An EPID-based radiotherapy dose real-time monitoring system

Abstract

The application discloses a kind of radiation therapy dose real-time monitoring systems based on EPID, it is related to radiation therapy technical field, including real-time signal acquisition module, configured to in the treatment of beam process real-time acquisition electron field image device output exit flux signal;Dose inversion module is connected with the prior knowledge coding module and the real-time signal acquisition module, is configured to with the prior feature representation and the exit flux signal as input, utilize differentiable forward mapping constraint to solve in-vivo dose distribution.The application is by for lung tissue area using electron imbalance compensation kernel, for bone tissue area using interface dose enhancement kernel, respectively correct low-density medium in secondary electron transverse transport imbalance and the dose accumulation effect at high-density medium interface of interface.Precision problem of insufficient of traditional algorithm at density mutation interface is effectively solved, to avoid the tumor target area dose deficiency or critical organ overexposure due to dose deviation.

Description

Technical Field

[0001] This invention relates to the field of radiotherapy technology, and more specifically to a real-time radiotherapy dose monitoring system based on EPID. Background Technology

[0002] In stereotactic body radiotherapy for thoracic and abdominal tumors such as lung cancer and esophageal cancer, the beam path must pass through an extremely non-homogeneous medium composed of low-density lung tissue and high-density bone tissue. This density abrupt change leads to severe secondary electron imbalance: due to the low density of the lung tissue region, the lateral transport distance of secondary electrons far exceeds the penumbra of the radiation field, making it difficult for traditional dose deposition models (such as pencil beam convolution) to accurately describe the energy deposition after the disruption of electron balance; at the lung-bone interface, the backscattering effect of high-density bone on secondary electrons causes dose accumulation, and the actual dose at the interface can reach more than 1.2 times the calculated value in a homogeneous medium.

[0003] Existing dose reconstruction algorithms generally have an error of 5%-10% in the lung-bone junction region. For a single 10-20 Gy dose of SBRT, this is equivalent to an absolute deviation of 0.5-2 Gy, which may lead to insufficient dose to the tumor target area or excessive irradiation of adjacent organs such as the spinal cord and esophagus, seriously threatening treatment safety.

[0004] More challenging is that this inversion of non-uniform media is essentially an ill-posed mathematical problem—different combinations of incident flux distribution and tissue density can produce almost identical outgoing flux signals. Existing technologies struggle to balance accuracy and speed within a real-time monitoring framework: Monte Carlo-based offline pre-computation methods, while highly accurate, cannot adapt to the dynamic changes in lung density caused by fractional respiratory movements, and their computational latency far exceeds real-time requirements; while purely data-driven deep learning methods, though fast inference speeds, lack physical model constraints, resulting in a sharp decline in generalization ability when anatomical structures deviate from the training data distribution, with errors exceeding 10% in the lung-bone junction region.

[0005] Therefore, how to accurately invert the in vivo dose distribution across the lung-bone junction region from the output flux signal during real-time EPID monitoring has become the core technical bottleneck restricting the clinical application of EPID dose monitoring systems. Summary of the Invention

[0006] The purpose of this invention is to provide a real-time radiotherapy dose monitoring system based on EPD to address the shortcomings of the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a real-time radiotherapy dose monitoring system based on EPD, comprising: The prior knowledge encoding module is configured to acquire individualized prior knowledge of the patient, wherein the prior knowledge includes at least the patient's image density distribution and treatment plan dose distribution, and outputs prior feature representations; The real-time signal acquisition module is configured to acquire the outgoing flux signal output by the electronic field imaging device in real time during the treatment beam exit process; The dose inversion module is connected to the prior knowledge encoding module and the real-time signal acquisition module, and is configured to use the prior feature representation and the outgoing flux signal as inputs to solve the in vivo dose distribution using differentiable positive mapping constraints. The output and early warning module is connected to the dose inversion module and is configured to output the real-time in vivo dose distribution and generate an early warning signal based on the dose deviation or uncertainty quantification result.

[0008] In a preferred embodiment, the dose inversion module includes a differentiable physical projection unit, which is implemented based on a deterministic dose calculation engine, and the deterministic dose calculation engine has built-in local scattering kernel correction units for the lung tissue region and the bone tissue region. The local scattering kernel correction unit includes: For the lung tissue region, an electron imbalance compensation kernel is configured to be used in a differentiable positive mapping, the mathematical expression of which is: =K(Δr)·(1+α( )·exp ); Where α( ) is a density-dependent compensation coefficient, ranging from 0.2 to 0.8, and is inversely proportional to lung tissue density; The penumbra width correction factor ranges from 2 mm to 8 mm and is inversely proportional to lung tissue density. For the skeletal tissue region, an interface dose enhancement kernel is configured to be used in a differentiable positive mapping, the mathematical expression of which is: (Δr, )=K(Δr)+β( )·δ(Δr); Where δ(Δr) is the interface-enhancing impact function, contributing an additional dose at the bone-soft tissue interface; β( The enhancement factor (F) ranges from 0.3 to 0.9 and is positively correlated with bone density.

