A radiotherapy process data stream scheduling method and system

Abstract

This invention relates to the field of data stream scheduling technology, specifically to a data stream scheduling method and system for radiotherapy processes. The method includes the following steps: constructing a device load matrix by collecting linear accelerator operating parameters and medical image detector transmission rates; obtaining real-time multimodal data from radiation target area image analysis to generate a data stream feature vector; calculating a nonlinear coupling weight by fusing the device load matrix and the data stream feature vector; and issuing data stream scheduling commands based on the nonlinear coupling weight and matching the communication channel. In this invention, the construction of the device load matrix, the generation of feature vectors from multimodal data from radiation target area image analysis, the calculation of specific nonlinear coupling weights by fusing the two data sets, and the direct matching of the corresponding communication channel and issuance of scheduling commands based on this weight completely changes the conventional static queue configuration's tendency to cause channel congestion when facing sudden large data volumes. It relies on real-time quantization and weight mapping mechanisms of the underlying state to allocate transmission channels.

Description

Technical Field

[0001] This invention relates to the field of data stream scheduling technology, and in particular to a data stream scheduling method and system for radiotherapy processes. Background Technology

[0002] The field of data stream scheduling technology involves the planning of transmission paths and dynamic resource allocation mechanisms for multi-source information among heterogeneous computing nodes. Among them, the traditional radiotherapy process data stream scheduling method refers to allocating transmission channels according to a preset time slice round-robin mechanism, configuring fixed priority queues, pushing accelerator control commands and real-time medical image data into corresponding buffer storage areas respectively, allocating network bandwidth in a fixed manner based on preset port mapping rules, and performing information packet forwarding and distribution operations in a first-in-first-out order.

[0003] The original scheduling rules rely heavily on the pre-defined static bandwidth allocation ratio. This fixed configuration lacks an adaptive adjustment mechanism when faced with drastic fluctuations in the status of radiotherapy equipment. This can easily lead to the accumulation of high-concurrency scanning image packets in the buffer storage area, causing continuous congestion and blockage of network communication ports. This disrupts the synchronous timing correspondence between high-frequency control signals and radiographic image feedback, resulting in severe lag and frame drops in the target tracking image. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and propose a method and system for scheduling data streams during radiotherapy.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a data stream scheduling method for radiotherapy, comprising the following steps:

[0006] S1: Call the communication interface to collect the waveguide temperature and position coordinates of the multi-leaf grating of the linear accelerator, obtain the reference transmission rate of the medical imaging detector through the velocity measurement protocol, and fuse the waveguide temperature, position coordinates and reference transmission rate to construct the device load matrix.

[0007] S2: Call the image acquisition module to obtain real-time scanning images of the radiation target area, perform pixel traversal to extract voxel gray values ​​and scattering interference intensity of the real-time scanning images, and map voxel gray values ​​and scattering interference intensity to generate data stream feature vectors.

[0008] S3: Extract the diagonal elements of the device load matrix, calculate the mean of the values ​​to generate the baseline load threshold, calculate the feature norm of the data flow feature vector, compare the feature norm with the baseline load threshold, and calculate the relative offset of the data flow feature vector from the baseline load threshold.

[0009] S4: Calculate the time series difference of waveguide temperature to generate nonlinear fluctuation gradient, fuse the relative offset and nonlinear fluctuation gradient to calculate nonlinear coupling weight, compare the nonlinear coupling weight with the baseline carrying threshold to screen communication channels, and issue data stream scheduling instructions to the communication channels.

[0010] As a further aspect of the present invention, step S1 specifically comprises:

[0011] S11: Call the communication interface to establish a bidirectional data link with the control bus, detect the physical operating status of the linear accelerator in real time according to the preset sampling frequency, extract and filter high-frequency noise signals, accurately collect the waveguide temperature and the three-dimensional spatial position coordinates of the multi-leaf grating of the linear accelerator under different radiation dose rates, and generate a set of state features.

[0012] S12: Send independent data probe packets to the medical imaging detector through the speed measurement protocol, record the delay time and channel packet loss rate parameters of the probe data packets during the round-trip transmission process, accurately calculate the current bandwidth redundancy based on the underlying network topology and historical communication logs, and obtain the baseline transmission rate by combining the round-trip delay time and packet loss rate parameters.

[0013] S13: Obtain the set of state features containing multi-dimensional parameters and the reference transmission rate generated by the front-end. Use the normalization algorithm to perform strict dimension unification processing on the waveguide temperature and position coordinates. According to the pre-calibrated matrix dimension, arrange and combine the dimension-unified values ​​and the reference transmission rate according to the physical mapping relationship of the communication nodes to establish the equipment load matrix.

[0014] As a further aspect of the present invention, step S2 specifically includes:

[0015] S21: Call the image acquisition module to activate the underlying photosensitive array element, and perform all-round multi-angle transmission scanning of the radiation target area within a specific imaging exposure window period, filter out artifact fluctuation noise caused by the patient's natural breathing movements, and acquire real-time scanning images.

[0016] S22: Execute the pixel traversal algorithm to calculate the spatial gradient of each independent pixel in the real-time scanned image, accurately identify the boundary contour features of complex anatomical structures, aggregate pixels with similar attributes through the region growth mechanism to extract voxel gray values, and estimate the scattering trajectory of particles inside the tissue based on the Monte Carlo model to extract the scattering interference intensity.

[0017] S23: Obtain voxel gray values ​​and scattering interference intensity, construct a deep neural mapping network containing nonlinear activation functions, use voxel gray values ​​as the main dimension input matrix and use scattering interference intensity to adjust the penalty weights of network parameters, and extract core spatial features through forward propagation operation of multilayer perceptron to generate data stream feature vectors.

[0018] As a further aspect of the present invention, step S3 specifically comprises:

[0019] S31: Extract all element values ​​on the main diagonal of the equipment load matrix, remove outliers that exceed three times the standard deviation statistical range, perform arithmetic mean summation on the remaining diagonal element values, and combine them with pre-calibrated hardware loss compensation coefficients for smooth scaling and nonlinear correction to generate baseline load threshold.

[0020] S32: Obtain the data flow feature vector in the high-dimensional space, call the microprocessor to calculate the sum of squares of each dimension of the vector in the Euclidean space and perform high-precision square root operation, assign independent adaptive weight parameters according to the importance of each dimension in the medical image reconstruction process, and multiply and fuse the sum of squares with the adaptive weight parameters to obtain the feature norm.

[0021] S33: Input the calculated feature norm into the comparator array, compare the feature norm with the baseline bearing threshold, calculate the absolute value difference between the feature norm and the baseline bearing threshold, divide the absolute difference by the baseline bearing threshold for dimensionless processing, and then introduce a dynamic scaling factor to perform nonlinear amplification mapping on the division result to obtain the relative offset.

[0022] As a further aspect of the present invention, step S4 specifically comprises:

[0023] S41: Call the memory management unit to retrieve the waveguide temperature corresponding to the continuous time window node in the historical cache queue, perform first-order difference calculation on the continuous temperature values ​​of adjacent time nodes to obtain the initial gradient variable, use the exponential moving average algorithm to smooth the transient thermal jitter phenomenon in the initial gradient sequence, and further calculate the higher-order derivative to fit the temperature change acceleration characteristics caused by the thermal energy accumulation effect, and generate a nonlinear fluctuation gradient.

[0024] S42: Integrating relative offset and nonlinear fluctuation gradient, based on the network redundancy allocation strategy of the underlying system, the two are cross-multiplied and exponentially decaying integrals are performed by a bivariate nonlinear function. The nonlinear coupling weight is calculated by comprehensively considering the superposition and amplification effect of thermodynamic risks and data flow congestion in the spatiotemporal dimension.

[0025] S43: Compare the nonlinear coupling weight with the baseline carrying threshold. When the weight value is determined to significantly exceed the upper limit of the threshold, forcibly remove the high-frequency interference band with low signal-to-noise ratio from the available channel resource pool. Sort the remaining communication links in descending order according to the actual remaining bandwidth margin and select the optimal communication channel with the highest transmission guarantee capability. Encapsulate the network physical address of the target device and the underlying scheduling control word, and issue a data flow scheduling instruction to the selected optimal communication channel.

[0026] As a further aspect of the present invention, the process of establishing the device load matrix specifically includes:

[0027] The system acquires the set of state features generated by the preprocessing stage, extracts the waveguide temperature and position coordinates after high-frequency noise reduction, and simultaneously accesses the monitoring module to read the reference transmission rate in a dynamically updated state. It also parses the physical mapping relationship of communication nodes associated with various data and identifies the network port matrix identifier of each node in the topology network.

[0028] Based on the rules of multidimensional spatial data mapping and vector product operation, a sequence of state variables is constructed and the covariance matrix between the waveguide temperature sequence and the reference transmission rate sequence is accurately calculated. The intrinsic relationship between physical parameters and communication parameters is quantified. The position coordinates in three-dimensional space are converted into a three-dimensional rotation matrix through a quaternion transformation algorithm and strictly embedded in the feature subspace of the covariance matrix.

