A Monte Carlo dose calculation denoising method, system and device based on denoising convolutional neural network

Abstract

The present invention discloses a Monte Carlo dose calculation denoising method, system, and device based on a denoising convolutional neural network. The method comprises the following steps: acquiring data, dividing the acquired data into datasets, and performing preliminary processing on the data; constructing a denoising convolutional neural network model; training a traditional UNet convolutional neural network model and the constructed denoising convolutional neural network model; performing denoising processing on the traditional UNet convolutional neural network model and the constructed denoising convolutional neural network model; extracting data from the obtained denoising results, performing multi-level comparison, and ultimately outputting the performance and computational efficiency of the two convolutional neural network models. The method improves the quality and convergence speed of subsequent network training, enhances the versatility of the denoising system, and significantly reduces the overall training time. The method can reduce computational complexity and alleviate the computational burden without reducing denoising performance.

Description

Technical Field

[0001] The present invention relates to the field of dose calculation for photon radiotherapy, and in particular to a Monte Carlo dose calculation denoising method, system and device based on a denoising convolutional neural network. Background Art

[0002] Accurate dose calculation is crucial to the success of modern radiotherapy techniques, and Monte Carlo simulation is considered the "gold standard" for dose calculation. However, because Monte Carlo simulation is a stochastic process, the resulting dose images contain inherent noise, whose variance is inversely proportional to the number of simulated events and, therefore, the simulation time. Although currently, parallel computing frameworks based on graphics processing units (GPUs) can accelerate a typical Monte Carlo simulation process to a few minutes, meeting the ultimate computational efficiency requirements of practical workflows, some areas of practical workflows require real-time computation, such as online adaptive therapy, in which a given treatment component must be completely replanned within minutes to account for the changing anatomy during the procedure. Since dose calculation is only a small part of the planning and reconstruction process, its duration must be shortened from minutes to seconds by accelerating GPU-based Monte Carlo dose calculation.

[0003] Although some research has used deep learning (DL) techniques based on convolutional neural networks (CNNs) to suppress noise generated by a small number of simulated events, this work is still far from achieving the real-time (sub-second) level required for clinical use without compromising denoising effectiveness. Furthermore, related work has not adequately addressed the learning curve for end users associated with the application of denoising systems across different hospitals, treatment machines, and treatment bundles. Summary of the Invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a Monte Carlo dose calculation denoising method, system and equipment based on a denoising convolutional neural network to solve the problems of long time, low efficiency and high transplantation cost in the image noise processing of Monte Carlo dose calculation.

[0005] Technical solution: The present invention provides a Monte Carlo dose calculation denoising method based on a denoising convolutional neural network, comprising the following steps:

[0006] (1) Acquire data, divide the acquired data into data sets and perform preliminary processing;

[0007] (2) Construct a denoising convolutional neural network model;

[0008] (3) The traditional UNet convolutional neural network model and the constructed denoising convolutional neural network model are trained. Specifically, the training data set and the verification data set obtained in step (1) are respectively input into two initialized convolutional neural network models. The two models are also trained using the Adam learning rate algorithm according to the weakly supervised learning framework to obtain two trained convolutional neural network models.

[0009] (4) Denoising the traditional UNet convolutional neural network model and the constructed denoising convolutional neural network model. Specifically, the trained UNet convolutional neural network model and the denoising convolutional neural network model are used to denoise the training dataset and the validation dataset of the noise image obtained by Monte Carlo rendering under low sampling number, and the denoising results are added to obtain the final denoised image.

[0010] (5) After extracting data from the obtained denoising results, a multi-level comparison is performed, and finally the performance and computational efficiency of the two convolutional neural network models are output.

[0011] Furthermore, the step (1) includes the following steps:

[0012] (1.1) Divide the dataset into training dataset and validation dataset;

[0013] (1.2) Perform data expansion on the preprocessed dataset;

[0014] (1.3) Perform data normalization on the preprocessed dataset.

