Method for correcting beta applicator uniformity based on deep learning scintillator detector
By using deep learning to correct crosstalk in the scintillator detector, the problem of low measurement accuracy of beta patch uniformity was solved, and high-precision beta particle incident position correction and uniformity distribution map generation were achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA INST FOR RADIATION PROTECTION
- Filing Date
- 2025-10-29
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies cannot effectively correct optical and electronic crosstalk in scintillator detectors, resulting in low measurement accuracy of β-applier uniformity.
By employing a deep learning-based approach, a scintillator array detector measurement system was built. Combining experiments and numerical simulations, a CNN network model was established to model optical crosstalk and electronic noise. The crosstalk correction model was then trained to achieve end-to-end signal correction.
The uniformity measurement accuracy of the β-applied device was improved, with the position error reduced from 2.1 mm to 0.4 mm and the crosstalk rate reduced from 32% to 12%, meeting the requirements of clinical measurement. The model has good generalization and strong real-time performance.
Smart Images
Figure CN121541247B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of nuclear medicine equipment calibration technology, and in particular to a method for calibrating the uniformity of a beta patch using a deep learning-based scintillator detector. Background Technology
[0002] Beta patchers are commonly used in clinical nuclear medicine to treat superficial skin lesions (such as hemangiomas and keloids). The uniformity of their radioactive distribution directly determines the accuracy of the clinical treatment dose—excessive uniformity deviation may lead to insufficient dose to the lesion area or excessive dose to normal tissue, causing treatment failure or complications. Currently, the measurement of beta patch uniformity mainly relies on scintillator array detectors, which typically consist of multiple small scintillator units arranged in a two-dimensional array. Each unit can be made of a different material (such as NaI(Tl), CsI(Tl), etc.) and is designed to respond independently to incident radiation. This structure significantly improves spatial resolution, enabling precise localization of radiation sources within a small area. When irradiated, each scintillator unit produces scintillation light, which is converted into an electrical signal by a photodetector (such as a photomultiplier tube or silicon photodetector). Due to the array design, multiple units can operate simultaneously, allowing the detector to process signals from multiple radiation events at the same time. This multi-channel parallel processing capability of scintillator array detectors greatly improves detection efficiency, especially in high count rate environments, effectively reducing signal overlap and dead time. Scintillator array detectors are typically equipped with advanced electronic systems, including high-gain amplifiers, analog-to-digital converters (ADCs), and digital signal processing units. These systems can process and analyze the received electrical signals in real time, enabling rapid acquisition of dose rate information within the target area.
[0003] Commonly used scintillator array detectors include SiPM scintillator detectors, which have the advantages of high sensitivity and fast response speed. However, they suffer from significant crosstalk problems in practical applications. On the one hand, optical crosstalk causes photons to leak from one SiPM pixel to adjacent pixels; on the other hand, electronic crosstalk causes signals to couple with each other in the readout circuit. These two types of crosstalk can cause a decrease in the position resolution of the detector output signal, energy estimation deviation, and thus measurement artifacts, resulting in distorted evaluation results of β-applier uniformity. Traditional crosstalk correction methods (such as reflective layer isolation and matrix inversion filtering) have obvious limitations: reflective layer isolation can only alleviate optical crosstalk and is ineffective against electronic crosstalk; matrix inversion is sensitive to noise, and its correction effect decreases significantly in high-density SiPM arrays (such as 4×4 and above), failing to meet the high-precision requirements of β-applier uniformity measurement. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a method for correcting the uniformity of a β-applier using a deep learning-based scintillator detector. This method solves the technical problem in existing technologies where optical and electronic crosstalk of radiation signals cannot be effectively corrected, resulting in low measurement accuracy of β-applier uniformity.
[0005] The technical solution adopted in this invention is as follows:
[0006] This invention provides a method for correcting the uniformity of a β-applier using a deep learning-based scintillator detector, comprising:
[0007] A scintillator array detector measurement system was constructed to obtain the radiation signal of the β patch radiation acting on the detector.
[0008] The detector was irradiated with a calibration source using an experimental method. Radiation signals from the calibration source were collected at a known location of the β-particle incident point. The detector was then irradiated with a β-patch, and radiation signals from different radioactive regions of the β-patch were collected as signals to be calibrated. The signals to be calibrated and the radiation signals from the calibration source were combined to form the original experimental signal.