[0029] The singular value decomposition or eigenvalue decomposition calculation module is invoked to extract the largest principal component vector with the highest variance contribution rate in the mixed feature subspace layer by layer. The extracted one-dimensional largest principal component vector is then deconstructed and reconstructed into a normalized square matrix form through the orthogonal basis transformation algorithm to establish the equipment load matrix.

[0030] As a further aspect of the present invention, the process of generating data stream feature vectors specifically includes:

[0031] Obtain voxel gray values ​​and scattering interference intensity, initialize the number of hidden layer nodes, activation function type, and backpropagation learning rate parameters of the multilayer perceptron in memory space, construct a deep neural mapping network with nonlinear approximation capability, and establish feature fusion calculation logic.

[0032] Based on the defined feature fusion calculation logic, the forward propagation mechanism is triggered and the underlying matrix multiplication library is called. The activation feature values ​​output by the network are calculated according to the defined operator operation process. The specific mapping formula used is as follows: ,in, The activation feature value represents the network output. Represents the Sigmoid non-linear activation function. This represents the total number of voxels extracted. Represents the voxel sequence index number. Representing the The weight coefficients of each neuron, Representing the The voxel gray value corresponding to each pixel. This represents the preset penalty control factor. Represents the intensity of the extracted scattered radiation interference;

[0033] Before the scattered interference intensity is substituted into the mapping formula for logarithmic calculation, it has been converted into a dimensionless relative value that meets the requirements of dimensional balance by introducing a reference interference constant and performing a phase ratio process.

[0034] Collect continuous activation feature values ​​from the terminal output of the deep neural mapping network, and concatenate and superimpose the feature values ​​of each component in the high-dimensional space according to the spatial coordinate sequence of the original ray. Perform principal component analysis or independent component analysis algorithm to remove secondary dimensions with variance redundancy contribution rate lower than the set safety threshold, and generate data stream feature vectors.

[0035] As a further aspect of the present invention, the process of obtaining the relative offset specifically includes:

[0036] Extract the feature norm and baseline carrying threshold, call the high-precision subtractor in the arithmetic logic unit, accurately calculate the absolute error and absolute difference of the feature norm of the current input flow relative to the baseline carrying threshold, and determine whether the obtained absolute difference is safely within the linear tolerance floating range preset by the system.

[0037] When the absolute difference is determined to be significantly beyond the system's preset linear tolerance range, triggering an overflow warning, the absolute difference is directly divided by the baseline carrying threshold to eliminate the influence of hardware scale and obtain a normalized initial deviation ratio. The frequency of short-term traffic mutations in the historical scheduling record repository is continuously monitored, and a logarithmic dynamic scaling factor with adaptive adjustment capability is constructed using the statistically obtained frequency of traffic mutations.

[0038] The standardized initial deviation ratio is losslessly input into a predefined and parameter-calibrated nonlinear amplification mapping function. The high-frequency oscillation component of the initial deviation ratio is nonlinearly weighted and amplified using the previously constructed logarithmic dynamic scaling factor to obtain the relative offset.

[0039] As a further aspect of the present invention, the process of calculating the nonlinear coupling weight specifically includes:

[0040] Extract the relative offset and nonlinear fluctuation gradient, access the underlying memory and query the global redundancy allocation strategy table pre-burned by the system, and combine the current network congestion flag to obtain the channel fading compensation coefficient and thermodynamic risk amplification coefficient applicable to the current physical moment.

[0041] The high-performance weight calculation engine inside the embedded microprocessor is invoked to perform deep fusion and numerical solution of the obtained correlation parameters based on a specially designed bivariate nonlinear function. The specific solution formula is as follows: ,in, This represents the calculated nonlinear coupling weights. Represents the channel fading compensation coefficient. Represents the relative offset of the input. Represents the natural index operation. Represents the thermodynamic risk amplification factor. Represents the gradient of nonlinear fluctuations. Represents the absolute value of the nonlinear fluctuation gradient;

[0042] Before performing square root calculation, the nonlinear wave gradient has been normalized by dividing it by the local standard environment reference gradient value, and converted into a dimensionless value that is used to adapt to the subsequent square root calculation and maintain the balance of the equation dimensions.

[0043] The nonlinear coupling weights in the original state obtained from the initial calculation engine are extracted, and the optimal state is estimated using the standard Kalman filter configured in the digital signal processor. Through iterative prediction and measurement update process, the high-frequency glitches and abnormal fluctuations caused by the sampling delay of the communication link are completely eliminated, and the nonlinear coupling weights are output.

[0044] A radiotherapy process data stream scheduling system, the system being used to implement the aforementioned radiotherapy process data stream scheduling method, the system comprising:

[0045] The equipment load coordination module is used to call the communication interface to collect the waveguide temperature and the position coordinates of the multi-leaf grating of the linear accelerator, obtain the reference transmission rate of the medical imaging detector through the velocity measurement protocol, and fuse the waveguide temperature, position coordinates and reference transmission rate to construct the equipment load matrix.

[0046] The feature dimensionality reduction mapping module is used to call the image acquisition module to obtain real-time scan images of the radiation target area, perform pixel traversal to extract voxel gray values ​​and scattering interference intensity of the real-time scan images, and map voxel gray values ​​and scattering interference intensity to generate data stream feature vectors.

[0047] The load threshold comparison module is used to extract the diagonal elements of the device load matrix, calculate the mean of the values ​​to generate the baseline load threshold, calculate the feature norm of the data flow feature vector, compare the feature norm with the baseline load threshold, and calculate the relative offset of the data flow feature vector from the baseline load threshold.

[0048] The instruction dynamic issuance module is used to calculate the time series difference of waveguide temperature to generate a nonlinear fluctuation gradient, fuse the relative offset and the nonlinear fluctuation gradient to calculate the nonlinear coupling weight, compare the nonlinear coupling weight with the baseline carrying threshold to screen the communication channel, and issue data stream scheduling instructions to the communication channel.

[0049] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0050] In this invention, real-time operating parameters such as waveguide temperature and multi-leaf grating position coordinates of the underlying physical equipment are collected. A specific equipment load matrix is ​​constructed by combining the transmission rate of the medical image detector. The voxel gray value and interference intensity of the radiation target area image are obtained to generate a data stream feature vector. The baseline carrying threshold of the equipment load matrix and the relative offset of the data stream feature vector are fused to calculate the nonlinear coupling weight. Based on this weight, the communication channel is dynamically matched and scheduling instructions are issued. This completely changes the conventional static queue configuration, which is prone to channel congestion when facing sudden large amounts of data. The transmission channel is accurately allocated based on the real-time quantization and weight mapping mechanism of the underlying hardware status, preventing delays and loss of synchronization between high-frequency radiation dose control information and image data during transmission. Attached Figure Description

[0051] Figure 1 This is a flowchart of the radiotherapy process data flow scheduling method of the present invention;

[0052] Figure 2 A flowchart for constructing the device load matrix for this invention;

[0053] Figure 3 Flowchart for generating data flow feature vectors for this invention;

[0054] Figure 4 This is a flowchart illustrating the calculation of relative offset in this invention;

[0055] Figure 5 This is a flowchart illustrating the data stream scheduling instruction issuance process of this invention. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this invention clearer, the software-based technical solution is described in detail below with reference to system architecture diagrams and embodiments. It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this invention and do not constitute a limitation on the scope of protection.

[0057] In the description of this invention, the system architecture relationships or data processing flows indicated by terms such as "layer," "module," "interface," "data flow," "client," and "server" are all defined based on the architecture diagram or flowchart corresponding to the embodiments. This way of describing is only used to clearly illustrate the logical relationships between the elements in the technical solution, and not to limit the physical deployment form. The term "multiple" includes two or more technical units, including but not limited to multiple data nodes, processing threads, service instances, or functional components and other scalable elements. The specific number is determined according to the actual business scenario and needs to be specifically specified.

[0058] Please see Figure 1 and Figure 2 This invention provides a technical solution: a method for scheduling data streams during radiotherapy, comprising the following steps:

[0059] S1: The communication interface is used to collect the waveguide temperature and position coordinates of the multi-leaf grating of the linear accelerator. The reference transmission rate of the medical imaging detector is obtained through a velocity measurement protocol. The waveguide temperature, position coordinates, and reference transmission rate are then fused to construct the device load matrix. The specific steps of S1 are as follows:

[0060] S11: Call the communication interface to establish a bidirectional data link with the control bus, detect the physical operating status of the linear accelerator in real time according to the preset sampling frequency, extract and filter high-frequency noise signals, accurately collect the waveguide temperature and the three-dimensional spatial position coordinates of the multi-leaf grating of the linear accelerator under different radiation dose rates, and generate a set of state features.

