X-ray beam light shutter and low dose cbct imaging method

Abstract

The application provides an X-ray beam light baffle and a low-dose CBCT imaging method. First, part of data is taken out from a training set, so that the number of remaining CBCT images and registered CT images is the same, and imbalance of training data is avoided. Secondly, the taken-out data is pre-trained in a self-supervised manner, so that the difficulty in network training caused by the random initialization of the Transformer added to the CycleGAN is avoided. Finally, the pre-trained Transformer is added to the CycleGAN to obtain a hybrid network, and the remaining data is used for unsupervised training, so that the advantages of the Transformer and the convolution are combined, and the ability of the network to enhance the CBCT image quality is improved.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology and relates to an X-ray beam baffle and a low-dose CBCT imaging method. Background Technology

[0002] Cone-beam computed tomography (CBCT) is an imaging technique that acquires images of the internal anatomical structures of the human body in a non-invasive manner. It offers advantages such as short scan time, low radiation dose, and high image resolution, and is widely used in image-guided radiotherapy to correct positioning errors and adjust radiotherapy plans. However, CBCT images contain a significant amount of artifacts and noise, limiting its further clinical applications, such as dose calculation and organ delineation.

[0003] Traditional CBCT image quality enhancement methods can be divided into two categories: hardware correction and software correction. Hardware correction methods, such as collimator methods, antiscattering grating methods, and air gap methods, suffer from severe thermal load and low quantum efficiency, while software correction methods, such as convolution methods, Monte Carlo simulation methods, and scattering estimation methods, suffer from long processing times and poor robustness.

[0004] In recent years, with the development of deep learning, an increasing number of supervised networks have been used to improve the quality of CBCT images, such as UNET and PIX2PIX. However, the training of supervised networks relies on hard-to-obtain paired data, such as CBCT images and registered CT images with pixel-by-pixel matching. CycleGAN, as the most classic unsupervised network, can be trained directly using unpaired data, attracting increasing attention from researchers. However, CycleGAN is essentially a convolutional neural network composed of convolutions, lacking the ability to model global dependencies, which to some extent limits its ability to improve the quality of CBCT images. Transformer, as the most general model in the field of natural language processing, is good at modeling global dependencies, but due to the lack of inductive biases inherent in convolutions, such as translational isovariance and local correlation, the network training usually requires more data. Therefore, directly adding Transformer to CycleGAN not only fails to effectively combine the advantages of both, but may also lead to network training difficulties. In addition, the imbalance of the training set, i.e., the large difference between the number of CBCT images and registered CT images, also leads to training difficulties for CycleGAN.

[0005] Based on the above problems, this invention proposes an X-ray beam baffle and a low-dose CBCT imaging method that can effectively combine the Transformer and CycleGAN methods. Summary of the Invention

[0006] In view of the above problems, the present invention proposes an X-ray beam baffle, which includes obtaining a CBCT image that is blocked;

[0007] A beam baffle is installed at the output end of the X-ray tube. The beam baffle consists of a lead part and a transparent part. The lead part is used to completely block the X-rays, and the transparent part is used to completely transmit the X-rays.

[0008] The lead elements are small lead balls or small lead blocks, evenly arranged inside the beam baffle, with a width-to-height ratio of W:H; the number of lead elements is the same as the number of learnable vectors l.

[0009] The transparent section includes two interlocking transparent plates, with a cavity between the two transparent plates for accommodating the lead portion. The size of the transparent section is the same as the size of the beam baffle.

[0010] The transverse and longitudinal sections of the transparent plate are labeled sequentially with Arabic numerals and English letters, respectively, forming a set Q. Based on a hybrid network with good training performance obtained using the low-dose CBCT imaging method, the learnable vector l in P is statistically analyzed. M =(p1,l,p3,...,l,l,p N By replacing the position in the set Q, the optimal position for placing the lead part in the receiving cavity is obtained.