[0009] Numerical simulation is used to model optical crosstalk and electronic noise, and to simulate the process of β particles incident on the measurement system under the same scenario as the experimental method, to obtain a simulated signal tensor containing β particle incident position and energy, crosstalk parameters, signal amplitude and time waveform information of each channel, including a crosstalk-free signal tensor corresponding to a crosstalk parameter value of zero and a crosstalk-inducing signal tensor corresponding to a crosstalk parameter value of non-zero.
[0010] The original experimental signals are preprocessed to obtain a dataset;
[0011] Add labels to the dataset to obtain training data; the labeling includes: using the signal amplitude and time waveform of each channel in the crosstalk-free signal tensor as the reference truth value, adding reference truth value labels to the corresponding original experimental signals; establishing the correspondence between the signal amplitude and the β particle incident position in the crosstalk-affected signal tensor, transferring the correspondence to the original experimental signal, and obtaining the β particle incident position corresponding to the original experimental signal as the true position label;
[0012] The CNN network is trained using the training data to obtain a crosstalk correction model;
[0013] The crosstalk correction model was used to correct the radiation signal of different radioactive regions of the target β patch, and the corrected signal and the incident position of β particles were obtained.
[0014] The target β-applied device is uniformly corrected based on the corrected signal.
[0015] The preferred technical solution is as follows:
[0016] The crosstalk parameters include the standard deviation of the noise and the optical crosstalk probability; the preprocessing of the original experimental signal includes normalization and data augmentation, wherein the data augmentation includes:
[0017] Gaussian noise is set according to the standard deviation of the noise set during the numerical simulation process and added to the original experimental signal;
[0018] Based on the optical crosstalk probability set during the numerical simulation, crosstalk between adjacent channels is randomly introduced into each channel of the original experimental signal to obtain signals with different crosstalk intensities.
[0019] Establishing the correspondence between the signal amplitude and the incident position of the β particle in the crosstalk signal tensor includes:
[0020] The corresponding relationship is established using the incident position of the β particle in the crosstalk signal tensor and the signal amplitude of each channel, based on the centroid method formula, which is as follows:
[0021]
[0022] in, The crosstalk signal tensor mentioned above is the first one. i The signal amplitude of the first channel, the first j The signal amplitude of each channel x , y These are the x and y coordinates of the incident position of the β particle in the crosstalk signal tensor, respectively.
[0023] The CNN network includes an encoder, a bottleneck layer, a decoder, and a position estimation head connected in sequence.
[0024] The encoder includes two levels of one-dimensional convolutional layers and a max pooling layer, used to extract local features from the original experimental signal and output the local features.
[0025] The bottleneck layer is a fully connected layer used to fuse the local features with global features to handle non-local crosstalk and output fused features.
[0026] The decoder includes two levels of one-dimensional deconvolution, which takes the fused features as input and outputs a corrected signal.
[0027] The position estimation head employs a fully connected layer to output the incident position of the β particle, using the corrected signal as input.
[0028] During the training of the CNN network, a multi-task loss function is used, the expression of which is as follows:
[0029]
[0030] in, These represent the loss function value, the mean square error between the corrected signal and the true reference value, and the mean square error between the estimated incident position and the true position of the β particle, respectively.
[0031] The process involves irradiating the detector with a calibration source, collecting the calibration source radiation signal at a known location of the β-particle incident point, and irradiating the detector with a β-patch, collecting radiation signals from different radioactive regions of the β-patch as signals to be calibrated. This includes:
[0032] use 137 The Cs calibration source collects calibration source radiation signals by irradiating different positions of the detector with a collimated beam;
[0033] The beta patch is fixed directly above the detector so that beta rays can directly irradiate the detector through the shortest path, and radiation signals from different radioactive regions of the beta patch are collected.
[0034] The uniformity correction of the target β-application based on the corrected signal includes:
[0035] Based on the β-particle incident position output by the crosstalk correction model, the detector measurement area is divided into several pixels, and the radioactivity density of each pixel is calculated to obtain the radioactivity density of different radioactive regions of the β-applicator. Then, a uniformity distribution map is generated to complete the uniformity correction.
[0036] During the experiment, the radiation signals of different radioactive regions of the β-applied device were collected using the baffle covering method.
[0037] The measurement system employs a SiPM scintillator detector measurement system, which includes a multilayer LSO or GAGG scintillator crystal, a 4×4 or larger SiPM array, a signal acquisition module, and a data processing module; the scintillator crystal and the SiPM array are coupled through an optical coupler; the signal acquisition module has a sampling rate >1 GHz.