[0061] S12: Send independent data probe packets to the medical imaging detector via the rate measurement protocol, record the delay time and channel packet loss rate parameters during the round-trip transmission of the probe data packets, accurately calculate the current bandwidth redundancy based on the underlying network topology and historical communication logs, and obtain the baseline transmission rate by combining the round-trip delay time and packet loss rate parameters; obtaining the baseline transmission rate specifically involves: reading the physical layer status register of the network node to obtain the maximum physical link bandwidth, combining the round-trip delay time and channel packet loss rate, and using a weighted moving average algorithm to deduct network protocol overhead and retransmission reserved bandwidth to obtain the baseline transmission rate.

[0062] S13: Obtain the state feature set containing multi-dimensional parameters and the reference transmission rate generated by the pre-processor. Use a normalization algorithm to strictly unify the dimensions of the waveguide temperature and position coordinates. Based on the pre-calibrated matrix dimensions, arrange and combine the unified values ​​and the reference transmission rate according to the physical mapping relationship of the communication nodes to establish the equipment load matrix. The process of establishing the equipment load matrix specifically includes:

[0063] The system acquires the set of state features generated by the preprocessing stage, extracts the waveguide temperature and position coordinates after high-frequency noise reduction, and simultaneously accesses the monitoring module to read the reference transmission rate in a dynamically updated state. It also parses the physical mapping relationship of communication nodes associated with various data and identifies the network port matrix identifier of each node in the topology network.

[0064] Based on the rules of multidimensional spatial data mapping and vector product operation, a sequence of state variables is constructed and the covariance matrix between the waveguide temperature sequence and the reference transmission rate sequence is accurately calculated. The intrinsic relationship between physical parameters and communication parameters is quantified. The position coordinates in three-dimensional space are converted into a three-dimensional rotation matrix through a quaternion transformation algorithm and strictly embedded in the feature subspace of the covariance matrix.

[0065] The singular value decomposition or eigenvalue decomposition calculation module is invoked to extract the largest principal component vector with the highest variance contribution rate in the mixed feature subspace layer by layer. The extracted one-dimensional largest principal component vector is then deconstructed and reconstructed into a normalized square matrix form through the orthogonal basis transformation algorithm to establish the equipment load matrix.

[0066] A full-duplex data link is established between the RS485 serial communication interface and the underlying control bus of the linear accelerator, with a baud rate of 115200bps and an interrupt trigger sampling frequency of 1000Hz set for the hardware timer. This link is used to send status read commands to the accelerator controller, continuously acquiring the waveguide temperature and the X, Y, and Z-axis three-dimensional spatial coordinates of the multi-leaf grating at a 6MV radiation dose rate. The acquired analog electrical signals are input to a 16-bit analog-to-digital converter (ADC) and converted into discrete digital signal sequences. One-dimensional state-space equations and observation equations are constructed. This digital signal sequence is then input to a discrete Kalman filter, with the diagonal elements of the system process noise covariance matrix set to 0.01 and the measurement noise covariance matrix set to 0.1. Through a two-step iterative operation of prediction and update, high-frequency thermal noise and electromagnetic interference signals above 50Hz are effectively filtered out, obtaining accurate waveguide temperature and multi-leaf grating position coordinates.

[0067] A custom user datagram protocol (UDP) probe packet of size 64 bytes is sent to the control motherboard of the medical imaging detector at a frequency of 100 times per second, continuously sending 500 probe packets. The round-trip time from sending to receiving an acknowledgment message for each probe packet is recorded at the application layer of the communication protocol stack, and the number of packets that do not receive an acknowledgment message is counted to calculate the channel packet loss rate. The gigabit Ethernet physical layer status register of the network node is read to obtain the maximum physical link bandwidth of 1000 Mbps. Combining the round-trip time and the channel packet loss rate, a weighted moving average algorithm is used to deduct network protocol overhead and retransmission reserved bandwidth to obtain the baseline transmission rate. The waveguide temperature, three-dimensional spatial coordinates, and baseline transmission rate are combined to form a multi-dimensional state feature vector. The minimum baseline temperature is set to 20℃, and the maximum extreme temperature is set to 60℃.

[0068] The boundary range of the position coordinates is set to -100mm to +100mm. Using a linear minimum-maximum normalization algorithm, each element in the aforementioned multidimensional state feature vector is mapped to a dimensionless interval of 0 to 1. The normalized waveguide temperature time series and the reference transmission rate time series are extracted, and the covariance of the two series over 100 sampling periods is calculated to construct a 3x3 covariance matrix, quantifying the physical interference of temperature drift on the communication rate. A quaternion rotation statistical algorithm is used, introducing a rotation factor w set to 0.9, and setting the parameters of the x, y, and z vector directions to 0.1, converting the three-dimensional spatial position coordinates into a 3x3 three-dimensional rotation matrix. A matrix adder is used to directly superimpose the three-dimensional rotation matrix into the eigenspace of the covariance matrix, forming a mixed-state matrix. The Jacobian eigenvalue solving algorithm is used to diagonalize the mixed-state matrix, and after 20 orthogonal transformation iterations, the first-order principal component vector corresponding to the largest eigenvalue is extracted. The elements of the first-order principal component vector are reshaped into a normalized 4x4 dimension device load matrix in row-major order.

[0069] Please see Figure 1 and Figure 3 S2: The image acquisition module is invoked to obtain real-time scanned images of the radiation target area. Pixel traversal is performed to extract voxel grayscale values ​​and scattering interference intensity from the real-time scanned images. The voxel grayscale values ​​and scattering interference intensity are mapped and calculated to generate a data stream feature vector. The specific steps of S2 are as follows:

[0070] S21: Call the image acquisition module to activate the underlying photosensitive array element, and perform all-round multi-angle transmission scanning of the radiation target area within a specific imaging exposure window period, filter out artifact fluctuation noise caused by the patient's natural breathing movements, and acquire real-time scanning images.

[0071] S22: Execute the pixel traversal algorithm to calculate the spatial gradient of each independent pixel in the real-time scanned image, accurately identify the boundary contour features of complex anatomical structures, aggregate pixels with similar attributes through the region growth mechanism to extract voxel gray values, and estimate the scattering trajectory of particles inside the tissue based on the Monte Carlo model to extract the scattering interference intensity.

[0072] S23: Obtain voxel gray values ​​and scattering interference intensity, construct a deep neural mapping network containing nonlinear activation functions, use voxel gray values ​​as the main dimension input matrix and adjust the network parameter penalty weights using scattering interference intensity, extract core spatial features through forward propagation operation of a multilayer perceptron, and generate a data stream feature vector. The process of generating the data stream feature vector specifically includes:

[0073] Obtain voxel gray values ​​and scattering interference intensity, initialize the number of hidden layer nodes, activation function type, and backpropagation learning rate parameters of the multilayer perceptron in memory space, construct a deep neural mapping network with nonlinear approximation capability, and establish feature fusion calculation logic.

[0074] Based on the established feature fusion calculation logic, voxel gray values ​​are used as the main dimension input matrix, and the penalty weights of the network parameters are adjusted using the intensity of scattering interference. This triggers the forward propagation mechanism and calls the underlying matrix multiplication library. The activation feature values ​​of the network output are calculated according to the established operator operation process. The specific mapping formula used is as follows: ,in, The activation feature value represents the network output. Represents the Sigmoid non-linear activation function. This represents the total number of voxels extracted. Represents the voxel sequence index number. Representing the The weight coefficients of each neuron, Representing the The voxel gray value corresponding to each pixel. This represents the preset penalty control factor. The extracted scattering interference intensity represents the feature fusion calculation logic, which refers to the feature dimensionality reduction and fusion mechanism based on deep neural mapping networks. Specifically, it uses the voxel gray values ​​of medical images as the main dimension input of the network, while introducing the scattering interference intensity in the physical environment as a dynamic penalty term. Through forward propagation and nonlinear activation formulas, it achieves mathematical decoupling and feature fusion of the main features and interference factors.

[0075] Before the scattered interference intensity is substituted into the mapping formula for logarithmic calculation, it has been converted into a dimensionless relative value that meets the requirements of dimensional balance by introducing a reference interference constant and performing a phase ratio process.

[0076] Collect continuous activation feature values ​​from the terminal output of the deep neural mapping network, and concatenate and superimpose the feature values ​​of each component in the high-dimensional space according to the spatial coordinate sequence of the original ray. Perform principal component analysis or independent component analysis algorithm to remove secondary dimensions with variance redundancy contribution rate lower than the set safety threshold, and generate data stream feature vectors.