[0015] Furthermore, the step (1.1) of preprocessing the data set further includes:

[0016] (1.1.1) Divide the dataset into training dataset and validation dataset;

[0017] (1.1.2) For the patient data in the training dataset, two independent low sampling number (1×106) Monte Carlo simulations were performed; for the patient data in the validation dataset, fifteen independent low sampling number (1×106) Monte Carlo simulations were performed.

[0018] Furthermore, the step (1.2) of performing data expansion on the preprocessed data set further includes: while expanding the edge of the valid area of ​​the Monte Carlo calculation output image (additionally filling 16 zeros outside each dimension), randomly exchanging the input and target noise images.

[0019] Furthermore, the step (1.3) of normalizing the preprocessed data set further includes: all dosage data are normalized and uniformly divided by the average prescription dose of patients in the training data set.

[0020] Furthermore, the step (2) includes the following steps:

[0021] (2.1) The denoising convolutional neural network model is constructed using a multi-channel input architecture;

[0022] (2.2) The denoising convolutional neural network model is constructed by using convolution operator decoupling and voxel shuffle / unshuffle operations;

[0023] (2.3) Initialize the parameters of the traditional UNet architecture convolutional neural network and the denoising convolutional neural network constructed in steps (2.1)-(2.2).

[0024] Furthermore, the multi-channel input architecture in the denoising convolutional neural network model architecture in step (2.1) further includes:

[0025] (2.1.1) Assume that the original input size is B×C×H×W×D, where B and C represent the batch size and number of channels, respectively, and H, W, and D represent the row, column, and slice dimensions, respectively.

[0026] (2.1.2) For the Monte Carlo dose denoising task, set C = 1;

[0027] (2.1.3) After the voxel splitting layer, the input is rearranged into a size of 8-channel input;

[0028] (2.1.4) In each channel, the subvolume is a downsampled version of the original input with a stride of 2 and a different starting point, i.e. X i,j,k =X[i∷2,j∷2,k∷2],i,j,k∈[0,1],X i,j,k Indicates the (4k+2j+i)th channel.

[0029] Furthermore, the step (2.2) of decoupling the convolution operator in the denoising convolutional neural network model architecture further includes:

[0030] (2.2.1) The conventional 3D volume convolution operator is decoupled into a 2D axial convolution operator and a 1D slice convolution operator, and the conventional convolution operator with the same kernel size K in each dimension is expressed as Among them C in and C out represent input and output channels respectively, K represents the kernel size;

[0031] (2.2.2) The decoupled convolution operator is designed to simulate the conventional three-dimensional convolution; The decoupled convolution operator is designed to simulate the conventional three-dimensional convolution; First, use the two-dimensional convolution operator To describe the axial spatial correlation; secondly, use the one-dimensional convolution operator To describe the slice correlation; each module consists of six operators, the first three operators are axial convolution normalization activation functions (CONV-normalization-ReLU), and the last three operators are slice convolution normalization activation functions.

[0032] Furthermore, the voxel shuffle / unshuffle operation used in the denoising convolutional neural network model in step (2.2) further includes:

[0033] (2.2.3) Propose voxel shuffling / de-shuffling operation;

[0034] (2.2.4) Attaching a voxel unshuffle layer and a voxel shuffle layer to the input layer and the output layer respectively. Another effect of using these operators to pre-process the input and post-process the output is that they use fewer downsampling and upsampling modules to achieve the same receptive field size because the resulting size is only half of the original size in each dimension.

[0035] Furthermore, the step (5) of comparing the two convolutional neural networks further includes: the step (5) of comparing the two convolutional neural networks further includes: performing simple processing on the denoising results of the two trained convolutional neural networks obtained in step (4), extracting the data, and comparing the original noisy image obtained with a low sampling number, the clean image obtained with a high sampling number, the denoised image of the UNet convolutional neural network and the denoised image of the denoising convolutional neural network, and finally outputting the comparison results of the model performance and computational efficiency of the two convolutional neural networks.

