A boron neutron capture therapy method and apparatus

Abstract

The application discloses a boron neutron capture therapy method, comprising the following steps: S1, obtaining a plurality of patient MRI initial scanning images and corresponding CT initial scanning images; S2, building a neural network, training and verifying the neural network by using the plurality of MRI initial scanning images and the corresponding CT initial scanning images, and obtaining a trained neural network, wherein the trained neural network can output a corresponding pseudo-CT image according to the MRI initial scanning image, and the similarity of the pseudo-CT image and the CT initial scanning image is within a limited error range; S3, using the pseudo-CT image to complete the delineation of a target region and an organ at risk, to plan a neutron beam relative position, to calculate a radiation dose, and to implement boron neutron capture therapy. The method can realize the plan design and dose calculation before radiotherapy without CT scanning before treatment; MRI can be combined with an accelerator to realize real-time positioning during radiotherapy; and MRI does not bring additional radiation and does not introduce metal artifacts.

Description

Technical Field

[0001] This application relates to the field of boron neutron capture therapy, and in particular, to a boron neutron capture therapy method and apparatus. Background Technology

[0002] As is well known, cancer has become the leading cause of death and health threats to humanity. Radiotherapy, as a localized and non-invasive treatment, has demonstrated excellent clinical efficacy and is currently one of the three major methods of cancer treatment. While photon therapy is currently the most mature radiotherapy technology, its energy deposition method inevitably leads to dose deposition on organs at risk within the target area, increasing the risk of secondary radiation-induced cancer. Therefore, new radiotherapy technologies require further research.

[0003] Boron neutron capture therapy (BNCT) is an emerging precision radiotherapy technique. It utilizes an externally irradiated neutron beam combined with boron-containing targeted drugs enriched in the tumor area. Through the reaction of 10B(n,α)₇Li, high-energy linear density heavy ions (α,₇Li) are generated, selectively killing cancer cells at the cellular scale while minimizing damage to normal cells that have not taken up the boron-containing drugs. This multi-particle synergistic radiotherapy technique offers advantages such as precise targeting, high biological efficacy, and short treatment duration. It has particular therapeutic advantages for large-scale diffuse malignant tumors and radiotherapy-recurrent tumors, such as primary brain tumors, recurrent head and neck tumors, and adenocarcinoma with liver metastases—cancers for which traditional radiotherapy is less effective. BNCT holds promise as a valuable supplement to traditional treatment techniques, achieving comprehensive cancer treatment coverage.

[0004] Currently, boron neutron capture therapy requires first using computed tomography (CT) images, and then using boron concentration information provided by MRI or PET images or offline boron concentration measurement methods, as well as anatomical information provided by CT scans, to calculate the radiation dose.

[0005] Existing conventional radiotherapy methods often employ image-guided techniques to achieve precise radiotherapy, taking into account anatomical errors during treatment. Commonly used methods include cone-beam CT (CBCT) image guidance, online CT guidance, and MRI guidance. Compared to CBCT, MRI has unparalleled advantages, such as superior soft tissue contrast, no additional radiation dose, and multi-sequence biological functional imaging.

[0006] Although MRI image guidance is safer and more effective than CT image guidance, this is only for photon therapy. There is currently no research on the application of MRI images in boron neutron capture therapy. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a novel boron neutron capture therapy method, comprising: S1, obtaining multiple initial MRI scan images and corresponding initial CT scan images of patients; S2, constructing a neural network, training and validating the neural network using the multiple initial MRI scan images and corresponding initial CT scan images to obtain a trained neural network, wherein the trained neural network can output corresponding pseudo-CT images based on the initial MRI scan images, and the similarity between the pseudo-CT images and the initial CT scan images is within a limited error range; S3, using the pseudo-CT images to delineate the target area and organs at risk, determine the relative position of the neutron beam, calculate the radiation dose, and implement boron neutron capture therapy.

