Rapid Monte Carlo dose calculation method based on statistical uncertainty perception noise reduction and application

Abstract

The invention provides a rapid Monte Carlo dose calculation method based on statistical uncertainty perception noise reduction and application, and the method comprises the steps: carrying out the Monte Carlo dose simulation of a low particle number, and obtaining the middle rough dose distribution containing random noise and a corresponding voxel-level statistical uncertainty distribution diagram; acquiring an anatomical structure image of a patient, performing normalization processing and registration processing on the anatomical structure image, the middle rough dose distribution and the corresponding voxel-level statistical uncertainty distribution diagram, and constructing three-channel input data; the three-channel input data is input into a pre-constructed and trained statistical uncertainty perception denoising network model, the network model adopts a dual-path cavity convolution fusion architecture, and local features and global context features are extracted and fused; and the network model outputs denoised high-precision dose distribution as a final Monte Carlo dose calculation result. According to the method, the efficiency and the accuracy of radiotherapy plan design are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image processing, and in particular to a fast Monte Carlo dose calculation method based on statistical uncertainty perception denoising and application. BACKGROUND

[0002] Accurate dose distribution calculation is the core of radiotherapy planning and clinical decision-making, which directly affects the treatment effect and patient prognosis. In the existing calculation methods, Monte Carlo (MC) simulation is recognized as the "gold standard" for radiotherapy dose calculation because it can accurately simulate the transport process of particles in non-uniform tissues. However, the accuracy of the MC method depends on the number of simulated particle histories. Due to the limited number of particles, the statistical uncertainty (SU) introduced by random noise reduces the reliability of the dose distribution. In order to reduce SU, a large number of particle histories (e.g., up to 10 9 orders of magnitude) need to be simulated, which leads to exponential growth of calculation time, and the huge calculation cost limits the real-time application of the MC method in clinical routine procedures.

[0003] In order to solve the contradiction between accuracy and efficiency, existing researches mainly focus on variance reduction techniques (VRT) and post-processing denoising algorithms. Traditional filtering denoising methods (such as Savitzky-Golay filtering, median filtering, etc.) are difficult to handle the irregular random characteristics of MC noise. In recent years, deep learning (DL) based methods (such as MCDNet, U-Net, etc.) have shown strong noise suppression ability. However, most of the existing DL denoising methods do not explicitly include "statistical uncertainty" in the network input, but only rely on noisy dose maps and CT images. In fact, SU fundamentally determines the voxel-level dose noise level, and ignoring this key physical information may lead to prediction bias and limit the model's generalization ability in different anatomical regions and different noise levels.

[0004] In summary, in order to recover high-fidelity MC dose distribution more quickly and accurately, it is urgent to develop a deep learning denoising framework that can explicitly fuse statistical uncertainty information. SUMMARY

[0005] Therefore, the present application provides a fast Monte Carlo dose calculation method based on statistical uncertainty perception denoising and application to solve the above problems.

[0006] This invention provides a fast Monte Carlo dose calculation method based on statistical uncertainty-aware denoising, comprising: performing a low-particle-number Monte Carlo dose simulation to obtain an intermediate coarse dose distribution containing random noise and a corresponding voxel-level statistical uncertainty distribution map; acquiring anatomical images of the patient, and performing normalization and registration processing on the anatomical images, the intermediate coarse dose distribution, and the corresponding voxel-level statistical uncertainty distribution map to construct three-channel input data; inputting the three-channel input data into a pre-constructed and trained statistical uncertainty-aware denoising network model, wherein the network model adopts a dual-path dilated convolutional fusion architecture to extract local features and global contextual features respectively and fuse them; and outputting a denoised high-precision dose distribution as the final Monte Carlo dose calculation result.

[0007] In another implementation of the present invention, the voxel-level statistical uncertainty distribution map is defined as the relative standard deviation at the voxel level, which is calculated based on the batch dose variance in the Monte Carlo simulation and characterizes the random noise level of the dose distribution under the current simulated particle number.

