A three-dimensional radiation field reconstruction method, device and equipment based on a residual network

Abstract

This invention relates to the field of radiation field reconstruction, and discloses a method, apparatus, and device for three-dimensional radiation field reconstruction based on residual networks. It acquires the deployment locations of a finite number of detectors in a three-dimensional radiation field, as well as the radiation dose data detected by these detectors. Based on the deployment locations of the finite detectors, the radiation dose data detected by the finite detectors, and a set low-resolution mesh size, low-resolution radiation distribution data is generated. This low-resolution radiation distribution data is input into a trained deep super-resolution reconstruction model based on a three-dimensional sub-voxel residual network. The deep super-resolution reconstruction model then performs shallow feature extraction, deep residual processing, and sub-voxel upsampling on the low-resolution radiation distribution data to generate high-resolution radiation distribution data. This invention can effectively reduce the computational cost and improve the reconstruction accuracy of radiation distribution data in three-dimensional radiation fields.

Description

Technical Field

[0001] This invention relates to the field of radiation field reconstruction, and in particular to a method, apparatus and equipment for three-dimensional radiation field reconstruction based on residual networks. Background Technology

[0002] High-precision reconstruction of the three-dimensional radiation field is a key aspect of the full life cycle management of nuclear facilities, directly affecting the implementation effectiveness of radiation protection optimization strategies, intelligent planning of work paths, and accurate assessment of shielding effectiveness.

[0003] Radiation fields are physical fields with extremely strong spatial inhomogeneity and nonlinearity. Their distribution is complexly coupled and modulated by radiation source parameters and environmental shielding structures, exhibiting highly variability. In engineering practice, acquiring measured data often faces significant challenges due to limitations in sensor deployment costs, complex on-site physical constraints, and the cumulative damage effects of high-radiation environments on electronic devices. The number of actual detection nodes deployed is extremely limited, and the usable sampling data accounts for a very low percentage (usually less than 0.01%) in three-dimensional space. This extreme sparsity and non-uniform distribution of data makes inverting the continuous dose field of the entire three-dimensional space from limited discrete observations a typical ill-conditioned inverse problem.

[0004] Related techniques can use deterministic interpolation to reconstruct radiation distribution data of three-dimensional radiation fields. Deterministic interpolation relies heavily on the assumption of spatial continuity of the data. In complex nuclear environments, the presence of strong shielding structures such as lead plates and concrete walls can cause abrupt, abrupt changes in radiation dose rate in space. Deterministic interpolation lacks the ability to perceive such physical shielding mechanisms and is prone to generating non-physical, smooth transition bands at the boundaries of shielding structures, thus introducing significant reconstruction errors and resulting in low reconstruction accuracy. Summary of the Invention

[0005] This invention provides a method, apparatus, and device for reconstructing three-dimensional radiation fields based on residual networks, in order to address the shortcomings of low accuracy in the reconstruction of three-dimensional radiation fields in related technologies and improve the accuracy of three-dimensional radiation field reconstruction.

[0006] In a first aspect, the present invention provides a method for reconstructing a three-dimensional radiation field based on a residual network, comprising:

[0007] The deployment locations of a finite number of detectors in a three-dimensional radiation field, as well as the radiation dose data detected by the finite number of detectors, are obtained. Based on the deployment location of the finite detectors, the radiation dose data detected by the finite detectors, and the set low-resolution grid division size, low-resolution radiation distribution data is generated. The low-resolution radiative distribution data is input into a trained deep super-resolution reconstruction model based on a three-dimensional sub-voxel residual network, so that the deep super-resolution reconstruction model performs shallow feature extraction, deep residual processing and sub-voxel upsampling on the low-resolution radiative distribution data to generate high-resolution radiative distribution data.

[0008] Optionally, the finite number of detectors includes a finite number of detectors deployed at different locations in the three-dimensional radiation field, and the radiation dose data detected by the finite number of detectors includes the radiation dose rate detected by each detector. Based on the deployment location of the finite detectors, the radiation dose data detected by the finite detectors, and the set low-resolution grid division size, low-resolution radiation distribution data is generated, including: Based on the deployment location of each detector, the radiation dose rate detected by each detector, and the low-resolution grid division size, sparse measurement point interpolation or Monte Carlo particle transport simulation is performed on the three-dimensional radiation field to obtain the initial radiation distribution data of the three-dimensional radiation field. The initial radiation distribution data is logarithmically transformed to compress it to a set range of values, resulting in compressed radiation distribution data. The compressed radiation distribution data is normalized to obtain normalized distribution data, which is then used as the low-resolution radiation distribution data.

[0009] Optionally, the depth super-resolution reconstruction model includes a feature extraction module, a residual module, and an upsampling module; The step of inputting the low-resolution radiative distribution data into a trained deep super-resolution reconstruction model based on a three-dimensional sub-voxel residual network, so that the deep super-resolution reconstruction model performs shallow feature extraction, deep residual processing, and sub-voxel upsampling on the low-resolution radiative distribution data to generate high-resolution radiative distribution data, includes: The low-resolution radiation distribution data is input into the feature extraction module for shallow feature extraction to obtain feature extraction data. The feature extraction data is input into the residual module for deep residual processing to obtain the processed data. The processed data is input into the upsampling module for subvoxel upsampling to obtain sampled data, which is then used as the high-resolution radiation distribution data.

[0010] Optionally, the feature extraction module includes a three-dimensional convolutional layer and an activation function layer; The step of inputting the low-resolution radiation distribution data into the feature extraction module for shallow feature extraction to obtain feature extraction data includes: The low-resolution radiative distribution data is input into a three-dimensional convolutional layer and an activation function layer for forward propagation to map the low-resolution radiative distribution data to a high-dimensional feature space and extract preliminary local features of the low-resolution radiative distribution data to obtain the feature extraction data.