[0009] In a preferred embodiment, the real-time signal acquisition module is configured to acquire the output flux signal of the electronic field imaging device in real time at a frame rate of acquisition time window ≤ 200ms, and the single inference delay of the dose inversion module is less than the acquisition time window. The single inference delay of the dose inversion module is ≤100ms, the acquisition time window of the real-time signal acquisition module is 50ms to 200ms, and the inference delay of the dose inversion module is less than 0.8 times the acquisition time window, so as to meet the real-time monitoring requirements in the non-smoothing mode.

[0010] In a preferred embodiment, the dose inversion module embeds a hidden neural representation network, which is represented as follows: D(r,t)= (r，t， c(t)) ; Where r represents the three-dimensional spatial coordinates and t represents time. The prior latent code output by the prior knowledge encoding module. c(t) For the real-time measurement context, the real-time measurement context includes at least the current output flux signal segment, rack angle, and multi-leaf grating status; The network It consists of 6 to 10 fully connected layers, with each hidden layer having a dimension of 256 to 512. The activation function used is either the sinusoidal activation function or the ReLU activation function. The network The parameters are pre-trained using offline simulation data, which includes data from no fewer than 1,000 virtual patients generated based on the Monte Carlo method, and are fixed during online inference.

[0011] In a preferred embodiment, the dose inversion module introduces physical consistency loss and prior regularization loss during the training phase; The physical consistency loss is a measure of the difference between the predicted outgoing flux output by the differentiable physical projection unit and the outgoing flux signal acquired in real time, and the mean square error loss function is used. The prior regularization loss is a measure of the difference between the in vivo dose distribution output by the implicit neural representation network and the treatment plan dose distribution within a stable anatomical region, and uses the L1 loss function. The stable anatomical region is determined based on the spatial gradient of the image density distribution or motion prior: the spatial gradient amplitude of the image density distribution is calculated, and the region with a gradient amplitude of less than 0.1 HU / mm is determined as a stable region; or based on the motion vector field of 4D-CT, the region with a motion amplitude of less than 2 mm is determined as a stable region.

[0012] In a preferred embodiment, the dose inversion module is configured to use different inversion confidence weights for the lung tissue region and the bone tissue region: For the lung tissue region, the weight in the loss of physical consistency was reduced to mitigate the measurement uncertainty caused by electronic imbalance. The reduced weight was 0.3 to 0.6 times that of the liver tissue region. This range was optimized and determined by grid search: using Monte Carlo simulated dose as the gold standard, the weight coefficients in the range of 0.1 to 1.0 times were tested. The weight interval that minimized the inversion error of the lung tissue region was 0.3 to 0.6 times. For the skeletal tissue region, the weights are enhanced in the prior regularization loss to maintain the physical rationality of dose deposition. The enhanced weights are 1.5 to 2.5 times that of the liver tissue region. This range is determined by optimization using a grid search method: the weight coefficients in the range of 1.0 to 3.0 times are tested, and the weight interval that minimizes the inversion error of the skeletal tissue region is 1.5 to 2.5 times. The liver tissue region refers to the liver parenchyma region with a CT value in the range of -50HU to +100HU.

[0013] In a preferred embodiment, the dose inversion module further includes an uncertainty quantification unit, which is configured as follows: Monte Carlo dropout is introduced into the implicit neural representation network. N forward inferences are performed on the same measurement context to obtain N sampling results of the in vivo dose distribution, and the statistical variance of the N sampling results is calculated, where N≥10. When the statistical variance exceeds a preset threshold, an uncertainty warning signal is generated. The preset threshold is 0.5 to 0.8 times that of the liver tissue region in the lung and bone tissue regions. This range is determined by the receiver operating characteristic curve analysis. With the optimization goal of uncertainty warning accuracy ≥90% and false alarm rate ≤5%, the optimal threshold range selected is 0.5 to 0.8 times.

[0014] In a preferred embodiment, the real-time signal acquisition module is further configured to: Acquire the synchronization timing signal of the treatment accelerator, wherein the synchronization timing signal includes at least the gantry angle, dose rate, and multi-leaf grating blade position; The emitted flux signal and the synchronization timing signal are aligned by timestamp to form a time-synchronized measurement context.

[0015] In a preferred embodiment, the dose inversion module is implemented based on a deterministic dose calculation engine, which employs an improved cone-beam convolution stacking algorithm, the improvement including: A density-dependent scaling factor is introduced into the cone-shaped beam core, which decreases exponentially with respect to the equivalent water depth along the beam path; Local scattering correction is achieved by modifying the high-frequency components of the convolution kernel, and the correction coefficient of the high-frequency components is inversely proportional to the lung tissue density.