[0008] In another implementation of the present invention, the anatomical structure image is a CT image; before the three-channel input data is input into the network model, the intermediate coarse dose distribution and the voxel-level statistical uncertainty distribution are subjected to 3D normalization processing based on the patient planning volume.

[0009] In another implementation of the present invention, the statistical uncertainty-aware denoising network model is constructed based on a batch renormalization denoising network architecture, which includes two parallel feature extraction branches. The first branch consists of multiple layers of "convolution + batch normalization + linear rectifier units" and is used to extract local anatomical and dose texture features of the image. The second branch consists of multiple layers of "dilated convolution + linear rectifier units" and expands the receptive field to capture long-distance dose-uncertainty correlation and global contextual information. The features extracted by the two branches are concatenated and fused at the end of the network, and the denoised dose residual or direct dose distribution is reconstructed and output through convolutional layers.

[0010] In another implementation of the present invention, the training process of the network model includes: constructing a training dataset containing different tumor sites and different simulated particle number gradients, using the low-uncertainty dose distribution generated by high particle number simulation as the gold standard, and using mean squared error as the loss function; and using the Adam optimizer to iteratively optimize the network parameters until the model converges.

[0011] In another implementation of the invention, the different simulated particle number gradients include those from 1 × 10⁻⁶. 5 Up to 5×10 7Multiple orders of magnitudes are used to train corresponding sub-models or a unified model for each noise level, in order to adapt to the denoising needs under different computation time constraints in clinical practice.

[0012] Another aspect of the present invention provides an application of a rapid Monte Carlo dose calculation method based on statistical uncertainty perception denoising, comprising: obtaining a high-precision Monte Carlo radiotherapy dose distribution by employing the method described in any one of the preceding claims, for inverse optimization of radiotherapy plans, dose assessment, or quality control.

[0013] In another aspect, the present invention provides a fast Monte Carlo dose calculation system based on statistical uncertainty-aware denoising, comprising: a data acquisition module: obtaining an intermediate coarse dose distribution containing random noise and a corresponding voxel-level statistical uncertainty distribution map by performing a low-particle-number Monte Carlo dose simulation; a data preprocessing module: acquiring anatomical images of the patient, normalizing and registering the anatomical images, the intermediate coarse dose distribution, and the corresponding voxel-level statistical uncertainty distribution map to construct three-channel input data; a denoising network module: inputting the three-channel input data into a pre-constructed and trained statistical uncertainty-aware denoising network model, wherein the network model adopts a dual-path dilated convolutional fusion architecture to extract local features and global contextual features respectively and fuse them; and a result output module: the network model outputs a denoised high-precision dose distribution as the final Monte Carlo dose calculation result.

[0014] In another aspect, the present invention provides an electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of a fast Monte Carlo dose calculation method based on statistical uncertainty-perceived noise reduction as described in any of the preceding claims. In another aspect, the present invention provides a computer storage medium storing a computer program that, when executed by a processor, implements the steps of a fast Monte Carlo dose calculation method based on statistical uncertainty-aware noise reduction as described in any of the preceding claims.

[0015] The rapid Monte Carlo dose calculation method based on statistical uncertainty perception denoising of the present invention has the following advantages and beneficial effects: Explicit fusion of physical information: By introducing statistical uncertainty (SU) as an independent channel into the network, the model can adaptively denoise based on the physical distribution characteristics of noise, which significantly reduces mean dose error (MDE) compared to a dual-channel model that only uses dose and CT.

[0016] Dual-path feature fusion: The dual-path architecture combines standard convolution and dilated convolution, which simultaneously ensures the fine preservation of local anatomical structures and the smoothness of global dose distribution.

[0017] High efficiency and accuracy coexist: It can shorten the high particle number simulation that takes tens of minutes to "second-level low particle number simulation + second-level denoising", significantly reducing the total time. Moreover, the dose distribution Gamma pass rate (3% / 3mm) after denoising can reach more than 99%, meeting the clinical accuracy requirements. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. By reading the detailed description of the embodiments below, the advantages and benefits of the solutions will become clear to those skilled in the art. The accompanying drawings are only for illustrating preferred embodiments and are not intended to limit the present invention. In the accompanying drawings: Figure 1 This is a schematic diagram of the overall process of a fast Monte Carlo dose calculation method based on statistical uncertainty perception denoising, according to an embodiment of the present invention.