[0011] A low-dose CBCT imaging method includes the following steps:

[0012] Step 1: Take a portion of the data from the training set to ensure that the number of remaining CBCT images is the same as the number of registered CT images, thus avoiding imbalanced training data.

[0013] Step 2: The extracted data is pre-trained in a self-supervised manner to avoid network training difficulties caused by adding a randomly initialized Transformer system to the CycleGAN system.

[0014] Step 3: Add the pre-trained Transformer system to the CycleGAN system to obtain a hybrid network, and use the remaining data for unsupervised training.

[0015] Furthermore, in step one, the training set includes training set A and training set B. The data in training set A consists of CBCT images, and the number of images is denoted as N. CBCT The training set B consists of registered CT images, denoted as N. CT When N CBCT >N CT At that time, N are randomly selected from the training set A. CBCT -N CT Zhang images are used to pre-train a Transformer system in a self-supervised manner; when N... CBCT<N CT At that time, N are randomly selected from the training set B. CT -N CBCT The Transformer system is pre-trained using images through a self-supervised approach.

[0016] Furthermore, the Transformer system contains N Transformer modules, and step two, the pre-training process of the Transformer system, includes:

[0017] Step 2.1) Convert the input CBCT image X∈R H×W Divide the sample into multiple non-overlapping square patches of size n×n, forming a set P = (p1, p2, p3, ..., pn). N-2 p N-1 p N ), where H and W represent the height and width of the image, respectively, and p i This represents each square patch, and N represents the number of square patches.

[0018] Step 2.2) Randomly and uniformly replace part of the square patch with a learnable vector l to obtain the occluded CBCT image, forming a set P. M =(p1,l,p3,...,l,l,p N )

[0019] Step 2.3) After downsampling, the image is input into the Transformer system for inter-patch information exchange, and finally upsampling is performed to output the predicted CBCT image:

[0020] I pCBCT =UP(T(DOWN(I mCBCT )) N (1)

[0021] In formula (1), I pCBCT I represents the CBCT image predicted by the self-supervised network. mCBCT The input to the self-supervised network is the occluded CBCT image, where UP(.) represents the upsampling function, DOWN(.) represents the downsampling function, T(.) represents the Transformer module, and N represents the number of modules.

[0022] Furthermore, the Transformer module consists of three parts: multi-head self-attention, multilayer perceptron, and normalization layer.

[0023] The calculation process of the Transformer module can be expressed as formula (2):

[0024]

[0025]

[0026] in

[0027] MSA=FC(Concat(head1,...,head h (3)

[0028]

[0029] MLP(.)=FC(σ(FC(.))) (5)

[0030]

[0031] In formula (2), Z l Indicates input features, Z represents intermediate features. l+1 In Equation (3), Concat(.) represents the concatenation operation, FC represents the fully connected layer, and head represents the output of self-attention. In Equation (4), Q, K, and V are obtained from the normalized features through different linear transformations, T represents the transpose, and d represents the output of self-attention. K The dimension of K is represented; in formula (5), σ represents the GELU activation function; in formula (6), x represents the input vector, u and σ represent the mean and standard deviation of x respectively, γ and τ are learnable parameters, representing the scaling factor and bias respectively.

[0032] Furthermore, the CycleGAN system includes two types of generators:

[0033] The GA generator takes a CBCT image as input and outputs an enhanced CBCT image, whose data distribution is similar to the registered CT images in training set B.

[0034] The GB generator takes a registered CT image as input and outputs a CT image with noise and artifacts, whose data distribution is similar to that of the CBCT images in training set A.

[0035] When the GA generator adds N CBCT >N CT After the Transformer system is pre-trained, a randomly initialized Transformer system is added to the GB generator to form the first hybrid network.

[0036] When GB generator adds N CBCT <N CT After the Transformer system is pre-trained, a randomly initialized Transformer system is added to the GA generator to form a second hybrid network.