[0036] The present invention also provides a denoising system for Monte Carlo dose calculation based on a denoising convolutional neural network, comprising:

[0037] Data processing module: mainly used to preprocess the acquired data set, and perform data expansion and normalization on the preprocessed data set;

[0038] Convolutional neural network denoising module: It is mainly used to complete the training of denoising convolutional neural network through the pre-built weakly supervised learning framework, using the pre-processed data obtained by the data processing module as the training dataset and validation dataset. The trained denoising convolutional neural network inputs the low-sampled Monte Carlo calculated image and outputs the denoised image.

[0039] Comparison output module: First, the denoising convolutional neural network is compared with the traditional UNet architecture convolutional neural network in terms of model performance and computational efficiency; secondly, the denoising convolutional neural network is compared with the high-sampling Monte Carlo calculation image based on MSE; the final output result meets the two checks in the comparison output module; if the above conditions are not met in this link, a prompt will be given, and specific comparison results in two aspects will be provided.

[0040] The present invention also provides a terminal device comprising: a processor, a memory, and a computer program stored in the memory and executable by the processor. When the processor executes the computer program, the steps of the aforementioned method for rapid analysis of prediction patterns are implemented. Alternatively, when the processor executes the computer program, the functions of the aforementioned systems and units are implemented.

[0041] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are as follows: (1) By performing importance sampling and preprocessing on the data set, low-importance data can be cleared and the data range can be converted, thereby improving the quality and convergence speed of subsequent network training; (2) The weakly supervised learning framework avoids the problem of providing a large number of dose images obtained under high sampling during training, improves the versatility of the denoising system, greatly reduces the entire training time, and is of great significance for the necessary fine-tuning of the model in clinical practice; (3) By using the voxel deconstruction layer to arrange the original input size into 8-channel input, the conventional three-dimensional volume convolution operator is decomposed into a two-dimensional axial convolution operator and a one-dimensional slice convolution operator, which can reduce the computational complexity and reduce the computational burden without reducing the denoising performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is the overall flow chart of the present invention;

[0043] Figure 2 This is a diagram of the denoising convolutional neural network architecture of the present invention;

[0044] Figure 3 Schematic diagram of the denoising system based on Monte Carlo dose calculation of the present invention. DETAILED DESCRIPTION

[0045] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0046] The Monte Carlo dose calculation denoising method based on a denoising convolutional neural network according to an embodiment of the present invention is as follows: Figure 1-2 As shown, the following steps are included:

[0047] (1) Acquire data, divide the acquired data into data sets and perform preliminary processing, including the following steps:

[0048] (1.1) Preprocess the dataset;

[0049] (1.1.1) Obtain a patient dataset and classify the patients in the dataset into a training dataset and a validation dataset;

[0050] The data set includes a number of noise images calculated and rendered by a fast Monte Carlo dose calculation engine at a low sampling number and a number of reference images obtained at a high sampling number;

[0051] (1.1.2) For training purposes, two independent low-sample noise Monte Carlo simulations were performed for each patient case in the training dataset, with 1×10 6 For validation purposes, 15 independent low-sampling-noise Monte Carlo simulations were performed for each patient case in the validation dataset, with 1×10 6 Multiple noisy Monte Carlo simulations were performed for each patient to show the stability of the proposed denoiser. In addition, for quantitative comparison, one Monte Carlo simulation with 1×10 9 A "clean" simulation with a sampling number of samples, serving as the basis for the true dose map;

[0052] (1.2) Dataset Expansion The effective area of ​​the Monte Carlo calculation output image is expanded (additional 16 zeros are added to each dimension) and the input and target noise images are randomly swapped. The effective area is expanded as follows: All dose maps are adjusted to 2.34×2.34×3.00mm 3 Then, a 256×256×64 volume containing the valid area was cropped from the original Monte Carlo dose map. To increase the training dataset, this cropped volume was padded with 16 zeros outside each dimension.

[0053] (1.3) In order to stabilize the training process, the fast Monte Carlo dose calculation results were processed and the doses of all patients were normalized by dividing them by the average prescribed dose of the patients in the training dataset.