[0016] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By employing an electron imbalance compensation core for the lung tissue region and an interface dose enhancement core for the bone tissue region, the imbalance of secondary electron lateral transport in low-density media and the dose accumulation effect at the interface of high-density media are corrected, respectively. This effectively solves the problem of insufficient accuracy of traditional algorithms at density abrupt interfaces, avoiding insufficient dose to the tumor target area or excessive irradiation of organs due to dose deviation.

[0017] By encoding individualized prior knowledge of patients into a differentiable continuous field representation, and combining this with a differentiable physical projection unit to construct a forward mapping from in vivo dose distribution to outgoing flux, the dose inversion module can utilize gradient optimization to solve for the in vivo dose distribution in real time. It can continuously output a three-dimensional dose distribution during treatment beam exit, and can promptly detect dose deviations and trigger warnings when fractional internal respiratory movements, tumor regression, or changes in lung density occur, achieving true real-time in vivo dose monitoring.

[0018] By employing continuous field representation and pre-training mechanisms through implicit neural representation networks, combined with a lightweight network structure and GPU acceleration, single inference latency is ≤100ms, and the acquisition time window is 50ms-200ms, fully meeting the real-time monitoring requirements of 1400-2400MU / min dose rates in FFF mode. Simultaneously, the physical model constraints of the differentiable physical projection unit ensure inversion accuracy, avoiding the problem of a sharp decline in generalization ability in purely data-driven methods when the anatomical structure deviates from the training data distribution, thus achieving synergistic optimization of accuracy and speed.

[0019] By introducing a quantitative index of uncertainty in the output dose distribution using the Monte Carlo dropout method and employing stricter warning thresholds in the lung and bone tissue regions, risk references are provided for clinical decision-making. Through regionally differentiated adaptive weighting, the weight of the lung tissue region is reduced in the physical consistency loss, while the weight of the bone tissue region is increased in the prior regularization loss, thereby optimizing the regional inversion accuracy. Through the hierarchical warning mechanism and threshold self-learning optimization of the output and warning modules, a closed-loop feedback is formed to continuously improve the warning accuracy and reduce the false alarm rate. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0021] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Differentiable continuous field representation: refers to encoding discrete CT density fields and planned dose fields into continuously differentiable functions through implicit neural representation networks, so that the corresponding density value and planned dose value can be queried for any spatial coordinate, and the function output is differentiable with respect to the input coordinates.

[0024] Real-time measurement context: refers to the combination of multi-source data aligned by timestamps during the treatment beam exit process, including EPD output flux signal segments, gantry angle, dose rate, and multi-leaf grating blade positions, to provide complete information characterizing the current irradiation state.

[0025] Stable anatomical region: refers to the region with a movement amplitude of less than 2 mm during the respiratory cycle, or the region with an image density spatial gradient of less than 0.1 HU / mm. The anatomical structures in this region are relatively fixed, and the dose distribution is less affected by movement.

[0026] Example 1, please refer to Figure 1 As shown in this embodiment, a real-time radiotherapy dose monitoring system based on EPD includes: The prior knowledge encoding module is configured to acquire individualized prior knowledge of the patient, wherein the prior knowledge includes at least the patient's image density distribution and treatment plan dose distribution, and outputs prior feature representations; The real-time signal acquisition module is configured to acquire the outgoing flux signal output by the electronic field imaging device in real time during the treatment beam exit process; The dose inversion module is connected to the prior knowledge encoding module and the real-time signal acquisition module, and is configured to use the prior feature representation and the outgoing flux signal as inputs to solve the in vivo dose distribution using differentiable positive mapping constraints. The output and early warning module is connected to the dose inversion module and is configured to output the real-time in vivo dose distribution and generate an early warning signal based on the dose deviation or uncertainty quantification result.

[0027] This invention combines a differentiable physical model with implicit neural representation and introduces differentiated scattering kernel correction in the lung-bone junction region, forming a triple synergistic mechanism of "physical constraints + continuous field representation + regional adaptation". Compared with pure data-driven methods, the differentiable physical projection unit embeds the physical laws of dose deposition (ray attenuation, scattering contribution) into the inversion process, ensuring that the inversion results always satisfy the radiation transport equation and maintain accuracy even when the anatomical structure deviates from the training data distribution. Compared with Monte Carlo offline pre-computation, the implicit neural representation network encodes the three-dimensional dose field as a continuous function, achieving fast inference through pre-training and a small number of iterations, thus resolving the contradiction between computational latency and real-time requirements in traditional methods. For the electron imbalance caused by low density in lung tissue and the interface dose accumulation caused by high density in bone tissue, electron imbalance compensation kernels and interface dose enhancement kernels are constructed respectively to correct the systematic errors of traditional convolutional algorithms in regions of abrupt density changes. The three techniques mentioned above are not simply superimposed, but rather interdependent and mutually reinforcing: the differentiable physical model provides physical constraints for the implicit neural representation, avoiding the generalization risk of purely data-driven approaches; the implicit neural representation provides a continuous-field representation for the differentiable physical model, enabling it to output at arbitrary resolution; and the differential kernel correction compensates for the inherent errors of the differentiable physical model in non-uniform media. The three techniques work together to achieve the technical effect of "high precision (≤3%) + high real-time performance (≤100ms) + adaptation to non-uniform media," which is something that existing technologies cannot achieve simultaneously.

