A Monte Carlo dose super-resolution denoising method for boron neutron capture therapy

Abstract

The present invention discloses a Monte Carlo dose super-resolution denoising method for boron neutron capture therapy, which recovers a noise-free high-resolution (HR) image from a noisy low-resolution (LR) image. The relationship between the HR image and its LR can be achieved through the final training output of a CNN-based network, by combining denoising and super-resolution to achieve the effect of real-time MC calculation. This scheme combines the high accuracy of the MC method with the advantages of accurate and fast prediction of the CNN, solving the time-consuming problem of the traditional MC method. At the same time, a physical energy conservation constraint is introduced in network training to solve the problem of non-physical solutions that may be caused by the lack of physical processes in pure data-driven DL methods, reducing measurement errors.

Description

Technical Field

[0001] The present invention relates to the technical field of radiotherapy planning systems, and specifically to a Monte Carlo dose super-resolution denoising method for boron neutron capture therapy. Background Art

[0002] Boron neutron capture therapy (BNCT) has shown important potential in the treatment of cancers such as head and neck tumors, gliomas, and malignant melanomas, where the treatment effects of traditional radiotherapy methods (such as photon radiotherapy) are poor, due to its strong targeting ability and small damage to normal tissues. In recent years, BNCT has developed rapidly in the world, and by the end of 2024, there are nearly 20 BNCT projects under construction in China. The treatment effect of BNCT directly depends on accurate dose calculation. The Monte Carlo (MC) simulation method is considered the "gold standard" for dose calculation. This method can obtain a high-precision dose distribution by simulating the transport and various physical processes of neutrons and secondary photons in the patient's body, and it is currently the only method that can calculate the dose distribution in a three-dimensional inhomogeneous medium. However, since MC simulation usually needs to simulate billions of particle events to achieve clinically acceptable accuracy (such as an uncertainty of less than 3%), this process requires several hours or even days of computing time, making it difficult to meet the scenarios in clinical practice that require real-time or near-real-time treatment planning, such as online adaptive treatment.

[0003] Although some studies have used deep learning (DL) techniques based on convolutional neural networks (CNNs) to suppress the Monte Carlo dose noise of photons, no one has applied this technology to boron neutron capture therapy with more complex physical processes. At the same time, in image processing, traditional methods focus on separately processing denoising or super-resolution tasks. Pure deep learning methods are prone to generating large dose deviations in key regions (such as the skin and inside the tumor) without considering the constraints of physical processes (such as basic physical laws like energy conservation), thus affecting the treatment safety and effect, and their results for clinical dose calculation in BNCT lack interpretability.

[0004] Most existing BNCT treatment planning systems (TPS) externally connect general Monte Carlo programs such as PHITS and MCNP for dose calculation. Usually, this method divides the calculation area into high-resolution grids and uses a large number of particle event simulations to reduce statistical errors and achieve accurate dose assessment of each organ (including tumors, skin, and other key structures). Although the calculation results of this method are accurate, due to the need to simulate hundreds of millions of particle events, its calculation time is long and it has high requirements for hardware resources, making it difficult to meet the clinical demand for rapid treatment plan generation.

[0005] Disadvantages of the prior art one:

[0006] 1. Low calculation efficiency:

[0007] Traditional Monte Carlo programs (such as PHITS and MCNP) are based on CPU architecture, and a single dose calculation takes 6-8 hours (Sumitomo Heavy Industries NeuCure BNCT system in Japan), which cannot meet the real-time requirements of treatment scenarios such as online adaptive radiotherapy (dynamic adjustment of plans during treatment).

[0008] 2. Poor compatibility between hardware and algorithms:

[0009] General Monte Carlo programs are not optimized for GPU parallel architecture and can generally only run on CPUs. The thread allocation and task batching strategies are extensive, and a large number of computing threads are required to achieve fast calculations, which is not friendly to medical institutions without high-performance computing servers, limiting the clinical use of Monte Carlo software.

[0010] The second existing technology proposes to use a pure deep learning method to achieve BNCT dose prediction. This solution constructs a deep learning model based on 3D U-Net, takes the patient's CT image as input, and directly outputs the three-dimensional dose distribution in the patient's body. By using the nonlinear mapping ability of convolutional neural networks, this method can complete dose calculation in a short time.