[0054] (2) Constructing a denoising convolutional neural network model, specifically: constructing a traditional UNet architecture convolutional neural network and a denoising convolutional neural network proposed in the present invention and initializing the parameters of the two networks to obtain two initialized convolutional neural network models, including the following steps:

[0055] (2.1) The denoising convolutional neural network model is constructed using a multi-channel input architecture, including the following steps:

[0056] (2.1.1) Assume that the original input size is B×C×H×W×D, where B and C represent the batch size and number of channels, respectively, and H, W, and D represent rows, columns, and slices.

[0057] (2.1.2) For the Monte Carlo dose denoising task, set C = 1;

[0058] (2.1.3) After the voxel splitting layer, the input is rearranged into a size of 8-channel input;

[0059] (2.1.4) In each channel, the subvolume is a downsampled version of the original input with a stride of 2 and a different starting point, i.e. X i,j,k =X[i∷2,j∷2,k∷2],i,j,k∈[0,1],X i,j,k represents the (4k+2j+i)th channel; this newly introduced multi-channel input still contains all the information, thus maintaining the same receptive field as the traditional single-channel-based input. At the same time, this multi-channel input can greatly improve the efficiency of the model because the input size is twice smaller in each dimension. Therefore, even if the number of channels is large, the increase in the associated computational burden is negligible;

[0060] (2.2) The denoising convolutional neural network model is constructed by using convolution operator decoupling and voxel shuffle / unshuffle operations, including the following steps:

[0061] (2.2.1) The conventional 3D volume convolution operator is decoupled into a 2D axial convolution operator and a 1D slice convolution operator, and the conventional convolution operator with the same kernel size K in each dimension is expressed as Among them C in and C out represent input and output channels respectively, K represents the kernel size;

[0062] (2.2.2) The decoupled convolution operator is designed to simulate the conventional three-dimensional convolution. First, the two-dimensional convolution operator is used. To describe the axial spatial correlation, and then use the one-dimensional convolution operator To describe the slice correlation. Therefore, in the network architecture of the present invention, each module consists of six operators. The first three operators are axial convolution normalization activation functions (CONV-normalization-ReLU), and the last three operators are slice convolution normalization activation functions;

[0063] (2.2.3) In order to reduce the main computational burden from the input and output layers, and to ensure that the information loss and upsampling performance degradation are acceptable, a voxel shuffling / de-scaling operation is proposed;

[0064] (2.2.4) Attach a voxel unshuffle layer and a voxel shuffle layer to the input and output layers respectively, and use these operators to pre-process the input and post-process the output. Another effect is that they use fewer downsampling and upsampling modules to achieve the same receptive field size, because the resulting size is only half of the original size in each dimension;

[0065] (2.3) Initialize the parameters of the traditional UNet architecture convolutional neural network and the denoising convolutional neural network constructed in steps (2.1)-(2.2).

[0066] (3) The traditional UNet convolutional neural network model and the constructed denoising convolutional neural network model are trained. Specifically, the training data set and the verification data set obtained in step (1) are respectively input into two initialized convolutional neural network models. The two models are also trained using the Adam learning rate algorithm according to the weakly supervised learning framework to obtain two trained convolutional neural network models, including the following steps:

[0067] (3.1) Construct a weakly supervised learning framework; let Y∈R be a noisy volume dose map. H×W×D , with H rows, W columns and D slices, is noised by ε∈R due to the undersampling MC dose calculation H×W×D The goal is to recover the basic clean signal from Y, X∈R H×W×D , which has the following relations:

[0068] Y=X+∈ (1)

[0069] Among them, Y represents the noisy signal, X represents the noise-free signal, and ∈ represents the noisy signal

[0070] And the cost function of the supervised learning framework is known:

[0071]

[0072] where φ W Represents the trained denoising network, X represents the noise-free signal, and ∈ represents the noise signal;

[0073] Relax the noise-free signal X in the cost function to its other noisy counterpart. That is, the network φ can be trained by using a strategy of using a noisy input Y1 = X + ε1 to another noisy target Y2 = X + ε2, rather than the traditional strategy of using a noisy input Y = X + ε to a clean target X, where the subscripts 1 and 2 represent the smaller number of samples calculated for the two different Monte Carlo doses. For this noise-to-noise training scheme, the cost function can be expressed as:

[0074]

[0075] where φ W Represents the trained denoising network, X represents the noise-free signal, ∈1 and ∈2 represent two different low-sampling Monte Carlo calculation results, i.e., noise signals.