[0011] Optionally, the residual module includes N residual blocks, N skip connection layers and a global fusion layer, where N is an integer greater than 1; The step of inputting the feature extraction data into the residual module for deep residual processing to obtain processed data includes: The feature extraction data is input into the first residual block for deep nonlinear feature extraction to obtain the first extracted data; the feature extraction data and the first extracted data are input into the first skip connection layer for element-wise addition to obtain the first added data; The first summed data is input into the second residual block for deep nonlinear feature extraction to obtain the second extracted data; the first summed data and the second extracted data are input into the second skip connection layer for element-wise addition to obtain the second summed data, until the Nth summed data output by the last skip connection layer is obtained; The Nth summed data and the feature extracted data are input into the global fusion layer for element-wise addition to obtain the processed data.

[0012] Optionally, each residual block includes two three-dimensional convolutional layers, two batch normalization layers, and an activation function; The data to be processed is input into the residual block for deep nonlinear feature extraction to obtain extracted data, including: The data to be processed is input into the first three-dimensional convolutional layer, the first batch normalization layer, the activation function, the second three-dimensional convolutional layer, and the second batch normalization layer of the residual block for forward propagation to extract deep nonlinear features from the data to be processed, thereby obtaining the extracted data.

[0013] Optionally, the upsampling module includes a dimension-expanded convolutional layer and a sub-voxel rearrangement layer; The step of inputting the processed data into the upsampling module for subvoxel upsampling to obtain sampled data includes: The processed data is input into the channel extreme expansion layer to amplify the number of channels of the processed data by a set factor, thereby obtaining the data after dimension expansion and convolution. The data after the dimension-expanded convolution is input into the sub-voxel rearrangement layer for data rearrangement to obtain the sampled data.

[0014] Optionally, before inputting the low-resolution radiative distribution data into the trained deep super-resolution reconstruction model based on a three-dimensional sub-voxel residual network, the method further includes: Multiple training datasets are constructed based on the three-dimensional radiation field of the sample, and each training dataset includes low-resolution radiation distribution data of the sample and corresponding high-resolution radiation distribution data of the sample. The low-resolution radiative distribution data of the samples in the training data is input into the reconstruction model to be trained for high-resolution reconstruction, thereby obtaining the high-resolution prediction data generated and output by the reconstruction model to be trained. Based on the difference between the high-resolution prediction data and the high-resolution radiative distribution data of the samples, the reconstruction model to be trained is updated to obtain the updated reconstruction model. The updated reconstruction model is then trained using other training data until a model that meets the set requirements is obtained, which is then used as the depth super-resolution reconstruction model.

[0015] Secondly, the present invention provides a three-dimensional radiation field reconstruction device based on a residual network, comprising: The acquisition unit is used to acquire the deployment location of the finite detectors in the three-dimensional radiation field, as well as the radiation dose data detected by the finite detectors; The first generation unit is used to generate low-resolution radiation distribution data based on the deployment location of the finite detectors, the radiation dose data detected by the finite detectors, and the set low-resolution grid division size. The second generation unit inputs the low-resolution radiation distribution data into a trained deep super-resolution reconstruction model based on a three-dimensional sub-voxel residual network, so that the deep super-resolution reconstruction model performs shallow feature extraction, deep residual processing, and sub-voxel upsampling on the low-resolution radiation distribution data to generate high-resolution radiation distribution data.

[0016] Thirdly, the present invention provides a computer device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the three-dimensional radiation field reconstruction method based on residual networks described in the first aspect or any corresponding embodiment thereof.

[0017] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the three-dimensional radiation field reconstruction method based on residual networks according to the first aspect or any corresponding embodiment described above.

[0018] This invention provides a method, apparatus, and device for reconstructing a three-dimensional radiation field based on residual networks. It acquires the deployment locations of finite detectors in a three-dimensional radiation field and the radiation dose data detected by these detectors. Based on the deployment locations of the finite detectors, the radiation dose data detected by the finite detectors, and a set low-resolution mesh size, low-resolution radiation distribution data is generated. This low-resolution radiation distribution data is then input into a trained deep super-resolution reconstruction model based on a three-dimensional sub-voxel residual network. The deep super-resolution reconstruction model performs shallow feature extraction, deep residual processing, and sub-voxel upsampling on the low-resolution radiation distribution data to generate high-resolution radiation distribution data. This invention can effectively reduce the computational cost and improve the reconstruction accuracy of radiation distribution data in a three-dimensional radiation field.

[0019] This invention improves radiation field super-resolution reconstruction by introducing a deep residual network. Compared with interpolation methods and neural network-based reconstruction methods in related technologies, it has the following significant advantages: 1. This invention significantly improves the geometric reconstruction accuracy and peak value estimation accuracy of radiation source terms: Shallow networks in related technologies, due to insufficient feature fitting capabilities, tend to smooth local abrupt changes, resulting in blurred point-like source terms and severely underestimating the central peak intensity. The model in this invention, through a residual learning strategy, focuses on capturing the local detail increments between low-resolution and high-resolution radiation fields. This invention can accurately reconstruct the geometry of source terms, preserving sharp peak features at the center of the source term. Compared to related technologies, the source term location deviation is significantly reduced, and the peak intensity estimation accuracy is greatly improved, thus providing a reliable data foundation for subsequent source term analysis.

[0020] 2. This invention achieves clear reconstruction of physical boundaries under complex shielding conditions: In nuclear environments, the dose rate behind a shield (such as a lead plate) often drops abruptly. Related techniques struggle to fit this steep gradient, often resulting in blurred edges. The model trained in this invention, thanks to its deep network architecture, possesses stronger nonlinear expression capabilities, enabling precise reconstruction of the step gradient at the shield edge in high-resolution space. This high-fidelity boundary recovery capability makes the reconstruction results physically more realistic, effectively eliminating artifacts and improving the accuracy of shield protection effectiveness assessment.