[0037] Furthermore, the GA generator and the GB generator have completely identical structures and are symmetrical. The generator consists of 4 convolutions, 2 deconvolutions, and 9 residual modules. The last convolution layer is followed only by the Tanh activation function, while the remaining convolutions and deconvolutions are followed by instance normalization and ReLU activation functions. The residual module contains two convolutions. The first convolution is followed by instance normalization and ReLU activation functions, while the second convolution is followed only by instance normalization. The convolutions are connected through residuals.

[0038] Furthermore, after incorporating the pre-trained Transformer system, the process of generating images is represented by formula (7):

[0039] I OUT =Conv(Up(T(I)+R(I))) (7)

[0040] in

[0041] I = Down(Conv(I) IN (8)

[0042] In formula (7), I OUT The image generated by the generator is represented by Conv(.), where Conv(.) represents a convolution operation with a kernel of 7×7 and a stride of 1, Up(.) represents upsampling consisting of two deconvolution operations with kernels of 3×3 and a stride of 2, T(.) represents the Transformer module, and R(.) represents 9 residual modules; in formula (8), Down(.) represents downsampling consisting of two convolution operations with kernels of 3×3 and a stride of 2, I IN This represents the image input to the generator.

[0043] Furthermore, the loss function of the hybrid network consists of three parts: adversarial loss L GAN Consistency loss L Identity Cyclic consistency loss L Cycle As shown in formula (9):

[0044] L = L GAN +αL Identity +βL Cycle (9)

[0045] in

[0046] L GAN (GA,D Y (x, y) = E x～pdata(x) [D Y (GA(x)) 2 ]+E y～pdata(y) [(D Y (y)-1) 2 ],

[0047] L GAN (GB, D) X (x, y) = E y～pdata(y) [D X (GB(y)) 2 ]+E x～pdata(x) [(D X (x)-1) 2 (10)

[0048] L Identity (GA, GB, x, y) = E x～pdata(x) [||GB(x)-x||1]+E y～pdata(y) [||GA(y)-y||1] (11)

[0049] L Cycle (GA, GB, x, y) = E x～pdata(x) [||GB(GA(x))-x||1]+E y～pdata(y) [||GA(GB(y))-y||1] (12)

[0050] In formula (9), α and β are weight coefficients with values ​​of 5 and 10 respectively, and L is the weight of the balanced network. GAN L Identity and L Cycle In formula (10), E represents the expectation function, pdata(x) and pdata(y) represent the data distributions of CBCT images and registered CT images in the training set, respectively, and x and y represent the CBCT images and registered CT images in the training set, respectively. X and D Y Denotes the discriminator in the hybrid network, which determines whether the data distribution of the images generated by the GB generator and the GA generator are consistent with the CBCT images and the registered CT images in the training set, respectively. Where D... X and D Y The structures are completely identical.

[0051] Furthermore, the discriminator contains 5 convolutions with 4×4 kernels and a stride of 2. The output channels are 64, 128, 256, 512, and 1, respectively. The first convolution is followed by a LeakyReLU activation function, while the second, third, and fourth convolutions are followed by instance normalization and a LeakyReLU activation function. Attached Figure Description

[0052] Figure 1 The diagram shown is a process diagram of the pre-trained Transformer system of the present invention;

[0053] Figure 2The diagram shown is a schematic representation of the self-supervised network of the present invention.

[0054] Figure 3 The diagram shown is a structural schematic of the Transformer module of this invention;

[0055] Figure 4 The diagram shown is a schematic of the generator of the present invention;

[0056] Figure 5 The diagram shown is a schematic of the hybrid network structure of the present invention;

[0057] Figure 6 The diagram shown is a schematic representation of the discriminator of the present invention.

[0058] Figure 7 The diagram shown is a structural schematic of the beam baffle of the present invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely 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 without creative effort are within the scope of protection of the present invention.