[0038] The range of β-particle energy used in the numerical simulation was adjusted according to the radiation energy range of the β-applier.
[0039] The technical solution of the present invention can achieve at least some of the following beneficial effects:
[0040] This invention utilizes deep learning to learn complex crosstalk patterns, constructing a crosstalk correction model within a CNN framework. This model achieves end-to-end correction of the crosstalk signal by learning the inverse mapping between the crosstalk signal (mixed with optical and electronic noise) and the true signal. The model exhibits good generalization, high reliability, and strong real-time performance, making it particularly suitable for real-time correction of radiation signals acquired by high-density arrays under optical and electronic crosstalk conditions. Furthermore, the model can directly output the incident position of β particles, providing accurate coordinates for generating uniformity distribution maps.
[0041] This invention employs a combination of simulation and experimentation to provide reliable data for model training. On one hand, simulation can directly provide crosstalk-free reference signals and precise location labels, solving the problem of not being able to directly obtain real signals in experiments. On the other hand, simulated data can provide controllable crosstalk parameters for the preprocessing of the original experimental signals, providing a standardized perturbation basis for data augmentation.
[0042] Other features and advantages of the invention will be set forth in the following description or may be learned by practicing the invention. Attached Figure Description
[0043] Figure 1 This is a flowchart illustrating the method of an embodiment of the present invention.
[0044] Figure 2 A schematic diagram of the structure of the β-application device according to an embodiment of the present invention.
[0045] Figure 3 This is a schematic diagram of the baffle structure used in the baffle covering method of this invention.
[0046] Explanation of reference numerals in the attached figures: 1. Inactive area; 2. Radioactive active area; 3. Unit area; 4. Baffle; 41. Hole. Detailed Implementation
[0047] The specific embodiments of the present invention are described below with reference to the accompanying drawings.
[0048] See Figure 1 This embodiment provides a method for correcting the uniformity of a β-applier using a deep learning-based scintillator detector, comprising:
[0049] S1. Construct a scintillator array detector measurement system to obtain the radiation signal of the β patch acting on the detector.
[0050] As a preferred embodiment, it includes a multilayer LSO or GAGG scintillator crystal (preferably a 6×6 array), a 4×4 or larger SiPM array (preferably a 6×6 array with 36 channels), a signal acquisition module, and a data processing module; the scintillator crystal and the SiPM array are coupled through an optical coupler to reduce additional crosstalk caused by interface reflection; the signal acquisition module has a sampling rate > 1 GHz. The thickness of the scintillator crystal is preferably 15 mm to accommodate β-applicators of different sizes.
[0051] As a preferred method, the β applicator is a ³²P applicator (active area 50×50mm); the β applicator is preferably fixed 10mm directly above the detector so that the β rays directly irradiate the detector with the shortest path.
[0052] S2. Using an experimental method, the detector is irradiated with a calibration source, and the radiation signal of the calibration source is collected at a known location of the β particle incident point; the detector is irradiated with a β patch, and the radiation signals of different radioactive regions of the β patch are collected as the signal to be calibrated; the signal to be calibrated and the radiation signal of the calibration source are combined to form the original experimental signal.
[0053] As a preferred method, use 137 A Cs calibration source (energy 662keV) illuminates different positions of the detector. The incident point of β particles at a known position is determined by a collimated beam (aperture preferably 0.5mm), and the radiation signal of the calibration source is collected.
[0054] As a preferred method, the radiation signals from different radioactive regions of the β-application were collected using a baffle covering method. (See also...) Figure 2 , Figure 2 (a) is a schematic diagram of the structure of the β patch, which includes two parts: the radioactive active region 2 and the inactive region 1. Figure 2 In diagram (b), the radioactive active region 2 is divided into several smaller unit regions 3, each corresponding to a radioactive area of the β-application device. The unit regions 3 in the diagram are in an array format and are for illustrative purposes only, not representing their actual shape. See also... Figure 3 The diagram shows the structure of the baffle 4. The baffle 4 has holes 41, the shape of which matches the outline of the divided unit region 3, and can be regular or irregular. When acquiring signals using the baffle covering method, each time the baffle 4 is used to cover the remaining areas and exposes one unit region 3 to be measured through the holes 41, the emissivity of that unit region 4 can be measured using the measurement system of this embodiment. By moving the baffle 4, different unit regions 3 are measured, thus allowing for a complete measurement of the entire radioactive active region 2. The baffle covering method can effectively reduce interference from other areas, ensuring the accuracy of the measurement results.