[0077] A trigger level is sent to the control terminal of the amorphous silicon flat panel image acquisition unit to activate the underlying 1024x1024 pixel photosensitive array element. The imaging exposure window is strictly controlled to 20ms, and the gantry is driven to perform a 360-degree omnidirectional transmission scan around the center of the field of view with a rotation step angle of 15 degrees. The initial two-dimensional transmission image sequence obtained from the scan is extracted, and the image is transformed from the spatial domain to the frequency domain using Fast Fourier Transform. Artifacts and fluctuation noise of the patient's natural breathing motion in the frequency band of 0.2Hz to 0.5Hz are identified and removed. Then, Inverse Fourier Transform is applied to reconstruct a high-fidelity real-time scan image. The Sobel operator is used to perform spatial gradient convolution calculations on the independent pixels of the real-time scan image in the horizontal and vertical directions. The gradient amplitude discrimination threshold is set to 50, and pixels with an amplitude greater than 50 are selected as the boundary contour features of the anatomical structure. A pixel with a gray value of 150 is randomly selected at the core position of the target area as the initial seed point for region growth, and the gray value tolerance range for region growth is set to ±20. The algorithm iterates through the 26 neighboring pixels, grouping those that meet the grayscale tolerance condition into a single set. It then calculates the arithmetic mean of all pixels within this set to extract the voxel grayscale value. A Monte Carlo computational geometry model for radioactive particle transport is established, with a total of 100,000 incident photons, simulating Compton scattering and the photoelectric effect in muscle and bone tissue. The number of scattered photons deviating more than 5 degrees from the initial incident trajectory is tracked and recorded, and their ratio to the total number of incident photons is quantified to extract the intensity of the scattered interference. The network topology of the multilayer perceptron is initialized in memory, constructing a deep neural mapping network containing one input layer, three hidden layers, and one output layer. The input layer has 1024 nodes to receive the voxel grayscale value sequence. The three hidden layers have 512, 256, and 128 nodes respectively, all using the ReLU nonlinear activation function. The output layer has 64 nodes and uses the Sigmoid nonlinear activation function. The loss function is a linear combination of mean squared error and L2 regularization, with a regularization penalty coefficient set to 0.005. The optimizer uses the Adam algorithm, with an initial learning rate of 0.001, a batch size of 32, and 100 training iterations. After backpropagation updates the network weight parameters, the forward propagation mechanism is triggered, calling the matrix multiplication operator from the underlying linear algebra subroutine library. The activation feature values ​​of the network output are calculated according to the set operator operation flow; the mapping formula is as follows: .

[0078] in, Represents continuous activation feature values. Represents the Sigmoid non-linear activation function. This represents the total number of input nodes in the network. Represents the index number of the input node. Representative effect on the first The neuron weight coefficients corresponding to each pixel. Representing the The voxel gray value corresponding to each pixel. This represents the preset penalty control factor applied to the interference term. This represents the logarithmic function with the natural constant as its base. Represents the dimensionless intensity of scattered radiation interference;

[0079] The non-numerical scattering interference intensity is quantified by dividing its absolute count value by the reference interference constant 500 to obtain the dimensionless scattering interference intensity I. The total number of input nodes N in the network is set to 1024. The weight coefficient ω_i of a specific neuron is extracted as 0.0000095, the voxel gray value V_i is extracted as 180, the preset penalty control factor λ is extracted as 0.05, and the dimensionless scattering interference intensity I is extracted as 0.12. Substituting these parameters into the aforementioned formula, firstly, the product of 0.0000095 and 180 is calculated and summed over all 1024 nodes to obtain the basic mapping value 1.75104; then, the sum of 1 and 0.12 is calculated and its natural logarithm is taken, multiplied by 0.05 to obtain the penalty term 0.00566; the penalty term is subtracted from the basic mapping value, and finally, the result is input into the Sigmoid function for nonlinear mapping to obtain the continuous activation feature value F as 0.85. The advantage of this formula lies in its effective suppression of blurring artifacts at the edges of medical images by introducing the intensity of scattering interference as a dynamic penalty term, thereby improving the spatial fidelity of the feature data. This result indicates that the extracted feature values ​​have sufficiently decoupled from scattering background noise and can be directly used for downstream scheduling. 64-dimensional continuous activation feature values ​​from the terminal outputs of the deep neural mapping network were collected, and principal component analysis was used to calculate the covariance matrix and eigenvalues ​​of this 64-dimensional feature matrix. The eigenvalues ​​were sorted in descending order of magnitude, and the variance contribution rate of each eigenvalue was calculated. A safety threshold of 95% for the variance redundancy contribution rate was set, and secondary dimensions with cumulative contribution rates below 95% were removed, retaining the first 16 principal component dimensions to generate the final 16-dimensional data stream feature vector.

[0080] Table 1 Physical Status and Communication Parameter Monitoring Table

[0081] Table 1 shows the operating status parameters of different underlying communication nodes at a certain physical moment, verifying the data diversity and accuracy of the high-frequency acquisition process.

[0082] The aforementioned region growing refers to a computer vision image segmentation technique that starts from a selected initial seed point and, based on a preset similarity criterion, continuously aggregates neighboring pixels that meet the conditions into the region where the seed point is located until no pixels that meet the conditions can be included.

[0083] Please see Figure 1 and Figure 4 S3: Extract the diagonal elements of the device load matrix, calculate the mean of the values ​​to generate the baseline load threshold, calculate the feature norm of the data flow feature vector, compare the feature norm with the baseline load threshold, and calculate the relative offset of the data flow feature vector from the baseline load threshold.

[0084] The specific steps for S3 are as follows:

[0085] S31: Extract all element values ​​on the main diagonal of the equipment load matrix, remove outliers that exceed three times the standard deviation statistical range, perform arithmetic mean summation on the remaining diagonal element values, and combine them with pre-calibrated hardware loss compensation coefficients for smooth scaling and nonlinear correction to generate baseline load threshold.

[0086] S32: Obtain the data flow feature vector in the high-dimensional space, call the microprocessor to calculate the sum of squares of each dimension of the vector in the Euclidean space and perform high-precision square root operation, assign independent adaptive weight parameters according to the importance of each dimension in the medical image reconstruction process, and multiply and fuse the sum of squares with the adaptive weight parameters to obtain the feature norm.

[0087] S33: Input the calculated feature norm into the comparator array, compare the feature norm with the baseline bearing threshold, calculate the absolute value difference between the feature norm and the baseline bearing threshold, divide the absolute difference by the baseline bearing threshold for dimensionless processing, and then introduce a dynamic scaling factor to perform nonlinear amplification mapping on the division result to obtain the relative offset.

[0088] The process of obtaining the relative offset specifically includes:

[0089] For each dimension's core contribution in the medical image reconstruction process, an independent adaptive weight parameter is assigned. The adaptive weight parameter is calculated comprehensively. The initial vector length obtained by taking the square root of the sum of squares of each dimension is multiplied and fused with the comprehensively calculated adaptive weight parameter to obtain the final feature norm.

[0090] Extract the feature norm and baseline carrying threshold, call the high-precision subtractor in the arithmetic logic unit, accurately calculate the absolute error and absolute difference of the feature norm of the current input flow relative to the baseline carrying threshold, and determine whether the obtained absolute difference is safely within the linear tolerance floating range preset by the system.

[0091] When the absolute difference is determined to be significantly beyond the system's preset linear tolerance range, triggering an overflow warning, the absolute difference is directly divided by the baseline carrying threshold to eliminate the influence of hardware scale and obtain a normalized initial deviation ratio. The frequency of short-term traffic mutations in the historical scheduling record repository is continuously monitored, and a logarithmic dynamic scaling factor with adaptive adjustment capability is constructed using the statistically obtained frequency of traffic mutations.

[0092] The standardized initial deviation ratio is losslessly input into a predefined and parameter-calibrated nonlinear amplification mapping function. The high-frequency oscillation component of the initial deviation ratio is nonlinearly weighted and amplified using the previously constructed logarithmic dynamic scaling factor to obtain the relative offset.

[0093] Scan the main diagonal of the 4x4 device load matrix and extract the values ​​of the four core diagonal elements: 0.85, 0.88, 0.82, and 0.90. Calculate the arithmetic mean of these four values ​​as 0.8625 and their standard deviation as 0.0303. The rule for identifying outliers is defined as deviations from the mean exceeding three times the standard deviation (i.e., greater than 0.9534 or less than 0.7716). After comparison, none of the four values ​​exceed the statistical range of three times the standard deviation, so no outlier removal is necessary. Perform an arithmetic mean summation on the four diagonal elements, keeping the mean value of 0.8625 unchanged. Call the system's preset parameter configuration file and read the hardware loss compensation coefficient, set to 1.05. Multiply the arithmetic mean of 0.8625 by the hardware loss compensation coefficient of 1.05 and scale the result to obtain a value of 0.9056. Inputting 0.9056 into the exponential smoothing function for nonlinear correction generates a final baseline carrying capacity threshold of 0.928.