[0028] In one embodiment, specifically, the prior knowledge encoding module performs the following operations: Data Acquisition and Preprocessing Unit: Acquires the patient's treatment-positioning CT images from the hospital's image archiving system. The spatial resolution of the CT images is no less than 1mm × 1mm × 1mm, the slice thickness is no more than 2mm, and the coverage area extends from 2cm above the diaphragm to 2cm below the lower edge of the liver. Acquires the patient's treatment plan dose distribution data from the radiotherapy planning system. The treatment plan dose distribution is three-dimensional grid data with the same grid resolution as the CT images. Each grid point records the planned prescription dose value. Performs HU value conversion on the CT images, mapping the CT values ​​to relative electron density values ​​to form an image density distribution field ρ(r), where r is the three-dimensional spatial coordinate.

[0029] Implicit Neural Representation Network Construction Unit: A multilayer perceptron-structured implicit neural representation network is constructed, consisting of 6 fully connected layers, each with a hidden layer dimension of 256. A sinusoidal activation function is used. The network input is spatial coordinate r, and the output is the feature encoding of the corresponding location. The image density distribution field ρ(r) is then compared with the treatment plan dose distribution. (r) serves as a supervisory signal, training the implicit neural representation network to fit the density distribution and planned dose distribution across the entire three-dimensional space. Upon completion of training, the network outputs a low-dimensional prior code. The prior latent code is a 128-dimensional vector representing the patient's individualized anatomical and planning characteristics. Since the network is differentiable, its output continuous field representation is inherently differentiable, allowing for seamless integration with the gradient backpropagation of subsequent modules.

[0030] Continuous Field Representation Generation Unit: The trained implicit neural representation network is used as a continuous field encoder. For any spatial coordinate r, the encoder outputs the density value corresponding to that position. (r) and planned dose value (r), that is (r), (r)= (r, The advantage of this continuous field representation is that: ① It is not limited by grid resolution and can query prior information at any spatial location; ② It possesses differentiability and can participate in subsequent differentiable physical projection calculations; ③ By compressing high-dimensional CT and planning data into low-dimensional latent codes, the computational complexity of subsequent modules is reduced.

[0031] In one embodiment, specifically, the real-time signal acquisition module performs the following operations: Flux signal acquisition unit: Acquires the output flux signal from the electron beam imaging device in real time at a frame rate with an acquisition time window ≤200ms. The electron beam imaging device uses an amorphous silicon flat panel detector with an effective detection area of ​​40cm×40cm, a pixel matrix of 1024×1024, and a pixel size of 0.39mm. During the treatment beam output process, the detector continuously acquires data at an integration time of 50ms to 200ms per frame, outputting a two-dimensional projection image sequence. (t), where each pixel value represents the X-ray flux intensity received at that location. For high dose rate irradiation (1400-2400 MU / min) in un-homogenized mode (FFF), the acquisition frame rate must be no less than 10 frames / second to ensure that the rapidly changing field shape can be captured.

[0032] Synchronous timing acquisition unit: acquires the synchronous timing signal of the treatment accelerator, the synchronous timing signal including at least the gantry angle. The signals are: dose rate (DR(t)) and multi-leaf grating blade position (MLC(t)). These signals are read in real-time via the accelerator control system interface at a sampling frequency of at least 100 Hz. The gantry angle accuracy is 0.1 degrees, the dose rate accuracy is 1 MU / min, and the multi-leaf grating blade position accuracy is 0.5 mm.

[0033] The time synchronization unit aligns the emitted flux signal with the synchronization timing signal according to timestamps, forming a time-synchronized measurement context. A high-precision timestamp synchronization mechanism is employed to match each frame of image acquired by EPID with the corresponding gantry angle, dose rate, and MLC position in the accelerator log, with a timestamp deviation not exceeding 10ms. For data with misaligned timestamps, linear interpolation is used for time alignment to ensure the timing consistency of the measurement context.

[0034] In one embodiment, specifically, the differentiable physical projection unit performs the following operations: Deterministic dose calculation engine unit: Employs an improved cone-beam convolution stacking algorithm to construct the output flux from the in vivo dose distribution D(r) to the EPD plane. The mapping. The mapping consists of two steps: The first step is to calculate the incident flux distribution from the accelerator source to the patient. Based on the beam parameters (gantry angle, MLC shape, dose rate) in the treatment plan, the incident flux at the patient's body surface is calculated. (u), where u is the body surface coordinate. This calculation uses an analytical pencil beam model, taking into account factors such as the geometric penumbra of the beam and the flatness of the field.