[0011] 1.3.2 Disadvantages of the Second Existing Technology

[0012] 1. Singleness of task:

[0013] Designed only for denoising tasks, it cannot simultaneously improve spatial resolution and requires additional interpolation steps, resulting in loss of high-frequency details; it cannot further improve the calculation speed

[0014] 2. Lack of physical constraints:

[0015] The purely data-driven mean squared error (MSE) loss function ignores physical laws such as energy conservation, and its prediction results often have large deviations in key areas (especially the skin and inside tumors). In addition, the results obtained by pure deep learning predictions are not clinically interpretable enough.

[0016] 3. Poor adaptability of BNCT:

[0017] The input only contains a single vector diagram, and does not integrate the multi-component doses of multiple secondary particles and patient anatomical information (such as HU value, binary mask of body contour, binary mask of skin area) involved in the BNCT process, making it difficult to model the complex dose superposition effect in BNCT. When encountering special cases that are not fully covered in the training data, unstable prediction results may occur, thus restricting its wide application in clinical practice. Summary of the invention

[0018] The object of the present invention is to provide a Monte Carlo dose super-resolution denoising method for boron neutron capture therapy, so as to solve the problems raised in the above-mentioned background art.

[0019] To achieve the above object, the present invention provides the following technical solutions:

[0020] A Monte Carlo dose super-resolution denoising method for boron neutron capture therapy includes the following steps:

[0021] Step 1: Obtain the CT, contour drawing, and radiotherapy plan data of a real patient, and use general MC software to calculate the dose;

[0022] Step 2: Divide the obtained data into data sets and perform preliminary processing such as cropping and normalization on the data;

[0023] Step 3: Construct a super-resolution denoising convolutional neural network model; train the traditional super-resolution convolutional neural network model SRCNN and the super-resolution denoising convolutional neural network model PDA-SRDNet constructed by the patent under the same training conditions to obtain two trained convolutional neural network models respectively;

[0024] Step 4: Input the test set data into the two trained convolutional neural networks respectively to obtain the results after super-resolution denoising, and perform multi-dimensional comparison;

[0025] Step 5: Perform post-processing operations on the predicted vector map to generate the final vector map; upload the final vector map to the radiotherapy planning system TPS for doctors and physicists to evaluate.

[0026] As a further technical solution of the present invention: in Step 1, use the general MC program TOPAS, set the parameters required for dose calculation, and calculate B-10 (boron dose, D B ), N-14 and H-1 (nitrogen dose and hydrogen dose, D N , D H ) and several absorbed doses and photon doses (D γ ) caused by the reaction with neutrons in the same patient human model; a total of 4 partial doses (D B , D H , D N , D γ ) are simulated for each patient. Each component dose simulates a low-resolution high-noise image and a high-resolution low-noise image respectively, which are used as training data and ground truth respectively.

[0027] As a further technical solution of the present invention: In step 2, first divide the data set into a training data set and a validation data set. In order to obtain sufficient and diverse training data, perform operations such as rotating each vector graph by 90°, 180°, 270° and flipping it up and down, or performing Gamma mapping, which not only effectively expands the scale of the data set, but also improves the robustness of the network to spatial transformation and noise distribution changes. After data augmentation, we perform cropping, edge padding, and data normalization on the augmented data, and divide it uniformly by the maximum value of the low-resolution and high-noise vector graph.

[0028] As a further technical solution of the present invention: In step 3, the MC super-resolution denoising network selects the improved lightweight convolutional neural network PDA-SRDNet mentioned in this patent, uses four-channel input, and the input includes the three-dimensional matrix of the CT value of the patient, the binary mask of the body contour, the binary mask of the skin area, and the high-noise coarse grid energy deposition matrix calculated by MC; initialize the parameters of the traditional super-resolution network SCRNN and PDA-SRDNet. Use the binary mask of the body contour, the binary mask of the skin area, and the high-noise coarse grid energy deposition matrix calculated by MC in the training set to train the super-resolution denoising network.

[0029] As a further technical solution of the present invention: In step 4, the comparison of the two convolutional neural networks further includes: performing the same simple processing on the denoising results of the two trained convolutional neural networks obtained in step 3. After extracting the data, compare the high-noise and low-resolution images obtained from the low-sampled data, the low-noise and high-resolution clean images obtained from the high-sampled data, the denoised images of the SRCNN convolutional neural network, and the denoised images of the PDA-SRDNet denoising convolutional neural network, and finally output the comparison results of the model performance and calculation efficiency of the two convolutional neural networks.