[0094] The 16-dimensional data stream feature vector, obtained after dimensionality reduction, is read. A 2.4GHz floating-point microprocessor is used to square the values ​​of each of the 16 dimensions of the vector in Euclidean space, and the sum of the 16 squares is achieved using a hardware multiplier-accumulator. The microprocessor's internal fast square root instruction set is then used to perform high-precision square root operations on the sum of squares, resulting in an initial vector length of 2.15. Independent adaptive weight parameters are assigned to the core contributions of these 16 dimensions in the medical image reconstruction process: the weight parameter for the core contour dimension is set to 1.2, and the weight parameter for the background detail dimension is set to 0.8. The initial vector length of 2.15 is multiplied and fused with the calculated adaptive weight parameter of 1.14 to obtain the final feature norm of 2.45. The calculated feature norm of 2.45 and the generated baseline carrying threshold of 0.928 are simultaneously input into a 64-bit high-precision digital comparator array configured within the FPGA. The high-precision subtractor within the arithmetic logic unit is invoked, and the baseline carrying threshold of 0.928 is subtracted from the feature norm of 2.45 to accurately calculate the absolute numerical difference of 1.522. The system security configuration file is read, and the system's pre-set linear tolerance fluctuation range is determined to be 0.05 to 0.50. Comparison reveals that the absolute numerical difference of 1.522 is significantly greater than the upper limit of 0.50, exceeding the linear tolerance fluctuation range and triggering the system's highest-level memory overflow interrupt warning. Dividing this absolute difference of 1.522 directly by the baseline carrying threshold of 0.928 eliminates the influence of the physical scale of the underlying hardware devices, resulting in a normalized initial deviation ratio of 1.64. A structured query statement is sent to the non-volatile memory where the historical scheduling record repository is located to continuously monitor the frequency of short-term traffic mutations within the past 60-second time window. The frequency of traffic mutations is statistically determined to be 15 times. Based on this mutation frequency of 15, the logarithm with the natural constant as the base is calculated, and a basic compensation constant of 0.5 is appended to the result to construct a logarithmic dynamic scaling factor of 3.2. The normalized initial deviation ratio of 1.64 is input into the calibrated nonlinear amplification mapping function. Inside this function, a logarithmic dynamic scaling factor of 3.2 is used to perform multiplicative weighting and exponential enhancement calculations on the high-frequency oscillation component in the initial deviation ratio of 1.64, ultimately calculating the relative offset to be 5.25.

[0095] Please see Figure 1 and Figure 5 S4: Calculate the time-series difference of the waveguide temperature to generate a nonlinear fluctuation gradient. Fuse the relative offset and the nonlinear fluctuation gradient to calculate the nonlinear coupling weight. Compare the nonlinear coupling weight with the baseline carrying threshold to screen communication channels. Issue data stream scheduling instructions to the communication channels. The specific steps of S4 are as follows:

[0096] S41: Call the memory management unit to retrieve the waveguide temperature corresponding to the continuous time window node in the historical cache queue, perform first-order difference calculation on the continuous temperature values ​​of adjacent time nodes to obtain the initial gradient variable, use the exponential moving average algorithm to smooth the transient thermal jitter phenomenon in the initial gradient sequence, and further calculate the higher-order derivative to fit the temperature change acceleration characteristics caused by the thermal energy accumulation effect, and generate a nonlinear fluctuation gradient.

[0097] S42: Integrating relative offset and nonlinear fluctuation gradient, based on the network redundancy allocation strategy of the underlying system, the two are cross-multiplied and exponentially decaying integrals are performed by a bivariate nonlinear function. The nonlinear coupling weight is calculated by comprehensively considering the superposition and amplification effect of thermodynamic risks and data flow congestion in the spatiotemporal dimension.

[0098] S43: Compare the nonlinear coupling weight with the baseline carrying threshold. When the weight value is determined to significantly exceed the upper limit of the threshold, forcibly remove the high-frequency interference band with low signal-to-noise ratio from the available channel resource pool. Sort the remaining communication links in descending order according to the actual remaining bandwidth margin and select the optimal communication channel with the highest transmission guarantee capability. Encapsulate the network physical address of the target device and the underlying scheduling control word, and issue a data flow scheduling instruction to the selected optimal communication channel.

[0099] The process of calculating the nonlinear coupling weights specifically includes:

[0100] Extract the relative offset and nonlinear fluctuation gradient, access the underlying memory and query the global redundancy allocation strategy table pre-burned by the system, and combine the current network congestion flag to obtain the channel fading compensation coefficient and thermodynamic risk amplification coefficient applicable to the current physical moment.

[0101] The high-performance weight calculation engine inside the embedded microprocessor is invoked to perform deep fusion and numerical solution of the obtained correlation parameters based on a specially designed bivariate nonlinear function. The specific solution formula is as follows: ,in, This represents the calculated nonlinear coupling weights. Represents the channel fading compensation coefficient. Represents the relative offset of the input. Represents the natural index operation. Represents the thermodynamic risk amplification factor. Represents the gradient of nonlinear fluctuations. Represents the absolute value of the nonlinear fluctuation gradient;

[0102] Before performing square root calculation, the nonlinear wave gradient has been normalized by dividing it by the local standard environment reference gradient value, and converted into a dimensionless value that is used to adapt to the subsequent square root calculation and maintain the balance of the equation dimensions.

[0103] The nonlinear coupling weights in the original state obtained from the initial calculation engine are extracted, and the optimal state is estimated using the standard Kalman filter configured in the digital signal processor. Through iterative prediction and measurement update process, the high-frequency glitches and abnormal fluctuations caused by the sampling delay of the communication link are completely eliminated, and the nonlinear coupling weights are output.

[0104] Data from the circular history cache queue in static random access memory is retrieved via the direct memory access channel of the memory management unit. The waveguide temperature sequences corresponding to five consecutive time windows are extracted: 40.1℃, 40.3℃, 40.6℃, 41.0℃, and 41.5℃. The arithmetic logic unit is invoked to perform first-order difference calculations on the temperature values ​​of adjacent time windows, obtaining an initial gradient variable sequence of 0.2, 0.3, 0.4, and 0.5. The smoothing factor coefficient of the exponential moving average algorithm is set to 0.3 to filter and smooth the transient thermal jitter in the initial gradient sequence, resulting in a smoothed gradient sequence of 0.23, 0.28, 0.35, and 0.42. The smoothed sequence is then subjected to another first-order difference to calculate higher-order derivatives, obtaining temperature change acceleration features of 0.05, 0.07, and 0.07. The maximum value of 0.07 is determined as the nonlinear fluctuation gradient. Access the pre-programmed global redundancy allocation strategy table in the underlying flash memory, read the applicable channel fading compensation coefficient α for the current physical moment and set it to 0.4, and read the thermodynamic risk amplification coefficient β and set it to 0.6. Obtain the local standard environment reference gradient value of 0.02, divide the nonlinear fluctuation gradient 0.07 by 0.02 to obtain the dimensionless normalized nonlinear fluctuation gradient G used for adapting the square root calculation, which is 3.5. Call the high-performance floating-point calculation engine inside the embedded microprocessor to perform numerical solution based on the specially designed bivariate nonlinear function. The solution formula is: .

[0105] in, Represents the nonlinear coupling weights in the original state. The channel fading compensation coefficient represents the factor acting on the exponential function term. This represents an exponential function with the natural constant as its base. Represents the relative offset. This represents the thermodynamic risk amplification factor acting on the fluctuation gradient term. Represents the absolute value of the normalized nonlinear fluctuation gradient. Represents the square root operation function. Represents the dimensionless normalized nonlinear wave gradient;

[0106] The relative offset O is extracted as 5.25 and input into the natural exponential calculation module to obtain the exponential amplification result. The square root of the absolute value of the dimensionless nonlinear fluctuation gradient, 3.5, is obtained to obtain the square root result. The channel fading compensation coefficient, 0.4, is multiplied with the exponential amplification result, and the thermodynamic risk amplification coefficient, 0.6, is multiplied with the square root result. Finally, the two multiplication results are fed into an adder for superposition, and the nonlinear coupling weight W in the original state is accurately calculated to be 77.35. The standard Kalman filter configured in the digital signal processor is extracted, and the nonlinear coupling weight in the original state is used as the measurement input. The state transition matrix is ​​set as the identity matrix, the initial value of the prediction covariance is 1.0, and the measurement noise variance is 0.5. After 10 measurement update processes, high-frequency glitches and abnormal fluctuations caused by communication link sampling delay are effectively filtered out, and the final nonlinear coupling weight is output as 76.0. The baseline carrying threshold is read as 0.928, and the upper limit of the weight allowed by the system is set to 50 times the baseline carrying threshold, i.e., 46.4. The comparator output status shows that the nonlinear coupling weight of 76.0 significantly exceeds the threshold limit of 46.4. The channel intervention mechanism is activated, scanning the eight frequency bands in the currently available channel resource pool. The signal-to-noise ratio (SNR) parameters of each frequency band are extracted, and the 2.4GHz and 2.5GHz bands with SNR below 20dB are directly isolated and removed from the available resource pool. The actual remaining bandwidth of the remaining six communication links is measured: 150Mbps, 300Mbps, 800Mbps, 500Mbps, 200Mbps, and 400Mbps, respectively. A quicksort algorithm is used to sort the actual remaining bandwidth in descending order, selecting the 5.8GHz band with the largest bandwidth margin (800Mbps) as the optimal communication channel with the highest transmission guarantee capability. The target device's media access control physical address and underlying scheduling control bytes are encapsulated into a standard data frame format, and a data stream scheduling command is issued to the optimal 5.8GHz communication channel through the network interface controller. With parameters α set to 0.4 and β set to 0.6, the system's data flow scheduling latency was reduced to 5ms, which objectively improved the response speed compared to the benchmark test method without dynamic scheduling.

