Machine learning based bragg peak probe dosimetry system and method

Abstract

The application discloses a kind of Bragg peak detector dose depth measurement system and method based on machine learning, it is characterized in that, including: Bragg peak detector array component, front-end readout electronics module, data acquisition and synchronization module and machine learning processing module;Bragg peak detector array component is used to convert received target particle beam into ionization signal;Front-end readout electronics module is used to respectively carry out charge sensitive amplification, shaping filtering and analog-digital conversion processing to the ionization signal output by each array component to obtain digital dose sequence;Data acquisition and synchronization module are used to synchronously collect the digital dose sequence output by front-end readout electronics module, and the spatial registration is carried out to the digital dose sequence collected, to obtain initial dose-depth distribution curve;Machine learning processing module is used to extract features and physical constraints to initial dose-depth distribution curve by preset depth learning model to obtain the dose depth measurement result of target particle beam.

Description

Technical Field

[0001] This invention relates to the field of particle beam radiation processing technology, and in particular to a machine learning-based Bragg peak detector dose depth measurement system and method. Background Technology

[0002] Particle radiotherapy, with its unique Bragg peak physics, can deposit the main dose onto the tumor target area while maximally protecting surrounding normal tissue. The precise location of the Bragg peak, i.e., the depth at which the dose maximum is located in the beam dose depth distribution, can provide reference information for clinical research. For example, physicians can use this information to determine whether the lesion target area can receive the expected prescription dose. Therefore, to provide a basis for clinical research, it is necessary to measure and monitor the Bragg peak position.

[0003] Currently, the Bragg peak region is typically approximated as a Gaussian function using nonlinear least-squares fitting to determine beam dose depth information. However, the shape of the Gaussian function differs significantly from the true distribution of the Bragg peak and is prone to getting trapped in local optima. Therefore, approximating the Gaussian function leads to low accuracy in determining beam measurement information such as dose depth. Furthermore, Gaussian fitting requires iterative solutions, which is time-consuming and labor-intensive. Summary of the Invention

[0004] This invention provides a Bragg peak detector dose depth measurement system and method based on machine learning, which mainly improves the measurement efficiency and accuracy of beam dose depth and other measurement information.

[0005] According to a first aspect of the present invention, a Bragg peak detector dose depth measurement system based on machine learning is provided, comprising: a Bragg peak detector array assembly, a front-end readout electronics module, a data acquisition and synchronization module, and a machine learning processing module; The Bragg peak detector array assembly is arranged sequentially along the target particle beam path along the target particle beam inlet direction, and is used to convert the received target particle beam into an ionization signal. The front-end readout electronics module is electrically connected to the Bragg peak detector array assembly and is used to sequentially perform charge-sensitive amplification, shaping filtering and analog-to-digital conversion processing on the ionization signals output by each array assembly to obtain a digital dose sequence. The data acquisition and synchronization module is electrically connected to the front-end readout electronics module and is used to synchronously acquire the digital dose sequence output by the front-end readout electronics module and perform spatial registration on the acquired digital dose sequence to obtain the initial dose-depth distribution curve sampled layer by layer along the target particle beam direction. The machine learning processing module is electrically connected to the data acquisition and synchronization module, and has a built-in preset deep learning model. It is used to extract features and physically constrain the initial dose-depth distribution curve through the preset deep learning model to obtain the dose-depth measurement result of the target particle beam. The dose-depth measurement result includes at least one of Bragg peak position, beam range, and dose-depth distribution curve.

[0006] Optionally, the machine learning processing module is configured to obtain the dose-depth measurement results of the target particle beam by performing the following steps: extracting features and applying physical constraints to the initial dose-depth distribution curve using the preset deep learning model. The preset deep learning model includes: a feature extraction layer, a physical constraint layer, a feature refinement layer, and a regression output layer; The initial dose-depth distribution curve is input into the preset deep learning model. Multiple convolutional kernels in the feature extraction layer sequentially extract features from the initial dose-depth distribution curve to obtain a high-dimensional feature vector representing the local variation pattern of the initial dose-depth distribution curve, output by the last convolutional kernel. The feature extraction process for each convolutional kernel is as follows: ,in, For the first Feature vectors output by the convolutional kernel For nonlinear activation functions, For the first Layer convolution kernel weights, For convolution operations, For the first Feature vectors output by the convolutional kernel , For the initial dose-depth distribution curve, For the first Bias term of the convolution kernel; The physical constraint layer determines the physical constraint loss value corresponding to the high-dimensional feature vector, and the physical constraint loss value is fed back to the feature extraction layer. The feature extraction layer then performs high-dimensional feature extraction on the initial dose-depth distribution curve based on the physical constraint loss value to obtain a new high-dimensional feature vector. The feature refinement layer extracts discriminative features that are highly correlated with the dose depth measurement results from the new high-dimensional feature vector. The dose depth measurement result of the target particle beam is obtained by performing dose depth prediction on the discriminative features through the regression output layer.

[0007] Optionally, the machine learning processing module is configured to determine the physical constraint loss value corresponding to the high-dimensional feature vector through the physical constraint layer by performing the following steps: The physical constraint layer is used to reconstruct the dose-depth distribution curve based on the high-dimensional feature vector to obtain the restored dose-depth distribution curve. The dose value monotonically increasing constraint function and the dose value monotonically decreasing constraint function in the physical constraint layer are used to predict the pre-peak region beam dose value monotonically increasing constraint loss value and the post-peak region beam dose value monotonically decreasing constraint loss value corresponding to the restored dose-depth distribution curve. Based on the predicted monotonically increasing constraint loss value and monotonically decreasing constraint loss value, the physical constraint loss value is determined and then fed back to the feature extraction layer. , , , For physical constraint loss value, For monotonically increasing constraint loss value, For monotonically decreasing constraint loss value, The hyperparameter weights for the monotonically increasing constraint loss value The hyperparameter weights for the monotonically decreasing constraint loss value, The dose-depth distribution curve of the reduction is in the first... Dose values ​​at each depth point The dose-depth distribution curve of the reduction is in the first... Dose values ​​at each depth point The end index of the region before the peak of the reduced dose-depth distribution curve. The starting index of the region after the peak of the reduced dose-depth distribution curve. The total length of the reduced dose-depth distribution curve, The dose-depth distribution curve of the reduction is in the first... Dose values ​​at each depth point The dose-depth distribution curve of the reduction is in the first... Dose values ​​at depth points.

[0008] Optionally, it may also include: a human-computer interaction and display module; The human-computer interaction and display module is electrically connected to the machine learning processing module and is used to obtain the expected dose depth result of the target particle beam. The expected dose depth result includes at least one of the expected Bragg peak position, expected beam range, and expected dose-depth distribution curve. Based on the expected dose depth result and the dose depth measurement result, the operating status of the target particle beam emitting device is determined, and the dose depth measurement result and the operating status are visualized.

[0009] Optionally, the human-computer interaction and display module is used to determine the operating status of the target particle beam emitting device based on the expected dose depth result and the dose depth measurement result by performing the following steps: The expected peak shape parameters and the expected beam dose at the Bragg peak position are determined based on the expected dose-depth distribution curve in the expected dose-depth results, and the peak shape parameters and the beam dose at the Bragg peak position are determined based on the dose-depth distribution curve in the dose-depth measurement results. Determine the peak shape difference information between the expected peak shape parameter and the peak shape difference information between the peak shape parameter, determine the dose difference information between the expected beam dose and the beam dose, and determine the range difference information between the expected beam range and the beam range in the dose depth measurement result; The operating status of the target particle beam emission device is determined based on the peak shape difference information, the dose difference information, and the range difference information.