[0030] As a further technical solution of the present invention: In step 5, the post-processing operation includes performing operations to exclude abnormal prediction regions and smooth the edges on the predicted contour. The operations of removing holes, excluding abnormal prediction regions, and smoothing the edges specifically include the following steps: using the flood-fill algorithm to remove abnormal hole information inside the boundary of the contour; finding the largest connected region containing the prediction information of the contour to remove other abnormal prediction regions; using the built-in library of OpenCV to perform edge smoothing on the contour. And input the processed denoised image into a commercial TPS for physicists to evaluate.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] (1) The image processing acceleration method effectively combines super-resolution and denoising, and proposes a dual-task joint optimization framework for accelerating Monte Carlo dose calculation. By combining the high accuracy of the MC method and the advantages of accurate and fast prediction of CNN, the time-consuming problem of the traditional MC method is solved;

[0033] (2) Strict physical constraints (energy conservation constraints) are introduced in network training, reducing dose errors and fundamentally solving the untrustworthy problem caused by the lack of physical processes in pure DL methods, which is of great significance for introducing neural network Monte Carlo denoising into clinical processes.

[0034] (3) Further, a physically constrained super-resolution denoising network composed of a spatial and channel integrated attention module, convolutional layers, pixel shuffling layers, and a bilinear residual mechanism is used. On the premise of ensuring denoising performance, a large amount of computational overhead is reduced to reduce the hardware burden, and the dose distribution of patients can be quickly recalculated in a short time, providing instant feedback for adaptive radiotherapy planning and being suitable for deployment in a clinical real-time computing environment.

[0035] (4) The present invention is not only applicable to dose calculation in BNCT treatment planning, but also its basic principle - that is, a statistical denoising method that combines MC simulation and a super-resolution denoising network - can be extended to other MC radiation transport simulation fields, such as photon treatment planning, proton and heavy ion treatment planning, CT dose assessment, radiation shielding design, and nuclear reactor physics calculations. This method follows a unified statistical and denoising principle and has broad cross-field application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 is the implementation flowchart established in the embodiment of the present invention;

[0037] Figure 2 is the schematic diagram of network training in the embodiment of the present invention;

[0038] Figure 3 is the schematic diagram of the PDA-SRDNet network in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0040] Please refer to Figures 1-3, A Monte Carlo dose super-resolution denoising method for boron neutron capture therapy, which can improve the spatial resolution of the obtained images and well remove the inherent noise of the Monte Carlo algorithm. We call this method the "super-resolution denoising network in Monte Carlo denoising research", and combine physical constraints to greatly accelerate the BNCT MC dose calculation while maintaining accuracy, which is of great significance for the application of this method to the clinic.

[0041] Step 1: Multi-component dose calculation of BNCT based on the general MC software TOPAS:

[0042] (1) Collect clinical data. Divide the data into data sets:

[0043] (1.1) Using the general Monte Carlo code TOPAS, set the B-10 concentrations in tumors, normal tissues, and skin to 60, 20, and 25 ppm respectively, and there is no B-10 in bone tissue;

[0044] (1.1.1) Calculate the several absorbed doses caused by the reactions of B-10 (boron dose, D B ), N-14, and H-1 (nitrogen dose and hydrogen dose, D H , D N ) with neutrons in the same patient phantom respectively;

[0045] (1.1.2) In addition, consider that the main gamma-ray dose from the neutron source and the secondary gamma-ray dose generated in the human body by the reaction with neutrons and hydrogen constitute the photon dose (D γ );

[0046] (1.2) Since each absorbed dose component causes different biological effects, the composite biological effectiveness factor (CBE) and relative biological effectiveness factor (RBE) are used, so a total of 4 parts of doses (D B , D H , D N , D γ ) are simulated for each patient.

[0047] (1.2.1) According to the latest publication of the International Atomic Energy Agency (IAEA), the "photon equivalent dose" (DisoE) that produces the same biological effect as photons can be calculated using Equation (1), which is expressed as the weighted sum of the main absorbed dose components:

[0048]

[0049] Where: represents the composite biological effectiveness factor, represents the B-10 vector map, Represents the relative biological effectiveness factor of H, represents the H-1 vector diagram; represents the relative biological effectiveness factor of N, represents the N-14 vector diagram represents the relative biological effectiveness factor of Gamma photons; represents the photon vector diagram.