[0060] Please see the appendix Figure 1 A low-dose CBCT imaging method includes the following steps:

[0061] Step 1: Take a portion of the data from the training set to ensure that the number of remaining CBCT images is the same as the number of registered CT images, thus avoiding imbalanced training data.

[0062] Step 2: The extracted data is pre-trained in a self-supervised manner to avoid network training difficulties caused by adding a randomly initialized Transformer system to the CycleGAN system.

[0063] Step 3: Add the pre-trained Transformer system to the CycleGAN system to obtain a hybrid network, and use the remaining data for unsupervised training.

[0064] Example 1: Pre-trained Transformer System

[0065] In step one, the training set includes training set A and training set B. The data in training set A consists of CBCT images, and the number of images is denoted as N. CBCT The training set B consists of registered CT images, denoted as N. CT When N CBCT >N CTAt that time, N are randomly selected from the training set A. CBCT -N CT Zhang images are used to pre-train a Transformer system in a self-supervised manner; when N... CBCT <N CT At that time, N are randomly selected from the training set B. CT -N CBCT The Transformer system is pre-trained using images through a self-supervised approach.

[0066] As attached Figure 2 As shown, the Transformer system contains 5 Transformer modules. It guides the training of the self-supervised network by calculating the mean squared error (MSE) of the input and output images. This process is expressed by the formula:

[0067] In the formula, F(X) i ) represents the network output image, F(.) represents a self-supervised network, which contains 3 convolutions, 3 deconvolutions and 5 Transformer modules. The stride of the convolutions and deconvolutions is 2. It is followed by instance normalization and ReLU activation function. The number of output channels is 64, 128, 256, 256, 128 and 1 respectively.

[0068] The process of pre-training a Transformer system includes:

[0069] Step 2.1) Convert the input CBCT image X∈R H×W Divide the sample into multiple non-overlapping square patches of size n×n, forming a set P = (p1, p2, p3, ..., pn). N-2 p N-1 p N ), where H and W represent the height and width of the image, respectively, and p i This represents each square patch, and N represents the number of square patches.

[0070] Step 2.2) Randomly and uniformly replace part of the square patch with a learnable vector l to obtain the occluded CBCT image, forming a set P. M =(p1,l,p3,...,l,l,p N )

[0071] Step 2.3) After downsampling, the image is input into the Transformer system for inter-patch information exchange, and finally upsampling is performed to output the predicted CBCT image:

[0072] I pCBCT =UP(T(DOWN(I mCBCT )) N (1)

[0073] In formula (1), I pCBCT I represents the CBCT image predicted by the self-supervised network. mCBCT The input to the self-supervised network is the occluded CBCT image, where UP(.) represents the upsampling function, DOWN(.) represents the downsampling function, T(.) represents the Transformer module, and N represents the number of modules.

[0074] The Transformer module consists of three parts: multi-head self-attention, multilayer perceptron, and normalization layer (see attached). Figure 3 (As shown).

[0075] The calculation process of the Transformer module can be expressed as formula (2):

[0076]

[0077]

[0078] in

[0079] MSA=FC(Concat(head1,...,head h (3)

[0080]

[0081] MLP(.)=FC(σ(FC(.))) (5)

[0082]

[0083] In formula (2), Z l Indicates input features, Z represents intermediate features. l+1 In Equation (3), Concat(.) represents the concatenation operation, FC represents the fully connected layer, and head represents the output of self-attention. In Equation (4), Q, K, and V are obtained from the normalized features through different linear transformations, T represents the transpose, and d represents the output of self-attention. K The dimension of K is represented; in formula (5), σ represents the GELU activation function; in formula (6), x represents the input vector, u and σ represent the mean and standard deviation of x respectively, γ and τ are learnable parameters, representing the scaling factor and bias respectively.