[0010] Optionally, the human-computer interaction and display module is used to determine the operating status of the target particle beam emitting device based on the peak shape difference information, the dose difference information, and the range difference information by performing the following steps: Determine the peak shape difference feature vector corresponding to the peak shape difference information, the dose difference feature vector corresponding to the dose difference information, and the range difference feature vector corresponding to the range difference information, respectively. The information entropy corresponding to the peak shape difference feature vector, the dose difference feature vector, and the range difference feature vector is determined respectively, and the initial weights corresponding to the peak shape difference feature vector, the dose difference feature vector, and the range difference feature vector are determined based on the information entropy; The vector magnitude and vector complexity corresponding to the peak shape difference feature vector, the dose difference feature vector, and the range difference feature vector are determined respectively. The initial weights are adjusted based on the vector magnitude and the vector complexity to obtain the difference weights. The peak shape difference feature vector, the dose difference feature vector, and the range difference feature vector are weighted and fused based on the difference weights. The working state of the target particle beam emission device is determined based on the weighted fusion result.

[0011] Optionally, the data acquisition and synchronization module is used to synchronously acquire the digital dose sequence output by the front-end readout electronics module by performing the following steps, and to spatially register the acquired digital dose sequence: The data acquisition and synchronization module includes: a trigger synchronization unit, a multi-channel data buffer unit, and a spatial registration unit. The trigger synchronization unit is used to respond to the level synchronization signal of the target particle beam and synchronously acquire the digital dose sequence output by the front-end readout electronics module based on the time in the level synchronization signal. The multi-channel data buffer unit is used to cache the digital dose sequence through a dual-port block random access memory; The spatial registration unit is used to determine the physical arrangement order of each electrode plate in the Bragg peak detector array assembly, and sort the cached digital dose sequence based on the physical arrangement order to obtain the initial dose-depth distribution curve sampled layer by layer along the target particle beam direction.

[0012] Optionally, The machine learning processing module is further configured to construct a preset initial deep learning model; acquire a sample dataset, wherein the sample dataset includes dose-depth distribution curves of sample particle beams with labeled information, the labeled information being the actual dose-depth measurement results of the sample particle beams; divide the sample dataset into a training set and a test set, train the preset initial deep learning model using the training set, and test the trained preset initial deep learning model using the test set, and finally use the trained preset initial deep learning model that meets the test conditions as the preset deep learning model.

[0013] Optionally, it also includes: a storage module: The storage module is used to store the dose depth measurement process data of the target particle beam and generate a dose depth measurement report of the target particle beam based on the dose depth measurement process data. The dose depth measurement process data includes at least one of the digital dose sequence, the initial dose-depth distribution curve, the restored dose-depth distribution curve, and the dose depth measurement result. The storage module is also used to respond to a query command for a dose depth measurement report, and to query and display the corresponding dose depth measurement report based on the target particle beam identifier carried by the query command.

[0014] According to a second aspect of the present invention, a machine learning-based method for measuring dose depth of a Bragg peak detector is provided, the method being applied to the aforementioned machine learning-based Bragg peak detector dose depth measurement system, the method comprising: The Bragg peak detector array assembly converts the received target particle beam into an ionization signal. The front-end readout electronics module sequentially performs charge-sensitive amplification, shaping filtering, and analog-to-digital conversion on the ionization signals output by each array component to obtain a digital dose sequence; The data acquisition and synchronization module synchronously acquires the digital dose sequence output by the front-end readout electronics module, and performs spatial registration on the acquired digital dose sequence to obtain the initial dose-depth distribution curve sampled layer by layer along the target particle beam direction; The machine learning processing module extracts features and performs physical constraints on the initial dose-depth distribution curve using the preset deep learning model to obtain the dose-depth measurement results of the target particle beam. The dose-depth measurement results include at least one of Bragg peak position, beam range, and dose-depth distribution curve.

[0015] According to the present invention, a Bragg peak detector dose depth measurement system and method based on machine learning is provided. Compared with the current method of approximating the Bragg peak region as a Gaussian function and performing nonlinear least squares fitting to determine beam dose depth information, the present invention uses a Bragg peak detector real-time dose depth measurement system with a built-in preset deep learning model to perform dose depth measurement. The preset deep learning model can learn any complex shape of the real Bragg peak, that is, the complex shapes such as the gentle leading edge and the steep trailing edge of the real Bragg peak can be accurately captured. Moreover, the model prediction process does not have the problem of local optima, thereby improving the accuracy of dose depth measurement. At the same time, the model prediction process avoids continuous iteration. This process improves the efficiency of dose depth measurement. The invention achieves continuous depth distribution acquisition in a single step through array-based layer-by-layer sampling of the Bragg peak detector array, eliminating time delays and positioning errors caused by mechanical scanning. The front-end readout electronics module sequentially performs charge-sensitive amplification, shaping filtering, and analog-to-digital conversion on the ionization signal, optimizing signal quality and resulting in a final digital dose sequence with high signal-to-noise ratio, low pulse stacking, and accurate timestamps. Synchronous acquisition and spatial registration by the data acquisition and synchronization modules ensure strict temporal and spatial alignment of data from all array channels, avoiding depth direction misalignment caused by beam pulse jitter. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This diagram illustrates the structure of a machine learning-based Bragg peak detector dose depth measurement system according to an embodiment of the present invention. Figure 2 This diagram illustrates a schematic structure of a Bragg peak detector array assembly according to an embodiment of the present invention. Figure 3This invention provides a schematic diagram of another machine learning-based Bragg peak detector dose depth measurement system according to an embodiment of the present invention. Figure 4 This diagram illustrates a human-computer interaction and measurement result display interface provided by an embodiment of the present invention. Detailed Implementation

[0017] The present invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the present application can be combined with each other.

[0018] Currently, the method of approximating the Bragg peak region as a Gaussian function and performing nonlinear least squares fitting to determine beam dose depth information has significant differences between the shape of the Gaussian function and the actual distribution of the Bragg peak, and it is easy to get trapped in local optima. Therefore, the approximation method of Gaussian function leads to low accuracy in determining beam measurement information such as dose depth. At the same time, Gaussian fitting requires iterative solution, which is time-consuming and labor-intensive.

[0019] To address the aforementioned problems, embodiments of the present invention provide a Bragg peak detector dose depth measurement system based on machine learning, such as... Figure 1 As shown, it includes: a Bragg peak detector array assembly, a front-end readout electronics module, a data acquisition and synchronization module, and a machine learning processing module; The Bragg peak detector array assembly is arranged sequentially along the target particle beam path along the target particle beam inlet direction, and is used to convert the received target particle beam into an ionization signal. The front-end readout electronics module is electrically connected to the Bragg peak detector array assembly and is used to sequentially perform charge-sensitive amplification, shaping filtering and analog-to-digital conversion processing on the ionization signals output by each array assembly to obtain a digital dose sequence. The data acquisition and synchronization module is electrically connected to the front-end readout electronics module and is used to synchronously acquire the digital dose sequence output by the front-end readout electronics module and perform spatial registration on the acquired digital dose sequence to obtain the initial dose-depth distribution curve sampled layer by layer along the target particle beam direction. The machine learning processing module is electrically connected to the data acquisition and synchronization module, and has a built-in preset deep learning model. It is used to extract features and physically constrain the initial dose-depth distribution curve through the preset deep learning model to obtain the dose-depth measurement result of the target particle beam. The dose-depth measurement result includes at least one of Bragg peak position, beam range, and dose-depth distribution curve.

