Synthetic CT generation method and system, storage medium and electronic equipment

Abstract

The invention provides a synthetic CT generation method and system, a storage medium and electronic equipment, and the method comprises the steps: 1, collecting unpaired CBCT and CT data, carrying out the preprocessing, and constructing a data set; step 2, sampling CBCT images from the data set, initializing a path, and randomly setting the number of interpolation steps; 3, for each interpolation step, predicting and synthesizing a CT image through a wavelet condition generator, and updating an intermediate state in combination with weighted mixing and time-related Gaussian noise; 4, respectively calculating three types of losses, and jointly optimizing a wavelet condition generator, a discriminator, an energy network and a feature extractor; 5, repeating the step 2 to the step 4 until the training cycle is completed; and step 6, fixing the number of interpolation steps, inputting a new CBCT image, recursively executing interpolation and noise injection through the trained wavelet condition generator, and outputting a final synthesized CT image, thereby solving the problem of high-fidelity generation from CBCT to sCT without paired data.

Description

Technical Field

[0001] This invention belongs to the field of synthetic CT generation technology, and specifically relates to a synthetic CT generation method, system, storage medium and electronic device. Background Technology

[0002] Radiation therapy is one of the core methods of cancer treatment, and its accuracy directly affects tumor control and the protection of healthy tissue. In image-guided radiotherapy, cone-beam computed tomography (CBCT) has become a key tool for daily patient localization and monitoring because it can be integrated with linear accelerators and enables real-time anatomical assessment. However, CBCT is affected by factors such as scattering artifacts, radiation hardening, and low-dose protocols, resulting in problems such as low soft tissue contrast and unreliable Huntsfield unit (HU) values. Therefore, it cannot be directly used for dose recalculation required for adaptive radiotherapy—its dose deviation from planning CT (pCT) can reach more than 10%, seriously jeopardizing treatment safety and effectiveness. To address the aforementioned issues, synthetic CT (sCT) generation technology has emerged. Its core principle is to establish a feature mapping between CBCT and the CT domain. By suppressing CBCT artifacts and restoring tissue density information, it generates high-quality images equivalent to pCT. Existing technologies are mainly divided into three categories: 1. Traditional methods based on physical correction (such as scattering modeling, Monte Carlo simulation, etc.): have high computational complexity and poor robustness, making it difficult to meet the real-time and accuracy requirements of clinical practice; 1. Supervised deep learning methods (such as U-Net, FCN, etc.): rely on strictly aligned paired CBCT-CT data, but in clinical practice, patients' respiratory movements and changes in body position can easily lead to data misalignment, which limits the model's generalization ability; 2. Unsupervised / semi-supervised methods: represented by GANs (such as CycleGAN) and traditional diffusion models (DMs). GANs have problems such as training instability, pattern collapse, and artifact residue. Traditional diffusion models are constrained by Gaussian noise priors, making it difficult to directly model the complex distribution mapping between CBCT and CT, and are prone to introducing physiologically unreasonable anatomical structures. Summary of the Invention

[0003] Based on this, embodiments of the present invention provide a synthetic CT generation method, system, storage medium, and electronic device, aiming to achieve high-precision and robust conversion from CBCT to sCT without paired data, and to provide reliable anatomical image support for adaptive radiotherapy.

[0004] A first aspect of this invention provides a method for generating synthetic CT scans, applied to wavelet neural Schrödinger bridge diffusion, the method comprising: Step 1: Collect unpaired CBCT and CT data, perform preprocessing, and construct the dataset; Step 2: Sample CBCT images from the dataset, initialize the path, and randomly set the number of interpolation steps; Step 3: For each interpolation step, the synthesized CT image is predicted by the wavelet conditional generator, and the intermediate state is updated by combining weighted mixing with time-related Gaussian noise. Step 4: Calculate the three types of losses separately, and jointly optimize the wavelet conditional generator, discriminator, energy network, and feature extractor; Step 5: Repeat steps 2 through 4 until the training cycle is complete; Step 6: Fix the number of interpolation steps, input a new CBCT image, and recursively perform interpolation and noise injection through a trained wavelet condition generator to output the final synthesized CT image.

[0005] Furthermore, the wavelet condition generator is used to provide additional noise embedding to the features of the wavelet residual through a time step, forming a conditional wavelet residual block for diffusion.

[0006] Furthermore, the step of updating the intermediate state by predicting the synthesized CT image through a wavelet conditional generator and combining it with weighted mixing of time-related Gaussian noise for each interpolation step includes: The input CBCT image is decomposed into low-frequency and high-frequency components by discrete wavelet transform to separate anatomical structures from artifacts. The time step is converted into a time embedding vector by sinusoidal embedding, and the domain information is converted into a domain embedding vector by linear transformation. The time embedding vector and the domain embedding vector are then integrated into the wavelet decomposition features by affine transformation to obtain the integrated result. The integrated results are then used to reconstruct the feature map through inverse wavelet transform, thus synthesizing a CT image. Based on the synthesized CT image, the intermediate state for the next round is calculated by combining the interpolation ratio and Gaussian noise, and then updated.