[0055] As a preferred approach, the detector temperature needs to be controlled (25±1℃) during the experimental data acquisition process to avoid temperature changes affecting the SiPM performance.
[0056] S3. A numerical simulation method is used to model optical crosstalk and electronic noise, simulating the process of β particles incident on the measurement system under the same scenario as the experimental method, and obtaining a signal tensor containing the incident position and energy of β particles, crosstalk parameters, signal amplitude and time waveform information of each channel, including a crosstalk-free signal tensor corresponding to a crosstalk parameter value of zero and a crosstalk-inducing signal tensor corresponding to a crosstalk parameter value of non-zero.
[0057] The signal tensor has dimensions (B, N, T), where B is the event batch size, N is the number of SiPM channels, and T is the waveform time point. Similarly, to standardize the data format, the original experimental signal obtained from the experiment is also converted into tensor form according to the above format.
[0058] Specifically, numerical simulation tools such as Geant4 or GATE are used to simulate the process of the β particle incident detector.
[0059] Specifically, the crosstalk parameters include the standard deviation of noise and the optical crosstalk probability, wherein the optical crosstalk probability is preferably 5%-18%, and the standard deviation of electronic noise is preferably 5%. (Generate 1.2×10) 6 The signal tensor for each event (B=128, N=36, T=120). Correspondingly, the calibration source radiation signal acquired in the experiment includes 8 × 10⁻⁶ events. 5 The original waveforms of each event, where each event can be understood as calibration source parameters including size, area, distribution, activity, etc.; radiation signals from different radioactive regions of the β-applied device were collected in the experiment, including 4 × 10⁻⁶. 5 The original waveform of each event.
[0060] As a specific method, the β-particle energy range used in the numerical simulation is based on clinically commonly used β-patch devices, such as... 32 The radiation energy range of the P-type applicator and the 90Sr-90Y applicator is adjusted to ensure that the simulated scenario matches the actual measurement scenario. In this embodiment, 0.12MeV is preferred.
[0061] S4. Preprocess the original experimental signal to obtain a dataset; the preprocessing includes normalization and data augmentation.
[0062] Specifically, the normalization process is as follows:
[0063] In the formula , These are the signals before and after normalization, respectively. , These are the channel mean and channel standard deviation, respectively.
[0064] Preferably, the data enhancement includes:
[0065] Gaussian noise is set according to the standard deviation of the noise set during the numerical simulation process and added to the original experimental signal;
[0066] Based on the optical crosstalk probability set during the numerical simulation, crosstalk between adjacent channels is randomly introduced into each channel of the original experimental signal to obtain signals with different crosstalk intensities.
[0067] As a preferred embodiment, the preprocessing also includes baseline correction: subtracting the waveform baseline to remove offset components in the signal, and feature extraction, calculating features such as the peak amplitude, integral charge, and rise time of the signal as auxiliary inputs.
[0068] S5. Add labels to the dataset to obtain training data; the addition of labels includes: using the signal amplitude and time waveform of each channel in the crosstalk-free signal tensor as the reference truth value, adding reference truth value labels to the corresponding original experimental signals; establishing the correspondence between the signal amplitude and the β particle incident position in the crosstalk-affected signal tensor, transferring the correspondence to the original experimental signal, and obtaining the β particle incident position corresponding to the original experimental signal as the true position label.
[0069] Specifically, establishing the correspondence between the signal amplitude and the incident position of the β particle in the crosstalk signal tensor includes:
[0070] The corresponding relationship is established using the incident position of the β particle in the crosstalk signal tensor and the signal amplitude of each channel, based on the centroid method formula, which is as follows:
[0071]
[0072] in, The crosstalk signal tensor mentioned above is the first one. i The signal amplitude of the first channel, the first j The signal amplitude of each channel x , y These are the x and y coordinates of the incident position of the β particle in the crosstalk signal tensor, respectively.
[0073] S6. Use the training data to train a CNN network (convolutional neural network) to obtain a crosstalk correction model.
[0074] Specifically, the CNN network combines UNet (for feature extraction and upsampling) and fully connected layers (for inter-channel fusion). This architecture is suitable for handling spatial correlations (inter-channel crosstalk is similar to pixel correlation in an image). The specific architecture is as follows:
[0075] It includes an encoder (downsampling path), a bottleneck layer, a decoder (upsampling path), and a position estimation head, which are connected in sequence.