[0020] This invention addresses the limitations of real-time dose depth measurement using Bragg peak detectors by proposing a machine learning-based Bragg peak detector dose depth measurement system. This system integrates a Bragg peak detector array assembly, a front-end readout electronics module, a data acquisition and synchronization module, and a machine learning processing module, all electrically connected sequentially. This achieves both high efficiency and accuracy in measuring the real-time dose depth of the Bragg peak detector. The Bragg peak detector array employs a multi-layered electrode plate stack structure, arranged sequentially along the beam inrush direction. The number of electrode plates can be configured according to the required detection accuracy. The thickness of each electrode plate group covers the complete attenuation region of the Bragg peak. When the particle beam enters the detector through the entrance window, it passes through each electrode plate sequentially, generating an ionization signal proportional to the deposited dose within each electrode plate. Each electrode plate outputs a signal independently. The front-end readout electronics module is electrically connected to the Bragg peak detector array assembly and is used to perform high-fidelity processing on the ionization signals output from each electrode plate of the detector. Since the Bragg peak detector contains multiple independent electrode channels, the front-end readout electronics system module has multi-channel parallel processing capabilities. It sequentially performs charge-sensitive amplification, shaping filtering, and analog-to-digital conversion on each ionization signal, converting the weak ionization signal generated by the electrode plates into a digital dose sequence, ensuring the signal of each channel is accurate and reliable. The amplitude exhibits a precise linear relationship with the deposition dose within the electrode plates. The data acquisition and synchronization module is electrically connected to the front-end readout electronics module. The core processing unit comprises a field-programmable gate array (FPGA) and a real-time processor, enabling synchronous acquisition and precise registration of multi-channel detection data (digital dose sequence). Based on the physical arrangement of the electrode plates (from the entrance window to the exit window), the multi-channel data in the digital dose sequence are spatially registered, forming a dose distribution array sampled layer by layer along the beam depth direction. This dose distribution array is used to construct the initial dose-depth distribution curve, providing standardized input data with clear physical meaning for the machine learning processing module. The machine learning processing module incorporates a pre-set deep learning model. This model extracts features and imposes physical constraints on the initial dose-depth distribution curve, outputting accurate Bragg peak positions and optimized dose-depth distribution information in real time. Due to the limited number of electrode plates in actual measurements, the initial dose-depth distribution curve exhibits discrete sampling intervals and measurement noise. The deep learning model effectively suppresses noise and compensates for sampling discreteness, achieving accurate peak positioning at sub-electrode spacing. This module possesses high-speed parallel computing capabilities, enabling rapid inference from data input to result output.

[0021] By applying the technical solution of this embodiment, a real-time dose depth measurement system using a Bragg peak detector with a built-in preset deep learning model is used for dose depth measurement. The preset deep learning model can learn any complex shape of the real Bragg peak, that is, complex shapes such as a gentle leading edge and a steep trailing edge of the real Bragg peak can be accurately captured, and the model prediction process does not have the problem of local optima, thereby improving the accuracy of dose depth measurement. At the same time, the model prediction process avoids the process of continuous iteration, thereby improving the measurement efficiency of dose depth. This invention achieves continuous depth distribution acquisition in one go through array-style layer-by-layer sampling of the Bragg peak detector array components, eliminating the time delay and positioning error caused by mechanical scanning. By performing charge-sensitive amplification, shaping filtering and analog-to-digital conversion processing on the ionization signal through the front-end readout electronics module, the signal quality can be optimized, so that the final output digital dose sequence has the characteristics of high signal-to-noise ratio, low pulse accumulation and accurate timestamp. Through the synchronous acquisition and spatial registration of the data acquisition and synchronization module, the data of all array channels are strictly aligned in time and space, avoiding depth direction misalignment caused by beam pulse jitter.

[0022] In this embodiment of the invention, the Bragg peak detector array assembly is the core execution component of the fast Bragg peak detector. It can be placed on the treatment bed, located in the particle beam path, for rapid verification of beam energy and range. The detector adopts a multi-layer electrode plate stacked structure, arranged sequentially along the beam inlet direction. The number of electrode plates can be configured according to the detection accuracy requirements (e.g., 128 groups, 200 groups, etc.), and each group of electrode plates has a uniform thickness (e.g., set to 2 mm), with the total thickness covering the complete attenuation region of the Bragg peak. When the particle beam enters the detector from the entrance window, it passes through each layer of electrode plates sequentially, generating an ionization signal proportional to the deposited dose within each layer of electrode plates. Each layer of electrode plates outputs a signal independently. Its multi-layer electrode structure directly realizes dose sampling in the beam depth direction without relying on secondary particle radiation field measurement, featuring direct signal, fast acquisition speed, and high measurement accuracy. Depending on actual needs, the detector type can be selected from ionization chamber arrays, semiconductor detector arrays, etc., to meet the requirements of different particle therapy devices (proton, carbon ion, etc.). The Bragg peak detector adopts a multi-layer parallel plate ionization chamber array structure, for example, referring to... Figure 2As shown. The detector includes: a front panel (101), a rear panel (102), a top plate (103), a bottom plate (104), a left side plate (105), a right side plate (106), and four horizontal tie rods (107). The front panel (101), rear panel (102), top plate (103), bottom plate (104), left side plate (105), and right side plate (106) together form a sealed cavity. Multiple sets (e.g., 200 sets) of parallel plate ionization chamber electrodes are arranged sequentially along the beam direction inside the cavity. Each set of electrodes is 4 mm thick, with a total thickness of 800 mm, covering the complete Bragg peak attenuation region of the proton beam energy range (e.g., 70~250 MeV). An entrance window (1011) is provided in the center of the front panel (101), and an exit window (1021) is provided in the center of the rear panel (102). The size of the entrance and exit windows can be set to 100mm×100mm, and the material is not limited to polyimide film. The thickness is set according to actual needs, such as 0.1mm, to reduce beam scattering and energy loss. Handles (108) are provided on the outer sides of the front panel (101) and the rear panel (102) respectively to facilitate the handling and placement of the detector. A positioning line is provided along the length of the top plate (103) of the detector housing for aligning the front and rear ends of the detector. Four cross lines are provided at the center line of the entrance window (1011) of the front panel (101), and four cross lines are provided at the center line of the exit window (1021) of the rear panel (102), for aligning the front and rear ends of the detector in the up, down, left and right directions respectively. The support bracket (1042) is folded when used in the horizontal beam terminal and unfolded when used in the 45° beam terminal, allowing the detector to be supported at a 45° angle. Five connector mounting holes are provided on the left side plate (105), including two signal connector mounting holes (1051), one temperature and humidity signal panel connector mounting hole (1052), and two high-voltage connection mounting holes (1053). The signal connectors are Micro-D type 25-pin connectors, the temperature and humidity signal connectors are LEMO-EGG type connectors, and the high-voltage connectors are LEMO-SWH type connectors. A positioning etched line is provided along the length of both the left side plate (105) and the right side plate (106) for aligning the detector's left and right sides vertically. Four horizontal tie rods (107) are made of 304 stainless steel, 8mm in diameter, and coated with a 0.1mm thick polyimide insulating film. They are arranged parallel to each other along the length of the cavity and are used to fix the front and rear end panels and install the electrode plate assembly. The front and rear ends of the horizontal tie rod (107) are provided with internal threads and are fastened to the front panel (101) and the rear panel (102) by screws. The electrode plates of each layer are positioned and stacked by the horizontal tie rod (107), and adjacent electrode plates are isolated by insulating gaskets to form 200 independent ionization chambers.

[0023] In this embodiment of the invention, the front-end readout electronics module is also responsible for providing a stable high-voltage bias to the detector electrodes, ensuring that the detector operates in the optimal ionization collection state, and providing a high signal-to-noise ratio and high linearity digital signal for subsequent modules. The front-end readout electronics module includes multiple parallel processing channels (the same number as the detector's electrode plates), each channel electrically connected to one layer of electrode plates. This module mainly consists of a multi-channel current-to-digital converter chip and peripheral configuration circuitry. The current-to-digital converter chip integrates charge-sensitive amplification, analog-to-digital conversion, and synchronous sampling functions, supporting multi-channel processing on a single chip. Each channel integrates a low-noise integrator and a synchronous sample-and-hold circuit, supporting high bit resolution and sampling rate, and employing a dead-zone-free design to ensure no charge signal loss. The chip configures registers via an SPI interface and outputs data via a clock serial interface, supporting cascaded synchronous operation to ensure strict synchronous acquisition of multi-channel signals. The front-end readout electronics module achieves high integration of multi-channel parallel processing, providing low-noise, high-precision digital signals for subsequent data acquisition. This module also includes a high-voltage power supply module, which integrates a high-voltage power supply to provide an adjustable high-voltage bias for the detector electrodes. The digital signals of each channel are transmitted to the data acquisition and synchronization module through the interface.