[0084] Example 2: Training a Hybrid Network

[0085] The CycleGAN system includes two generators (as shown in the appendix) Figure 4 As shown):

[0086] The GA generator takes a CBCT image as input and outputs an enhanced CBCT image with a data distribution similar to the registered CT images in the training set; it is used to generate high-quality pseudo-CBCT images (i.e., images obtained after suppressing noise and artifacts in CBCT images).

[0087] The GB generator takes a registered CT image as input and outputs a CT image with noise and artifacts. Its data distribution is similar to that of the CBCT images in the training set. It is used to generate PCBCT images with noise and artifacts (this process is equivalent to the reverse process of the GA generator removing noise and artifacts from the CBCT image to generate a pseudo PCBCT. It can be seen that the two generators have completely identical structures and are symmetrical).

[0088] The Transformer system in the pre-training phase uses CBCT images for self-supervised training, which has more prior information about CBCT images compared to conventional training. This effectively avoids the undesirable effect of random initial weights in the model. Since the CycleGAN system's performance in training with unpaired data is related to the difference in data, when training with paired data, if the number of images in the two sets of data (assuming training set A consists of CBCT images and training set B consists of CBCT images) differs significantly, the training performance of the CycleGAN system will become very poor. Therefore, this invention adds the pre-trained Transformer system to the CycleGAN system to obtain a hybrid network with good training performance. The pre-training process involves taking the excess data from training set A (where A has more images than B) or training set B (where A has fewer images than B) and using it to train the GB generator or GA generator of the Transformer system that will be added to the CycleGAN system separately. Then, the remaining dataset is used for unsupervised training, thus solving the above problem.

[0089] The specific steps are as follows:

[0090] If N CBCT >N CT GA generator adds N CBCT >N CT After the Transformer system is pre-trained, a randomly initialized Transformer system (not pre-trained) is added to the GB generator to form the first hybrid network, which directly enhances the network's ability to improve the quality of CBCT images.

[0091] If N CBCT <N CT GB generator adds N CBCT <N CTAfter the Transformer system is pre-trained, a randomly initialized Transformer system (not pre-trained) is added to the GA generator to form a second hybrid network, which indirectly enhances the network's ability to improve the quality of CBCT images.

[0092] The generator consists of 4 convolutions, 2 deconvolutions, and 9 residual modules (see attached). Figure 5 As shown in the diagram, the last convolutional layer has only the Tanh activation function, while the remaining convolutions and deconvolutions have instance normalization and ReLU activation functions. The residual module contains two convolutions: the first convolution has instance normalization and ReLU activation functions, while the second convolution has only instance normalization. The convolutions are connected through residuals.

[0093] After adding the pre-trained Transformer system, the process of generating images is represented by formula (7):

[0094] I OUT =Conv(Up(T(I)+R(I))) (7)

[0095] in

[0096] I = Down(Conv(I) IN (8)

[0097] In formula (7), I OUT The image generated by the generator is represented by Conv(.), where Conv(.) represents a convolution operation with a kernel of 7×7 and a stride of 1, Up(.) represents upsampling consisting of two deconvolution operations with kernels of 3×3 and a stride of 2, T(.) represents the Transformer module, and R(.) represents 9 residual modules; in formula (8), Down(.) represents downsampling consisting of two convolution operations with kernels of 3×3 and a stride of 2, I IN This represents the image input to the generator.

[0098] The loss function of a hybrid network consists of three parts: adversarial loss L GAN Consistency loss L Identity Cyclic consistency loss L Cycle As shown in formula (9):

[0099] L = L GAN +αL Identity +βL Cycle (9)

[0100] in

[0101] L GAN (GA,D Y (x, y) = E x～pdata(x)[D Y (GA(x)) 2 ]+E y～pdata(y) [(D Y (y)-1) 2 ],

[0102] L GAN (GB, D) X (x, y) = E y～pdata(y) [D X (GB(y)) 2 ]+E x～pdata(x) [(D X (x)-1) 2 (10)