[0019] Figure 2 This is a detailed flowchart illustrating a fast Monte Carlo dose calculation method based on statistical uncertainty-aware noise reduction, according to an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram comparing the noise reduction effect with the gold standard in head and neck, brain, and lung cases according to an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram comparing the denoising error of an embodiment of the present invention with that of a model without introduced statistical uncertainty (SU-ignored). Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art should fall within the protection scope of the present invention.

[0023] Figure 1This is a schematic diagram of the overall process of a rapid Monte Carlo dose calculation method based on statistical uncertainty-aware denoising provided in an embodiment of the present invention. The method mainly includes two stages: a low-particle-number Monte Carlo (MC) simulation stage to obtain a coarse dose distribution for the radiotherapy plan, and a stage where the dose distribution is input into a constructed deep learning dose denoising model. Using these two stages sequentially yields an accurate and efficient calculated dose distribution for the radiotherapy plan. Stage 1 can be performed using any open-source or validated self-developed Monte Carlo dose calculation package for radiotherapy planning. The number of particles required for the low-particle-number simulation can be set according to the actual clinical situation, such as... , , , , , This provides an initial coarse dose distribution for subsequent dose denoising. Stage 2 is the core dose denoising model of this invention. The method of this invention can restore the coarse dose simulated in low particle number simulations to the accuracy level of high particle number simulations in a very short time, significantly improving the efficiency and accuracy of radiotherapy planning. Figure 1 As shown, this embodiment mainly includes: S101. By performing a low-particle-number Monte Carlo dose simulation, an intermediate coarse dose distribution containing random noise and the corresponding voxel-level statistical uncertainty distribution are obtained.

[0024] For example, simulations are performed using a high-efficiency Monte Carlo (MC) dosing engine based on a GPU architecture.

[0025] Simulation settings: To achieve fast calculations, the number of simulated particles is set to a low level (e.g., 1 × 10⁻⁶ in this embodiment). 5 Up to 5 × 10 7 (Different gradients within the range), compared to the 10 typically required by the clinical gold standard. 9 The particle number setting can significantly reduce simulation time.

[0026] Simulated data: Noisy Dose: denoted as D noisy Due to insufficient particle count, this distribution contains significant random statistical noise.

[0027] Statistical uncertainty distribution map (SU Map): denoted as SU, its calculation is based on the statistical standard deviation of voxel-level dose.

[0028] S102. Obtain the anatomical structure image of the patient, and perform normalization and registration processing on the anatomical structure image, the intermediate coarse dose distribution and the corresponding voxel-level statistical uncertainty distribution to construct three-channel input data.

[0029] For example, localization CT images of the patient are acquired, ensuring that the CT images and coarse dose distribution D noisy The statistical uncertainty map (SU) is spatially strictly registered to ensure that it has the same voxel size and resolution.

[0030] To accelerate network convergence, the data from the three channels were subjected to 3D normalization based on patient volume, including: For the dose distribution D, the maximum value normalization is used: Dmean = D / Dmax.

[0031] For the SU plot, it is also normalized to the [0,1] interval.

[0032] For CT images, truncate to the range [-1000, 2000] HU and normalize.

[0033] The processed data is combined into a tensor of size (H, W, 3) (for 2D slice processing mode), where the three channels are: normalized CT, normalized coarse dose D, etc. noisy Normalized SU.

[0034] S103. Input the three-channel input data into a pre-built and trained statistical uncertainty-aware denoising network model. The network model adopts a dual-path dilated convolutional fusion architecture to extract local features and global context features respectively and fuse them.

[0035] S104. The network model outputs a denoised high-precision dose distribution as the final Monte Carlo dose calculation result.

[0036] For example, low-particle-number simulation data of the patient to be calculated is input into a trained model, which outputs a denoised dose distribution within milliseconds. This distribution can be directly used for calculating the inverse optimization objective function of clinical radiotherapy planning, or for final dose assessment.