[0021] 3. This invention addresses the challenge of training neural networks in related technologies, achieving a balance between performance and efficiency: In 3D voxel computation, simply increasing network depth leads to vanishing training gradients and difficulty in convergence. The skip connections introduced in this invention effectively solve the degradation problem of deep networks, enabling the model to maintain fast and stable convergence while significantly increasing the number of network layers to obtain better reconstruction performance. Furthermore, combined with sub-voxel upsampling technology, most convolutional operations of the network are still performed in the low-resolution space, ensuring that while significantly improving reconstruction accuracy, inference time and memory usage remain at a low level. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in this invention or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a three-dimensional radiation field reconstruction method based on residual networks, provided as an embodiment of the present invention; Figure 2 A flowchart of another three-dimensional radiation field reconstruction method based on residual networks provided in an embodiment of the present invention; Figure 3 A schematic diagram of a three-dimensional radiation field reconstruction device based on a residual network provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0025] With the development of artificial intelligence technology, radiation field reconstruction techniques based on convolutional neural networks have been increasingly applied to radiation field reconstruction. However, these techniques generally employ shallow end-to-end network architectures (such as super-resolution convolutional neural networks and their variants), which are inadequate for processing radiation field data with high dynamic range. On the one hand, shallow networks struggle to extract deep nonlinear spatial features and cannot accurately recover the geometric position of the radiation source when faced with extremely sparse inputs. On the other hand, due to the limited feature fitting ability of shallow networks, driven by the training objective of minimizing the overall mean square error, the network tends to smooth local abrupt signals. This smoothing phenomenon, when processing high-radiation point sources or hotspot regions, leads to a decrease in the peak intensity at the source term center and an expansion and blurring of the high-value region. This lack of local extremum features and edge gradient information significantly reduces the fidelity of the reconstructed field.

[0026] Specifically, the relevant technologies have the following limitations when dealing with highly dynamic and drastically changing radiation characteristics: The neural networks in related technologies suffer from significant distortions in the geometric reconstruction of radiation source terms and biases in peak intensity estimation. Radiation source terms (source terms, high-intensity hotspots) exhibit high-intensity local extremum features with specific geometric structures in three-dimensional space. The shallow network architectures of these technologies, due to insufficient feature extraction depth, struggle to accurately fit the steep edge gradients of the source term regions, leading to inaccurate geometric reconstruction of the source terms in the results. This is accompanied by a severe underestimation of peak intensity (i.e., the highest point of the reconstructed value is significantly lower than the true physical value). This dual distortion of form and intensity reduces the reliability of hotspot region perception.

[0027] Neural networks in related technologies face training degradation and convergence difficulties in radiation field feature extraction: Theoretically, to accurately fit the complex nonlinear spatial mapping relationships in the radiation field, it is necessary to construct neural networks with greater depth. However, convolutional networks in related technologies are prone to gradient vanishing or gradient exploding as the number of layers increases, causing network performance to decrease with increasing depth. Given the high computational cost of 3D voxel computation, constructing a deep network architecture that can deeply extract abstract physical features while maintaining efficient and stable convergence is a key bottleneck in improving reconstruction accuracy.

[0028] Related technologies suffer from blurred physical boundaries and insufficient gradient fidelity in radiation fields under complex shielding conditions: In nuclear environments, the radiation dose rate behind strong shielding structures such as lead plates and concrete walls often exhibits abrupt, abrupt changes. The limited nonlinear representation capabilities of related technologies make it difficult to reconstruct such steep gradient changes in high-resolution space, often resulting in non-physically smooth transition bands at the shielding edges. This boundary blurring not only reduces the clarity of the reconstructed field but also affects the accurate assessment of the actual protective effectiveness of the shielding structure.

[0029] The following is combined with Figures 1-2 The present invention describes a three-dimensional radiation field reconstruction method based on residual networks.

[0030] like Figure 1 As shown, this embodiment proposes a first three-dimensional radiation field reconstruction method based on residual networks, which may include the following steps: S101. Obtain the deployment location of the finite detectors in the three-dimensional radiation field, as well as the radiation dose data detected by the finite detectors.

[0031] In this context, a finite number of detectors refers to a limited number of detectors deployed in a three-dimensional radiation field. Each detector will detect the radiation dose rate at its corresponding deployment location, and the radiation dose rate detected by each detector constitutes the radiation dose data detected by the finite number of detectors.

[0032] Specifically, the deployment location of each detector in the three-dimensional radiation field can be different.

[0033] S102. Based on the deployment location of the limited detectors, the radiation dose data detected by the limited detectors, and the set low-resolution grid division size, generate low-resolution radiation distribution data.

[0034] The low-resolution mesh size is the three-dimensional size used to coarsely mesh the three-dimensional geometric model corresponding to the three-dimensional radiation field, such as 40×10×14.

[0035] Specifically, in this embodiment, after obtaining the deployment location of each detector in a limited number of detectors and the radiation dose rate detected by each detector, the radiation distribution data of the three-dimensional radiation field can be reconstructed at low resolution based on the deployment location of each detector, the radiation dose rate detected by each detector, and the low-resolution grid division size, to generate a low-resolution three-dimensional radiation field tensor, i.e., low-resolution radiation distribution data.

[0036] Optionally, in other three-dimensional radiation field reconstruction methods based on residual networks proposed in this embodiment, the finite detectors include a finite number of detectors deployed at different locations in the three-dimensional radiation field, and the radiation dose data detected by the finite detectors includes the radiation dose rate detected by each detector. In this case, step S102 includes: Based on the deployment location of the finite detectors, the radiation dose data detected by the finite detectors, and the set low-resolution grid size, low-resolution radiation distribution data is generated, including: Based on the deployment location of each detector, the radiation dose rate detected by each detector, and the low-resolution grid size, sparse measurement point interpolation or Monte Carlo particle transport simulation is performed on the three-dimensional radiation field to obtain the initial radiation distribution data of the three-dimensional radiation field. The initial radiation distribution data is logarithmically transformed to compress it to a set range of values, resulting in compressed radiation distribution data. The compressed radiation distribution data is normalized to obtain normalized distribution data, which is then used as low-resolution radiation distribution data.