[0103] L Identity (GA, GB, x, y) = E x～pdata(x) [||GB(x)-x||1]+E y～pdata(y) [||GA(y)-y||1] (11)

[0104] L Cycle (GA, GB, x, y) = E x～pdata(x) [||GB(GA(x))-x||1]+E y～pdata(y) [||GA(GB(y))-y||1] (12)

[0105] In formula (9), α and β are weight coefficients with values ​​of 5 and 10 respectively, and L is the weight of the balanced network. GAN L Identity and L Cycle In formula (10), E represents the expectation function, pdata(x) and pdata(y) represent the data distributions of CBCT images and registered CT images in the training set, respectively, and x and y represent the CBCT images and registered CT images in the training set, respectively. X and D Y This represents the discriminator in the network, which determines whether the data distribution of the images generated by the GB generator and the GA generator is consistent with the CBCT images and the registered CT images in the training set, respectively. Where D... X and D Y The structures are completely identical.

[0106] The discriminator contains 5 convolutions (as shown in the attached diagram). Figure 6 As shown in the figure, the convolution kernel is 4×4, the stride is 2, and the output channels are 64, 128, 256, 512, and 1 respectively. The first convolution is followed by the LeakyReLU activation function, and the second, third, and fourth convolutions are followed by instance normalization and the LeakyReLU activation function.

[0107] Example 3

[0108] like Figure 7 As shown, the present invention provides an X-ray beam baffle to obtain an obscured CBCT image.

[0109] To reduce the X-ray dose received by patients, a beam baffle is installed at the output end of the X-ray tube. The beam baffle consists of a lead part and a transmission part. The lead part is used to completely block X-rays, and the transmission part is used to completely transmit X-rays.

[0110] The lead elements are small lead balls or small lead blocks, evenly arranged inside the beam baffle. The aspect ratio of the beam baffle is W:H, and the number of lead elements is the same as the number of learnable vectors l.

[0111] The transparent section includes two interlocking transparent plates, with a cavity between the two transparent plates for accommodating the lead section. The size of the transparent section is the same as the size of the beam baffle.

[0112] To facilitate labeling and positioning, the horizontal and vertical axes of the transparent plate are labeled with Arabic numerals and English letters respectively, forming a set Q. Based on the well-trained hybrid network, the learnable vector l in P is statistically analyzed. M =(p1,l,p3,...,l,l,p N By replacing the position in the set Q, the optimal position of the cavity for placing the lead part is obtained.

[0113] Compared to supervised learning methods, this invention employs unsupervised learning, avoiding the difficulty of pairing CBCT images with registered CT images. This invention utilizes the CycleGAN network to enhance CBCT images (suppressing noise and artifacts in CBCT images). However, CycleGAN has two problems: 1. It is essentially a convolutional neural network and lacks the ability to model global dependencies. 2. When the number of images in training set A (CBCT image set) and training set B (registered PCT image set) differs significantly, it hinders the training of CycleGAN.

[0114] To address the first problem, this invention enhances CycleGAN's ability to model global dependencies by adding a transformer system to the two generators, GA and GB. However, since the transformer requires a large amount of data to train and function effectively, this invention extracts the extra images from training set A or B and pre-trains the transformer system in a self-supervised manner. Then, the pre-trained transformer system is added to GA (to generate enhanced CBCT images) or GB (to generate CT images with noise and artifacts). Finally, unsupervised training is performed using the remaining dataset, thus simultaneously solving the second problem mentioned above.

[0115] The present invention provides an X-ray beam baffle that can be adjusted by adding or removing lead sections according to the dose requirements of different scenarios, thereby adjusting the output of X-ray dose and greatly reducing the X-ray dose received by patients under the same image acquisition conditions.

[0116] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0117] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is limited by the appended claims and their equivalents.