[0008] Furthermore, the neural network is constructed and trained as follows: an artificial intelligence network based on a convolutional neural network is selected to construct the neural network; an attention mechanism is introduced into the neural network, and the neural network includes attention weights and convolutional network parameters. When outputting each pseudo-CT image, the attention weights are calculated using the attention mechanism, and the attention weights and convolutional network parameters with the smallest error range are selected to form the trained neural network.

[0009] Further, the method for calculating the attention weights and outputting the pseudo-CT image is as follows: the attention mechanism includes a channel attention mechanism, and the initial MRI scan image includes an initial feature F; the channel attention mechanism processes the three-dimensional initial MRI scan image into two one-dimensional channel images through global average pooling and max pooling methods, and then passes them through a multilayer perceptron (MLP) to obtain two output features. The two output features are added together and then passed through an activation function σ to obtain the attention weight coefficient Vc. The product of the attention weight coefficient Vc and the initial feature F is the pseudo-CT image.

[0010] Vc(F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0011] In the formula, Max Pool and Avg Pool are the max pooling and average pooling operations, respectively, MLP is a multilayer perceptron, σ is the activation function, and Vc is the attention weight coefficient.

[0012] Further, the method for calculating the attention weights and outputting the pseudo-CT image is as follows: the attention mechanism includes a spatial attention mechanism, and the initial MRI scan image includes an initial feature F; the spatial attention mechanism processes the three-dimensional initial MRI scan image into two two-dimensional channel images through average pooling and max pooling methods in one channel dimension, stitches the two two-dimensional channel images together according to the channels, and obtains the attention weight coefficient Vs through the activation function σ; the product of the attention weight coefficient Vs and the initial feature F is the pseudo-CT image.

[0013] V s (F)=σ(f([AvgPool(F), MaxPool(F)]))

[0014] In the formula, Max Pool and Avg Pool are the max pooling and average pooling operations, respectively, MLP is a multilayer perceptron, σ is the activation function, and Vs is the attention weight coefficient.

[0015] Furthermore, the method for selecting the attention weights with the smallest defined error range is as follows: calculate the average absolute error or mean square error of the similarity between the pseudo CT image and the initial CT image, select the weight with the smallest average absolute error or mean square error, and set it as the weight value in the neural network.

[0016] Furthermore, the initial MRI scan image includes boron concentration distribution information. Based on this boron concentration distribution information, the radiation dose is calculated using a Monte Carlo program.

[0017] Furthermore, it also includes: using the maximum class variance method to divide the above-mentioned initial MRI scan image into two parts, background and foreground, according to the grayscale characteristics of the image; selecting the MRI input image of the foreground part; and using rigid iterative registration to rigidly register the above-mentioned initial MRI scan image of the foreground part with the corresponding initial CT scan image of the same patient, so as to obtain the registered MRI input image as the initial MRI image input to the above-mentioned neural network.

[0018] Furthermore, the above rigid registration method is as follows: the rigid iterative registration is performed using the SimpleITK toolkit, wherein the histogram classification number is set to 50, the sampling percentage is 1%, and the interpolator is sitkLinear.

[0019] This application also relates to an offline boron neutron capture therapy apparatus, comprising: an MRI scanner for performing MRI scan imaging of a patient's lesion site; a treatment bed for supporting the patient and sliding within the MRI scanner or particle accelerator; a processor for performing any of the methods described above based on the MRI scan imaging, outputting a pseudo-CT image, calculating the boron neutron radiation dose, and sending instructions to the particle accelerator; and a particle accelerator for receiving the instructions and emitting boron neutrons to the lesion site for boron neutron capture therapy.

[0020] This application also relates to an online boron neutron capture therapy device, comprising: an MRI-particle accelerator integrated machine for performing MRI scans on the lesion site of a patient, outputting boron concentration distribution and anatomical structure information, and emitting a beam for treatment after receiving instructions; a treatment bed for supporting the patient and sliding within the aforementioned MRI-particle accelerator integrated machine; and a processor for implementing any of the above methods based on the aforementioned MRI scans, outputting pseudo-CT images, calculating the boron neutron radiation dose, formulating a treatment plan based on the aforementioned boron concentration distribution and anatomical structure information and the aforementioned radiation dose, and sending instructions to the aforementioned MRI-particle accelerator integrated machine.