[0076] The encoder input is the preprocessed raw experimental signal. The encoder includes two levels of one-dimensional convolutional layers (Conv1D) and max-pooling layers (MaxPooling1D). The two Conv1D levels are used to extract local features (such as waveform peaks) from the original experimental signal and output the local features. After processing with the ReLU activation function, the signal is downsampled by MaxPooling1D to reduce the data dimensionality. Preferably, the first-level Conv1D layer has a kernel size of 3, padding of 1, and 32 output channels; the second-level Conv1D layer has 64 output channels; and the MaxPooling1D layer has a pooling size of 2.
[0077] The bottleneck layer employs a fully connected layer to fuse local features with global features to handle non-local crosstalk and output fused features. Preferably, the fully connected layer has an input dimension of 64×(N / 4), where N is the number of channels (36), and an output dimension of 128.
[0078] The decoder includes two levels of one-dimensional deconvolution (Deconv1D) to restore the data dimension and perform upsampling. Preferably, the first-level Deconv1D layer has 64 input channels and 32 output channels, and the second-level Deconv1D layer has 1 output channel. High-resolution features are preserved by skipping connections to the corresponding encoder layers, improving the model's ability to utilize detailed information. The decoder takes the fused features as input and outputs the corrected signal. .
[0079] The position estimation head employs a fully connected layer (36 inputs, 2 outputs) to output the β-particle incident position (x, y) from the corrected signal as input.
[0080] Specifically, during the training of the CNN network, a multi-task loss function is used, the expression of which is as follows:
[0081]
[0082] in, These represent the loss function value, the mean square error between the corrected signal and the true reference value, and the mean square error between the estimated incident position and the true position of the β particle, respectively.
[0083] Specifically, the CNN network training process involves dividing the training data into an 80% training set, a 10% validation set, and a 10% test set; using the Adam optimizer (initial learning rate 0.001, decaying by 0.96 every 10 epochs), a batch size of 32,128, and 50,200 training epochs, while introducing time-dependent physical constraints. , t i For subsequent firing timing, t 0 For initial timing, a i For coefficients, t c The time constant is used, an early stopping strategy is adopted (the patience value of loss is verified to be 10 epochs), and GPU is used to accelerate training.
[0084] As a specific approach, during model training, a pre-trained PET attenuation correction model (such as UNet) can be introduced for fine-tuning to accelerate convergence (convergence speed is improved by 40%).
[0085] Specifically, the trained model is exported to an FPGA or ASIC in ONNX format to achieve real-time correction of the uniformity measurement of the β-applier, with a single event processing time of <1ms.
[0086] S7. The crosstalk correction model is used to correct the radiation signal of different radioactive regions of the target β patch to obtain the corrected signal and the incident position of the β particles.
[0087] Specifically, the measurement system of this embodiment measures the radiation signal of different radioactive regions of the target β patch, and inputs it into the trained crosstalk correction model to obtain the corrected signal and the incident position of the β particles.
[0088] S8. Perform uniformity correction on the target β applicator based on the corrected signal.
[0089] As a preferred method, based on the incident position of the β particle output by the crosstalk correction model, the detector measurement area is divided into several pixels (e.g., 100×100 pixels), the radioactivity density of each pixel is calculated, thus obtaining the radioactivity density of different radioactive regions of the β patch, and then a uniformity distribution map is generated to complete the uniformity correction.
[0090] Specifically, when evaluating uniformity, multiple measurements (≥5 times) are required to take the average value, further reducing the impact of random noise.
[0091] Specifically, the evaluation indicators for the uniformity of the β-applier include surface source uniformity (deviation ≤5%) and regional activity coefficient of variation (CV ≤8%), both of which are calculated based on the corrected signal.