[0035] The second step involves calculating the outgoing flux after passing through the patient using convolutional superposition. The patient's CT density distribution is converted into an electron density field, the attenuation coefficient μ(r) along the beam path is calculated, and an improved cone-beam convolutional superposition algorithm is used to calculate the outgoing flux reaching the EPD plane. (v)= ·exp ·K du; Where K(vu) is the cone-shaped beam nucleus, describing the contribution of scattered photons to the outgoing flux.

[0036] Local scattering kernel correction unit: Built-in local scattering kernel correction for lung and bone tissue regions improves computational accuracy in non-homogeneous media. Specifically includes: For the lung tissue region, an electronic imbalance compensation core is used. Correction of the pyramidal tract nucleus. Low lung tissue density (0.3-0.4 g / cm³). 3 The secondary electron balance is severely disrupted, leading to large errors in traditional convolution kernels. The electron imbalance compensation kernel is constructed based on the correlation function between lung tissue density and the penumbra of the radiation field, and its form is as follows: =K(Δr)·(1+α( )·exp ); Where α( ) is the density-dependent compensation coefficient. The smaller the value, the larger α becomes; This is the penumbra width correction factor. The smaller the value, the larger the value of σ, and the wider the range of the compensation effect.

[0037] For bone tissue regions, an interface dose-enhanced core was used. Correction of the pyramidal fasciculus. High bone density (1.5-1.8 g / cm³). 3 A dose accumulation effect exists at the bone-soft tissue interface. The interface dose enhancement kernel is constructed based on the correlation function between bone density and secondary electron transport distance, and its form is: (Δr, )=K(Δr)+β( )·δ(Δr); Where δ(Δr) is the interface-enhancing impact function, contributing an additional dose at the bone-soft tissue interface; β( () is the enhancement coefficient, which is positively correlated with bone density.

[0038] Differentiability Implementation Unit: Each step of the above forward mapping operation is designed as a differentiable operator. A deterministic dose calculation engine is implemented using an automatic differentiation framework (such as PyTorch or JAX) to ensure the accuracy of the output flux from the in vivo dose distribution D(r). The mapping of (v) is completely differentiable. This means that the gradient of the loss function with respect to the in vivo dose distribution can be calculated through backpropagation, and the in vivo dose distribution can be optimized during the inversion process.

[0039] The specific process for achieving differentiability: The PyTorch framework is used to implement the differentiability of the deterministic dose calculation engine. The specific steps are as follows: ① Convert the patient's CT density distribution ρ(r) into a PyTorch tensor and set... =True; ② Define the cone-shaped convolution kernel K(Δr) as a differentiable convolutional layer, and fix the kernel parameters through pre-calculation; ③ Implement the forward propagation function =F(D), which calculates the incident injection, the attenuation path integral, and the convolution superposition in sequence; ④ Loss function L=MSE( , Automatically construct computational graphs; ⑤ The gradient of the in vivo dose distribution D(r) with respect to the loss is automatically calculated using loss.backward(); ⑥ Update D(r) using the Adam optimizer (lr=0.001), and output the result after 3-5 iterations.

[0040] In one embodiment, specifically, the dose inversion module performs the following operations: Implicit neural representation network unit: Embedded implicit neural representation network, the network being represented as: D(r,t)= (r，t， c(t)) ; Where r is the spatial coordinate (three-dimensional) and t is the time (one-dimensional). The prior knowledge encoding module outputs a 128-dimensional prior latent code, and c(t) represents the real-time measurement context. The real-time measurement context c(t) includes at least a segment of the emitted flux signal at the current moment. (t) (flattened into an N-dimensional vector), frame angle (t) (1D), Multi-leaf grating leaf position MLC(t) ( The construction method of c(t) is as follows: the EPD signal sequence (e.g., 10 frames) within a 200ms time window is flattened into a one-dimensional vector, and concatenated with the current rack angle and MLC blade position to form a context vector of dimension M. The network... The network consists of 8 fully connected layers, each with a hidden layer dimension of 512. It uses a sinusoidal activation function and outputs the in vivo dose value D(r, t) at the current spatial location r and the current time t. The parameters are pre-trained using offline simulated data and fixed during online inference.

[0041] Specifically, the EPD detector has a resolution of 1024×1024, but to reduce computational complexity, it is downsampled to 256×256 resolution. Ten frames are acquired within a 200ms time window, and after flattening, a 256×256×10=655360-dimensional vector is obtained. The frame angle is a 1-dimensional scalar, the number of multi-leaf grating blades is 120 (60 pairs), and the MLC position is a 120-dimensional vector. The dimension of the stitched context vector c(t) is M=655360+1+120=655481. To adapt to the network input dimension, principal component analysis is used to reduce the dimension of c(t) to 512, or a convolutional neural network is used to extract features from the EPD image sequence before stitching.