[0050] (1.3) For training purposes, low-sampling-number noise Monte Carlo simulations at low resolution are performed for each patient case in the training dataset, with the voxel set to (32×32×24, 9.76×9.76×7.81 mm 3 ) and 5×10 6 neutron event numbers per simulation as the training data. For verification purposes, independent high-sampling-number "clean" Monte Carlo simulations at high resolution are performed for each patient case, with the voxel set to (128×128×96, 2.44×2.44×1.95 mm 3 ) and 1×10 9 neutron event numbers per simulation as the ground truth vector diagram.

[0051] (1.4) Predicting high resolution from low resolution is because: Although the spatial resolution is reduced, it improves the statistical efficiency. In other words, fewer particles are required for the result to converge faster. This is because a larger sampling cross-section means that more particle tracks are scored in each grid, thus reducing the statistical uncertainty. At the same time, predicting a 1×10 6 vector diagram from a 5×10 9 vector diagram is a process of image denoising.

[0052] (1.5) The purpose of using a CNN-based method to solve the denoising and super-resolution problems is to establish a mapping F from the LR image to its HR reconstructed image, and this mapping F makes the prediction result as similar as possible to the ground truth HR image. Let's represent the LR image as Y. Our goal is to recover the image F(Y) from Y, as shown in Equation (2). In our proposed method, the goal of PDA-SRDNet is to recover the mapping F of the image F(Y) from Y. The image denoising problem and the SR problem are similar because both problems need to process the high-frequency part while retaining most other information. Although the noise in the obtained Monte Carlo vector diagram is caused by uncertainty, it can still be regarded as a kind of Gaussian noise. Therefore, only one CNN-based deep network is used to solve these two problems. This means that there is only one mapping F in our method.

[0053]

[0054] (2) We crop the vector map obtained in step (1) with the body contour as the boundary to eliminate irrelevant regions, such as air, and at the same time perform edge filling on the effective region (filling 0 to the image edge in each dimension).

[0055] (2.1) To obtain sufficient and diverse training data, each vector map is rotated by 90°, 180°, 270° and flipped vertically, etc., or Gamma mapping is performed. This not only effectively expands the scale of the dataset but also improves the network's robustness to spatial transformation and noise distribution changes.

[0056] (2.2) To stabilize the training process, the Monte Carlo dose calculation results are processed. The doses of all patients are normalized by dividing by the maximum value of the low-resolution and high-noise vector map and linearly normalized to between [0, 1]. Our calculation method is as shown in Equation (3):

[0057]

[0058] (3) Construct and train a super-resolution neural network model: As Figure 1 shown, the data preprocessed in steps 1 - 2 is input into a neural network using the SRCNN structure (or the improved lightweight convolutional network we proposed, such as PDA - SRDNet) to achieve the mapping from low-particle-number and low-resolution dose to high-particle-number and high-resolution dose.

[0059] (3.1) As Figure 2 shown, the model is constructed using an architecture with multi-channel input. The input end includes: a three-dimensional matrix of the patient's HU value; a binary mask of the body contour; a binary mask of the skin area; a low-particle and coarse-grid energy deposition matrix calculated by MC;

[0060] (3.2) For the SRCNN network, simple convolutional layer connections are used. SRCNN first uses bicubic interpolation to enlarge the low-resolution image to the target size, and then extracts the features of the image patches through a three-layer convolutional network. The kernel sizes of the convolutional layers are 9×9, 1×1, and 5×5 respectively; then the features are non-linearly mapped, and finally the high-resolution image result is reconstructed.

[0061] (3.2.1) The SRCNN network uses MSE as the loss function and uses the Adam learning rate algorithm for training according to the supervised learning algorithm. The formula is as shown in Equation (4)

[0062]

[0063] Where: represents the network prediction result, Denote the low-noise high-resolution vector map, and \(F\) denote the mapping from the high-noise low-resolution vector map to the high-noise high-resolution vector map.

[0064] (3.3) For the PDA-SRDNet network, as Figure 3 shown, it consists of a spatial and channel integration attention module, convolutional layers, pixel shuffle layers, and a bilinear mechanism. The network contains 6 convolutional layers with dual attention modules, and combines 1 global residual learning (GRL) and 6 local residual learning (LRL).