[0092] Cross-validation was used to compare the results of the deep learning-based method and the baseline (uncorrected) method for evaluating the uniformity of the β-application. The results showed that the method in this embodiment reduced the positional error from 2.1 mm to 0.4 mm and the crosstalk rate from 32% to 12%. Furthermore, the surface source uniformity deviation was 3.2%, and the coefficient of variation of region activity was 6.5%, demonstrating significant correction effects that meet clinical measurement requirements. In addition, the method in this embodiment offers strong real-time performance: the model parameter scale is only 10. 4 ~10 5 After exporting to the FPGA, the single event processing time is <1ms, adapting to the rapid measurement requirements of β-applicators. This embodiment demonstrates good generalization: it supports different sizes of SiPM arrays (4×4 and above) and β-applicators (…). 32 (e.g., P, 90Sr-90Y), which can be fine-tuned to adapt to different measurement scenarios. The method in this embodiment has high reliability: it adopts a training method that integrates simulation and experimental data, combined with physical constraints, to avoid model overfitting and ensure strong stability of measurement results.
[0093] The inventive concept of the method in this embodiment is further explained below.
[0094] First, the core invention lies in using deep learning to learn complex crosstalk patterns, constructing a crosstalk correction model using a CNN framework: assuming the original signal vectors of the detector's N channels are... Crosstalk causes the true signal to be confused, therefore In the formula, M For crosstalk matrix, n Noise, real signal CNNs learn inverse mappings This enables end-to-end correction of crosstalk signals. Simultaneously, the model integrates a position estimation head, directly outputting the incident position of the β-particles, providing accurate coordinates for generating the uniformity distribution map.
[0095] Secondly, a combination of simulation and experimentation is used to provide reliable data for model training.
[0096] On the one hand, simulation can directly provide crosstalk-free reference signals and precise position labels, solving the problem of not being able to directly obtain real signals in experiments. Modeling parameters that disable optical crosstalk and electronic noise can be set (e.g., relevant parameters can be set to 0), directly outputting a crosstalk-free real signal tensor, including the pure signal amplitude and time waveform of each channel. This signal serves as the "reference truth value" for labeling the original experimental signal during preprocessing. Aligning the original experimental signal with the simulated crosstalk-free signal determines the deviation caused by crosstalk in the experimental signal, providing a reference for subsequent "signal correction direction." Precise position labels are generated: the simulated signal tensor contains the "known incident position" parameter of the β particle. During simulation, the particle incident coordinates can be preset, and combined with the centroid method formula, the correspondence between "signal amplitude distribution - incident position" can be established by calculating the signal amplitude of each channel in the simulated signal tensor. Transferring this relationship to the original experimental signal allows the derivation of the true incident position label of the β particle based on the amplitude distribution of the experimental signal, solving the problem of limited accuracy in collimated beam calibration positions in experiments.
[0097] On the other hand, in the preprocessing of the original experimental signals, simulated data can provide controllable crosstalk parameters, providing a standardized perturbation basis for data augmentation. Guiding Gaussian noise addition: When modeling simulated data, a standard deviation range for electronic noise (e.g., 5%) is preset. This range is determined based on the actual electronic crosstalk characteristics of the detector. When augmenting the original experimental signals, 5%-10% Gaussian noise can be added directly, referring to the noise parameters of the simulated data—ensuring that the noise perturbation matches the noise level of the real detector while avoiding signal distortion due to excessive noise, ensuring that the augmented data still closely resembles the actual measurement scenario. Guiding random crosstalk introduction: The signal tensor of the simulated data contains "crosstalk parameter" information, such as an optical crosstalk probability of 0-20%, which corresponds to the proportion of signal leakage between adjacent channels. When preprocessing the original experimental signals, 5%-15% of crosstalk between adjacent channels can be randomly introduced into the experimental signal according to the crosstalk coefficient range of the simulated data. Part of the signal amplitude from channel 1 is superimposed onto channel 2 to simulate scenarios with different crosstalk intensities, compensating for the limitation of only obtaining signals with a fixed crosstalk level in the experiment.