[0076] (3.2) To train the network using the noise-noise weakly supervised learning framework, the input and target are set as two noise simulations, each with 1×10 6 It should be noted that during the training process, the input and target are randomly exchanged between the two noise simulations; the initial learning rate is 1×10 -4 The Adam optimizer is implemented on a Tesla K80 GPU with 2×10 5 iterations to train the network. 5 and 1.5×10 5 When the learning rate is reduced by 10 times, corresponding to 1×10 -5 and 1×10 -6 The hyperparameters associated with the Adam optimizer are β1 = 0.9 and β2 = 0.999.

[0077] (4): The traditional UNet convolutional neural network model and the constructed denoising convolutional neural network model are subjected to denoising. Specifically, the trained UNet convolutional neural network model and the denoising convolutional neural network model are used to denoise the training dataset and the validation dataset of the noise image obtained by Monte Carlo rendering under low sampling number, and the denoising results are added to obtain the final denoised image.

[0078] (5): After extracting data from the obtained denoising results, perform multi-level comparisons and finally output the performance and computational efficiency of the two convolutional neural network models. Specifically, the denoising results of the two trained convolutional neural networks obtained in step (4) are simply processed. After extracting data, the original noisy image obtained with a low sampling number, the clean image obtained with a high sampling number, the denoised image of the UNet convolutional neural network, and the denoised image of the denoised convolutional neural network are compared. Finally, the comparison results of the model performance and computational efficiency of the two convolutional neural networks are output, which includes the following steps:

[0079] (5.1): Comparison of model performance. First, the performance is qualitatively compared by showing the dose difference between the denoised and "clean" Monte Carlo dose maps and plotting representative one-dimensional profiles. Then, the mean square error (MSE) is used to quantify the difference between the denoised dose map and the baseline dose map X. GT The difference between.

[0080]

[0081] Among them, X denoised represents the dose map after denoising, X GT Representative baseline dose map.

[0082] At the same time, use the denoised and X GT The absolute difference in area between the two related DVHs is used to quantify the denoising performance.

[0083]

[0084] Where DVH Error is the absolute difference in area between the two DVHs, ΔD i is the width of the i-th DVH bin, and The values ​​of the two histograms representing the bin.

[0085] Other clinically relevant metrics were also used to show the performance of the algorithm. First, D# was defined as the minimum dose received by #% of the planning target volume (PTV). This number # was defined as a percentage of the prescribed dose. PTV D95, PTV D98, and PTV D99 were compared as clinical evaluation criteria. Their relative differences were also called where D#denoised and D#GT correspond to the denoised and baseline target doses, respectively, greater than the target volume dose. Gamma transmission rate is another widely used metric in clinical practice. In this paper, a 2% / 2mm standard is used to calculate the gamma transmission rate for each volume dose and compare it with the baseline volume dose.

[0086] The isodose coverage similarity is also compared, which is one of the important indicators to indicate the similarity between two plans. The isodose volume is defined as a binary mask, which has a value of 1 if the voxel contains a dose value exceeding a specified threshold and 0 otherwise. #％ISO Expressed as the volume of the #% isodose region. Then, the two dose maps X denoised and X GT The similarity of isodose coverage between dose levels of #% can be defined as the Dice coefficient between the relevant isodose volumes.

[0087]

[0088] in, Representative dose graph X denoised The volume of the isodose region with a medium dose level of #%, Represents X GT The volume of the isodose region at the mid-dose level is #%.

[0089] For a more comprehensive comparison, six different dose levels (10%, 30%, 50%, 70%, 80%, and 90%) were used to calculate the mean isodose similarity.

[0090] It should be noted that in order to eliminate the influence of different prescription doses for different patients, the dose map needs to be normalized first and divided by PTV D95.