[0021] The method described in this application acquires anatomical images using magnetic resonance imaging (MRI) and then converts these images to CT images using deep learning technology. The BNCT irradiation dose is calculated by combining boron concentration data obtained using MRI, PET, or other boron concentration measurement methods, and then used for subsequent boron neutron capture therapy. This method solves the problem of inaccurate registration between boron concentration distribution and CT images, simplifying the treatment process while significantly improving the accuracy of dose calculation.

[0022] Furthermore, MRI offers advantages that CT / CBCT cannot match, such as: superior soft tissue contrast, absence of bone artifacts; no additional radiation dose; 3D data, tomographic imaging in any orientation; multi-sequence biological functional imaging; and no need for surface markers. Moreover, the high entropy of MRI images indicates the existence of a surjective transformation from MRI to CT images.

[0023] Based on this method, this application also discloses a device for boron neutron capture therapy, comprising: an MRI-particle accelerator integrated machine for performing MRI scans on the lesion site of the patient, outputting boron concentration distribution anatomical structure information, and emitting a beam for treatment after receiving instructions; a treatment bed for supporting the patient and sliding within the aforementioned MRI-particle accelerator integrated machine; and a processor for implementing any of the above methods based on the aforementioned MRI scans, outputting pseudo-CT images, calculating the boron neutron radiation dose, formulating a treatment plan based on the aforementioned boron concentration distribution anatomical structure information and the aforementioned radiation dose, and sending instructions to the aforementioned particle accelerator integrated machine.

[0024] This application also relates to an apparatus for boron neutron capture therapy, comprising: an MRI scanner for performing MRI scan imaging on a patient's lesion site; a treatment bed for supporting the patient and sliding within the MRI scanner or particle accelerator; a processor for performing any of the above methods based on the MRI scan imaging, outputting a pseudo-CT image, calculating the boron neutron radiation dose, and sending instructions to the particle accelerator; and a particle accelerator for receiving the instructions and emitting boron neutrons to the lesion site for boron neutron capture therapy.

[0025] The advantages of using MRI to guide boron neutron capture therapy (BNCT) through deep learning are summarized as follows: 1) Before treatment, the CT values ​​of the corresponding voxels can be obtained using deep learning networks, enabling pre-radiotherapy planning and dose calculation without CT scans; 2) MRI can be combined with accelerators to achieve real-time localization during radiotherapy; 3) MRI has excellent soft tissue resolution; 4) MRI does not introduce additional radiation; 5) MRI does not introduce metal artifacts; 6) For BNCT, MRI images can provide boron concentration information.

[0026] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0027] Figure 1 This is a flowchart of the boron neutron capture therapy method described in this application.

[0028] Figure 2 This is a schematic diagram of the boron neutron capture therapy method described in this application.

[0029] Figure 3 This diagram illustrates how an artificial intelligence network using a convolutional neural network processes an initial MRI image into a pseudo-CT image.

[0030] Figure 4 A schematic diagram of the offline mode boron neutron capture therapy method of this application is shown.

[0031] Figure 5 A schematic diagram of the online mode boron neutron capture therapy method of this application is shown. Detailed Implementation

[0032] The specific embodiments of the present invention will be described in more detail below with reference to the accompanying drawings and examples, so as to better understand the solution of the present invention and its advantages in various aspects. However, the specific embodiments and examples described below are for illustrative purposes only and are not intended to limit the present invention.

[0033] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0034] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection, an electrical connection, or a connection that allows for communication; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0035] The following disclosure provides many different embodiments or examples for implementing various structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. These are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. In addition, examples of various specific processes and materials are provided in this invention, but those skilled in the art will recognize the application of other processes and / or the use of other materials.

[0036] Figure 1 This is a flowchart of the boron neutron capture therapy method based on neural networks proposed in this application.