[0107] A radiotherapy process data stream scheduling system, the system being used to execute the above-mentioned radiotherapy process data stream scheduling method, the system comprising:

[0108] The device load coordination module is activated, and the RS485 serial data transceiver interface configured inside the high-speed field-programmable gate array (FPGA) is enabled. Driven by a 1000Hz hardware clock, the transceiver continuously listens for and receives data packets from the underlying sensor bus of the linear accelerator. The received analog waveforms are sent to a 16-bit analog-to-digital converter chip, which converts them into discrete digital signal streams representing the waveguide temperature and the position coordinates of the multileaf grating. The digital signal stream is directly pushed into a first-in-first-out (FIFO) hardware buffer queue with direct memory access to prevent data frame overflow and loss during high-speed concurrency. The digital signal processor reads the data from the buffer queue through the internal high-speed bus and calls the built-in hardware multiply-accumulate unit to execute the one-dimensional Kalman filter algorithm. By iterating the values ​​of the prediction covariance register and the Kalman gain register, high-frequency electromagnetic interference noise attached to the original data is removed, and the cleaned waveguide temperature and multileaf grating position coordinates are stored in a designated memory block of static random access memory (SRAM). The gigabit Ethernet physical layer transceiver is activated, and 64-byte network probe data packets are continuously sent to the medical imaging detector network card via a user datagram protocol socket. The network protocol stack's timer module records the machine clock cycle emitted by the probe and records the end of the clock cycle when the network interface card (NIC) interrupt controller receives an acknowledgment packet. The CPU's arithmetic unit calculates the time difference to obtain the round-trip delay, and the network counter counts the number of unacknowledged packets to calculate the packet loss rate. The arithmetic unit, combined with the maximum link bandwidth status word indicated in the physical layer register, executes a weighted moving average algorithm to calculate the baseline transmission rate, which is then stored in the shared memory pool. The microprocessor reads the temperature, coordinates, and rate data from the shared memory pool, calls the hardware divider to perform min-max normalization scaling, and unifies all data to the floating-point range of 0 to 1. It then calls the matrix operation coprocessor to feed the normalized data into a cross-multiplier array to calculate the covariance matrix. Finally, the hardware quaternion rotation calculation unit converts the position coordinates into a three-dimensional rotation matrix and embeds it into the covariance matrix using matrix addition instructions. The singular value decomposition hardware acceleration core is activated. After multiple pipeline iterations of the Jacobian rotation circuit, the principal component vector corresponding to the largest eigenvalue is output. Then, the data rearrangement circuit formats this one-dimensional vector into a 4x4 device load matrix and latches it into a high-speed register group for subsequent module calls.

[0109] The feature reduction mapping module is activated, connecting the high-bandwidth video memory channel within the graphics processor (GPU). A synchronous exposure pulse signal is sent to the image acquisition interface of the amorphous silicon flat panel detector, controlling the CMOS photosensitive array element to acquire X-ray transmitted photons and convert them into charge signals within a 20ms hard real-time exposure window. The charge signals are converted into a high-resolution raw two-dimensional transmission image matrix via a low-voltage differential signal transmission line and directly written to the GPU's frame buffer. The GPU invokes the unified computing device architecture core, initiates the Fast Fourier Transform (FFT) thread block, transfers the image matrix into the frequency domain buffer, and removes artifact fluctuation noise reflecting breathing motion by configuring the stopband filter with a lower stopband limit of 0.2Hz and an upper stopband limit of 0.5Hz. Then, the Inverse Fourier Transform (IFT) thread block is initiated to reconstruct and output the real-time scan image. The GPU's texture picking unit performs a two-dimensional sliding window convolution operation on the image memory block using a Sobel operator convolution kernel. A digital comparator determines the gradient magnitude of the convolution output and pushes the pixel coordinates with magnitudes greater than 50 into the boundary feature queue. The parallel computing kernel for region growing is activated. A pixel with a grayscale value of 150 is selected as the initial growth point in the image's core memory region, and the grayscale tolerance range register is set to ±20. Pixels meeting the tolerance are clustered and written to a specific memory pool by traversing the 26-neighborhood addresses in memory. The mean is calculated using a reduction and summation instruction, and the voxel grayscale value is output. The Monte Carlo random number generation circuit is activated, injecting 100,000 photon trajectory calculation tasks into the ray tracing engine. The number of scattered photons deviating more than 5 degrees from the preset trajectory in memory is counted and marked. The dimensionless scattered ray interference intensity is calculated and output using a divider. A pre-trained deep neural mapping network weight file is loaded into the GPU's constant memory. The network structure includes an input layer with 1024 input channels, three hidden layers with 512, 256, and 128 nodes respectively, and an output layer with 64 output channels. The tensor computation core reads the voxel grayscale value sequence from the video memory pool and performs a large-scale dot product multiplication and accumulation operation with the weight matrix in constant memory. The operation result is mapped through a hardware lookup table configured with ReLU nonlinear activation logic and passed layer by layer to the output. The floating-point arithmetic unit is then invoked, according to the formula... The video memory is read with N configured as 1024, ω_i configured as 0.0000095, V_i configured as 180, the preset penalty factor λ configured as 0.05, and the dimensionless interference intensity I configured as 0.12.

[0110] in, Represents continuous activation feature values. Represents the Sigmoid non-linear activation function. This represents the total number of input nodes in the network. Represents the index number of the input node. Representative effect on the first The neuron weight coefficients corresponding to each pixel. Representing the The voxel gray value corresponding to each pixel. This represents the penalty control factor acting on the interference penalty term. This represents the logarithmic function with the natural constant as its base. Represents the dimensionless intensity of scattered radiation interference;

[0111] The floating-point multiplier and adder tree calculate the basic mapping value, while the logarithmic operation core calculates the natural logarithm of the interference intensity and the penalty term. After the subtractor performs the subtraction operation, the result is sent to the Sigmoid hardware lookup table to output the continuously active feature values. The principal component analysis coprocessor is called to calculate the covariance and eigenvalues ​​of the continuously active feature values. The sorting circuit removes secondary dimension register data with a variance redundancy contribution rate of less than 95%, retains the floating-point data of the first 16 dimensions, and outputs a 16-dimensional data stream feature vector to the next stage of the pipeline.

[0112] Table 2 Threshold Parameter Settings and Offset Statistics 1 0.86 2.45 1.64 1 2 0.88 2.50 1.70 1 3 0.85 1.05 0.20 0 4 0.87 3.10 2.30 1

[0113] The load threshold comparison module is activated, and the address generator is configured to scan the 4x4 dimension device load matrix latched in the front-end high-speed register group. The address generator sequentially outputs the memory physical addresses of the diagonal elements, controlling the multiplexer to extract the data of the four diagonal elements with values ​​of 0.85, 0.88, 0.82, and 0.90 into the temporary accumulator of the arithmetic logic unit. The arithmetic logic unit performs an arithmetic mean operation using a shift register and an adder to obtain a mean of 0.8625. Subsequently, the standard deviation hardware calculation core is activated to obtain a standard deviation of 0.0303. The comparator is controlled to determine whether each element deviates from the mean by more than 3 times the standard deviation threshold network. After verifying that all elements are safe, the mean of 0.8625 is latched into the main arithmetic register. The preset hardware loss compensation coefficient constant of 1.05 in the system read-only memory is retrieved, and the mean of 0.8625 is multiplied by 1.05 using a hardware multiplier, outputting the result 0.9056 to the exponential smoothing circuit. After low-pass filtering and mapping, the smoothing circuit generates a baseline carrying threshold of 0.928 at the output port and sends it to the left reference port of the shared digital comparator array. The direct memory access controller is enabled, and the reduced 16-dimensional data stream feature vector is read in parallel from the video memory output channel into the vector inner product processor. The processor's 16 multiplier channels are simultaneously activated, multiplying the floating-point data of each dimension by themselves to calculate the square. A multi-stage pipelined addition tree cascades and sums the 16 squared results. The summation result enters the hardware fast square root evaluation circuit, which converges after 5 clock cycles using Newton's iteration method, outputting an initial vector length of 2.15. The processor reads the weight register file, obtains the adaptive weight of 1.2 for the core contour dimension and the adaptive weight of 0.8 for the background detail dimension, and calculates the comprehensive weight parameter of 1.14. The floating-point multiplier multiplies the initial vector length of 2.15 with the comprehensive weight of 1.14 to obtain the feature norm of 2.45, and sends it to the right input port of the shared digital comparator array. The comparator array synchronously receives the characteristic norm of 2.45 and the baseline bearing threshold of 0.928. An internal 64-bit high-precision floating-point subtractor is activated, subtracting 0.928 from 2.45 to obtain an absolute numerical difference of 1.522. The upper limit value of 0.50 configured in the limiting register is read. The digital comparator detects that 1.522 is greater than 0.50 and immediately pulls the processor's overflow interrupt pin high, triggering the highest-priority hardware interrupt routine. In the interrupt service routine, the floating-point divider is called to divide 1.522 by the baseline bearing threshold of 0.928, calculating the normalized initial deviation ratio of 1.64. The microcontroller accesses the historical access log of non-volatile memory via the external memory bus, and the internal counter of the calculation module accumulates the mutation records of the last 60 seconds, outputting a mutation frequency count of 15. The logarithm unit receives the count value of 15, performs a logarithmic mapping, and adds an addition operation with a base compensation constant of 0.5. The register outputs a logarithmic dynamic scaling factor of 3.2.The initial deviation ratio of 1.64 is fed into the hardware multiplier of the nonlinear amplification mapping module and multiplied with the dynamic scaling factor of 3.2 for enhanced calculation. The final result of the processed value of 5.25 is written into the relative offset result register.