[0024] In this embodiment of the invention, the data acquisition and synchronization module uses the beam trigger signal as the synchronization reference to strictly align the digital signals of each electrode plate channel in time, ensuring that the dose signal acquired by each layer of electrode plates corresponds one-to-one with the physical process of the beam passing through the detector. Simultaneously, the module spatially registers the multi-channel data according to the physical arrangement order of the electrode plates (from the entrance window to the exit window), forming a dose distribution array sampled layer by layer along the beam depth direction. This array directly characterizes the initial dose-depth distribution curve of the dose deposited by the beam within the detector, providing standardized input data with clear physical meaning for the machine learning processing unit.

[0025] In an embodiment of the present invention, optionally, a data acquisition and synchronization module is used to synchronously acquire the digital dose sequence output by the front-end readout electronics module by performing the following steps, and to spatially register the acquired digital dose sequence: the data acquisition and synchronization module includes: a trigger synchronization unit, a multi-channel data buffer unit, and a spatial registration unit; the trigger synchronization unit is used to synchronously acquire the digital dose sequence output by the front-end readout electronics module based on the time in the level synchronization signal of the target particle beam in response to the level synchronization signal of the target particle beam; the multi-channel data buffer unit is used to cache the digital dose sequence through a dual-port block random access memory; the spatial registration unit is used to determine the physical arrangement order of each electrode plate in the Bragg peak detector array assembly, and sort the cached digital dose sequence based on the physical arrangement order to obtain an initial dose-depth distribution curve sampled layer by layer along the direction of the target particle beam.

[0026] Specifically, the trigger synchronization unit receives the beam extraction synchronization signal from the particle therapy device. Using this signal as a time reference, it simultaneously latches the signals of each channel to ensure the time alignment accuracy of each channel signal. The multi-channel data buffer unit uses a dual-port block random access memory to cache the data depth of multiple channel sampling points. The spatial registration unit sorts the digital dose sequence composed of 200 channels of data according to the physical arrangement order of the electrode plates, forming a one-dimensional dose sampling array with a length equal to the number of channels. This array directly constitutes the initial dose-depth distribution curve of the dose deposited by the beam in the detector. The data acquisition and synchronization module transmits the initial dose-depth distribution curve to the machine learning processing module through a network interface. This embodiment of the invention, through data synchronization, buffering, and registration operations, ensures from the data source that the dose-depth curve input to the machine learning model is jitter-free, error-free, and noise-free, thereby improving the prediction accuracy of subsequent beam dose depth.

[0027] In this embodiment of the invention, optionally, the machine learning processing module adopts an embedded server platform and has a pre-trained preset deep learning model built in. To ensure the prediction accuracy of the preset deep learning model, the machine learning processing module is also used to construct a preset initial deep learning model; obtain a sample dataset, wherein the sample dataset includes dose-depth distribution curves of sample particle beams with annotation information, and the annotation information is the actual dose-depth measurement result of the sample particle beams; divide the sample dataset into a training set and a test set, use the training set to train the preset initial deep learning model, and use the test set to test the trained preset initial deep learning model, and finally use the trained preset initial deep learning model that meets the test conditions as the preset deep learning model.

[0028] Specifically, during model training, a pre-defined initial deep learning model is first constructed, followed by the acquisition of a sample dataset. The dataset must contain all necessary files. The data is then converted to a format understandable by the pre-defined initial deep learning model. Finally, the model is trained and tested. Specifically, the dataset can be divided first: using randomness or a specific strategy (such as stratified sampling), the sample dataset is divided into training and test sets. The model is then trained using the training set, and tested using the test set to evaluate its performance on unseen data. Metrics such as precision and recall on the test set are calculated and recorded. If the model performance does not meet requirements, the training phase can be returned for further iterations or adjustments. This process yields a pre-defined deep learning model that meets the requirements.

[0029] In this embodiment of the invention, optionally, when acquiring the sample dataset, a detector geometric model can be established using Monte Carlo simulation software: such as a 200-layer parallel-plate ionization chamber, each layer with a thickness of 4 mm, the electrode material being aluminum (0.1 mm thick), and the gas medium being air (standard temperature and pressure). The simulated proton beam energy range is 70–250 MeV, with a step size of 1 MeV, totaling 181 energies. At each energy, different beam spot sizes (σ = 2 mm, 3 mm, 4 mm, 5 mm), different current intensities (0.5 nA, 1 nA, 2 nA, 5 nA), and different noise levels (signal-to-noise ratio 20 dB, 30 dB, 40 dB, 50 dB) are simulated, with each parameter combination simulating 1000 events. For each event, the deposition dose of the 200 electrode layers is recorded as the input feature (D); the true Bragg peak depth (ptrue) and 800-point high-precision dose curves (obtained through 1 mm step size simulation) are recorded as beam dose depth results as output labels. Multiple sets of data samples were generated. These could be divided into training, validation, and test sets in an 8:1:1 ratio. Preprocessing of the sample dataset included: maximum value normalization, dividing the 200-point dose sequence of each sample by the maximum value of that sample to eliminate the influence of flow intensity differences; and Z-score standardization, calculating the mean and standard deviation of the sample dataset to standardize the dose sequences. Data augmentation employed an online method, randomly adding Gaussian noise (mean 0, standard deviation between 0.01 and 0.05), randomly simulating electrode response non-uniformity (multiplying some channels by random coefficients between 0.95 and 1.05), and random smoothing filtering (kernel size 3 to 5). The augmentation factor was approximately 5 times, resulting in approximately 58 million equivalent training samples. Finally, the pre-processed sample dataset was used to train and construct a pre-defined deep learning model. If the model was built, training was performed on a GPU server. The batch size was set to 256. An optimizer was used with an initial learning rate of 0.001. A cosine annealing learning rate scheduling strategy was employed, with the learning rate decaying to 0.1 times the previous cycle's rate every 50 epochs. Training was performed for 200 epochs, with evaluation on the validation set every 10 epochs. The model with the smallest validation loss was saved as the final model. After training, the model was evaluated on the test set: the root mean square error of Bragg peak prediction was 0.28 mm, the maximum absolute error was 0.32 mm, and the root mean square error of R90 prediction was 0.15 mm. After training, the model was exported to a specific format, and the quantizer and compiler converted it into an executable instruction file. Finally, the compiled model was loaded. After the model was deployed, parameters were updated every fixed period.

[0030] In an embodiment of the present invention, optionally, the machine learning processing module is configured to obtain the dose-depth measurement result of the target particle beam by performing the following steps: feature extraction and physical constraint of the initial dose-depth distribution curve using the preset deep learning model; the preset deep learning model includes: a feature extraction layer, a physical constraint layer, a feature refinement layer, and a regression output layer; the initial dose-depth distribution curve is input into the preset deep learning model, and features are extracted from the initial dose-depth distribution curve sequentially through multiple convolutional kernels in the feature extraction layer to obtain a high-dimensional feature vector representing the local change pattern of the initial dose-depth distribution curve output by the last convolutional kernel; the feature extraction process of each convolutional kernel is as follows: ,in, For the first Feature vectors output by the convolutional kernel For nonlinear activation functions, For the first Layer convolution kernel weights, For convolution operations, For the first Feature vectors output by the convolutional kernel , For the initial dose-depth distribution curve, For the first The bias term of the convolutional kernel is used; the physical constraint loss value corresponding to the high-dimensional feature vector is determined by the physical constraint layer, and the physical constraint loss value is fed back to the feature extraction layer. The feature extraction layer re-extracts high-dimensional features from the initial dose-depth distribution curve based on the physical constraint loss value to obtain a new high-dimensional feature vector; the feature refinement layer extracts discriminative features that are highly correlated with the dose-depth measurement results from the new high-dimensional feature vector; the regression output layer performs dose-depth prediction on the discriminative features to obtain the dose-depth measurement results of the target particle beam.