[0037] like Figure 1 As shown, the boron neutron capture therapy method based on neural networks in this application mainly involves the following steps:

[0038] Step S1: Obtain multiple initial MRI scan images and corresponding initial CT scan images of patients. Save all initial MRI images and their corresponding initial CT scan images as image pairs for subsequent neural network training and validation. Divide all images into three parts, which will serve as training samples, validation samples, and test samples for the MRI-CT neural network.

[0039] Furthermore, after step S1, the obtained initial MRI scan images can be further processed:

[0040] 1) The initial MRI scan image is divided into background and foreground parts according to the image's grayscale characteristics using the Otsu's algorithm, also known as the maximum inter-class variance method. The foreground portion of the MRI input image is selected. In this embodiment, the Otsu's algorithm is chosen as the Otsu's algorithm. This algorithm performs fixed-threshold binarization on the MRI image, maximizing the inter-class variance. Based on the image's grayscale characteristics, the image is divided into background and foreground parts. Segmentation with the maximum inter-class variance means the minimum probability of misclassification. The foreground portion is selected as the region of interest.

[0041] The selected foreground portion can be further processed.

[0042] 2) Using rigid iterative registration, the initial MRI scan image of the foreground portion is rigidly registered with the corresponding initial CT scan image of the same patient to obtain the registered MRI input image, which is then used as the initial MRI image input to the neural network. Rigid iterative registration is then used to complete the rigid registration of MRI and CT data for the same patient.

[0043] This application uses the SimpleITK toolkit to perform rigid iterative registration of MRI and CT images, setting the histogram classification number to 50 and the sampling percentage to 1%. The interpolator used is sitkLinear.

[0044] Step S2: Build a neural network. Train and validate the neural network using multiple initial MRI scan images and corresponding initial CT scan images to obtain a trained neural network. The trained neural network can obtain the corresponding pseudo-CT image based on the initial MRI scan image, and the similarity between the pseudo-CT image and the initial CT image is within a limited error range.

[0045] Step S2 first includes step S21, which involves building the neural network:

[0046] In this process, the neural network can be built using the following methods. The artificial intelligence algorithms used can be one or more of the following frameworks: DeepLearnToolbox, Caffe, CNTK, TensorFlow, Theano, Torch, Keras, Lasagne, DSSTNE, MXNet, DeepLearning4J, ConvNetJs, Chainer, Scikit-Learn, and Matlab. The artificial intelligence network model can be one or more of the following: deep belief network, deep neural network, convolutional neural network, stacked autoencoder, or convolutional autoencoder.

[0047] This application selects an artificial intelligence network based on a convolutional neural network, and introduces an attention mechanism into this neural network. Specifically, this application employs the Unet artificial intelligence network (SARUnet) with a self-attention mechanism. Patient MRI samples are fed into the SARUnet artificial intelligence network framework for training to obtain pseudo-CT images.

[0048] Step S2 also includes step S22, which trains the neural network:

[0049] The neural network contains approximately 18M parameters to be trained (including attention weights and other convolutional network parameters), and is trained using stochastic gradient descent. The attention weights and other convolutional network parameters are selected to minimize the maximum difference (MAE) between the pseudo-CT image and the initial CT image within a specified range; this selected MAE represents the trained neural network.

[0050] More specifically, the attention mechanism is a special structure embedded in a machine learning model to automatically learn and calculate the contribution of input data to output data. The attention mechanism obtains weights by taking a weighted average of the hidden states across all time steps of the encoder. The decoder adjusts these weights—the attention weights—at each time step, allowing it to focus on different parts of the input sequence at different time steps.

[0051] This method introduces an attention mechanism into the Unet artificial neural network (i.e., SARUnet) for MRI-CT image conversion. For example... Figure 2 and Figure 3 , Figure 2 This is a schematic diagram of the boron neutron capture therapy method described in this application. Figure 3This diagram illustrates how an artificial intelligence network using a convolutional neural network processes initial MRI images into pseudo-CT images. The SARUnet neural network is continuously trained using multiple sets of initial MRI and CT scan images. Two main methods exist for calculating attention weights and outputting pseudo-CT images using attention mechanisms: channel attention and spatial attention.