Claims

1. An X-ray beam baffle, characterized in that, include Obtain the occluded CBCT image; A beam baffle is provided at the output end of the X-ray tube. The beam baffle consists of a lead part and a transparent part. The lead part is used to completely block the X-rays, and the transparent part is used to completely transmit the X-rays. The lead elements are small lead balls or small lead blocks, evenly arranged inside the beam baffle, and the aspect ratio of the beam baffle is W:H; the number of lead elements is the same as the number of learnable vectors l. The transparent portion includes two interlocking transparent plates, with a cavity for accommodating the lead portion between the two transparent plates. The size of the transparent portion is the same as the size of the beam baffle. The transparent plates are sequentially marked with Arabic numerals and English letters in the transverse and longitudinal directions, respectively, to form a set Q, a mixed network with good training effect obtained according to a low-dose CBCT imaging method, and a replaceable position of a learnable vector l in P M =(p1,l,p3,…,l,l,p N ) is obtained, and the set Q is marked to obtain an optimal position of a containing cavity in which the lead part is placed. The low-dose CBCT imaging method includes the following steps: Step 1: Take a portion of the data from the training set to ensure that the number of remaining CBCT images is the same as the number of registered CT images, thus avoiding imbalanced training data. Step 2: The extracted data is pre-trained in a self-supervised manner to avoid network training difficulties caused by adding a randomly initialized Transformer system to the CycleGAN system. Step 3: Add the pre-trained Transformer system to the CycleGAN system to obtain a hybrid network, and use the remaining data for unsupervised training; The Transformer system contains N Transformer modules. Step two, the pre-training of the Transformer system, includes: Step 2.1) Divide the input CBCT image X e R H×W into a set of non-overlapping square patches of size n x n, denoted as P = (p1, p2, p3,..., p N-2 N-1 N N, where H and W denote the height and width of the image respectively, p i denotes each square patch, and N denotes the number of square patches.​​ Step 2.2) Replace some of the square patches with the learnable vectors l uniformly at random to get occluded CBCT images, which form a set P M = (p1, l, p3,..., l, l, p N ) Step 2.3) After downsampling, the image is input into the Transformer system for inter-patch information exchange, and finally upsampling is performed to output the predicted CBCT image: I pCBCT = UP(T(DOWN(I mCBCT )) N ) (1) In formula (1), I pCBCT denotes the CBCT image predicted by the self-supervised network, I mCBCT denotes the CBCT image input to the self-supervised network, UP(.) denotes an up-sampling function, DOWN(.) denotes a down-sampling function, T(.) denotes a Transformer module, and N denotes the number of modules.

2. The X-ray beam baffle according to claim 1, characterized in that, The training set mentioned in step one includes training set A and training set B. The data in training set A consists of CBCT images, and the number of images is denoted as N. CBCT The training set B consists of registered CT images, and its number is denoted as N. CT When N CBCT >N CT At that time, N are randomly selected from the training set A. CBCT -N CT Zhang images are used to pre-train a Transformer system in a self-supervised manner; when N... CBCT <N CT At that time, N are randomly selected from the training set B. CT -N CBCT The Transformer system is pre-trained using a self-supervised approach with a number of images.

3. An X-ray beam baffle according to claim 1, characterized in that, The Transformer module consists of three parts: multi-head self-attention, multilayer perceptron, and normalization layer. The calculation process of the Transformer module can be expressed as formula (2): in MSA=FC(Concat(head1,…,head h )) (3) MLP(.)=FC(σ(FC(.))) (5) In formula (2), Z k Indicates input features, Z represents intermediate features. l+1 In Equation (3), Concat(.) represents the concatenation operation, FC represents the fully connected layer, and head represents the output of self-attention. In Equation (4), Q, K, and V are obtained from the normalized features through different linear transformations, T represents the transpose, and d represents the output of self-attention. K The dimension of K is represented; in formula (5), σ represents the GELU activation function; in formula (6), x represents the input vector, u and σ represent the mean and standard deviation of x respectively, γ and τ are learnable parameters, representing the scaling factor and bias respectively.