[0037] Specifically, in this embodiment, sparse measurement point interpolation or Monte Carlo particle transport simulation can be performed on the three-dimensional radiation field based on the deployment location of each detector, the radiation dose rate detected by each detector, and the low-resolution grid division size, so as to reconstruct the initial radiation distribution data of the three-dimensional radiation field and obtain the initial radiation distribution data.

[0038] It should be noted that the initial radiation distribution data is a three-dimensional radiation field tensor. In this embodiment, the radiation dose rate recorded at each location in the initial radiation distribution data can be logarithmically transformed to compress it to a reasonable numerical range. Then, normalization processing is performed to make all data distributed between 0 and 1, effectively reducing the amount of subsequent network computation.

[0039] S103. Input the low-resolution radiation distribution data into the trained deep super-resolution reconstruction model based on the three-dimensional sub-voxel residual network, so that the deep super-resolution reconstruction model can perform shallow feature extraction, deep residual processing and sub-voxel upsampling on the low-resolution radiation distribution data to generate high-resolution radiation distribution data.

[0040] Specifically, this embodiment is based on the radiation field reconstruction of a three-dimensional sub-voxel residual network. By constructing a deep residual learning network, a nonlinear mapping relationship between a low-resolution coarse radiation field and a high-resolution fine radiation field can be established. The residual structure is used to focus on recovering the local extremum features (source term peaks) and edge gradient information that are smoothed away by the low-resolution grid, and accurate source term inversion is achieved based on the reconstructed high-fidelity field.

[0041] like Figure 2 As shown, the deep super-resolution reconstruction model includes a feature extraction module, a residual module, and an upsampling module. Step S103 includes: Low-resolution radiation distribution data is input into the feature extraction module for shallow feature extraction to obtain feature-extracted data. The feature extraction data is input into the residual module for deep residual processing to obtain the processed data. The processed data is input into the upsampling module for subvoxel upsampling to obtain sampled data, which is then used as high-resolution radiation distribution data.

[0042] Specifically, Figure 2 The shallow feature extraction module in the example is the feature extraction module of this embodiment. Figure 2The deep residual learning module in this embodiment is the residual module, and the sub-voxel upsampling module is the upsampling module. This embodiment can input low-resolution radiation distribution data into the feature extraction module, residual module, and upsampling module of the deep super-resolution reconstruction model to generate a high-resolution three-dimensional radiation field tensor, i.e., high-resolution radiation distribution data.

[0043] Optionally, the feature extraction module includes a 3D convolutional layer and an activation function layer. The low-resolution radiative distribution data described above is input into the feature extraction module for shallow feature extraction, resulting in feature-extracted data, including: Low-resolution radiative distribution data is input into a three-dimensional convolutional layer and activation function layer for forward propagation to map the low-resolution radiative distribution data to a high-dimensional feature space and extract preliminary local features of the low-resolution radiative distribution data, thus obtaining feature extraction data.

[0044] like Figure 2 As shown, the shallow feature extraction module includes a 3D convolutional layer Conv3D and an activation function layer ReLU.

[0045] Optionally, the residual module includes N residual blocks, N skip connection layers, and one global fusion layer, where N is an integer greater than 1. The feature extraction data is input into the residual module for deep residual processing to obtain the processed data, including: The feature extraction data is input into the first residual block for deep nonlinear feature extraction to obtain the first extracted data; the feature extraction data and the first extracted data are input into the first skip connection layer for element-wise addition to obtain the first added data; The first summed data is input into the second residual block for deep nonlinear feature extraction to obtain the second extracted data; the first summed data and the second extracted data are input into the second skip connection layer for element-wise addition to obtain the second summed data, until the Nth summed data is obtained from the output of the last skip connection layer; The data after the Nth summation is input into the global fusion layer along with the feature extraction data for element-wise summation to obtain the processed data.

[0046] Optionally, each residual block includes two 3D convolutional layers, two batch normalization layers, and one activation function. In this case, the data to be processed is input into the residual block for deep nonlinear feature extraction, resulting in extracted data, including: The data to be processed is input into the first 3D convolutional layer, the first batch normalization layer, the activation function, the second 3D convolutional layer, and the second batch normalization layer of the residual block for forward propagation to extract deep nonlinear features from the data to be processed, thus obtaining the extracted data.

[0047] like Figure 2As shown, each residual block includes two 3D convolutional layers (Conv3D), two batch normalization layers (BN), and an activation function layer (RELU).

[0048] Optionally, the upsampling module includes dimension-expanded convolutional layers and sub-voxel rearrangement layers. The processed data is input into the upsampling module for sub-voxel upsampling to obtain sampled data, including: The processed data is input into the dimension-expanding convolutional layer for extreme channel expansion, which increases the number of channels of the processed data by a set factor to obtain the dimension-expanded convolutional data. The data after dimension expansion and convolution is input into the sub-voxel rearrangement layer for data rearrangement to obtain the sampled data.

[0049] like Figure 2 As shown, the subvoxel upsampling module includes a dimension-expanded convolutional layer Conv3D and a subvoxel rearrangement layer VoxelShuffle.

[0050] The proposed three-dimensional radiation field reconstruction method based on residual networks in this embodiment obtains the deployment locations of a finite number of detectors in the three-dimensional radiation field, as well as the radiation dose data detected by these detectors. Based on the deployment locations of the finite detectors, the radiation dose data detected by the detectors, and a set low-resolution mesh size, low-resolution radiation distribution data is generated. This low-resolution radiation distribution data is then input into a trained deep super-resolution reconstruction model based on a three-dimensional sub-voxel residual network. The deep super-resolution reconstruction model performs shallow feature extraction, deep residual processing, and sub-voxel upsampling on the low-resolution radiation distribution data to generate high-resolution radiation distribution data. This embodiment can effectively reduce the computational cost and improve the reconstruction accuracy of radiation distribution data in a three-dimensional radiation field.