[0052] Channel attention mechanism: Global average pooling and max pooling are used to utilize high-level and low-level information respectively. When the input is an H×W×C feature map, it first undergoes spatial global average pooling and max pooling to obtain two 1×1×C channel images. These are then fed into a multilayer perceptron. The two output features are then summed and passed through a sigmoid activation function σ to obtain weight coefficients. Finally, the weight coefficients Vc are multiplied by the original feature F to obtain a scaled new feature map.

[0053] Vc(F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0054] In the formula, Max Pool and Avg Pool are the max pooling and average pooling operations, respectively, MLP is a multilayer perceptron, σ is the activation function, Vc is the attention weight coefficient, and F is the initial feature.

[0055] Spatial attention mechanism: Similar to channel attention, given a 3D H×W×C feature map, first perform average pooling and max pooling on each channel to obtain two 2D H×W×1 images, and then concatenate these two images along their channels. Next, pass the concatenation through an activation function σ to obtain weight coefficients Vs. Finally, multiply the weight coefficients Vs by the initial feature map F to obtain the scaled new feature.

[0056] V s (F)=σ(f([AvgPool(F), MaxPool(F)]))

[0057] In the formula, Max Pool and Avg Pool are the max pooling and average pooling operations, respectively, MLP is a multilayer perceptron, σ is the activation function, and Vs is the attention weight coefficient.

[0058] Step S2 also includes step S23, which selects weight values:

[0059] The error (MAE, mean absolute error or MSE, standard deviation) between the output pseudo-CT image and the initial CT scan image used for comparison is calculated. The weight that minimizes the mean absolute error or standard deviation is selected and set as the weight value in the neural network.

[0060] The method for calculating the mean absolute error (MAE) or mean squared error is as follows: using the previously obtained validation set data, the similarity between the pseudo CT images predicted by the AI ​​network and the actual initial CT scan images is analyzed, focusing on the two parameters: mean absolute error (MAE) and mean squared error (ME). By comparing the training set error and the validation set error, it is determined whether the network has overfitted (the training set error continuously decreases while the validation set error remains unchanged or increases). If significant overfitting has occurred or the network has reached the predetermined training batch size, network training is stopped, and the weights of the batch with the smallest mean absolute error (MAE) on the validation set and the corresponding parameters of other convolutional networks are selected as the parameter settings for the trained network.

[0061] MAE (mean absolute error) is defined as the average of the absolute values ​​of errors, as shown in the following formula, where yi is the pseudo-CT image. Real CT images:

[0062]

[0063] It can represent the average distance between a fake CT image and a real CT image.

[0064] The root mean square error (RMSE) is defined as follows: Assuming the CT image size is m*n, calculate the sum of squares of the errors for each pixel, take the mean, and then take the square root. Where y... i This is a pseudo-CT image. For real CT images:

[0065]

[0066] In summary, the training process of a neural network can be described as follows: SARUnet is an artificial intelligence network based on convolutional networks, primarily capable of image-to-image prediction. For example, by inputting a T1-weighted MRI image of a patient's brain into SARUnet, SARUnet generates a predicted pseudo-CT image based on that image. The loss function during training is the L1 norm of the real CT image and the pseudo-CT image. As the number of training iterations increases, gradient descent is used to update the network parameters. The pseudo-CT image generated by SARUnet becomes increasingly closer to the real CT image; when the training batch is reached or the error requirement is met, the network training can be considered complete.

[0067] The loss function maps the values ​​of a random event or its related random variables to non-negative real numbers to represent the "risk" or "loss" of that random event. It is typically used as a learning criterion in relation to optimization problems; that is, the model is solved and evaluated by minimizing the loss function. The L1 loss function, also known as the minimum absolute value deviation (LAD) or absolute value loss function (LAE), generally minimizes the sum of the absolute differences between the target value and the estimated value. This method uses the L1 error (MAE) as the loss function for network optimization.

[0068]

[0069] In the formula, L is the norm error, yi is the real CT image, f(xi) is the pseudo CT image, xi is the input of the neural network, and the f() function represents the function mapped by the neural network.