[0042] Physical constraint unit: During the training phase, physical consistency loss and prior regularization loss are introduced. The physical consistency loss is the predicted outgoing flux output by the differentiable physical projection unit. (v) and the real-time acquired output flux signal The difference between (v) is measured using the mean squared error loss function: = ; Where K is the number of pixels on the EPD plane. The prior regularization loss is a measure of the difference between the in vivo dose distribution D(r) output by the implicit neural representation network and the treatment plan dose distribution D_plan(r) within the stable anatomical region, using the L1 loss function: = ; in To stabilize the anatomical region, spatial gradients or motion priors based on image density distribution are used for determination.

[0043] Specifically, calculate the spatial gradient magnitude of the image density distribution. ρ(r)| is used to classify regions with gradient magnitudes less than a threshold (e.g., 0.1 HU / mm) as stable regions; or, based on the motion vector field of 4D-CT, regions with motion amplitudes less than 2 mm are classified as stable regions. The total loss function is: = · + · ; in and These are the weighting coefficients, set to 1.0 and 0.5 respectively.

[0044] The method for extracting the 4D-CT motion vector field is as follows: 4D-CT image sequences are acquired during the patient's free breathing state. The end-expiratory phase is used as the reference phase. A non-rigid registration algorithm (such as the Demons algorithm or the B-spline free deformation model) is employed to register CT images from other phases to the reference phase. The displacement vector of each voxel in three-dimensional space is calculated to form the motion vector field. The spatial resolution of the motion vector field is 1mm × 1mm × 1mm, and each voxel records the maximum displacement amplitude at that location within the respiratory cycle.

[0045] Region Adaptive Weighting Unit: Configured to apply different inversion confidence weights to lung and bone tissue regions. In the physical consistency loss, the weight of the lung tissue region is reduced to mitigate the measurement uncertainty caused by electronic imbalance. Specifically, the equivalent thickness of lung tissue traversed by the ray path corresponding to a pixel on the EPD plane is calculated, and the physical consistency loss weight of that pixel is set to 0.3 to 0.6 times that of the liver tissue region. In the prior regularization loss, the weight of the bone tissue region is enhanced to maintain the physical plausibility of dose deposition. Specifically, the prior regularization loss weight of the bone tissue region is set to 1.5 to 2.5 times that of the liver tissue region. The liver tissue region refers to the liver parenchyma region with CT values ​​in the range of -50 HU to +100 HU, where the dose distribution is relatively stable and serves as the weighting benchmark.

[0046] Uncertainty Quantification Unit: A Monte Carlo dropout method is introduced into the implicit neural representation network. N forward inferences are performed on the same measurement context to obtain N sampled results of the in vivo dose distribution, and the statistical variance of these N sampled results is calculated. Specifically, the dropout layer of the network is kept active during the inference phase, and N=20 forward propagations are performed on the same input to obtain 20 sampled results of the in vivo dose distribution. (r). Calculate the mean μ(r) and variance σ at each spatial location. 2 (r). When the variance exceeds a preset threshold, an uncertainty warning signal is generated. The preset threshold is 0.5 to 0.8 times that of the liver tissue region in the lung and bone tissue regions, meaning that a stricter warning threshold is applied to non-homogeneous media regions to ensure the safety of high-risk areas.

[0047] The Monte Carlo dropout method is implemented as follows: a dropout layer is added after each fully connected layer of the implicit neural representation network, with a dropout rate of 0.1. During training, the dropout layer is normally activated; during inference, the dropout layer remains activated, and N=20 forward propagations are performed on the same input. In each forward propagation, 10% of the neurons are randomly dropped, resulting in 20 intra-individual dose distribution samples. The mean μ(r) and variance σ at each spatial location are calculated. 2 (r), variance σ 2(r) is the quantification value of the uncertainty of the dose distribution.

[0048] Inversion inference process: During online inference, the dose inversion module performs the following steps: ① Receive the prior latent code output by the prior knowledge encoding module ; ② Receive the real-time measurement context c(t) output by the real-time signal acquisition module; ③ Receive the differentiable positive mapping operator output by the differentiable physical projection unit; ④ Implicit neural network The output is taken as the in vivo dose distribution D(r,t) to be determined; ⑤ Calculate and predict the outgoing flux using a differentiable physical projection unit. =F(D); ⑥ Calculate the difference between the predicted and measured outgoing flux, and iteratively optimize the network output using the gradient descent method (since the network is differentiable, backpropagation can be performed) to minimize the difference; ⑦ Output the optimized in vivo dose distribution D(r,t).

[0049] Since the network is pre-trained and only requires a small number of iterations (usually 3-5 steps) during inference, the latency of a single inference is less than the acquisition time window. In actual tests on an NVIDIA A100 GPU, the latency of a single inference is ≤100ms, which meets the real-time monitoring requirements under the non-uniform mode.