[0114] The dynamic instruction delivery module is activated, and the dynamic memory mapping mechanism of the memory management unit is enabled to locate the circular history buffer in the static random access memory. The data extraction circuit pops the waveguide temperature sequence of the most recent 5 physical time nodes (40.1℃, 40.3℃, 40.6℃, 41.0℃, and 41.5℃) in timestamp order and sends it to the processor's pipeline register. The first-stage adder / subtractor of the pipeline performs a first-order difference operation on the sequence, generating an initial gradient variable sequence of 0.2, 0.3, 0.4, and 0.5. The second stage of the pipeline activates the hardware-implemented exponential moving average filter, reads the smoothing factor constant of 0.3 from the configuration register, and iteratively filters the initial gradient variables. The buffer outputs the smoothed gradient sequence of 0.23, 0.28, 0.35, and 0.42. The third stage of the pipeline performs a difference operation again on the smoothed sequence to extract acceleration features of 0.05, 0.07, and 0.07. The peak detection circuit locks the maximum value of 0.07 and stores it in the nonlinear fluctuation gradient register. The microprocessor accesses the global configuration flash memory chip via the internal integrated circuit bus. From this, it parses the static value of the channel fading compensation coefficient α (0.4) and the static value of the thermodynamic risk amplification coefficient β (0.6) for the current physical moment. The microprocessor reads the system's basic environmental constant register to obtain the reference gradient value of 0.02. The floating-point divider executes a division instruction, dividing the nonlinear fluctuation gradient 0.07 by 0.02 to generate a normalized nonlinear fluctuation gradient G of 3.5 for dimension alignment. The microcontroller then activates the parallel computing coprocessor in the high-performance floating-point computing engine, based on the encoded bivariate nonlinear function... Perform hardware-level calculations.

[0115] in, Represents the nonlinear coupling weights in the original state. The channel fading compensation coefficient represents the factor acting on the exponential function term. This represents an exponential function with the natural constant as its base. Represents the relative offset. This represents the thermodynamic risk amplification factor acting on the fluctuation gradient term. Represents the absolute value of the normalized nonlinear fluctuation gradient. Represents the square root operation function. Represents the dimensionless normalized nonlinear wave gradient;

[0116] The coprocessor reads the value 0 (5.25) from the relative offset register and sends it to the Taylor series expansion circuit to calculate the natural exponent. Simultaneously, it sends the value G (3.5) to a dedicated square root hardware module to calculate the square root. Two multipliers multiply the compensation coefficient 0.4 with the exponent result and the risk coefficient 0.6 with the square root result, respectively. The adder tree aggregates and sums the two products, locking the original nonlinear coupling weight W (77.35) in the result register. The coprocessor unit of the digital signal processor takes over this original weight and imports it into the built-in standard Kalman filter circuit. The initial covariance register of the filter circuit is set to 1.0, and the measurement noise variance register is configured to 0.5. Prediction and measurement update operations are performed through the matrix multiplication and inversion hardware module. After 10 iteration clock cycles of convergence, the output buffer of the filter circuit latches the final nonlinear coupling weight (76.0). The comparator logic gate compares this weight 76.0 with the baseline carrying threshold upper limit 46.4 stored in the threshold register, outputting a high level to trigger a channel switching interrupt. The interrupt controller activates the RF monitoring module of the network management chip, polling and scanning the signal quality indicator registers of the eight physical frequency bands. The monitoring module forcibly removes routing entries for the 2.4GHz and 2.5GHz bands with a signal-to-noise ratio below the 20dB threshold from the available channel resource forwarding table. The bandwidth detector extracts the actual remaining bandwidth register values ​​for the remaining six channels. The hardware sorting tree sorts the six values—150Mbps, 300Mbps, 800Mbps, 500Mbps, 200Mbps, and 400Mbps—in descending order. The network processing unit locks the physical port address of the 5.8GHz band corresponding to the 800Mbps bandwidth at the top of the sort. The media access control layer protocol stack assembles the target device's MAC address and scheduling control word into the frame header of the Ethernet data frame, pushes it into the transmit first-in-first-out queue of the 5.8GHz physical port via the data bus, and transmits the data stream scheduling command.

[0117] The above embodiments illustrate preferred embodiments of the present invention. Any equivalent adjustments to the technical solution based on software engineering methods are within the scope of protection, including but not limited to: implementing algorithm logic using different programming languages, refactoring functional modules into services, adjusting data interaction protocols, and optimizing resource scheduling strategies. Any implementation scheme derived from reasonable modifications to the data processing flow, service call chain, or system architecture layer without departing from the core technology of the present invention should be considered within the protection scope defined by the technical solution of the present invention.

Claims

1. A method for scheduling data streams during radiotherapy, characterized in that, Includes the following steps: S1: Call the communication interface to collect the waveguide temperature and position coordinates of the multi-leaf grating of the linear accelerator, obtain the reference transmission rate of the medical imaging detector through the velocity measurement protocol, and fuse the waveguide temperature, position coordinates and the reference transmission rate to construct the device load matrix; S2: Call the image acquisition module to obtain real-time scanning images of the radiation target area, perform pixel traversal to extract the voxel gray values ​​and scattering interference intensity of the real-time scanning images, and map the voxel gray values ​​and scattering interference intensity to generate a data stream feature vector. S3: Extract the diagonal elements of the device load matrix, calculate the mean of the values ​​to generate a baseline load threshold, calculate the feature norm of the data flow feature vector, compare the feature norm with the baseline load threshold, and calculate the relative offset of the data flow feature vector from the baseline load threshold. S4: Calculate the time series difference of the waveguide temperature to generate a nonlinear fluctuation gradient, fuse the relative offset with the nonlinear fluctuation gradient to calculate the nonlinear coupling weight, compare the nonlinear coupling weight with the baseline carrying threshold to screen the communication channel, and send a data stream scheduling command to the communication channel.

2. The radiotherapy process data stream scheduling method according to claim 1, characterized in that, The specific steps of S1 are as follows: S11: Call the communication interface to establish a bidirectional data link with the control bus, detect the physical operating status of the linear accelerator in real time according to the preset sampling frequency, extract and filter high-frequency noise signals, accurately collect the waveguide temperature and the three-dimensional spatial position coordinates of the multi-leaf grating of the linear accelerator under different radiation dose rates, and generate a set of state features. S12: Send independent data probe packets to the medical image detector through the speed measurement protocol, record the delay time and channel packet loss rate parameters of the probe data packets during the round-trip transmission process, accurately calculate the current bandwidth redundancy based on the underlying network topology and historical communication logs, and obtain the reference transmission rate by combining the round-trip delay time and packet loss rate parameters. S13: Obtain the set of state features containing multi-dimensional parameters generated by the pre-processor and the reference transmission rate; use a normalization algorithm to perform strict dimension unification processing on the waveguide temperature and position coordinates; and arrange and combine the dimension-unified values ​​and the reference transmission rate according to the physical mapping relationship of the communication nodes based on the pre-calibrated matrix dimension to establish the device load matrix.

3. The radiotherapy process data stream scheduling method according to claim 1, characterized in that, The specific steps of S2 are as follows: S21: The image acquisition module is invoked to activate the underlying photosensitive array element, and a full-range multi-angle transmission scan of the radiation target area is performed within a specific imaging exposure window period to filter out artifact fluctuation noise caused by the patient's natural breathing movements and acquire the real-time scan image. S22: Execute the pixel traversal algorithm to calculate the spatial gradient of each independent pixel in the real-time scan image, accurately identify the boundary contour features of complex anatomical structures, aggregate pixels with similar attributes through the region growth mechanism to extract the voxel gray value, and estimate the scattering trajectory of particles in the tissue based on the Monte Carlo model to extract the scattering interference intensity. S23: Obtain the voxel gray value and the scattering interference intensity, construct a deep neural mapping network containing a nonlinear activation function, use the voxel gray value as the main dimension input matrix and use the scattering interference intensity to adjust the penalty weights of the network parameters, extract the core spatial features through forward propagation operation of a multilayer perceptron, and generate the data stream feature vector.

