-energy experimental data is more than 88%, providing data support for the study of detector limit response.

[0287] Overall workflow and coordination mechanism:

[0288] The output module dynamically loads corresponding sub-modules according to task instructions during runtime, and completes: structured reconstruction of signal and noise data; automatic enhancement of task-related physical characteristics; seamless switching and parameter sharing between multiple tasks.

[0289] Through interaction with the space-time-noise collaborative controller and the noise decoupler, the output module can still maintain energy conservation constraints and noise distribution consistency during the data export stage, ensuring that the physical characteristics of the signal are not damaged.

[0290] Task-level customization and structured reconstruction of the signal output layer are achieved; output data can be directly used for hardware optimization, system calibration, and extreme event research; data generation efficiency is improved to complete within about 2 hours per million events, significantly reducing research and testing costs.

[0291] Technical effects of embodiment four:

[0292] Compared with embodiments one to three, the present embodiment further realizes the following innovations and improvements:

[0293] Comprehensive enhancement of task adaptability: the output data format can be automatically reconstructed according to the task type, compatible with laboratory testing and industrialization verification scenarios;

[0294] Information enhancement and multi-dimensional labeling: through noise component, energy deposition, and timing feature labeling, the output data has complete physical traceability;

[0295] Simulation-test closed loop: the output signal can be directly imported into the electronics simulation software of the real PET system, realizing virtual-real interactive testing;

[0296] High efficiency and high scalability: supporting batch signal generation and multi-GPU parallel output, meeting the needs of industrial-level signal simulation.

[0297] The above is only a preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solutions falling within the scope of the present application should be considered within the protection scope of the present application. It should be noted that for ordinary technical personnel in the technical field, some improvements and refinements without departing from the principles of the present application should also be considered within the protection scope of the present application.

Claims

1. A method for simulating and generating PET detector signals based on artificial intelligence, characterized in that: Signal generation steps: Obtain real PET coincidence event dataset and detector hardware parameters, input them into a trained two-photon signal generator to generate initial two-photon signal pairs; the two-photon signal generator adopts a neural network architecture that integrates a Transformer encoder and a U-Net decoder. The Transformer encoder learns the spatiotemporal correlation features between two photons, and the U-Net decoder generates initial signal pairs containing mixed noise based on this. Its training process ensures that the output signal meets the preset energy and time difference physical range through initial coincidence constraints; Noise decoupling steps: The initial two-photon signal pair is input to the coincident event noise decoupler. Based on the pre-built PET noise physics prior library, the mixed noise is decoupled into multiple independent physical noise components, and the two-photon clean signal and noise component spectrum are output. The decoupler adopts a multi-branch decoupling network, with each branch corresponding to a type of physical noise. The distribution of the same type of noise between the two photons is constrained by the cross-decoupling attention mechanism to satisfy the annihilation energy conservation. Coordinated regulation steps: Receive the target spatiotemporal parameters and target noise parameters set by the user, and adjust the two-photon pure signal and physical noise components in a coordinated manner through a spatiotemporal-noise co-regulator; the adjustment includes adjusting the timestamp and spatial coordinates according to the target spatiotemporal parameters, scaling the amplitude of the specified noise component according to the intensity adjustment coefficient in the target noise parameters, and synchronously adjusting the noise distribution of the two photons after scaling to maintain energy conservation; Signal output steps: Based on the specified downstream task type, the regulated customized two-photon signal pair is adapted and output as a data format that matches the task.

2. The method for simulating and generating PET detector signals based on artificial intelligence according to claim 1, characterized in that: The physical noise prior library for PET noise includes at least three of the following physical noise components: SiPM dark current noise, ADC quantization noise, Compton scattering noise, crystal afterglow noise, power supply noise, and crosstalk noise.

3. The method for simulating and generating PET detector signals based on artificial intelligence according to claim 1, characterized in that: When performing noise decoupling, the coincident event noise decoupler first decomposes the initial two-photon signal pair into different frequency bands through wavelet transform, and then enables each branch network to extract and decouple the noise components in the corresponding frequency band according to the power spectrum characteristics of the corresponding noise in the PET noise physical prior library.

4. The method for simulating and generating PET detector signals based on artificial intelligence according to claim 1, characterized in that: The noise decoupling step is implemented through a unified, end-to-end deep learning model; The deep learning model is trained using a total loss function consisting of consistency loss and noise reasonableness loss. The consistency loss is calculated using the mean absolute error to determine the spatiotemporal deviation between the generated signal and the real signal, and the formula is as follows: ； in, and These represent the energy values ​​of the first and second photons generated, respectively. This represents the time difference between the generated two-photon signals. This represents the average two-photon time difference in the dataset of truly matching events. This is the time weighting coefficient; The noise rationality loss is calculated using KL divergence to differentiate the distribution of generated noise from that of actual noise. The formula is as follows: ； in, This represents the generated i-th type of noise distribution. This represents the true noise distribution from the aforementioned noise prior library. This indicates the loss of reasonableness of the noise.