[0070] S3 uses the trained neural network from S2 to predict pseudo-CT images based on arbitrary patient MRI images, completes the delineation of target areas and organs at risk, determines the relative position of the neutron beam, calculates the radiation dose, and implements boron neutron capture therapy.

[0071] In detail, the method of using a pre-trained neural network to convert MRI-CT images for boron neutron capture therapy in patients is as follows: Figure 4 A schematic diagram of the offline boron neutron capture therapy method based on neural networks in this application:

[0072] P1, The patient on the treatment bed is scanned using an MRI scanner to obtain anatomical information on the distribution of boron concentration;

[0073] P2, using the deep learning model trained in steps S1 and S2 above, achieves rapid conversion from MRI to a model that can be used for dose calculation;

[0074] P3, using the MRI image sequence obtained in step P1, a professional radiation physicist delineates the patient's target area and organs at risk.

[0075] P4. The relative positions of the radiation source or neutron beam are determined using prior information about the patient's lesions.

[0076] In step P5, using the dose calculation model obtained in step P2, combined with the relative position of the radiation source or neutron beam determined in step P4, the dose is calculated using the Monte Carlo algorithm based on the boron concentration distribution anatomical information and the source position. A treatment plan is then formulated, and the actual BNCT treatment is implemented.

[0077] The Monte Carlo method, also known as the random sampling method or statistical experimental method, is a branch of computational mathematics that can realistically simulate the physical processes of actual particle transport and solve practical problems. Optionally, the Monte Carlo program is used to construct a model based on the patient's CT images and the material composition of different tissues, and to determine its response to different particles. The Monte Carlo program is a general software package based on the Monte Carlo method for calculating neutron, photon, electron, or multi-particle transport problems in three-dimensional complex geometries. The Monte Carlo program can use one or more of MCNP, GEANT4, TOPAS, GATE, FLUKA, SuperMC, or Phits. The Monte Carlo program is used to simulate the energy deposition and particle transport processes under different treatment regimens to obtain characteristic transient gamma ray yield distribution maps and three-dimensional dose distribution maps in the target area under different tumor shapes, tumor types, tumor locations, and boron concentration distributions. According to the calculation methods of energy deposition and dose, the information of different dose components is obtained according to formulas (1) to (3), and the total biological dose distribution is obtained according to formula (4).

[0078] Dose boron = (Energy) α +Energy7 Li ) / m (1),

[0079] Dose neutron =Energ yproton / m (2),

[0080] Dose gamma =Energy gamma / m (3),

[0081] Dose REE =RBE boron Dose boron +RBE neutron Dose neutron +RBE gamm a·Dose gamma (4).

[0082] Among them, Dose boron For boron dosage, Dose neutron Neutron dose, Dose gamma For gamma dose, Energy α Energy is the energy deposition of alpha particles within voxels. 7 Li The energy deposition of 7Li particles within a voxel, where m is the mass of the voxel. proton Energy is the energy deposited by protons within voxels. gammaDose is the energy deposition of photons within voxels. REB RBE is the total biological dose within the voxel. boron This represents the relative biological effect value of boron dosage; for tumor voxels, this value is 3.8, for normal tissue voxels, it is 1.4, and for skin voxels, it is 2.5. (RBE) neutron This represents the relative biological effect value of the neutron dose; for tumor voxels, normal tissues, and skin tissues, this value is 3.2 (RBE). gamma This represents the relative biological effect value of the gamma dose; for tumor voxels, normal tissue, and skin tissue, this value is 1. The Monte Carlo method for calculating radiation dose will not be explained in detail here.

[0083] Finally, the dose calculation results obtained in step P5 can be used to verify whether the treatment dose meets the treatment requirements (IAEA); in this embodiment of the method, the following parameters are used to evaluate the performance of the neutron beam dose:

[0084] 1. Irradiation time ≤ 30 min (the proton current used in the accelerator is 10 mA)

[0085] 2. 30 Gy can treat depths ≥7 cm

[0086] 3. Maximum tumor dose ≥60 Gy

[0087] 4. Maximum dose to normal brain tissue ≤12.5 Gy

[0088] 5. Maximum skin dose ≤11Gy

[0089] If the above requirements 1-5 are not met, repeat steps P4-5, i.e., adjust the position of the radiation source or neutron beam, recalculate the dose, etc., until the treatment plan meets the treatment requirements.