[0065] (3.3.1) The core component of the network is the lightweight dual attention module (Double Attention), which integrates spatial and efficient bidirectional channel attention mechanisms. The spatial attention block captures the pixel-by-pixel relationships within the spatial dimension of the feature map, while the bidirectional channel attention module effectively collects the inter-channel attention information from these feature maps, taking into account both forward and backward dependencies.

[0066] (3.3.2) Residual learning means that the input and output are largely similar. Therefore, in our method, the residual is defined as where most values may be zero or very small. In PDA-SRDNet, \(r\) is the residual pixel of \(x\) and \(y\). Therefore, we can make the training of the entire network easier and can further make the entire network deeper than normal. And both types of such residuals, namely global residual learning and local residual learning, are used in this network. Such a structure can reduce a large amount of computational overhead while maintaining accuracy to ensure the balance between denoising accuracy and computational efficiency.

[0067] (3.3.3) The PDA-SRDNet network not only uses the MSE function in Equation (4), but also incorporates strict physical constraints into the loss function, introducing the physical energy conservation constraint, that is, requiring the energy deposition within each low-resolution unit to be consistent with the sum of the energies of its corresponding \(N\) high-resolution units, so as to ensure that the network prediction satisfies the physical laws. The loss function used is shown in Equation (5). And the Adam learning rate algorithm is used for training according to the supervised learning algorithm.

[0068]

[0069] Where: represents the network prediction result, represents the high-noise low-resolution vector map, and \(N\) represents the number of high-noise high-resolution vector maps corresponding to each voxel of the high-noise low-resolution vector map.

[0070] (4)Super-resolution denoising prediction and dose distribution reconstruction: After training, the high-noise low-resolution dose distribution is input into the trained neural network, and the network will quickly output the low-noise fine-grid dose distribution, realizing dose reconstruction from coarse to fine and from high noise to low noise.

[0071] (4.1)And the output results of different components are weighted according to Equation (1) to obtain the final predicted total dose DisoE.

[0072] (4.2)Evaluation metrics include: using dose deviation (D_dev) to statistically analyze the prediction errors of each voxel;

[0073] (4.2.1)Using DVH to visually compare the dose distributions of key organs (such as tumors, skin, and other radiation-sensitive tissues);

[0074] (4.2.2)Calculating the average root mean square error RMSE, GIPR, to quantify the matching degree of the predicted dose and the MC simulation results in terms of space and dose values.

[0075] (4.2.3)GIPR is based on two other dose distribution comparison tools: dose difference (DD) and distance to agreement (DTA). Dose deviation uses the dose difference at corresponding points of the reference image and the test image as a measure of similarity. If the dose difference at a certain point is less than the user-specified value (e.g., 3% of the maximum reference dose), then that point passes the test. Dose deviation is more suitable for comparing regions with small dose gradients, but its reliability is not high in regions with sharp dose changes. The distance to agreement measures similarity based on the closest distance between points in the reference image and points with the same dose in the test image. If the distance to agreement is less than the user-specified value (e.g., 3 mm), then that point passes the test. In contrast to dose deviation, the distance to agreement is very efficient in regions with large dose gradients and less efficient in regions with gentle doses. The gamma index combines the advantages of both, and its core calculation method is:

[0076]

[0077] In the formula, and are the coordinates of points in the reference image and the test image respectively, is the gamma index of point , is point and point the distance between them. is point and point the dose difference between them. is the threshold of dose deviation, is the threshold of the coincidence distance. If ≤ 1, the point passes the gamma test. When actually calculating the gamma passing rate, low-dose regions without clinical significance are usually ignored, and the commonly used threshold is the maximum dose of 10% of the reference image.

[0078] (5) Perform post-processing operations on the predicted vector map to obtain the final result. The post-processing operations include removing holes, excluding abnormal prediction regions, and edge smoothing, etc., to obtain the final dose prediction result and upload it to the radiotherapy planning system TPS for doctors and physicists to evaluate.

[0079] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, in any regard, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be encompassed within the present invention. Any reference signs in the claims should not be construed as limiting the claims involved.