[0037] The rapid Monte Carlo dose calculation method based on statistical uncertainty perception denoising of the present invention has the following advantages and beneficial effects: Explicit fusion of physical information: By introducing statistical uncertainty (SU) as an independent channel into the network, the model can adaptively denoise based on the physical distribution characteristics of noise, which significantly reduces mean dose error (MDE) compared to a dual-channel model that only uses dose and CT.

[0038] Dual-path feature fusion: The dual-path architecture combines standard convolution and dilated convolution, which simultaneously ensures the fine preservation of local anatomical structures and the smoothness of global dose distribution.

[0039] High efficiency and accuracy coexist: It can shorten the high particle number simulation that takes tens of minutes to "second-level low particle number simulation + second-level denoising", significantly reducing the total time. Moreover, the dose distribution Gamma pass rate (3% / 3mm) after denoising can reach more than 99%, meeting the clinical accuracy requirements.

[0040] In another implementation of the present invention, the voxel-level statistical uncertainty distribution map is defined as the relative standard deviation at the voxel level, which is calculated based on the batch dose variance in the Monte Carlo simulation and characterizes the random noise level of the dose distribution under the current simulated particle number.

[0041] For example, the MC engine records the dose deposition of each voxel in different batches during the simulation, calculates its standard deviation σ and mean μ, and then the statistical uncertainty at voxel (i,j,k) is defined as the relative standard deviation: SU i,j,k = σ i,j,k / D i,j,k Therefore, the SU plot can accurately reflect the spatial distribution characteristics of noise. For example, the SU value is usually higher at the interface between low-density tissue (such as lungs) and high-density tissue (such as bone), or in deep regions with fewer particles.

[0042] In another implementation of the present invention, the anatomical structure image is a CT image; before the three-channel input data is input into the network model, the intermediate coarse dose distribution and the voxel-level statistical uncertainty distribution are subjected to 3D normalization processing based on the patient planning volume.

[0043] In another implementation of the present invention, the statistical uncertainty-aware denoising network model is constructed based on a batch renormalization denoising network architecture, which includes two parallel feature extraction branches. The first branch consists of multiple layers of "convolution + batch normalization + linear rectifier units" and is used to extract local anatomical and dose texture features of the image. The second branch consists of multiple layers of "dilated convolution + linear rectifier units" and expands the receptive field to capture long-distance dose-uncertainty correlation and global contextual information. The features extracted by the two branches are concatenated and fused at the end of the network, and the denoised dose residual or direct dose distribution is reconstructed and output through convolutional layers.

[0044] For example, such as Figure 2 As shown, an improved Batch Renormalization Denoising Network (BRDNet) can be used as the core architecture.

[0045] Network architecture design: The network is designed as a dual-path structure to balance the extraction of local details and the expansion of the global receptive field.

[0046] Upper Path (Local Feature Extraction): This path consists of a series of stacked convolutional layers (Conv), batch renormalization layers (BRN), and activation functions (ReLU). The convolutional kernel size is set to 3×3 to extract tissue edge features and local texture of dose in CT images. Compared to traditional batch normalization, batch renormalization provides more stable gradient estimation during mini-batch training.

[0047] The next path (global context-aware): This path mainly consists of "dilated convolutional layers + ReLU activation functions". The dilation rate of the dilated convolution is set sequentially to r=2, 4, 8, ..., exponentially expanding the receptive field while maintaining the feature map resolution. This path utilizes the noise distribution prior provided by the SU map to identify and smooth high-noise regions globally, while preserving details in low-noise regions.

[0048] Feature fusion: The feature maps of the upper and lower paths are concatenated along the channel dimension, then fused through a convolutional layer, and finally output as a residual map.

[0049] Final output: The residual map predicted by the network and the rough dose D of the input. noisy Add them together to obtain the high-precision dose D after noise reduction. pred .

[0050] In another implementation of the present invention, the training process of the network model includes: constructing a training dataset containing different tumor sites and different simulated particle number gradients, using the low-uncertainty dose distribution generated by high particle number simulation as the gold standard, and using mean squared error as the loss function; and using the Adam optimizer to iteratively optimize the network parameters until the model converges.