[0091] (5.2): Computational efficiency comparison. The number of floating point operations (FLOPs) involved in the data flowing through the network is used to theoretically quantify the amount of computation required. FLOPs can be defined as the number of multiplications and summations. For example, given a kernel size of k h ×k w ×k d The convolution operator, given a size of 1×C in×H×W×D If the output size is 1×C out ×H out ×W out ×D out , then FLOPs can be calculated as:

[0092] FLOPs = 2xC in ×k h ×k w ×k d ×C out ×H out ×W out ×D out (7)

[0093] Among them, C in Represents input and output channels, H out 、W out 、D out Represents the output channels of rows, columns, and slices, k h 、k w 、k d Convolution operators representing rows, columns, and slices

[0094] like Figure 3 As shown, an embodiment of the present invention further provides a denoising system based on Monte Carlo dose calculation, including a data processing module, a convolutional neural network denoising module, and a comparison output module.

[0095] Data processing module: mainly used to preprocess the acquired data set, and perform data expansion and normalization on the preprocessed data set;

[0096] Convolutional neural network denoising module: It is mainly used to complete the training of denoising convolutional neural network through the pre-built weakly supervised learning framework, using the pre-processed data obtained by the data processing module as the training dataset and validation dataset. The trained denoising convolutional neural network inputs the low-sampled Monte Carlo calculated image and outputs the denoised image.

[0097] Comparison output module: First, the denoising convolutional neural network is compared with the traditional UNet architecture convolutional neural network in terms of model performance and computational efficiency; secondly, the denoising convolutional neural network is compared with the high-sampling Monte Carlo calculation image based on MSE; the final output result meets the two checks in the comparison output module; if the above conditions are not met in this link, a prompt will be given, and specific comparison results in two aspects will be provided.

[0098] An embodiment of the present invention also provides a terminal device comprising: a processor, a memory, and a computer program stored in the memory and executable by the processor. When the processor executes the computer program, the steps of the aforementioned method for rapid analysis of prediction patterns are implemented. Alternatively, when the processor executes the computer program, the functions of the aforementioned systems and units are implemented.

Claims

1. A Monte Carlo dose calculation denoising method based on a denoising convolutional neural network, characterized in that: The following steps are involved: (1) Acquire data, divide the acquired data into data sets and perform preliminary processing; (2) Constructing a denoising convolutional neural network model; including the following steps: (2.1) The denoising convolutional neural network model is constructed using a multi-channel input architecture; (2.2) The constructed denoising convolutional neural network model is constructed by using convolution operator decoupling and voxel shuffling / de-scaling operations; the convolution operator decoupling in the denoising convolutional neural network model architecture further includes: (2.2.1) Decouple the conventional 3D volume convolution operator into a 2D axial convolution operator and a 1D slice convolution operator, and express the conventional convolution operator with the same kernel size K in each dimension as in and represent input and output channels respectively, K represents the kernel size; (2.2.2) The decoupled convolution operator is designed to simulate the conventional three-dimensional convolution; first, use the two-dimensional convolution operator To describe the axial spatial correlation; secondly, use the one-dimensional convolution operator To describe the slice correlation; each module consists of six operators, the first three operators are axial convolution normalization activation functions, and the last three operators are slice convolution normalization activation functions; (2.3) Initialize the parameters of the traditional UNet architecture convolutional neural network and the denoising convolutional neural network constructed in steps (2.1)-(2.2); (3) The traditional UNet convolutional neural network model and the constructed denoising convolutional neural network model are trained. Specifically, the training data set and the verification data set obtained in step (1) are respectively input into two initialized convolutional neural network models. The two models are also trained using the Adam learning rate algorithm according to the weakly supervised learning framework to obtain two trained convolutional neural network models. (4) Denoising the traditional UNet convolutional neural network model and the constructed denoising convolutional neural network model. Specifically, the trained UNet convolutional neural network model and the denoising convolutional neural network model are used to denoise the training dataset and the validation dataset of the noise image obtained by Monte Carlo rendering under low sampling number, and the denoising results are added to obtain the final denoised image. (5) After extracting data from the obtained denoising results, a multi-level comparison is performed, and finally the performance and computational efficiency of the two convolutional neural network models are output.