[0031] Specifically, the feature extraction layer employs a one-dimensional convolutional neural network. The input layer receives the initial dose-depth distribution curve. The network contains three convolutional layers: for example, the first layer has a kernel size of 5, 32 channels, and a stride of 1; the second layer has a kernel size of 3, 64 channels, and a stride of 1; and the third layer has a kernel size of 3, 128 channels, and a stride of 1. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function. The initial dose-depth distribution curve is sequentially input from the first convolutional layer to the last convolutional layer. Feature extraction is performed sequentially in each convolutional layer. Finally, the features output from the last convolutional layer are processed by the batch normalization layer and the ReLU activation function to obtain a high-dimensional feature vector representing the local variation pattern of the initial dose-depth distribution curve.

[0032] After determining the high-dimensional feature vector, to ensure the extraction accuracy of the high-dimensional feature vector, it is necessary to verify the high-dimensional feature vector using a physical constraint layer. Therefore, in this embodiment of the invention, optionally, the machine learning processing module is used to determine the physical constraint loss value corresponding to the high-dimensional feature vector through the physical constraint layer by performing the following steps: The physical constraint layer reconstructs the dose-depth distribution curve based on the high-dimensional feature vector to obtain a restored dose-depth distribution curve; the physical constraint layer uses the dose value monotonically increasing constraint function and the dose value monotonically decreasing constraint function in the physical constraint layer to predict the pre-peak region beam dose value monotonically increasing constraint loss value and the post-peak region beam dose value monotonically decreasing constraint loss value corresponding to the restored dose-depth distribution curve; based on the predicted monotonically increasing constraint loss value and monotonically decreasing constraint loss value, the physical constraint loss value is determined, and the physical constraint loss value is fed back to the feature extraction layer. , , , For physical constraint loss value, For monotonically increasing constraint loss value, For monotonically decreasing constraint loss value, The hyperparameter weights for the monotonically increasing constraint loss value The hyperparameter weights for the monotonically decreasing constraint loss value, The dose-depth distribution curve of the reduction is in the first... Dose values ​​at each depth point The dose-depth distribution curve of the reduction is in the first... Dose values ​​at each depth point The end index of the region before the peak of the reduced dose-depth distribution curve. The starting index of the region after the peak of the reduced dose-depth distribution curve. The total length of the reduced dose-depth distribution curve, The dose-depth distribution curve of the reduction is in the first... Dose values ​​at each depth point The dose-depth distribution curve of the reduction is in the first... Dose values ​​at depth points.

[0033] Specifically, the physical constraint layer consists of a curve generation subnetwork and a physical constraint mechanism. The curve generation subnetwork is composed of a series of fully connected layers. High-dimensional feature vectors are input into the curve generation subnetwork, and the data flows through one or more fully connected layers. Each layer performs a nonlinear transformation on the features, progressively decoding the abstract features into specific curve shape information (restoring the dose-depth distribution curve). The restored dose-depth distribution curve is then input into the physical constraint mechanism, which includes a monotonically increasing dose value constraint function and a monotonically decreasing dose value constraint function. This function introduces a physical prior regularization term. Subsequently, based on the physical characteristics of the Bragg peak, the physical constraint layer divides the depth domain into a pre-peak region (index 1 to K-1) and a post-peak region (index M to L-1). For the pre-peak region, a monotonically increasing constraint loss value is calculated using a dose value monotonically increasing constraint function. Specifically, this involves iterating through each point before the peak; if the dose value at the previous moment is greater than the dose value at the next moment (i.e., a non-monotonic increase occurs), the difference is accumulated as a penalty. For the post-peak region, a monotonically decreasing constraint loss value is calculated using a dose value monotonically decreasing constraint function. Specifically, this involves iterating through each point after the peak; if the dose value at the next moment is greater than the dose value at the previous moment (i.e., a non-monotonic decrease occurs), the difference is accumulated as a penalty. Finally, the physical constraint mechanism obtains the physical constraint loss value by weighted summation of the monotonically increasing and monotonically decreasing constraint loss values ​​based on hyperparameter weights. The machine learning processing module uses the calculated physical constraint loss value as a feedback signal and sends it back to the feature extraction layer. In this embodiment, the backpropagation process can be implemented by the backpropagation algorithm, which transmits the gradient information of the physical constraint loss value to the convolutional kernel weights of the feature extraction layer.

[0034] Furthermore, after receiving the physical constraint loss value, the feature extraction layer adjusts the weight parameters of the convolutional neural network and re-extracts high-dimensional features from the initial dose-depth distribution curve. During the adjustment of weight parameters, if the physical constraint loss value is large, it indicates that the currently extracted feature vector contains noise or erroneous patterns that violate physical monotonicity. The feature extraction layer will suppress the feature channels that cause these erroneous patterns. If the physical constraint loss value is small, it indicates that the feature extraction conforms to physical laws, and the weight parameters are preserved or enhanced. After the above iterations or single optimization, the feature extraction layer outputs a new high-dimensional feature vector. This new feature vector not only contains the amplitude information of the dose distribution but also implicitly encodes the physical topology of monotonically increasing before the peak and monotonically decreasing after the peak in the feature space.

[0035] Furthermore, the new high-dimensional feature vector is input into the feature refinement layer, which employs three fully connected layers. Each fully connected layer is followed by a Dropout layer and a ReLU activation function. The feature refinement layer extracts discriminative features step by step through the three fully connected layers: First, the high-dimensional feature vector is flattened and input into the first fully connected layer for feature mapping and nonlinearity is introduced. Then, some neurons are randomly dropped to prevent overfitting. Next, the second fully connected layer further abstracts and combines the features, again introducing nonlinearity and regularization. Finally, the third fully connected layer condenses the features into a low-dimensional discriminative vector. Each layer uses a combination of linear transformation, nonlinear activation, and Dropout regularization to gradually select key features highly correlated with dose-depth measurement as discriminative features.

[0036] Furthermore, the regression output layer employs a fully connected layer to map the discriminative features to dose-depth measurement results. The output includes: Bragg peak depth location, the location of the distal 90% dose point (R90), beam range, and a high-resolution dose-depth distribution curve. The high-resolution curve is generated by a transposed convolutional decoder in the regression output layer, containing two transposed convolutional layers with upsampling factors of 2 and 2, respectively.

[0037] In this embodiment of the invention, the physical constraint layer calculates the monotonicity loss of the dose distribution (increasing before the peak and decreasing after the peak) and feeds the loss back to the feature extraction layer. This forces the model to learn features that conform to the physical laws of the Bragg peak, thereby improving the extraction accuracy of high-dimensional feature vectors and consequently enhancing the prediction accuracy of subsequent dose depth measurements. This embodiment of the invention uses a real-time dose depth measurement system with a Bragg peak detector that incorporates a pre-set deep learning model. This pre-set deep learning model can learn any complex shape of the real Bragg peak, accurately capturing complex shapes such as a gentle leading edge and a steep trailing edge. Furthermore, the model prediction process avoids the problem of local optima, thus improving the accuracy of dose depth measurement. Simultaneously, the model prediction process avoids continuous iteration, thereby improving the efficiency of dose depth measurement.