[0051] based on Figure 1 This embodiment proposes a second method for reconstructing a three-dimensional radiation field based on residual networks. Before step S103, this method may further include: Multiple training datasets are constructed based on the three-dimensional radiation field of the samples. Each training dataset includes low-resolution radiation distribution data of the samples and corresponding high-resolution radiation distribution data of the samples. The low-resolution radiative distribution data of the samples in the training data is input into the reconstruction model to be trained for high-resolution reconstruction, and the high-resolution prediction data generated and output by the reconstruction model to be trained is obtained. Based on the difference between the high-resolution prediction data and the high-resolution radiative distribution data of the samples, the reconstruction model to be trained is updated to obtain the updated reconstruction model. The updated reconstruction model is trained again using other training data until a model that meets the set requirements is obtained, which is then used as the depth super-resolution reconstruction model.

[0052] It should be noted that this embodiment can establish a nonlinear mapping relationship between a low-resolution coarse radiation field and a high-resolution fine radiation field by constructing a deep residual learning network. It can focus on recovering the local extreme value features (source term peaks) and edge gradient information that are smoothed away by the low-resolution grid, and achieve accurate source term inversion based on the reconstructed high-fidelity field.

[0053] Specifically, this embodiment can construct a multi-resolution radiation field dataset based on Monte Carlo simulation. Specifically, this embodiment can utilize Monte Carlo particle transport simulation software to construct multi-resolution ground truth data pairs under the same physical scenario. In a single particle transport simulation calculation, two sets of grid counters with different precisions are simultaneously set. The first set is a coarse grid, used to statistically analyze low-resolution radiation field data as input to the network; the second set is a fine grid (resolution multiplied in each dimension), used to statistically analyze high-resolution radiation field data as training labels (ground truth) for the network. This approach ensures that the low-resolution and high-resolution data correspond to the exact same physical process. Subsequently, the extracted dose field data undergoes logarithmic transformation and normalization to construct paired three-dimensional tensor datasets to accommodate the large dynamic range characteristics of the radiation field.

[0054] Subsequently, this embodiment can construct a deep super-resolution reconstruction model based on a three-dimensional sub-voxel residual network. Its core lies in directly learning the difference between the low-resolution and high-resolution radiation fields using a residual learning strategy. This network sequentially includes a feature extraction module, a residual module, and an upsampling module.

[0055] The feature extraction module can map low-resolution inputs to a high-dimensional feature space through three-dimensional convolutional layers.

[0056] Specifically, the residual module can be composed of multiple cascaded stacked 3D residual blocks. Each residual block contains two consecutive 3D convolutional layers connected by batch normalization and activation functions, with skip connections introduced between the block's input and output. This structure allows the network to focus on learning residuals (i.e., the spike increments at the center of the source term and the sharpness compensation at the edges of the shield), thus addressing the degradation problem of deep networks. Furthermore, global residual fusion is introduced at the end of the residual module, adding deep features to shallow features. This design constructs a direct transmission channel from input to output, ensuring the lossless transfer of the basic distribution information of the radiation field, forcing the network to focus on learning local detailed residual features, thereby significantly reducing the difficulty of model optimization and accelerating training convergence.

[0057] The upsampling module can use extended convolutional layers to expand the number of feature map channels to a cube multiple of the target magnification factor. Then, through periodic tensor rearrangement operations (sub-voxel rearrangement), the channel dimension data is reorganized into the spatial dimension to directly generate a high-resolution fine radiation field.

[0058] Specifically, this embodiment employs supervised learning, using mean squared error as the loss function and an optimizer to train the model until the network converges. In practical applications, the low-resolution radiation field to be reconstructed is input into the trained model, directly outputting a high-fidelity, high-resolution radiation field. This reconstruction result clearly displays the fine three-dimensional structure of the radiation field, effectively recovering local extremum features and edge gradient information smoothed away by the low-resolution mesh.

[0059] To better illustrate the above technical solutions, the following examples are provided in this embodiment.

[0060] This embodiment proposes a radiation field reconstruction method based on a three-dimensional sub-voxel residual network. The overall implementation process mainly includes three steps: multi-resolution dataset construction, three-dimensional sub-voxel residual network model construction, and model training and application, specifically: Step 1: Construct a multi-resolution radiation field dataset based on Monte Carlo simulation, and use Monte Carlo particle transport simulation software to generate the data pairs required for training.

[0061] 1. Scene modeling: Construct a three-dimensional geometric model that includes radiation sources (such as point sources and volume sources), shielding structures (such as concrete walls and lead plates), and air regions.

[0062] 2. Synchronous grid counting: In the same particle transport simulation calculation, two sets of grid counters with the same spatial range but different resolutions are set up.

[0063] Low-resolution mesh: used as input data for the model. For example, the space is divided into coarser mesh cells (e.g., mesh side length of 25 cm), and the average dose rate at this scale is calculated.

[0064] High-resolution mesh: used as training labels (ground values) for the model. For example, dividing the space into fine mesh units (e.g., mesh side length of 5 cm), which is a 5x magnification relative to the coarse mesh in each of the length, width, and height dimensions.

[0065] 3. Data Preprocessing: Due to the extremely large dynamic range of radiation field data, direct training can easily lead to network non-convergence. Therefore, logarithmic transformation is performed on the extracted low-resolution and high-resolution dose data to compress them into a reasonable numerical range, followed by normalization to ensure that all data are distributed between 0 and 1.

[0066] Step 2: Construct a deep super-resolution reconstruction model based on a 3D sub-voxel residual network. The network structure mainly consists of three parts: shallow feature extraction, deep residual learning, and sub-voxel upsampling.

[0067] 1. Shallow Feature Extraction: The network receives low-resolution 3D radiation field data as input (single channel). First, a 3D convolutional layer maps the input data to a high-dimensional feature space (e.g., expanded to 64 feature channels) to extract preliminary local features of the radiation field.

[0068] 2. Deep residual learning (core part): This is the main body of the network, consisting of 16 cascaded "residual blocks".