[0098] It will be understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for correcting the uniformity of a β-applier using a deep learning-based scintillator detector, characterized in that, include: A scintillator array detector measurement system was constructed to obtain the radiation signal of the β patch radiation acting on the detector. The detector was irradiated with a calibration source using an experimental method. Radiation signals from the calibration source were collected at a known location of the β-particle incident point. The detector was then irradiated with a β-patch, and radiation signals from different radioactive regions of the β-patch were collected as signals to be calibrated. The signals to be calibrated and the radiation signals from the calibration source were combined to form the original experimental signal. Numerical simulation is used to model optical crosstalk and electronic noise, and to simulate the process of β particles incident on the measurement system under the same scenario as the experimental method, to obtain a simulated signal tensor containing β particle incident position and energy, crosstalk parameters, signal amplitude and time waveform information of each channel, including a crosstalk-free signal tensor corresponding to a crosstalk parameter value of zero and a crosstalk-inducing signal tensor corresponding to a crosstalk parameter value of non-zero. The original experimental signals are preprocessed to obtain a dataset; Labels are added to the dataset to obtain training data; The tagging process includes: using the signal amplitude and time waveform of each channel in the crosstalk-free signal tensor as the reference truth value, adding reference truth value tags to the corresponding original experimental signals; establishing the correspondence between the signal amplitude and the β particle incident position in the crosstalk-affected signal tensor, transferring the correspondence to the original experimental signals, and obtaining the β particle incident position corresponding to the original experimental signals as the true position tag. The CNN network is trained using the training data to obtain a crosstalk correction model; The crosstalk correction model was used to correct the radiation signal of different radioactive regions of the target β patch, and the corrected signal and the incident position of β particles were obtained. The target β-applied device is uniformly corrected based on the corrected signal.
2. The method according to claim 1, characterized in that, The crosstalk parameters include the standard deviation of the noise and the optical crosstalk probability; the preprocessing of the original experimental signal includes normalization and data augmentation, wherein the data augmentation includes: Gaussian noise is set according to the standard deviation of the noise set during the numerical simulation process and added to the original experimental signal; Based on the optical crosstalk probability set during the numerical simulation, crosstalk between adjacent channels is randomly introduced into each channel of the original experimental signal to obtain signals with different crosstalk intensities.
3. The method according to claim 1 or 2, characterized in that, Establishing the correspondence between the signal amplitude and the incident position of the β particle in the crosstalk signal tensor includes: The corresponding relationship is established using the incident position of the β particle in the crosstalk signal tensor and the signal amplitude of each channel, based on the centroid method formula, which is as follows: in, The crosstalk signal tensor mentioned above is the first one. i The signal amplitude of the first channel, the first j The signal amplitude of each channel x , y These are the x and y coordinates of the incident position of the β particle in the crosstalk signal tensor, respectively.
4. The method according to claim 1, characterized in that, The CNN network includes an encoder, a bottleneck layer, a decoder, and a position estimation head connected in sequence. The encoder includes two levels of one-dimensional convolutional layers and a max pooling layer, used to extract local features from the original experimental signal and output the local features. The bottleneck layer is a fully connected layer used to fuse the local features with global features to handle non-local crosstalk and output fused features. The decoder includes two levels of one-dimensional deconvolution, which takes the fused features as input and outputs a corrected signal. The position estimation head employs a fully connected layer to output the incident position of the β particle, using the corrected signal as input.
5. The method according to claim 1, wherein during the training of the CNN network, a multi-task loss function is used, the expression of which is as follows: in, These represent the loss function value, the mean square error between the corrected signal and the true reference value, and the mean square error between the estimated incident position and the true position of the β particle, respectively.
6. The method according to claim 1, characterized in that, The process involves irradiating the detector with a calibration source, collecting the calibration source radiation signal at a known location of the β-particle incident point, and irradiating the detector with a β-patch, collecting radiation signals from different radioactive regions of the β-patch as signals to be calibrated. This includes: use 137 The Cs calibration source collects calibration source radiation signals by irradiating different positions of the detector with a collimated beam; The beta patch is fixed directly above the detector so that beta rays can directly irradiate the detector through the shortest path, and radiation signals from different radioactive regions of the beta patch are collected.
7. The method according to claim 1, characterized in that, The uniformity correction of the target β-application based on the corrected signal includes: Based on the β-particle incident position output by the crosstalk correction model, the detector measurement area is divided into several pixels, and the radioactivity density of each pixel is calculated to obtain the radioactivity density of different radioactive regions of the β-applicator. Then, a uniformity distribution map is generated to complete the uniformity correction.
8. The method according to claim 1, characterized in that, During the experiment, the radiation signals of different radioactive regions of the β-applied device were collected using the baffle covering method.
9. The method according to claim 1, characterized in that, The measurement system employs a SiPM scintillator detector measurement system, which includes a multilayer LSO or GAGG scintillator crystal, a 4×4 or larger SiPM array, a signal acquisition module, and a data processing module; the scintillator crystal and the SiPM array are coupled through an optical coupler; the signal acquisition module has a sampling rate >1 GHz.
10. The method according to claim 1, characterized in that, The range of β-particle energy used in the numerical simulation was adjusted according to the radiation energy range of the β-applier.