[0080] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A Monte Carlo dose super-resolution denoising method for boron neutron capture therapy, characterized in that It includes the following steps: Step 1: Obtain the CT, contouring, and radiotherapy plan data of real patients, and perform dose calculation using general MC software. Specifically: Using the general Monte Carlo code TOPAS, set the B-10 concentrations in tumors, normal tissues, and skin to 60, 20, and 25 ppm, respectively; Calculate the absorbed doses caused by the reactions of B-10, N-14, and H-1 with neutrons in the same patient phantom; The photon dose D is composed of the primary gamma-ray dose from the neutron source and the secondary gamma-ray dose generated in the human body by the reaction of neutrons and H-1 γ ; A total of 4 partial doses D were simulated for each patient B , D H , D N , D γ ; Use Equation (1) to calculate the photon isodose that produces an equivalent biological effect to photons, which is expressed as a weighted sum of absorbed dose components: ； Wherein: represents the composite biological effectiveness factor, represents the B-10 dose, represents the relative biological effectiveness factor of H-1, represents the H-1 dose; represents the relative biological effectiveness factor of N-14, represents the N-14 dose, represents the relative biological effectiveness factor of Gamma photons; represents the photon dose; Perform low-sampling-number noise Monte Carlo simulations at low resolution for each patient case in the training dataset, with the voxel set to 32×32×24, 9.76×9.76×7.81mm 3 , with 5×10 6 neutron event numbers for each simulation, as training data. For validation purposes, perform independent high-sampling-number Monte Carlo simulations at high resolution for each patient case, with the voxel set to 128×128×96, 2.44×2.44×1.95mm 3 , with 1×10 9 neutron event numbers for each simulation, as the ground truth vector map. Predicting the vector map with 1×10 6 neutron event numbers from the vector map with 5×10 9 neutron event numbers is an image denoising process; Step 2: Divide the obtained data into datasets and perform operations such as data cropping, edge padding, and data normalization; Step 3: Construct a super-resolution denoising convolutional neural network model; train the super-resolution convolutional neural network model SRCNN and the constructed super-resolution denoising convolutional neural network model PDA-SRDNet under the same training conditions to obtain two trained convolutional neural network models. The lightweight convolutional neural network PDA-SRDNet uses a four-channel input, including the three-dimensional matrix of the patient's CT values, the binary mask of the body contour, the binary mask of the skin area, and the high-noise coarse-grid energy deposition matrix calculated by MC; Step 4: Input the test set data into the two trained convolutional neural networks respectively to obtain the super-resolution denoised results, and perform multi-dimensional comparisons on the results; Step 5: Perform post-processing operations on the predicted vector map to generate the final vector map; upload the final vector map to the radiotherapy planning system TPS for doctors and physicists to evaluate.

2. A Monte Carlo dose super-resolution denoising method for boron neutron capture therapy according to claim 1, wherein In Step 2, first divide the dataset into a training dataset and a validation dataset, perform operations such as rotating each vector map by 90°, 180°, 270°, and flipping it vertically, or perform Gamma mapping to achieve data augmentation, and then perform data cropping, edge padding, and data normalization on the augmented data, and divide by the maximum value of the low-resolution high-noise vector map uniformly.

3. A Monte Carlo dose super-resolution denoising method for boron neutron capture therapy according to claim 1, characterized in that, In Step 3, the MC super-resolution denoising network selects the improved lightweight convolutional neural network PDA-SRDNet, initialize the parameters of the super-resolution networks SCRNN and PDA-SRDNet, and train the super-resolution denoising network using the binary mask of the body contour, the binary mask of the skin area, and the high-noise coarse-grid energy deposition matrix calculated by MC in the training set.

4. A Monte Carlo dose super-resolution denoising method for boron neutron capture therapy according to claim 1, characterized in that In Step 4, the multi-dimensional comparison of the results further includes: performing the same processing on the denoising results of the two trained convolutional neural networks obtained in Step 3. After extracting the data, compare the high-noise low-resolution images obtained from low-sampled data, the low-noise high-resolution clean images obtained from high-sampled data, the denoised images of the SRCNN convolutional neural network, and the denoised images of the PDA-SRDNet denoising convolutional neural network, and finally output the comparison results of the model performance and calculation efficiency of the two convolutional neural networks.

5. A Monte Carlo dose super-resolution denoising method for boron neutron capture therapy according to claim 1, characterized in that, In the said step 5, the post-processing operation includes performing operations to exclude abnormal prediction regions and smooth the edges on the predicted contour, specifically including the following steps: adopting a flood algorithm to remove abnormal hole information within the boundary of the contour; finding the largest connected domain containing the prediction information of the contour to remove other abnormal prediction regions; performing an edge smoothing operation on the contour using the built-in library of OpenCV; and inputting the processed denoised image into a commercial TPS for evaluation by a physicist.

Images