[0090] Using this method, this application discloses an offline boron neutron capture therapy device, comprising: an MRI scanner for performing MRI scanning imaging of the patient's lesion site and outputting boron concentration distribution anatomical structure information; a treatment bed for supporting the patient and sliding within the MRI to adjust the angle for boron neutron incidence; and a processor for implementing the above method based on the MRI scan imaging, outputting pseudo-CT images, calculating the boron neutron radiation dose by combining the boron concentration information obtained from the MRI, and formulating a treatment plan based on the boron concentration distribution anatomical structure information and the radiation dose.

[0091] Furthermore, using the MRI-CT image conversion technology mentioned in this application, boron neutron capture therapy can also be performed online, such as... Figure 5 As shown, this application presents a schematic diagram of the online mode boron neutron capture therapy method based on neural networks:

[0092] P1' The patient on the treatment bed is scanned using an MRI-particle accelerator integrated machine to obtain the patient's real-time positioning information. Combined with the patient's lesion information, the patient's positioning and placement are completed by moving the treatment bed.

[0093] The movement of the treatment bed is as follows: for example, when lying flat, the patient positioning information obtained in step P1' is transmitted to the signal processing system to calculate the location of the lesion. Since the location does not match the neutron incident angle of the neutron emitter, the treatment bed controller controls the treatment bed to deflect or rotate, and the inertial sensor in the treatment bed continuously measures and reflects the angle information to the treatment bed controller until the lesion positioning information matches the neutron incident angle.

[0094] P2', the MRI-particle accelerator integrated machine emits boron neutrons, and boron neutron capture therapy is performed;

[0095] The treatment process is as follows: an MRI-particle accelerator integrated machine performs MRI scans on the patient's lesion; a treatment bed supports the patient and slides within the MRI-particle accelerator integrated machine, adjusting to a suitable boron neutron incident angle; a processor, based on the MRI scan, performs the above-described process, outputs pseudo-CT images, calculates the boron neutron radiation dose, and sends instructions to the particle accelerator; the particle accelerator receives the instructions and emits boron neutrons to the lesion site for boron neutron capture therapy.

[0096] Applying the above-described online boron neutron capture therapy method, this application also discloses an online boron neutron capture therapy device, such as... Figure 5 As shown, it includes: an integrated MRI-particle accelerator for performing MRI scans on the patient's lesion; a treatment bed that supports the patient and slides within the integrated MRI-particle accelerator; a processor that, based on the MRI scan, performs the above-described method to output pseudo-CT images, calculates the boron neutron radiation dose by combining the boron concentration information obtained from the MRI, and sends instructions to the integrated MRI-particle accelerator; and the integrated MRI-particle accelerator receives the instructions and emits boron neutrons to the lesion site for boron neutron capture therapy.

[0097] The device may also include a radiation detection device and a control device. The radiation detection device is used to detect the instantaneous gamma rays emitted after neutron beam irradiation in real time. Simultaneously, the detection device will also monitor the patient's positioning information in step P1' in real time. If the target area deviation exceeds a specified threshold, a command to stop beam emission from the accelerator will be sent. The device also includes radiation shielding devices, etc. The processor performs image processing and neutron beam control between the host and slave computers via fiber optic communication; this is standard industry practice and will not be explained in detail here.

[0098] The boron neutron capture therapy method described in this application avoids the negative effects of multiple CT scans. By utilizing pseudo-CT values ​​obtained through a deep learning network, pre-radiotherapy planning and dose calculation can be performed without CT scans. Furthermore, MRI can be combined with an accelerator for real-time localization during radiotherapy. Moreover, BNCT boron neutron capture therapy requires boron concentration information, which can be directly obtained from MRI images, facilitating the capture therapy process.