[0069] Residual block structure: Each residual block contains two consecutive 3D convolutional layers (Conv3D), with batch normalization (BN) layers and activation functions (ReLU) placed between the convolutional layers.

[0070] Skip connections: A "bypass" is introduced into each residual block, where the input of that residual block is directly added element-wise to the output of the convolutional branch. Through this addition operation, the network only needs to learn the "difference" between the input and output. In the radiation field reconstruction task, this "difference" corresponds to the microstructural details that low-resolution meshes cannot describe, such as the amount of compensation for the height of the spikes pulled down at the center of the source term, and the amount of gradient correction for the blurring of the shield edges.

[0071] Global fusion: At the end of the stack of all residual blocks, a convolutional layer extracts deep features and adds them element-wise with the shallow features again to achieve global residual learning. This design constructs a direct transmission channel from input to output, ensuring the lossless transfer of the basic distribution information of the radiation field, forcing the network to focus on learning local detailed residual features, thereby significantly reducing the difficulty of model optimization and accelerating training convergence.

[0072] 3D Sub-Vocet Upsampling: Sub-voxel convolution is used at the network's end to improve spatial resolution. First, a dimension-expanding convolutional layer significantly increases the number of channels in the feature map. The increase is equal to the cube of the target magnification factor. For example, to magnify the spatial resolution by 5 times, the number of channels needs to be expanded to 125 times the original. Sub-voxel rearrangement is then performed. This operation does not involve numerical calculations but rearranges the channel-dimensional data according to a specific rule into the spatial dimensions (depth, height, and width). Through this rearrangement, the low-resolution high-dimensional feature map is directly converted into a high-resolution single-channel radiation field output.

[0073] Step 3: Model training and reconstruction.

[0074] 1. Network initialization: The network weights are randomly initialized using a normal distribution.

[0075] 2. Loss Function Calculation: Low-resolution data is input into the network to obtain high-resolution predicted values. The difference between the predicted values ​​and the high-resolution ground truth values ​​(e.g., mean squared error) is calculated. To enhance the recovery of source terms, a weighted strategy can be adopted, that is, assigning greater loss weights to high-dose regions, forcing the network to prioritize optimizing peak regions.

[0076] 3. Parameter Update: Update the network parameters using the optimizer based on the loss gradient. Repeat training until the loss function converges, and save the optimal model parameters.

[0077] 4. Practical application: Input the low-resolution radiation field obtained on site (which can be obtained through sparse measurement point interpolation or coarse mesh simulation) into the trained model, and output a high-fidelity high-resolution radiation field.

[0078] This embodiment improves the super-resolution reconstruction of the radiation field by introducing a deep residual network. Compared with interpolation methods and neural network-based reconstruction methods in related technologies, it has the following significant advantages: 1. Significantly improves the geometric reconstruction accuracy and peak value estimation accuracy of radiation source terms: Shallow networks in related techniques, due to insufficient feature fitting ability, tend to smooth local abrupt changes, resulting in blurred point-like source terms and severely underestimating the central peak intensity. The model in this embodiment, through a residual learning strategy, focuses on capturing the local detail increments between low-resolution and high-resolution radiation fields. This embodiment can accurately reconstruct the geometry of source terms, preserving sharp peak features at the center of the source term. Compared to related techniques, the source term location deviation is significantly reduced, and the peak intensity estimation accuracy is greatly improved, thus providing a reliable data foundation for subsequent source term analysis.

[0079] 2. Achieved clear reconstruction of physical boundaries under complex shielding conditions: In nuclear environments, the dose rate behind shielding bodies (such as lead plates) often drops abruptly. Related techniques struggle to fit such steep gradients, frequently resulting in blurred edges. The model trained in this embodiment, thanks to its deep network architecture, possesses stronger nonlinear expression capabilities, enabling accurate reconstruction of the step gradient at the shielding body edge in high-resolution space. This high-fidelity boundary recovery capability makes the reconstruction results physically more realistic, effectively eliminating artifacts and improving the accuracy of shielding effectiveness assessment.

[0080] 3. This invention addresses the challenge of training neural networks in related technologies, achieving a balance between performance and efficiency: In 3D voxel computation, simply increasing network depth leads to vanishing training gradients and difficulty in convergence. The skip connections introduced in this embodiment effectively solve the degradation problem of deep networks, enabling the model to maintain fast and stable convergence while significantly increasing the number of network layers to obtain better reconstruction performance. Furthermore, combined with sub-voxel upsampling technology, most of the network's convolutional operations remain in the low-resolution space, ensuring that while significantly improving reconstruction accuracy, inference time and memory usage remain at a low level.

[0081] like Figure 3 As shown, this embodiment proposes a three-dimensional radiation field reconstruction device based on residual networks, comprising: The acquisition unit 301 is used to acquire the deployment location of the finite detectors in the three-dimensional radiation field, as well as the radiation dose data detected by the finite detectors; The first generation unit 302 is used to generate low-resolution radiation distribution data based on the deployment location of the limited detectors, the radiation dose data detected by the limited detectors, and the set low-resolution grid division size. The second generation unit 303 inputs low-resolution radiation distribution data into a trained deep super-resolution reconstruction model based on a three-dimensional sub-voxel residual network, so that the deep super-resolution reconstruction model can perform shallow feature extraction, deep residual processing and sub-voxel upsampling on the low-resolution radiation distribution data to generate high-resolution radiation distribution data.

[0082] It should be noted that the processing procedures of the acquisition unit 301, the first generation unit 302, and the second generation unit 303, and their beneficial effects, can be referred to respectively. Figure 1 Steps S101 to S103 are not described in detail here.

[0083] Optionally, the finite number of detectors may be deployed at different locations in the three-dimensional radiation field, and the radiation dose data detected by the finite number of detectors may include the radiation dose rate detected by each detector.