4. The radiotherapy process data stream scheduling method according to claim 1, characterized in that, The specific steps in S3 are as follows: S31: Extract all element values ​​on the main diagonal of the device load matrix, remove outliers that exceed three times the standard deviation statistical range, perform an arithmetic mean summation on the remaining diagonal element values, and perform smooth scaling and nonlinear correction in combination with the pre-calibrated hardware loss compensation coefficient to generate the baseline load threshold. S32: Obtain the feature vector of the data stream in the high-dimensional space, call the microprocessor to calculate the sum of squares of each dimension of the vector in the Euclidean space and perform high-precision square root operation, assign independent adaptive weight parameters according to the importance of each dimension in the medical image reconstruction process, and multiply and fuse the sum of squares result with the adaptive weight parameters to obtain the feature norm. S33: Input the calculated feature norm into the comparator array, compare the feature norm with the baseline bearing threshold, calculate the absolute value difference of the feature norm relative to the baseline bearing threshold, divide the absolute difference by the baseline bearing threshold for dimensionless processing, and then introduce a dynamic scaling factor to perform nonlinear amplification mapping on the division result to obtain the relative offset.

5. The radiotherapy process data stream scheduling method according to claim 1, characterized in that, The specific steps of S4 are as follows: S41: Call the memory management unit to retrieve the waveguide temperature corresponding to the continuous time window node in the historical cache queue, perform first-order difference calculation on the continuous temperature values ​​of adjacent time nodes to obtain the initial gradient variable, use the exponential moving average algorithm to smooth the transient thermal jitter phenomenon in the initial gradient sequence, and further calculate the higher-order derivative to fit the temperature change acceleration characteristics caused by the thermal energy accumulation effect, and generate the nonlinear fluctuation gradient. S42: Integrate the relative offset and the nonlinear fluctuation gradient, and based on the network redundancy allocation strategy fixed in the underlying system, perform cross-multiplication and exponential decay integral operation on the two through a bivariate nonlinear function, and comprehensively consider the superposition and amplification effect of thermodynamic risks and data flow congestion in the spatiotemporal dimension to calculate the nonlinear coupling weight. S43: Compare the nonlinear coupling weight with the baseline carrying threshold. When the weight value is determined to significantly exceed the upper limit of the threshold, forcibly remove the high-frequency interference band with low signal-to-noise ratio from the available channel resource pool. Sort the remaining communication links in descending order according to the actual remaining bandwidth margin and select the optimal communication channel with the highest transmission guarantee capability. Encapsulate the network physical address of the target device and the underlying scheduling control word, and issue the data stream scheduling instruction to the selected optimal communication channel.

6. The radiotherapy process data stream scheduling method according to claim 2, characterized in that, The process of establishing the device load matrix specifically includes: The state feature set generated by the preprocessing stage is obtained, and the temperature and position coordinates of the waveguide after high-frequency noise reduction are extracted from it. At the same time, the monitoring module is accessed in parallel to read the reference transmission rate in a dynamically updated state, the physical mapping relationship of communication nodes associated with various data is analyzed, and the network port matrix identifier of each node in the topology network is identified. Based on the rules of multidimensional spatial data mapping and vector product operation, a sequence of state variables is constructed and the covariance matrix between the waveguide temperature sequence and the reference transmission rate sequence is accurately calculated. The intrinsic relationship between physical parameters and communication parameters is quantified. The position coordinates in three-dimensional space are converted into a three-dimensional rotation matrix through a quaternion transformation algorithm and strictly embedded in the feature subspace of the covariance matrix. The singular value decomposition or eigenvalue decomposition calculation module is invoked to extract the largest principal component vector with the highest variance contribution rate in the mixed feature subspace layer by layer. The extracted one-dimensional largest principal component vector is then deconstructed and reconstructed into a normalized square matrix form through the orthogonal basis transformation algorithm to establish the device load matrix.

7. The radiotherapy process data stream scheduling method according to claim 3, characterized in that, The process of generating the data stream feature vector specifically includes: The voxel gray value and the scattering interference intensity are obtained. The number of hidden layer nodes, activation function type and backpropagation learning rate parameters of the multilayer perceptron are initialized in the memory space. A deep neural mapping network with nonlinear approximation capability is constructed, and feature fusion calculation logic is established. Based on the defined feature fusion calculation logic, the forward propagation mechanism is triggered and the underlying matrix multiplication library is called. The activation feature values ​​output by the network are calculated according to the defined operator operation process. The specific mapping formula used is as follows: ,in, The activation feature value represents the network output. Represents the Sigmoid non-linear activation function. This represents the total number of voxels extracted. Represents the voxel sequence index number. Representing the The weight coefficients of each neuron, Representing the The voxel gray value corresponding to each pixel. This represents the preset penalty control factor. This represents the intensity of the extracted scattered radiation interference. Before the scattered interference intensity is substituted into the mapping formula for logarithmic calculation, it has been converted into a dimensionless relative value that meets the requirements of dimensional balance by introducing a reference interference constant and performing a phase ratio process. Collect continuous activation feature values ​​output from the terminal of the deep neural mapping network, and concatenate and superimpose the feature values ​​of each component in the high-dimensional space according to the spatial coordinate sequence of the original ray. Perform principal component analysis or independent component analysis algorithm to remove secondary dimensions with variance redundancy contribution rate lower than the set safety threshold, and generate the data stream feature vector.

8. The radiotherapy process data stream scheduling method according to claim 4, characterized in that, The process of obtaining the relative offset specifically includes: Extract the feature norm and the baseline carrying threshold, call the high-precision subtractor in the arithmetic logic unit, accurately calculate the absolute error and absolute difference of the feature norm of the current input flow relative to the baseline carrying threshold, and determine whether the obtained absolute difference is safely within the linear tolerance floating range preset by the system. When it is determined that the absolute difference significantly exceeds the system's preset linear tolerance range and triggers an overflow warning, the absolute difference is directly divided by the baseline carrying threshold to eliminate the influence of hardware scale and obtain a normalized initial deviation ratio. The frequency of short-term traffic mutations in the historical scheduling record repository is continuously monitored, and a logarithmic dynamic scaling factor with adaptive adjustment capability is constructed using the statistically obtained frequency of traffic mutations. The standardized initial deviation ratio is losslessly input into a predefined and parameter-calibrated nonlinear amplification mapping function. The high-frequency oscillation component of the initial deviation ratio is nonlinearly weighted and amplified using the previously constructed logarithmic dynamic scaling factor to obtain the relative offset.

9. The radiotherapy process data stream scheduling method according to claim 5, characterized in that, The process of calculating the nonlinear coupling weights specifically includes: Extract the relative offset and the nonlinear fluctuation gradient, access the underlying memory and query the global redundancy allocation strategy table pre-burned by the system, and combine the current network congestion flag bit to obtain the channel fading compensation coefficient and thermodynamic risk amplification coefficient applicable to the current physical moment. The high-performance weight calculation engine inside the embedded microprocessor is invoked to perform deep fusion and numerical solution of the obtained correlation parameters based on a specially designed bivariate nonlinear function. The specific solution formula is as follows: ,in, The nonlinear coupling weights are represented by the calculated values. Represents the channel fading compensation coefficient. The relative offset representing the input. Represents the natural index operation. Represents the thermodynamic risk amplification factor. Represents the nonlinear fluctuation gradient. This represents the absolute value of the nonlinear fluctuation gradient; Before the square root calculation, the nonlinear fluctuation gradient has been normalized by dividing it by the local standard environment reference gradient value, and converted into a dimensionless value that is used to adapt to the subsequent square root calculation and maintain the balance of the equation dimensions. The nonlinear coupling weights in the original state initially obtained by the computing engine are extracted, and the optimal state is estimated using a standard Kalman filter configured in the digital signal processor. Through iterative prediction and measurement update processes, high-frequency glitches and abnormal fluctuations caused by the sampling delay of the communication link are completely eliminated, and the nonlinear coupling weights are output.

10. A data stream scheduling system for radiotherapy processes, characterized in that, The system is used to implement the radiotherapy process data stream scheduling method according to any one of claims 1-9, and the system includes: The equipment load coordination module is used to call the communication interface to collect the waveguide temperature and the position coordinates of the multi-leaf grating of the linear accelerator, obtain the reference transmission rate of the medical imaging detector through the velocity measurement protocol, and fuse the waveguide temperature, position coordinates and the reference transmission rate to construct the equipment load matrix. The feature dimensionality reduction mapping module is used to call the image acquisition module to obtain real-time scanning images of the radiation target area, perform pixel traversal to extract the voxel gray values ​​and scattering interference intensity of the real-time scanning images, and map the voxel gray values ​​and scattering interference intensity to generate the data stream feature vector. The load threshold comparison module is used to extract the diagonal elements of the device load matrix, calculate the mean value to generate a baseline load threshold, calculate the feature norm of the data stream feature vector, compare the feature norm with the baseline load threshold, and calculate the relative offset of the data stream feature vector from the baseline load threshold. The instruction dynamic issuance module is used to calculate the time series difference of the waveguide temperature to generate a nonlinear fluctuation gradient, fuse the relative offset and the nonlinear fluctuation gradient to calculate the nonlinear coupling weight, compare the nonlinear coupling weight with the baseline carrying threshold to screen the communication channel, and issue the data stream scheduling instruction to the communication channel.

Images