2. The Monte Carlo dose calculation denoising method based on a denoising convolutional neural network according to claim 1, characterized in that: The step (1) comprises the following steps: (1.1) Preprocess the dataset; (1.2) performing data expansion on the preprocessed data set, including: expanding the edge of the valid area of ​​the Monte Carlo calculation output image while randomly exchanging the input and target noise images; (1.3) Perform data normalization on the preprocessed dataset, including: all dosage data are normalized and uniformly divided by the average prescription dose of patients in the training dataset.

3. The Monte Carlo dose calculation denoising method based on a denoising convolutional neural network according to claim 2, characterized in that: The step (1.1) of preprocessing the data set further includes: (1.1.1) Divide the dataset into training dataset and validation dataset; (1.1.2) For the patient data in the training dataset, two independent low-sampling (1×10 6 ) Monte Carlo simulation; for the patient data in the validation dataset, fifteen independent low-sampling (1×10 6 )Monte Carlo simulation.

4. The Monte Carlo dose calculation denoising method based on a denoising convolutional neural network according to claim 1, characterized in that: The multi-channel input architecture in the denoising convolutional neural network model architecture of step (2.1) further includes: (2.1.1) Assume that the original input has the shape B×C×H×W×D, where B and C represent the batch size and number of channels, respectively, and H, W, and D represent the row, column, and slice dimensions, respectively. (2.1.2) For the Monte Carlo dose denoising task, set C = 1; (2.1.3) After the voxel splitting layer, the input is rearranged into a size of 8-channel input; (2.1.4) In each channel, the subvolume is a downsampled version of the original input with a stride of 2 and a different starting point, i.e. , , Indicates the (4k+2j+i)th channel.

5. The Monte Carlo dose calculation denoising method based on a denoising convolutional neural network according to claim 1, characterized in that: The voxel shuffling / de-scrambling operation employed in the denoising convolutional neural network model in step (2.2) further comprises: (2.2.3) Propose voxel shuffling / de-shuffling operation; (2.2.4) Attaching a voxel unshuffling layer and a voxel shuffling layer to the input layer and the output layer respectively. Another effect of using these operators to pre-process the input and post-process the output is that fewer downsampling and upsampling modules are used to achieve the same receptive field size.

6. The Monte Carlo dose calculation denoising method based on a denoising convolutional neural network according to claim 1, characterized in that: The step (5) of comparing the two convolutional neural networks further includes: performing simple processing on the denoising results of the two trained convolutional neural networks obtained in step (4), extracting the data, and comparing the original noisy image obtained with a low sampling number, the clean image obtained with a high sampling number, the denoised image of the UNet convolutional neural network, and the denoised image of the denoising convolutional neural network, and finally outputting the comparison results of the model performance and computational efficiency of the two convolutional neural networks.

7. A denoising system for Monte Carlo dose calculation based on a denoising convolutional neural network, applied to any method of claims 1-6, characterized in that: include: Data processing module: mainly used to preprocess the acquired data set, and perform data expansion and normalization on the preprocessed data set; Convolutional neural network denoising module: It is mainly used to complete the training of denoising convolutional neural network through the pre-built weakly supervised learning framework, using the pre-processed data obtained by the data processing module as the training dataset and validation dataset. The trained denoising convolutional neural network inputs the low-sampled Monte Carlo calculated image and outputs the denoised image. Comparison output module: First, the denoising convolutional neural network is compared with the traditional UNet architecture convolutional neural network in terms of model performance and computational efficiency; secondly, the denoising convolutional neural network is compared with the high-sampling Monte Carlo calculation image based on MSE; the final output result meets the two checks in the comparison output module; if the above conditions are not met in this link, a prompt will be given, and specific comparison results in two aspects will be provided.

8. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the Monte Carlo dose calculation denoising method based on a denoising convolutional neural network are implemented as described in any one of claims 1 to 6.

Images