[0084] The first generating unit 302 is also used for: Based on the deployment location of each detector, the radiation dose rate detected by each detector, and the low-resolution grid size, sparse measurement point interpolation or Monte Carlo particle transport simulation is performed on the three-dimensional radiation field to obtain the initial radiation distribution data of the three-dimensional radiation field. The initial radiation distribution data is logarithmically transformed to compress it to a set range of values, resulting in compressed radiation distribution data. The compressed radiation distribution data is normalized to obtain normalized distribution data, which is then used as low-resolution radiation distribution data.

[0085] Optionally, the deep super-resolution reconstruction model includes a feature extraction module, a residual module, and an upsampling module; the second generation unit 303 is also used for: Low-resolution radiation distribution data is input into the feature extraction module for shallow feature extraction to obtain feature-extracted data. The feature extraction data is input into the residual module for deep residual processing to obtain the processed data. The processed data is input into the upsampling module for subvoxel upsampling to obtain sampled data, which is then used as high-resolution radiation distribution data.

[0086] Optionally, the feature extraction module includes a 3D convolutional layer and an activation function layer; The second generating unit 303 is also used for: Low-resolution radiative distribution data is input into a three-dimensional convolutional layer and activation function layer for forward propagation to map the low-resolution radiative distribution data to a high-dimensional feature space and extract preliminary local features of the low-resolution radiative distribution data, thus obtaining feature extraction data.

[0087] Optionally, the residual module includes N residual blocks, N skip connection layers and a global fusion layer, where N is an integer greater than 1; The second generating unit 303 is also used for: The feature extraction data is input into the first residual block for deep nonlinear feature extraction to obtain the first extracted data; the feature extraction data and the first extracted data are input into the first skip connection layer for element-wise addition to obtain the first added data; The first summed data is input into the second residual block for deep nonlinear feature extraction to obtain the second extracted data; the first summed data and the second extracted data are input into the second skip connection layer for element-wise addition to obtain the second summed data, until the Nth summed data is obtained from the output of the last skip connection layer; The data after the Nth summation is input into the global fusion layer along with the feature extraction data for element-wise summation to obtain the processed data.

[0088] Optionally, each residual block includes two 3D convolutional layers, two batch normalization layers, and an activation function; The data to be processed is input into the residual block for deep nonlinear feature extraction, and the extracted data is set as follows: The data to be processed is input into the first 3D convolutional layer, the first batch normalization layer, the activation function, the second 3D convolutional layer, and the second batch normalization layer of the residual block for forward propagation to extract deep nonlinear features from the data to be processed, thus obtaining the extracted data.

[0089] Optionally, the upsampling module includes dimension-expanded convolutional layers and subvoxel rearrangement layers; The second generating unit 303 is also used for: The processed data is input into the dimension-expanding convolutional layer for extreme channel expansion, which increases the number of channels of the processed data by a set factor to obtain the dimension-expanded convolutional data. The data after dimension expansion and convolution is input into the sub-voxel rearrangement layer for data rearrangement to obtain the sampled data.

[0090] Optionally, the above-mentioned device may also include a training unit; Training units are used for: Before inputting low-resolution radiative distribution data into a pre-trained deep super-resolution reconstruction model based on a 3D sub-voxel residual network, multiple training datasets are constructed based on the 3D radiative field of the samples. Each training dataset includes low-resolution radiative distribution data of the samples and corresponding high-resolution radiative distribution data of the samples. The low-resolution radiative distribution data of the samples in the training dataset is input into the reconstruction model to be trained for high-resolution reconstruction, resulting in high-resolution prediction data generated and output by the reconstruction model. Based on the difference between the high-resolution prediction data and the high-resolution radiative distribution data of the samples, the reconstruction model to be trained is updated, resulting in an updated reconstruction model. The updated reconstruction model is then trained using other training data until a model that meets the set requirements is obtained, which is then used as the deep super-resolution reconstruction model.

[0091] The three-dimensional radiation field reconstruction device based on residual networks proposed in this embodiment acquires the deployment locations of a finite number of detectors in the three-dimensional radiation field, as well as the radiation dose data detected by these detectors. Based on the deployment locations of the finite detectors, the radiation dose data detected by the finite detectors, and the set low-resolution grid size, low-resolution radiation distribution data is generated. This low-resolution radiation distribution data is then input into a trained deep super-resolution reconstruction model based on a three-dimensional sub-voxel residual network. The deep super-resolution reconstruction model performs shallow feature extraction, deep residual processing, and sub-voxel upsampling on the low-resolution radiation distribution data to generate high-resolution radiation distribution data. This embodiment can effectively reduce the computational cost and improve the reconstruction accuracy of radiation distribution data in a three-dimensional radiation field.

[0092] In this embodiment, the three-dimensional radiation field reconstruction device based on residual networks is presented in the form of functional units. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0093] This invention also provides a computer device having the above-described features. Figure 3 The three-dimensional radiation field reconstruction device based on residual networks is shown.

[0094] Please see Figure 4 The present invention provides a schematic diagram of the structure of a computer device according to an optional embodiment. The computer device includes one or more processors 10, a memory 20, and interfaces for connecting the various components, including high-speed interfaces and low-speed interfaces. The various components are interconnected via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on an external input / output device (such as a display device coupled to the interface). In some optional embodiments, multiple processors and / or multiple buses can be used with multiple memories, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 4 Take a processor 10 as an example.

[0095] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GPA), or any combination thereof.

[0096] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0097] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function. The data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, which can be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0098] Memory 20 may include volatile memory, such as random access memory. Memory may also include non-volatile memory, such as flash memory, hard disk, or solid-state drive. Memory 20 may also include combinations of the above types of memory.

[0099] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0100] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for three-dimensional radiation field reconstruction based on a residual network, characterized in that, include: The deployment locations of a finite number of detectors in a three-dimensional radiation field, as well as the radiation dose data detected by the finite number of detectors, are obtained. Based on the deployment location of the finite detectors, the radiation dose data detected by the finite detectors, and the set low-resolution grid division size, low-resolution radiation distribution data is generated. The low-resolution radiative distribution data is input into a trained deep super-resolution reconstruction model based on a three-dimensional sub-voxel residual network, so that the deep super-resolution reconstruction model performs shallow feature extraction, deep residual processing and sub-voxel upsampling on the low-resolution radiative distribution data to generate high-resolution radiative distribution data.