5. The method for simulating and generating PET detector signals based on artificial intelligence according to claim 4, characterized in that: The total loss function is composed of the conformity loss, the noise rationality loss, and a newly added noise distribution stability loss, and its expression is configured as follows: ； in, Let the total loss function be... For the noise distribution stability loss, , , These are the learnable weight coefficients for each loss, which are dynamically updated using gradient descent in the early stages of training and fixed in the later stages of training. It is the hyperbolic tangent function. The Gaussian error function is... It is the natural logarithm function; The scaling factor is used to account for the noise distribution stability loss and to balance the input dynamic range of the tanh function. and These are the target expected value and the tolerance standard deviation of the noise reasonableness loss, respectively, used to construct the expected constraint for the noise loss; The shape adjustment factor is used to adjust the stability loss. The noise distribution stability loss is defined as the negative log-likelihood of the variance of each noise component in the generated noise component spectrum, and its expression is: ,in This represents the variance of the i-th type of noise component. It is a very small positive number that prevents numerical overflow.

6. The method for simulating and generating PET detector signals based on artificial intelligence according to claim 4, characterized in that: In the coordinated control step, when adjusting the timestamp and spatial coordinates according to the target spatiotemporal parameters, an interpolation and coordinate mapping algorithm under physical constraints is adopted. This process specifically includes: Time interpolation adjustment: Based on the target time difference and the original time difference, the first photon timestamp and the second photon timestamp are synchronously adjusted through a linear interpolation function so that the adjusted time difference is equal to the absolute value of the difference between the first photon timestamp and the second photon timestamp, thus maintaining the causality of the photon time series during the time interpolation process; Spatial coordinate mapping: Based on the target spatial distance and the original spatial distance, the first photon detection position and the second photon detection position are synchronously adjusted through a conformal transformation algorithm to ensure that the distance between the two points after adjustment is equal to the target spatial distance, and the mapped coordinates are still located within the effective crystal array area of ​​the PET detector.

7. The method for simulating and generating PET detector signals based on artificial intelligence according to claim 4, characterized in that: The intensity adjustment coefficient in the target noise parameter is defined as Its value range is ,in It is 0.

5. It is 2.0; When the amplitude of the specified noise component is scaled, the coordinated control step ensures that the amplitude of the same type of noise between the two photons changes synchronously in proportion through the cross-decoupling attention mechanism.

8. The method for simulating and generating PET detector signals based on artificial intelligence according to claim 7, characterized in that: In the coordinated control step, when scaling the amplitude of the specified noise component according to the intensity adjustment coefficient, a noise-energy balance constraint mechanism is introduced. The noise-energy balance constraint mechanism is configured to: scale the amplitude of the i-th type of noise component After multiplying, the single-photon energy deviation caused by this type of noise change is calculated, and the pure signal energy of the two-photon is reverse-compensated accordingly, so that the absolute value of the deviation between the total energy of the two-photon after compensation and the total energy before compensation is not greater than a preset threshold.

9. The method for simulating and generating PET detector signals based on artificial intelligence according to claim 6, characterized in that: In the signal generation step, the Transformer encoder extracts the spatiotemporal correlation features between two photons through a multi-head self-attention mechanism, specifically including: Spatiotemporal feature embedding: The first photon energy value, the first photon timestamp, and the first photon detection position are encoded into a first feature vector, the second photon energy value, the second photon timestamp, and the second photon detection position are encoded into a second feature vector, and position encoding information representing the photon sequence number and the detector crystal ring position is incorporated into the first feature vector and the second feature vector; Correlation feature extraction: The first feature vector and the second feature vector are concatenated and input into the multi-head self-attention layer of the Transformer encoder; the self-attention layer is configured to calculate the interaction attention weight between the first feature vector and the second feature vector to explicitly model the spatiotemporal correlation of the two-photon signal pair and output the correlation feature vector; the correlation feature vector includes at least the implicit representation of the adjusted time difference and the implicit representation of the vector connecting the spatial probe points; Signal Decoding: The associated feature vector is input to the U-Net decoder, which reconstructs the initial two-photon signal pair that conforms to the initial conformity constraint based on the associated feature vector and the underlying information transmitted by the skip connections.

10. The method for simulating and generating PET detector signals based on artificial intelligence according to claim 1, characterized in that: The signal output step, based on the downstream task type, performs dynamic data structure reconstruction and information enhancement on the customized two-photon signal pair, specifically including: When the downstream task is component-level optimization, the output module deconstructs the customized two-photon signal pair to generate a noise-signal comparison dataset indexed by a single detector component. The dataset separates the various physical noise components after decoupling from the corresponding clean signals and associates them with the detector hardware parameters to form a two-dimensional mapping table for component performance analysis. When the downstream task is a system-level test, the output module sorts the 100,000 to 1 million customized two-photon signal pairs by timestamp, synthesizes a continuous coincidence event stream, and generates a corresponding system timing graph. The system timing graph includes at least a histogram of the time difference distribution of each coincidence event and a curve of the coincidence count rate changing with time, which is used for the calibration of the system dead time and coincidence resolution time. When the downstream task is extreme event assessment, the output module selects signals from the customized two-photon signal pairs whose energy values ​​are greater than the preset upper energy threshold or less than the lower energy threshold, or whose spatial scattering angle is greater than the upper dispersion angle, to construct an extreme event signal library, and labels each signal in the library with its corresponding physical parameters, including energy deposition, scattering angle and noise category.

Images