[0051] For example, the model training strategy includes: Dataset construction: Taking the collection of 69 clinical IMRT plans as an example (including 39 cases of head and neck cancer, 15 cases of lung cancer and 15 cases of brain cancer).

[0052] Gold Standard (Ground Truth): Using 5 × 10 9 Taking a high-precision MC simulated dose of a certain number of particles as an example for training labels.

[0053] Loss function: Minimize the difference between the predicted dose and the gold standard using Mean Squared Error (MSE). L=

[0054] Optimizer: The Adam optimizer is used with an initial learning rate of 0.001, and a gradual decay strategy is adopted as the training loss converges.

[0055] Cross-validation: To ensure the generalization ability of the model, a 5-fold cross-validation strategy is adopted, in which approximately 75% of the cases are randomly selected each time for training and validation, and the remaining 25% are used for testing.

[0056] In another implementation of the invention, the different simulated particle number gradients include those from 1 × 10⁻⁶. 5 Up to 5×10 7 Multiple orders of magnitudes are used to train corresponding sub-models or a unified model for each noise level, in order to adapt to the denoising needs under different computation time constraints in clinical practice.

[0057] Example 1 To verify the effectiveness of the method of this invention, extensive quantitative and qualitative experiments were conducted. Fifteen independent test cases (covering three sites: head and neck, lungs, and brain) were selected. For each test case, six different simulated particle number levels were set as inputs: , , , , , For each level, a corresponding sub-model is trained. The model evaluation results are based on the mean dose error (MDE), gamma pass rate (3% / 3mm standard), and dose volume histogram (DVH) differences between the planned dose distribution reconstructed by the network and the gold standard dose distribution.

[0058] like Figure 3 As shown, the simulated particle number is Taking this intermediate noise level as an example: Head and Neck (H&N): Mean Dose Error (MDE) before denoising was After processing by the model of this invention, the value is reduced to .

[0059] Brain: MDE from Reduce to .

[0060] Lung: MDE from Reduce to .

[0061] Under the 3% / 3mm standard, when the number of input particles reaches At and above these levels, the Gamma pass rate of the method of the present invention exceeds 97%, wherein... The pass rate for the input was 99.86%.

[0062] Example 2 To verify the necessity of introducing a statistical uncertainty (SU) channel, the three-channel model (SU-aware) of this invention was compared with a two-channel model (SU-agnostic) that only includes CT and coarse dose. Experimental results show... Figure 4 As shown, in all test cases, the MDE of the three-channel model was significantly lower than that of the two-channel model (P<0.05). Specifically, for H&N: three-channel 0.84% ​​vs two-channel 1.12%; for Brain: three-channel 0.85% vs two-channel 1.21%; and for Lung: three-channel 0.36% vs two-channel 0.53%. The results fully demonstrate that explicitly introducing SU information enables the network to more intelligently identify noise distribution, thereby smoothing noise more effectively while protecting the edges.

[0063] Meanwhile, this invention achieves high-precision dose calculation at the "second" level, with a total process time of only about 4-5 seconds. Among these, based on... The low-particle-count GPU simulation took approximately 4.06 seconds, while deep learning denoising network inference took only milliseconds (almost negligible). This is in contrast to the approximately 10 minutes (5 × 10⁻⁶) required to achieve equivalent accuracy. 9 Compared with traditional Monte Carlo simulations of particles, this invention improves computational efficiency by two orders of magnitude, successfully meeting the needs of time-critical clinical applications such as online adaptive radiotherapy.

[0064] In summary, this invention provides a feasible technical approach to effectively address the long-standing challenge of balancing accuracy and efficiency in Monte Carlo dose calculation by constructing a deep learning framework based on statistical uncertainty perception and dual-path feature fusion.

[0065] Another aspect of the present invention provides an application of a fast Monte Carlo dose calculation method based on statistical uncertainty-aware noise reduction, comprising: By employing any of the methods described above, a high-precision Monte Carlo radiotherapy dose distribution can be obtained for inverse optimization of radiotherapy plans, dose assessment, or quality control.