[0050] The gradient descent inversion employs the Adam optimizer with a learning rate of 0.001 and 3 to 5 iterations. In each iteration, the mean square error between the predicted and measured output flux is calculated, and the gradient of the implicit neural network parameters is calculated via backpropagation to update the network's output dose distribution. In actual testing on an NVIDIA A100 GPU, a single iteration takes approximately 20ms, and the total time for 5 iterations is ≤100ms, meeting the requirements for real-time monitoring.

[0051] In one embodiment, specifically, the output and warning module performs the following operations: Dose output unit: Outputs real-time in vivo dose distribution in a 3D visualization format, overlaid on the patient's CT images, displaying the radiation dose to the target area (GTV, CTV, PTV) and organs at risk (lungs, spinal cord, esophagus, heart). Supports real-time updating and display of dose curves and dose-volume histograms.

[0052] Deviation monitoring unit: Calculates the dose deviation between the real-time in vivo dose distribution and the planned treatment dose distribution. Gamma analysis is used to calculate the dose deviation. γ(r)=sqrt

[0053] Where ΔD represents the dose difference. The reference dose is (usually a percentage of the prescribed dose), and Δd represents the distance difference. This is the allowable distance (usually 3mm). The gamma pass standard is 3% / 3mm, and a deviation warning is triggered when the pass rate is below 95%.

[0054] Warning generation unit: When the dose deviation exceeds a preset deviation threshold or the uncertainty quantification result exceeds a preset variance threshold, a graded warning signal is generated. The warnings are divided into three levels: Level 1 Warning (Red): Dosage deviation exceeds 10%, or uncertainty variance exceeds 1.5 times the preset threshold, indicating serious deviation, and it is recommended to immediately interrupt treatment; Level 2 warning (orange): Dosage deviation exceeds 8% but is less than 10%, or uncertainty variance exceeds the preset threshold but is less than 1.5 times, indicating moderate deviation. It is recommended to suspend treatment and check. Level 3 warning (yellow): The dose deviation exceeds 5% but is less than 8%, or the uncertainty variance is close to the preset threshold, indicating a slight deviation. Close monitoring is recommended.

[0055] Early Warning Optimization Unit: Dynamically adjusts preset deviation thresholds and preset variance thresholds based on historical early warning response results. It collects early warning records from the past 30 days, statistically analyzing early warning accuracy (the percentage of early warnings that actually resulted in dose deviation) and false alarm rate. Reinforcement learning algorithms (such as PPO) are used to optimize threshold parameters, aiming to increase early warning accuracy to ≥90% and reduce the false alarm rate to ≤5%.

[0056] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A real-time radiotherapy dose monitoring system based on EPID, characterized in that, include: The prior knowledge encoding module is configured to acquire individualized prior knowledge of the patient, the prior knowledge including at least the patient's image density distribution and treatment plan dose distribution, and output a prior feature representation, the prior feature representation including a prior latent code and / or continuous field representation; The real-time signal acquisition module is configured to acquire the outgoing flux signal output by the electronic field imaging device in real time during the treatment beam exit process; The dose inversion module is connected to the prior knowledge encoding module and the real-time signal acquisition module, and is configured to use the prior feature representation and the outgoing flux signal as inputs to solve the in vivo dose distribution using differentiable positive mapping constraints. The output and early warning module is connected to the dose inversion module and is configured to output the real-time in vivo dose distribution and generate an early warning signal based on the dose deviation or uncertainty quantification result.

2. The real-time radiotherapy dose monitoring system based on EPID according to claim 1, characterized in that: The dose inversion module includes a differentiable physical projection unit, which is implemented based on a deterministic dose calculation engine. The deterministic dose calculation engine has built-in local scattering kernel correction units for the lung tissue region and the bone tissue region. The local scattering kernel correction unit includes: For the lung tissue region, an electron imbalance compensation kernel is configured to be used in a differentiable positive mapping, the mathematical expression of which is: =K(Δr)·(1+α( )·exp ); Where α( ) is a density-dependent compensation coefficient, ranging from 0.2 to 0.8, and is inversely proportional to lung tissue density; The penumbra width correction factor ranges from 2 mm to 8 mm and is inversely proportional to lung tissue density. For the skeletal tissue region, an interface dose enhancement kernel is configured to be used in a differentiable positive mapping, the mathematical expression of which is: (Δr, )=K(Δr)+β( )·δ(Δr); Where δ(Δr) is the interface-enhancing impact function, contributing an additional dose at the bone-soft tissue interface; β( The enhancement factor (F) ranges from 0.3 to 0.9 and is positively correlated with bone density.