2. The method of claim 1, wherein, The finite number of detectors includes a finite number of detectors deployed at different locations in the three-dimensional radiation field, and the radiation dose data detected by the finite number of detectors includes the radiation dose rate detected by each detector. Based on the deployment location of the finite detectors, the radiation dose data detected by the finite detectors, and the set low-resolution grid division size, low-resolution radiation distribution data is generated, including: Based on the deployment location of each detector, the radiation dose rate detected by each detector, and the low-resolution grid division size, sparse measurement point interpolation or Monte Carlo particle transport simulation is performed on the three-dimensional radiation field to obtain the initial radiation distribution data of the three-dimensional radiation field. The initial radiation distribution data is logarithmically transformed to compress it to a set range of values, resulting in compressed radiation distribution data. The compressed radiation distribution data is normalized to obtain normalized distribution data, which is then used as the low-resolution radiation distribution data.

3. The method according to claim 1, characterized in that, The deep super-resolution reconstruction model includes a feature extraction module, a residual module, and an upsampling module. The step of inputting the low-resolution radiative distribution data into a trained deep super-resolution reconstruction model based on a three-dimensional sub-voxel residual network, so that the deep super-resolution reconstruction model performs shallow feature extraction, deep residual processing, and sub-voxel upsampling on the low-resolution radiative distribution data to generate high-resolution radiative distribution data, includes: The low-resolution radiation distribution data is input into the feature extraction module for shallow feature extraction to obtain feature extraction data. The feature extraction data is input into the residual module for deep residual processing to obtain the processed data. The processed data is input into the upsampling module for subvoxel upsampling to obtain sampled data, which is then used as the high-resolution radiation distribution data.

4. The method according to claim 3, characterized in that, The feature extraction module includes a three-dimensional convolutional layer and an activation function layer; The step of inputting the low-resolution radiation distribution data into the feature extraction module for shallow feature extraction to obtain feature extraction data includes: The low-resolution radiative distribution data is input into a three-dimensional convolutional layer and an activation function layer for forward propagation to map the low-resolution radiative distribution data to a high-dimensional feature space and extract preliminary local features of the low-resolution radiative distribution data to obtain the feature extraction data.

5. The method according to claim 3, characterized in that, The residual module includes N residual blocks, N skip connection layers and a global fusion layer, where N is an integer greater than 1; The step of inputting the feature extraction data into the residual module for deep residual processing to obtain processed data includes: The feature extraction data is input into the first residual block for deep nonlinear feature extraction to obtain the first extracted data; the feature extraction data and the first extracted data are input into the first skip connection layer for element-wise addition to obtain the first added data; The first summed data is input into the second residual block for deep nonlinear feature extraction to obtain the second extracted data; the first summed data and the second extracted data are input into the second skip connection layer for element-wise addition to obtain the second summed data, until the Nth summed data output by the last skip connection layer is obtained; The Nth summed data and the feature extracted data are input into the global fusion layer for element-wise addition to obtain the processed data.

6. The method according to claim 5, characterized in that, Each residual block includes two 3D convolutional layers, two batch normalization layers, and an activation function; The data to be processed is input into the residual block for deep nonlinear feature extraction to obtain extracted data, including: The data to be processed is input into the first three-dimensional convolutional layer, the first batch normalization layer, the activation function, the second three-dimensional convolutional layer, and the second batch normalization layer of the residual block for forward propagation to extract deep nonlinear features from the data to be processed, thereby obtaining the extracted data.

7. The method according to claim 3, characterized in that, The upsampling module includes a dimension-expanded convolutional layer and a subvoxel rearrangement layer; The step of inputting the processed data into the upsampling module for subvoxel upsampling to obtain sampled data includes: The processed data is input into the dimension-expanded convolutional layer for extreme channel expansion, thereby increasing the number of channels of the processed data by a set factor to obtain dimension-expanded convolutional data. The data after the dimension-expanded convolution is input into the sub-voxel rearrangement layer for data rearrangement to obtain the sampled data.

8. The method according to claim 1, characterized in that, Before inputting the low-resolution radiative distribution data into the trained deep super-resolution reconstruction model based on a three-dimensional sub-voxel residual network, the method further includes: Multiple training datasets are constructed based on the three-dimensional radiation field of the sample, and each training dataset includes low-resolution radiation distribution data of the sample and corresponding high-resolution radiation distribution data of the sample. The low-resolution radiative distribution data of the samples in the training data is input into the reconstruction model to be trained for high-resolution reconstruction, thereby obtaining the high-resolution prediction data generated and output by the reconstruction model to be trained. Based on the difference between the high-resolution prediction data and the high-resolution radiative distribution data of the samples, the reconstruction model to be trained is updated to obtain the updated reconstruction model. The updated reconstruction model is then trained using other training data until a model that meets the set requirements is obtained, which is then used as the depth super-resolution reconstruction model.

9. A three-dimensional radiation field reconstruction device based on residual networks, characterized in that, include: The acquisition unit is used to acquire the deployment location of the finite detectors in the three-dimensional radiation field, as well as the radiation dose data detected by the finite detectors; The first generation unit is used to generate low-resolution radiation distribution data based on the deployment location of the finite detectors, the radiation dose data detected by the finite detectors, and the set low-resolution grid division size. The second generation unit inputs the low-resolution radiation distribution data into a trained deep super-resolution reconstruction model based on a three-dimensional sub-voxel residual network, so that the deep super-resolution reconstruction model performs shallow feature extraction, deep residual processing, and sub-voxel upsampling on the low-resolution radiation distribution data to generate high-resolution radiation distribution data.

10. A computer device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the three-dimensional radiation field reconstruction method based on residual networks as described in any one of claims 1 to 8.

Images