[0099] Obviously, the above embodiments are merely examples for clearly illustrating the present invention and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. An apparatus for off-line boron neutron capture therapy, characterized by , comprising: an MRI scanner for MRI scanning imaging of a lesion site of a patient; a treatment bed for bearing the patient and sliding in the MRI scanner or a particle accelerator; a processor for implementing a processing method according to the MRI scanning imaging, outputting a pseudo-CT image, calculating a boron neutron capture therapy radiation dose, and sending an instruction to the particle accelerator; a particle accelerator for receiving the instruction and emitting neutrons to the lesion site for boron neutron capture therapy; the processing method comprises, S1, obtaining a plurality of MRI initial scanning images and corresponding CT initial scanning images of patients; S2, building a neural network, training and verifying the neural network by using the plurality of MRI initial scanning images and the corresponding CT initial scanning images, obtaining a trained neural network, wherein the trained neural network can output a corresponding pseudo-CT image according to the MRI initial scanning image, and the similarity of the pseudo-CT image and the CT initial scanning image is within a limited error range; wherein an artificial intelligence network based on a convolutional neural network is selected to build the neural network; an attention mechanism is introduced into the neural network, and the neural network includes attention weights and convolution network parameters, when each pseudo-CT image is output, the corresponding attention weight is calculated by using the attention mechanism, and the attention weight and the corresponding convolution network parameter with the minimum limited error range are selected to form the trained neural network; the attention mechanism includes a channel attention mechanism, and the MRI initial scanning image includes an initial feature F; the channel attention mechanism processes the three-dimensional MRI initial scanning image into two one-dimensional channel images through global average pooling and maximum pooling methods, then obtains two output features through a multi-layer perceptron (MLP), and obtains an attention weight coefficient Vc through an activation function σ after the two output features are added, and the product of the attention weight coefficient Vc and the initial feature F is the pseudo-CT image; ， wherein Max Pool and Avg Pool are maximum pooling and average pooling operations respectively, MLP is a multi-layer perceptron, σ is an activation function, and Vc is an attention weight coefficient; or the attention mechanism includes a spatial attention mechanism, and the MRI initial scanning image includes an initial feature F; the spatial attention mechanism processes the three-dimensional MRI initial scanning image into two two-dimensional channel images through an average pooling and maximum pooling method in a channel dimension, splices the two two-dimensional channel images together, obtains an attention weight coefficient Vs through an activation function σ, and the product of the attention weight coefficient Vs and the initial feature F is the pseudo-CT image; ， wherein Max Pool and Avg Pool are maximum pooling and average pooling operations respectively, MLP is a multi-layer perceptron, σ is an activation function, and Vs is an attention weight coefficient; S3, using the pseudo-CT image, completing the delineation of a target region and an organ at risk, determining a relative position of a neutron beam, calculating a radiation dose, and implementing boron neutron capture therapy.

2. The off-line boron neutron capture therapy device of claim 1, wherein the method for selecting the attention weight with the minimum limited error range is The average absolute error or mean square error of the similarity of the pseudo CT image and the CT initial scan image is calculated, and the weight when the average absolute error or mean square error is the smallest is selected and set as the weight value in the neural network.

3. The off-line boron neutron capture therapy device of claim 1, wherein The MRI initial scan image includes boron concentration distribution information, and the radiation dose is calculated by using a Monte Carlo program in combination with the boron concentration distribution information.

4. The off-line boron neutron capture therapy device of claim 1, wherein The method further comprises: dividing the MRI initial scan image into a background and a foreground according to the gray scale characteristics of the image by using a maximum inter-class variance method, selecting an MRI input image of the foreground, and performing rigid registration on the MRI initial scan image of the foreground and a CT initial scan image of the same patient to obtain a registered MRI input image as the MRI initial image input into the neural network.

5. The off-line boron neutron capture therapy device of claim 4, wherein The method for rigid registration is to perform the rigid iterative registration by using a SimpleITK tool package, wherein the number of histogram classifications is set to 50, the sampling percentage is set to 1%, and the interpolator is set to sitkLinear.

Images