[0038] In embodiments of the present invention, optionally, such as Figure 3 As shown, the Bragg peak detector dose-depth measurement system based on machine learning further includes: a human-computer interaction and display module; the human-computer interaction and display module is electrically connected to the machine learning processing module, and is used to acquire the expected dose-depth result of the target particle beam, wherein the expected dose-depth result includes at least one of the expected Bragg peak position, expected beam range, and expected dose-depth distribution curve; the operating state of the target particle beam emitting device is determined based on the expected dose-depth result and the dose-depth measurement result, and the dose-depth measurement result and the operating state are visualized.

[0039] Specifically, the human-computer interaction and display module uses a graphical interface to display the dose-depth measurement results predicted by the model, the Bragg peak position (equivalent water thickness value) optimized by the model, and the expected dose-depth distribution curve imported from the treatment planning system. It overlays and compares the information from the model-predicted dose-depth measurement results and the imported expected dose-depth distribution curve to intuitively display key indicators such as range deviation and dose difference. The graphical interface of the human-computer interaction and display module includes a curve display area: this area simultaneously displays the model-predicted dose-depth measurement results and the measured dose-depth measurement results in a dual-axis format, overlaid and compared with the imported expected dose-depth distribution curve. The curve colors can be differentiated, for example, the curve colors can be blue (measured), red (model predicted), and green (expected). The graphical interface of the human-computer interaction and display module also includes a parameter display area: this area displays key parameters such as the measured Bragg peak position, the expected Bragg peak position, deviation, the distal 90% dose point position R90, and the peak dose. The graphical interface in the human-computer interaction and display module also includes a judgment indicator area: this area automatically determines whether the target particle beam emitting device displays a green "qualified" indicator, a yellow "warning" indicator, or a red "unqualified" indicator based on the deviation. The graphical interface in the human-computer interaction and display module also includes an operation control area: this area contains function buttons such as "Start Measurement," "Save Report," "Historical Query," and "System Settings," and clicking the corresponding function button will activate the corresponding function. This embodiment of the invention enables the visualization of dose depth measurement results through the human-computer interaction and display module. The expected dose depth result refers to the ideal dose distribution data calculated by the particle therapy system based on the specific conditions of the irradiated object.

[0040] In an embodiment of the present invention, optionally, the human-computer interaction and display module is configured to determine the operating state of the target particle beam emitting device based on the expected dose-depth result and the dose-depth measurement result by performing the following steps: determining the expected peak shape parameter and the expected beam dose at the Bragg peak position based on the expected dose-depth distribution curve in the expected dose-depth result; determining the peak shape parameter and the beam dose at the Bragg peak position based on the dose-depth distribution curve in the dose-depth measurement result; determining the peak shape difference information between the expected peak shape parameter and the peak shape parameter; determining the dose difference information between the expected beam dose and the beam dose; determining the range difference information between the expected beam range and the beam range in the dose-depth measurement result; and determining the operating state of the target particle beam emitting device based on the peak shape difference information, the dose difference information, and the range difference information.

[0041] Among them, the expected peak shape parameters and peak shape parameters include, but are not limited to, peak width, asymmetry, and rising slope; the expected beam dose and beam dose include, but are not limited to, peak dose, pre-peak dose, and post-peak dose; the expected beam range and beam range include, but are not limited to, range value, range distribution of particles with different energies, and the degree of range variation with time or different beam pulses.

[0042] Specifically, the degree of difference between the expected dose depth result and the corresponding parameters in the dose depth measurement result is determined. Based on this degree of difference, the operating status of the target particle beam emission device is determined. For example, if the range difference is large, it indicates that the beam energy of the target particle beam emission device is drifting; if the peak shape difference is large, it indicates that there is a problem with the beam focusing of the target particle beam emission device; if the dose difference is large, it indicates that there is a problem with the beam intensity, and at this time, it is determined that the ion source stability of the target particle beam emission device is poor. The above difference information and operating status are displayed in a visual form (such as curve comparison graphs, difference heat maps, status indicator lights, etc.) on the human-computer interaction interface, allowing operators to quickly identify problems and take adjustment measures. This embodiment of the invention achieves accurate and rapid determination of the operating status of the particle beam emission device through real-time difference analysis of multiple parameters (peak shape, dose, range), solving the problem of misjudgment caused by traditional single-parameter monitoring.

[0043] In an embodiment of the present invention, optionally, the human-computer interaction and display module is used to determine the working state of the target particle beam emission device based on the peak shape difference information, the dose difference information, and the range difference information by performing the following steps: determining the peak shape difference feature vector corresponding to the peak shape difference information, the dose difference feature vector corresponding to the dose difference information, and the range difference feature vector corresponding to the range difference information; determining the information entropy corresponding to the peak shape difference feature vector, the dose difference feature vector, and the range difference feature vector, respectively; determining the initial weights corresponding to the peak shape difference feature vector, the dose difference feature vector, and the range difference feature vector based on the information entropy; determining the vector magnitude and vector complexity corresponding to the peak shape difference feature vector, the dose difference feature vector, and the range difference feature vector, respectively; adjusting the initial weights based on the vector magnitude and the vector complexity to obtain difference weights; performing weighted fusion of the peak shape difference feature vector, the dose difference feature vector, and the range difference feature vector based on the difference weights; and determining the working state of the target particle beam emission device based on the weighted fusion result.

[0044] Specifically, feature extraction models, such as CNN models, can be used to extract peak shape difference feature vectors, dose difference feature vectors, and range difference feature vectors, respectively. For each difference feature vector, its information entropy is calculated. The information entropies of all difference feature vectors are then normalized to obtain the initial weights for each difference feature vector, ensuring that the sum of the initial weights for each difference feature vector is 1. For each difference feature vector, the vector magnitude is determined according to the following formula:

[0045] in, For differential feature vectors vector magnitude, For differential feature vectors Total number of elements For differential feature vectors The Middle The values ​​of each element are calculated. Simultaneously, for each differential feature vector, its corresponding total variation, zero-crossing rate, and uncertainty are determined, and the vector complexity of the corresponding differential feature vector is determined based on the total variation, zero-crossing rate, and uncertainty. Specifically, the total variation of each differential feature vector is determined according to the following formula:

[0046] in, The total variation of the i-th differential eigenvector, For the (j+1)th element in the i-th difference feature vector, The j-th element of the i-th differential feature vector, where n is the vector length of the i-th differential feature vector, or the total number of elements. Further, for each differential feature vector, determine the number of positive-to-negative transitions between adjacent elements, and use the ratio of the number of positive-to-negative transitions to the total number of adjacent element pairs as the zero-crossing rate of the corresponding differential feature vector. Further, determine the uncertainty of each differential feature vector according to the following formula:

[0047] in, The uncertainty of the i-th differential eigenvector Let be the frequency of the element with value k in the i-th differential feature vector, i.e., the ratio of the number of elements with value k to the total number of elements in the vector, and m be the different element values. Further, weight coefficients for total variation, zero crossover rate, and uncertainty are determined according to actual needs. Based on these weight coefficients, a weighted sum is performed on the total variation, zero crossover rate, and uncertainty. The vector complexity of the corresponding differential feature vector is determined based on the weighted sum; a larger weighted sum results in a higher vector complexity for the corresponding differential feature vector. Further, determine the magnitude ratio factor for each differential eigenvector, i.e., determine the sum of the magnitudes of each differential eigenvector, and use the ratio of the magnitude of each differential eigenvector to the sum of its magnitudes as the magnitude ratio factor for the corresponding differential eigenvector. Simultaneously, determine the complexity ratio factor for each differential eigenvector, i.e., determine the sum of the vector complexities of each differential eigenvector, and use the ratio of the vector complexity of each differential eigenvector to the sum of its vector complexities as the vector complexity ratio factor for the corresponding differential eigenvector. Then, for each differential eigenvector, determine the comprehensive adjustment factor corresponding to the initial weight of that vector based on the magnitude ratio factor and the vector complexity ratio factor. The comprehensive adjustment factor for the initial weight of each differential eigenvector is determined according to the following formula:

[0048] in, The comprehensive adjustment factor for the initial weights of the i-th differential eigenvector. The strength coefficient is adjusted according to actual needs. A balance coefficient between module length and complexity set according to actual needs. The modulus of the i-th differential eigenvector is the proportion factor. The modulus ratio factor is the ratio of the magnitudes of all differential feature vectors when they are uniformly distributed. In this embodiment of the invention, there are 3 differential feature vectors. , The vector complexity ratio factor of the i-th differential feature vector. The vector complexity factor when all differential feature vectors are uniformly distributed, in this embodiment of the invention, includes 3 differential feature vectors. The adjusted weights are then obtained by multiplying the comprehensive adjustment factor by the corresponding initial weights. The adjusted weights of each differential feature vector are then normalized so that the sum of all weights is 1, thus obtaining the differential weight for each differential feature vector. Based on these differential weights, the peak shape difference feature vector, dose difference feature vector, and range difference feature vector are then weighted and fused to obtain a weighted fusion result. Finally, the operating state of the target particle beam emission device is determined based on the weighted fusion result. For example, the weighted fusion result is a weighted fusion vector, which is a three-dimensional vector that integrates the differential information of peak shape, dose, and range. If the first component of the weighted fusion vector reflects peak shape difference information, the second component reflects dose difference information, and the third component reflects range difference information, then a preset threshold corresponding to each vector component is pre-determined according to actual needs. Different vector components are compared with their corresponding preset thresholds to determine the operating state of the particle beam emission device. If all components are less than their corresponding preset thresholds, the particle beam emission device is determined to be in normal working condition. If the component representing peak shape difference is greater than its corresponding preset threshold, the beam focusing of the emission device is determined to be abnormal. If the component representing dose difference is greater than its corresponding preset threshold, the emission device is determined to have beam intensity fluctuations, such as unstable ion source output. If the component representing range difference is greater than its corresponding preset threshold, the emission device is determined to have beam energy drift, such as accelerator radio frequency field phase deviation. This embodiment of the invention determines the working state of the particle beam emission device by comprehensively analyzing multi-dimensional differences, which improves the accuracy of determining the working state. The above difference information and working state are then visualized through a human-computer interaction and display module. Figure 4 The following is a list of information displayed on a human-computer interaction interface.

[0049] In an embodiment of the present invention, optionally, the Bragg peak detector dose depth measurement system based on machine learning further includes: a storage module: the storage module is used to store the dose depth measurement process data of the target particle beam, and generate a dose depth measurement report of the target particle beam based on the dose depth measurement process data, wherein the dose depth measurement process data includes at least one of the digital dose sequence, the initial dose-depth distribution curve, the restored dose-depth distribution curve, and the dose depth measurement result; the storage module is also used to respond to a query command for the dose depth measurement report, and query and display the corresponding dose depth measurement report based on the target particle beam identifier carried by the query command.

[0050] Specifically, the storage module generates a dose depth measurement report of the target particle beam based on the dose depth measurement process data, and authorized users can query this report. The dose depth measurement process data can be stored in a local database, and the report can be exported in multiple formats, supporting integration with various systems via interfaces. This embodiment of the invention achieves traceability of the dose measurement process through the data storage function of the storage module.

[0051] In one embodiment, a machine learning-based Bragg peak detector dose-depth measurement method is provided. This method is applied to a machine learning-based Bragg peak detector dose-depth measurement system as described in any of the above embodiments. The method includes: a Bragg peak detector array component converting a received target particle beam into an ionization signal; a front-end readout electronics module sequentially performing charge-sensitive amplification, shaping filtering, and analog-to-digital conversion processing on the ionization signals output by each array component to obtain a digital dose sequence; a data acquisition and synchronization module synchronously acquiring the digital dose sequence output by the front-end readout electronics module and spatially registering the acquired digital dose sequence to obtain an initial dose-depth distribution curve sampled layer by layer along the target particle beam direction; and a machine learning processing module performing feature extraction and physical constraints on the initial dose-depth distribution curve using a preset deep learning model to obtain the dose-depth measurement result of the target particle beam. The dose-depth measurement result includes at least one of the following: Bragg peak position, beam range, and dose-depth distribution curve.

[0052] Specifically, the Bragg peak detector array receives the target particle beam and outputs the ionization signal generated by the interaction between the particles and the detector medium to the front-end readout electronics module. The front-end readout electronics module sequentially performs charge-sensitive amplification (improving the signal-to-noise ratio of weak signals), shaping filtering (suppressing high-frequency noise and optimizing signal waveforms), and analog-to-digital conversion (quantizing analog signals into digital dose sequences) on the ionization signals output by each array component to obtain a high-precision digital dose sequence. The data acquisition and synchronization module synchronously acquires the digital dose sequence output by the front-end readout electronics module. Through time synchronization and spatial location correspondence (such as spatial arrangement mapping of the detector array), it performs spatial registration on the acquired data to generate an initial dose-depth distribution curve sampled layer by layer along the particle beam direction. The machine learning processing module calls a preset deep learning model to extract features (such as peak shape features, dose gradient features, range features, etc.) from the initial dose-depth distribution curve and optimizes the feature representation by combining physical constraints (such as dose conservation, range consistency, prior knowledge of Bragg peak position, etc.). Finally, it outputs the dose-depth measurement results of the target particle beam, including key parameters such as Bragg peak position, beam range, and complete dose-depth distribution curve.

[0053] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0054] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A machine learning-based Bragg peak detector dose depth measurement system, characterized in that, include: Bragg peak detector array assembly, front-end readout electronics module, data acquisition and synchronization module, and machine learning processing module; The Bragg peak detector array assembly is arranged sequentially along the target particle beam path along the target particle beam inlet direction, and is used to convert the received target particle beam into an ionization signal. The front-end readout electronics module is electrically connected to the Bragg peak detector array assembly and is used to sequentially perform charge-sensitive amplification, shaping filtering and analog-to-digital conversion processing on the ionization signals output by each array assembly to obtain a digital dose sequence. The data acquisition and synchronization module is electrically connected to the front-end readout electronics module and is used to synchronously acquire the digital dose sequence output by the front-end readout electronics module and perform spatial registration on the acquired digital dose sequence to obtain the initial dose-depth distribution curve sampled layer by layer along the target particle beam direction. The machine learning processing module is electrically connected to the data acquisition and synchronization module, and has a built-in preset deep learning model. It is used to extract features and physically constrain the initial dose-depth distribution curve through the preset deep learning model to obtain the dose-depth measurement result of the target particle beam. The dose-depth measurement result includes at least one of Bragg peak position, beam range, and dose-depth distribution curve.

2. The Bragg peak detector dose depth measurement system based on machine learning according to claim 1, characterized in that, The machine learning processing module is used to obtain the dose-depth measurement results of the target particle beam by performing the following steps: extracting features and applying physical constraints to the initial dose-depth distribution curve using the preset deep learning model. The preset deep learning model includes: a feature extraction layer, a physical constraint layer, a feature refinement layer, and a regression output layer; The initial dose-depth distribution curve is input into the preset deep learning model. Multiple convolutional kernels in the feature extraction layer sequentially extract features from the initial dose-depth distribution curve to obtain a high-dimensional feature vector representing the local variation pattern of the initial dose-depth distribution curve, output by the last convolutional kernel. The feature extraction process for each convolutional kernel is as follows: ,in, For the first Feature vectors output by the convolutional kernel For nonlinear activation functions, For the first Layer convolution kernel weights, For convolution operations, For the first Feature vectors output by the convolutional kernel , For the initial dose-depth distribution curve, For the first Bias term of the convolution kernel; The physical constraint layer determines the physical constraint loss value corresponding to the high-dimensional feature vector, and the physical constraint loss value is fed back to the feature extraction layer. The feature extraction layer then performs high-dimensional feature extraction on the initial dose-depth distribution curve based on the physical constraint loss value to obtain a new high-dimensional feature vector. The feature refinement layer extracts discriminative features that are highly correlated with the dose depth measurement results from the new high-dimensional feature vector. The dose depth measurement result of the target particle beam is obtained by performing dose depth prediction on the discriminative features through the regression output layer.