[0066] Another aspect of the present invention provides a fast Monte Carlo dose calculation system based on statistical uncertainty-aware noise reduction, comprising: Data acquisition module: By performing Monte Carlo dose simulations with low particle numbers, an intermediate coarse dose distribution containing random noise and the corresponding voxel-level statistical uncertainty distribution are obtained.

[0067] Data preprocessing module: Acquires anatomical images of the patient, normalizes and registers the anatomical images, the intermediate coarse dose distribution, and the corresponding voxel-level statistical uncertainty distribution, and constructs three-channel input data.

[0068] Denoising network module: The three-channel input data is input into a pre-built and trained statistical uncertainty-aware denoising network model. The network model adopts a dual-path dilated convolutional fusion architecture to extract local features and global context features respectively and then fuse them.

[0069] Output module: The network model outputs a denoised, high-precision dose distribution as the final Monte Carlo dose calculation result.

[0070] The fast Monte Carlo dose calculation system based on statistical uncertainty perception denoising of the present invention has the following advantages and beneficial effects: Explicit fusion of physical information: By introducing statistical uncertainty (SU) as an independent channel into the network, the model can adaptively denoise based on the physical distribution characteristics of noise, which significantly reduces mean dose error (MDE) compared to a dual-channel model that only uses dose and CT.

[0071] Dual-path feature fusion: The dual-path architecture combines standard convolution and dilated convolution, which simultaneously ensures the fine preservation of local anatomical structures and the smoothness of global dose distribution.

[0072] High efficiency and accuracy coexist: It can shorten the high particle number simulation that takes tens of minutes to "second-level low particle number simulation + second-level denoising", significantly reducing the total time. Moreover, the dose distribution Gamma pass rate (3% / 3mm) after denoising can reach more than 99%, meeting the clinical accuracy requirements.

[0073] In another aspect of the present invention, the electronic device includes: a processor, a memory, and a communication bus and a communication interface.

[0074] in: The processor, memory, and communication interface communicate with each other via a communication bus.

[0075] A communication interface is used to communicate with other electronic devices or servers.

[0076] The processor is used to execute programs, specifically, to perform any of the steps of the fast Monte Carlo dose calculation method based on statistical uncertainty-aware noise reduction in the above embodiments.

[0077] Specifically, the program may include program code, which includes computer operation instructions.

[0078] The processor may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. The one or more processors included in the smart device may be processors of the same type, such as one or more CPUs; or they may be processors of different types, such as one or more CPUs and one or more ASICs.

[0079] Memory is used to store programs. Memory may include high-speed RAM, and may also include non-volatile memory, such as at least one disk storage device.

[0080] Specifically, the program can be used to cause the processor to execute the steps of any of the fast Monte Carlo dose calculation methods based on statistical uncertainty-aware denoising described in the embodiments. The specific implementation of each step in the program can be found in the corresponding descriptions of the steps and units executed in any of the fast Monte Carlo dose calculation methods based on statistical uncertainty-aware denoising described above, and will not be repeated here. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the devices and modules described above can be referred to the corresponding process descriptions in the foregoing method embodiments.

[0081] An exemplary embodiment of this application also provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to perform the methods of various embodiments of this application.

[0082] The methods described above according to embodiments of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0083] Specific embodiments of the present invention have now been described. Other embodiments are within the scope of the appended claims. In some cases, the actions described in the claims can be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result.

[0084] It should be noted that all directional indications (such as up, down, left, right, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship between the components in a certain order (as shown in the figure). If the specific order changes, the directional indication will also change accordingly.

[0085] In the description of this invention, the terms "first" and "second" are used only for convenience in describing different components or names, and should not be construed as indicating or implying a sequential relationship, relative importance, or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" and "second" may explicitly or implicitly include at least one of that feature.

[0086] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0087] It should be noted that although specific embodiments of the present invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of the present invention. Various modifications and variations that can be made by those skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of the present invention.