4. An X-ray beam baffle according to claim 1, characterized in that, The CycleGAN system includes two types of generators: The GA generator takes a CBCT image as input and outputs an enhanced CBCT image, whose data distribution is similar to the registered CT images in training set B. The GB generator takes a registered CT image as input and outputs a CT image with noise and artifacts, whose data distribution is similar to that of the CBCT images in training set A. When the GA generator adds N CBCT >N CT After the Transformer system is pre-trained, a randomly initialized Transformer system is added to the GB generator to form the first hybrid network. When the GB generator is added to N CBCT <N CT After the Transformer system is pre-trained, a randomly initialized Transformer system is added to the GA generator to form a second hybrid network.

5. An X-ray beam baffle according to claim 4, characterized in that, The GA generator and the GB generator have completely identical structures and are symmetrical. The generator consists of 4 convolutions, 2 deconvolutions, and 9 residual modules. The last convolution layer is followed only by the Tanh activation function, while the remaining convolutions and deconvolutions are followed by instance normalization and ReLU activation functions. The residual module contains two convolutions. The first convolution is followed by instance normalization and ReLU activation functions, while the second convolution is followed only by instance normalization. The convolutions are connected through residuals.

6. An X-ray beam baffle according to claim 5, characterized in that, After adding the pre-trained Transformer system, the process of generating images is represented by formula (7): I OUT =Conv(Up(T(I)+R(I))) (7) in I=Down(Conv(I IN )) (8) In formula (7), I OUT The image generated by the generator is represented by Conv(.), where Conv(.) represents a convolution operation with a kernel of 7×7 and a stride of 1, Up(.) represents upsampling consisting of two deconvolution operations with kernels of 3×3 and a stride of 2, T(.) represents the Transformer module, and R(.) represents 9 residual modules; in formula (8), Down(.) represents downsampling consisting of two convolution operations with kernels of 3×3 and a stride of 2, I IN This represents the image input to the generator.

7. An X-ray beam baffle according to claim 6, characterized in that, The loss function of the hybrid network consists of three parts: adversarial loss L GAN Consistency loss L Identity Cyclic consistency loss L Cycle As shown in formula (9): L=L GAN +αL Identity +βL Cycle (9) in L GAN (GA,D Y ,x,y)=E x～pdata(x) [D Y (GA(x)) 2 ]+E y～pdata(y) [(D Y (y)-1) 2 ], L GAN (GB,D X ,x,y)=E y～pdata(y) [D X (GB(y)) 2 ]+E x～pdata(x) [(D X (x)-1) 2 ] (10) L Identity (GA,GB,x,y)=E x～pdata(x) [‖GB(x)-x‖1]+E y～pdata(y) [‖GA(y)-y‖1] (11) L Cycle (GA,GB,x,y)=E x～pdata(x) [||GB(GA(x))-x||1]+E y～pdata(y) In formula (9), α and β are weight coefficients with values ​​of 5 and 10 respectively. The L of the balanced network is... GAN L Identity and L Cycle In formula (10), E represents the expectation function, pdata(x) and pdata(y) represent the data distributions of CBCT images and registered CT images in the training set, respectively, and x and y represent the CBCT images and registered CT images in the training set, respectively. X and D Y The discriminator in the hybrid network determines whether the data distribution of the images generated by the GB generator and the GA generator is consistent with the CBCT images and the registered CT images in the training set, where D X and D y The structures are completely identical.

8. An X-ray beam baffle according to claim 7, characterized in that, The discriminator contains 5 convolutions with a kernel size of 4×4 and a stride of 2. The output channels are 64, 128, 256, 512, and 1, respectively. The first convolution is followed by a LeakyReLU activation function, and the second, third, and fourth convolutions are followed by instance normalization and a LeakyReLU activation function.

Images