3. The Bragg peak detector dose depth measurement system based on machine learning according to claim 2, characterized in that, The machine learning processing module is used to determine the physical constraint loss value corresponding to the high-dimensional feature vector through the physical constraint layer by performing the following steps: The physical constraint layer is used to reconstruct the dose-depth distribution curve based on the high-dimensional feature vector to obtain the restored dose-depth distribution curve. The dose value monotonically increasing constraint function and the dose value monotonically decreasing constraint function in the physical constraint layer are used to predict the pre-peak region beam dose value monotonically increasing constraint loss value and the post-peak region beam dose value monotonically decreasing constraint loss value corresponding to the restored dose-depth distribution curve. Based on the predicted monotonically increasing constraint loss value and monotonically decreasing constraint loss value, the physical constraint loss value is determined and then fed back to the feature extraction layer. , , , For physical constraint loss value, For monotonically increasing constraint loss value, For monotonically decreasing constraint loss value, The hyperparameter weights for the monotonically increasing constraint loss value The hyperparameter weights for the monotonically decreasing constraint loss value, The dose-depth distribution curve of the reduction is in the first... Dose values ​​at each depth point The dose-depth distribution curve of the reduction is in the first... Dose values ​​at each depth point The end index of the region before the peak of the reduced dose-depth distribution curve. The starting index of the region after the peak of the reduced dose-depth distribution curve. The total length of the reduced dose-depth distribution curve, The dose-depth distribution curve of the reduction is in the first... Dose values ​​at each depth point The dose-depth distribution curve of the reduction is in the first... Dose values ​​at depth points.

4. The Bragg peak detector dose depth measurement system based on machine learning according to claim 1, characterized in that, Also includes: Human-computer interaction and display module; The human-computer interaction and display module is electrically connected to the machine learning processing module and is used to obtain the expected dose depth result of the target particle beam. The expected dose depth result includes at least one of the expected Bragg peak position, expected beam range, and expected dose-depth distribution curve. Based on the expected dose depth result and the dose depth measurement result, the operating status of the target particle beam emitting device is determined, and the dose depth measurement result and the operating status are visualized.

5. The Bragg peak detector dose depth measurement system based on machine learning according to claim 4, characterized in that, The human-computer interaction and display module is used to determine the operating status of the target particle beam emitting device based on the expected dose depth result and the dose depth measurement result by performing the following steps: The expected peak shape parameters and the expected beam dose at the Bragg peak position are determined based on the expected dose-depth distribution curve in the expected dose-depth results, and the peak shape parameters and the beam dose at the Bragg peak position are determined based on the dose-depth distribution curve in the dose-depth measurement results. Determine the peak shape difference information between the expected peak shape parameter and the peak shape difference information between the peak shape parameter, determine the dose difference information between the expected beam dose and the beam dose, and determine the range difference information between the expected beam range and the beam range in the dose depth measurement result; The operating status of the target particle beam emission device is determined based on the peak shape difference information, the dose difference information, and the range difference information.

6. The Bragg peak detector dose depth measurement system based on machine learning according to claim 5, characterized in that, The human-computer interaction and display module is used to determine the operating status of the target particle beam emission device based on the peak shape difference information, the dose difference information, and the range difference information by performing the following steps: Determine the peak shape difference feature vector corresponding to the peak shape difference information, the dose difference feature vector corresponding to the dose difference information, and the range difference feature vector corresponding to the range difference information, respectively. The information entropy corresponding to the peak shape difference feature vector, the dose difference feature vector, and the range difference feature vector is determined respectively, and the initial weights corresponding to the peak shape difference feature vector, the dose difference feature vector, and the range difference feature vector are determined based on the information entropy; The vector magnitude and vector complexity corresponding to the peak shape difference feature vector, the dose difference feature vector, and the range difference feature vector are determined respectively. The initial weights are adjusted based on the vector magnitude and the vector complexity to obtain the difference weights. The peak shape difference feature vector, the dose difference feature vector, and the range difference feature vector are weighted and fused based on the difference weights. The working state of the target particle beam emission device is determined based on the weighted fusion result.

7. The Bragg peak detector dose depth measurement system based on machine learning according to claim 1, characterized in that, The data acquisition and synchronization module is used to synchronously acquire the digital dose sequence output by the front-end readout electronics module by performing the following steps, and to spatially register the acquired digital dose sequence: The data acquisition and synchronization module includes: a trigger synchronization unit, a multi-channel data buffer unit, and a spatial registration unit. The trigger synchronization unit is used to respond to the level synchronization signal of the target particle beam and synchronously acquire the digital dose sequence output by the front-end readout electronics module based on the time in the level synchronization signal. The multi-channel data buffer unit is used to cache the digital dose sequence through a dual-port block random access memory; The spatial registration unit is used to determine the physical arrangement order of each electrode plate in the Bragg peak detector array assembly, and sort the cached digital dose sequence based on the physical arrangement order to obtain the initial dose-depth distribution curve sampled layer by layer along the target particle beam direction.

8. The Bragg peak detector dose depth measurement system based on machine learning according to claim 1, characterized in that, The machine learning processing module is further configured to construct a preset initial deep learning model; acquire a sample dataset, wherein the sample dataset includes dose-depth distribution curves of sample particle beams with labeled information, the labeled information being the actual dose-depth measurement results of the sample particle beams; divide the sample dataset into a training set and a test set, train the preset initial deep learning model using the training set, and test the trained preset initial deep learning model using the test set, and finally use the trained preset initial deep learning model that meets the test conditions as the preset deep learning model.

9. The Bragg peak detector dose depth measurement system based on machine learning according to any one of claims 1-8, characterized in that, Also includes: storage module: The storage module is used to store the dose depth measurement process data of the target particle beam and generate a dose depth measurement report of the target particle beam based on the dose depth measurement process data. The dose depth measurement process data includes at least one of the digital dose sequence, the initial dose-depth distribution curve, the restored dose-depth distribution curve, and the dose depth measurement result. The storage module is also used to respond to a query command for a dose depth measurement report, and to query and display the corresponding dose depth measurement report based on the target particle beam identifier carried by the query command.

10. A method for measuring dose depth using a Bragg peak detector based on machine learning, characterized in that, The method is applied to the machine learning-based Bragg peak detector dose depth measurement system as described in any one of claims 1-9, and the method includes: The Bragg peak detector array assembly converts the received target particle beam into an ionization signal. The front-end readout electronics module sequentially performs charge-sensitive amplification, shaping filtering, and analog-to-digital conversion on the ionization signals output by each array component to obtain a digital dose sequence; The data acquisition and synchronization module synchronously acquires the digital dose sequence output by the front-end readout electronics module, and performs spatial registration on the acquired digital dose sequence to obtain the initial dose-depth distribution curve sampled layer by layer along the target particle beam direction; The machine learning processing module extracts features and applies physical constraints to the initial dose-depth distribution curve using the preset deep learning model to obtain the dose-depth measurement results of the target particle beam. The dose-depth measurement results include at least one of the Bragg peak position, beam range, and dose-depth distribution curve.

Images