[0088] The examples of the embodiments of the present invention are intended to concisely illustrate the technical features of the embodiments of the present invention, so that those skilled in the art can intuitively understand the technical features of the embodiments of the present invention, and are not intended to be an improper limitation of the embodiments of the present invention.

[0089] 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 rapid Monte Carlo dose calculation method based on statistical uncertainty-aware denoising, characterized in that, include: By performing Monte Carlo dose simulations with low particle numbers, an intermediate coarse dose distribution containing random noise and the corresponding voxel-level statistical uncertainty distribution were obtained. The anatomical structure image of the patient is acquired, and the anatomical structure image, the intermediate coarse dose distribution and the corresponding voxel-level statistical uncertainty distribution are normalized and registered to construct three-channel input data. The three-channel input data is input into a pre-built and trained statistical uncertainty-aware denoising network model. The network model adopts a dual-path dilated convolutional fusion architecture to extract local features and global context features respectively and then fuse them. The network model outputs a denoised, high-precision dose distribution as the final Monte Carlo dose calculation result.

2. The method according to claim 1, characterized in that, The voxel-level statistical uncertainty distribution is defined as the relative standard deviation at the voxel level. It is calculated based on the batch dose variance in the Monte Carlo simulation and characterizes the random noise level of the dose distribution under the current simulated particle number.

3. The method according to claim 2, characterized in that, The anatomical images are CT images; Before being input into the network model, the three-channel input data undergoes 3D normalization processing on the intermediate coarse dose distribution and the voxel-level statistical uncertainty distribution based on the patient planning volume.

4. The method according to claim 1, characterized in that, The statistical uncertainty-aware denoising network model is built on a batch renormalization denoising network architecture and includes two parallel feature extraction branches. The first branch consists of multiple layers of "convolution + batch normalization + linear rectified units", which are used to extract local anatomical and dose texture features of the image; The second branch consists of multiple layers of "dilated convolution + linear rectifier units", which expands the receptive field to capture long-range dose-uncertainty correlations and global contextual information; The features extracted from the two branches are spliced ​​and fused at the end of the network, and the output denoised dose residuals or direct dose distribution are reconstructed through convolutional layers.

5. The method according to claim 4, characterized in that, The training process of the network model includes: We constructed a training dataset containing different tumor sites and different simulated particle number gradients, using the low-uncertainty dose distribution generated by high particle number simulation as the gold standard, and the mean squared error as the loss function. The Adam optimizer is used to iteratively optimize the network parameters until the model converges.

6. The method according to claim 5, characterized in that, The different simulated particle number gradients include those from 1×10⁻⁶. 5 Up to 5×10 7 Multiple orders of magnitudes are used to train corresponding sub-models or a unified model for each noise level, in order to adapt to the denoising needs under different computation time constraints in clinical practice.

7. An application of a fast Monte Carlo dose calculation method based on statistical uncertainty perception denoising, characterized in that, include: The method described in any one of claims 1 to 6 can be used to obtain a high-precision Monte Carlo radiotherapy dose distribution for inverse optimization of radiotherapy planning, dose assessment, or quality control.

8. A fast Monte Carlo dose calculation system based on statistical uncertainty-aware denoising, characterized in that, include: Data acquisition module: By performing Monte Carlo dose simulations with low particle numbers, the intermediate coarse dose distribution containing random noise and the corresponding voxel-level statistical uncertainty distribution are obtained; Data preprocessing module: acquires anatomical images of the patient, normalizes and registers the anatomical images, the intermediate coarse dose distribution, and the corresponding voxel-level statistical uncertainty distribution, and constructs three-channel input data; Denoising network module: The three-channel input data is input into a pre-built and trained statistical uncertainty-aware denoising network model. The network model adopts a dual-path dilated convolutional fusion architecture to extract local features and global context features respectively and fuse them. Output module: The network model outputs a denoised, high-precision dose distribution as the final Monte Carlo dose calculation result.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the fast Monte Carlo dose calculation method based on statistical uncertainty-aware noise reduction as described in any one of claims 1 to 6.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the steps in the rapid Monte Carlo dose calculation method based on statistical uncertainty-aware noise reduction as described in any one of claims 1 to 6.

Images