3. The real-time radiotherapy dose monitoring system based on EPID according to claim 1, characterized in that: The real-time signal acquisition module is configured to acquire the output flux signal of the electronic field imaging device in real time at a frame rate of acquisition time window ≤ 200ms, and the single inference delay of the dose inversion module is less than the acquisition time window. The single inference delay of the dose inversion module is ≤100ms, the acquisition time window of the real-time signal acquisition module is 50ms to 200ms, and the inference delay of the dose inversion module is less than 0.8 times the acquisition time window, so as to meet the real-time monitoring requirements in the non-smoothing mode.

4. The real-time radiotherapy dose monitoring system based on EPID according to claim 2, characterized in that: The dose inversion module embeds a hidden neural representation network, which is represented as follows: D(r,t)= (r，t， c(t)) ； Where r represents the three-dimensional spatial coordinates and t represents time. The prior latent code output by the prior knowledge encoding module. c (t) For the real-time measurement context, the real-time measurement context includes at least the current output flux signal segment, rack angle, and multi-leaf grating status; The network It consists of 6 to 10 fully connected layers, with each hidden layer having a dimension of 256 to 512. The activation function used is either the sinusoidal activation function or the ReLU activation function. The network The parameters are pre-trained using offline simulation data, which includes data from no fewer than 1,000 virtual patients generated based on the Monte Carlo method, and are fixed during online inference.

5. The real-time radiotherapy dose monitoring system based on EPD according to claim 4, characterized in that: The dose inversion module introduces physical consistency loss and prior regularization loss during the training phase; The physical consistency loss is a measure of the difference between the predicted outgoing flux output by the differentiable physical projection unit and the outgoing flux signal acquired in real time, and the mean square error loss function is used. The prior regularization loss is a measure of the difference between the in vivo dose distribution output by the implicit neural representation network and the treatment plan dose distribution within a stable anatomical region, and uses the L1 loss function. The stable anatomical region is determined based on the spatial gradient of the image density distribution or motion prior: the spatial gradient amplitude of the image density distribution is calculated, and the region with a gradient amplitude of less than 0.1 HU / mm is determined as a stable region; or based on the motion vector field of 4D-CT, the region with a motion amplitude of less than 2 mm is determined as a stable region.

6. The real-time radiotherapy dose monitoring system based on EPID according to claim 5, characterized in that: The dose inversion module is configured to use different inversion confidence weights for the lung tissue region and the bone tissue region: For the lung tissue region, the weight in the loss of physical consistency was reduced to mitigate the measurement uncertainty caused by electronic imbalance. The reduced weight was 0.3 to 0.6 times that of the liver tissue region. This range was optimized and determined by grid search: using Monte Carlo simulated dose as the gold standard, the weight coefficients in the range of 0.1 to 1.0 times were tested. The weight interval that minimized the inversion error of the lung tissue region was 0.3 to 0.6 times. For the skeletal tissue region, the weights are enhanced in the prior regularization loss to maintain the physical rationality of dose deposition. The enhanced weights are 1.5 to 2.5 times that of the liver tissue region. This range is determined by optimization using a grid search method: the weight coefficients in the range of 1.0 to 3.0 times are tested, and the weight interval that minimizes the inversion error of the skeletal tissue region is 1.5 to 2.5 times. The liver tissue region refers to the liver parenchyma region with a CT value in the range of -50HU to +100HU.

7. The real-time radiotherapy dose monitoring system based on EPID according to claim 1, characterized in that: The dose inversion module further includes an uncertainty quantization unit, which is configured as follows: Monte Carlo dropout is introduced into the implicit neural representation network. N forward inferences are performed on the same measurement context to obtain N sampling results of the in vivo dose distribution, and the statistical variance of the N sampling results is calculated, where N≥10. When the statistical variance exceeds a preset threshold, an uncertainty warning signal is generated. The preset threshold is 0.5 to 0.8 times that of the liver tissue region in the lung and bone tissue regions. This range is determined by the receiver operating characteristic curve analysis. With the optimization goal of uncertainty warning accuracy ≥90% and false alarm rate ≤5%, the optimal threshold range selected is 0.5 to 0.8 times.

8. The real-time radiotherapy dose monitoring system based on EPD according to claim 1, characterized in that: The real-time signal acquisition module is also configured to: Acquire the synchronization timing signal of the treatment accelerator, wherein the synchronization timing signal includes at least the gantry angle, dose rate, and multi-leaf grating blade position; The emitted flux signal and the synchronization timing signal are aligned by timestamp to form a time-synchronized measurement context.

9. A real-time radiotherapy dose monitoring system based on EPID according to claim 2, characterized in that: The dose inversion module is implemented based on a deterministic dose calculation engine, which employs an improved cone-beam convolution stacking algorithm. The improvements include: A density-dependent scaling factor is introduced into the cone-shaped beam core, which decreases exponentially with respect to the equivalent water depth along the beam path; Local scattering correction is achieved by modifying the high-frequency components of the convolution kernel, and the correction coefficient of the high-frequency components is inversely proportional to the lung tissue density.

Images