[0007] Furthermore, the wavelet condition generator includes a convolutional layer, a wavelet residual block group, and an upsampling network connected in sequence. The convolutional layer includes a first sub-convolutional layer, a second sub-convolutional layer, and a third sub-convolutional layer. The first sub-convolutional layer and the third sub-convolutional layer each include a 7×7 convolutional unit, a batch normalization unit, and a corrected linear unit. The second sub-convolutional layer includes a 3×3 convolutional unit, a batch normalization unit, and a corrected linear unit. The wavelet residual block group comprises three cascaded wavelet residual blocks. In each wavelet residual block, the input is the feature map of the current level, the time step of the diffusion process, and the domain condition. The feature map of the current level is processed by wavelet convolution + instance normalization + activation. The time step of the diffusion process and the domain condition are encoded into vectors of the same dimension by linear layers. The encoded feature map of the current level, the time step of the diffusion process, and the domain condition are added together. During the addition process, affine modulation is performed through scaling factors and offset factors to guide the features with time / domain information. Finally, the modulated features are processed again by wavelet convolution + instance normalization + activation, and finally added to the original input feature map of the current level to output the updated features.

[0008] Furthermore, in the step of calculating the next round's intermediate state based on the synthesized CT image, combined with the interpolation ratio and Gaussian noise, for updating, the expression for calculating the next round's intermediate state is: ; ; in, Let s be the intermediate state corresponding to step i+1, s be the normalized interpolation ratio, and t be the intermediate state. i Let t be the time step parameter for the i-th step. i+1 For the (i+1)th step time step parameter, This represents the current intermediate state corresponding to step i. The image is a synthetic CT image predicted by a wavelet conditional generator, where ε represents Gaussian noise, introducing randomness into the diffusion process to ensure path diversity. Let I represent a standard multivariate normal distribution with a mean of 0 and a covariance matrix that is the identity matrix.

[0009] Furthermore, in the step of calculating the three types of losses separately and jointly optimizing the wavelet conditional generator, discriminator, energy network, and feature extractor, the three types of losses include adversarial loss, Schrödinger bridge loss, and regularization loss. The expression for the adversarial loss is: ; To counteract the loss, CE is the cross-entropy loss. Let x1 be the probability that the discriminator considers x1 to be a real sample. The discriminator considers It is the probability of being a real sample; The expression for the Schrödinger bridge loss is: ; The Schrödinger bridge loss is represented by t, which is the time step parameter of the diffusion process, ranging from [0,1], and E represents the energy network. For the input sample pairs of the energy network, x t This is an intermediate state in the diffusion process. The synthesized CT image is predicted by the wavelet conditional generator. For negative samples, The contrast energy term for negative samples. This is the L2 loss term; The expression for the regularization loss is: ; The loss is the regularization loss, and L is the total number of layers in the feature extractor. To estimate the loss for information noise contrast, Let q be the l-th layer feature extracted by the feature extractor F from the source image, and let q be the query. Let k be the l-th layer feature extracted by the feature extractor F from the generated image; ; For the total loss, , , User-defined balancing weights.

[0010] Furthermore, in the step of calculating the three types of losses separately and jointly optimizing the wavelet conditional generator, discriminator, energy network, and feature extractor, the total loss is used to optimize the wavelet conditional generator and feature extractor, while the adversarial loss is used to update the discriminator, and the Schrödinger bridge loss is used to optimize the energy network.

[0011] A second aspect of this invention provides a synthetic CT generation system for implementing the synthetic CT generation method provided in the first aspect, the system comprising: The building module is used to acquire unpaired CBCT and CT data, and after preprocessing, to build the dataset; A module is used to sample CBCT images from the dataset, initialize the path, and randomly set the number of interpolation steps; The prediction module is used to predict the synthesized CT image for each interpolation step using a wavelet conditional generator, and update the intermediate state by combining weighted mixing with time-related Gaussian noise. The optimization module is used to calculate the three types of loss separately and jointly optimize the wavelet conditional generator, discriminator, energy network and feature extractor. The training module is used to repeatedly use the module to the prediction module until the training cycle is complete; The input module is used to fix the number of interpolation steps. It takes a new CBCT image as input, recursively performs interpolation and noise injection through a trained wavelet condition generator, and outputs the final synthesized CT image.

[0012] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the synthetic CT generation method provided in the first aspect.

[0013] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the program to implement the synthetic CT generation method provided in the first aspect.

[0014] This invention provides a synthetic CT generation method, system, storage medium, and electronic device. The method comprises the following steps: Step 1: Acquiring unpaired CBCT and CT data, preprocessing them, and constructing a dataset; Step 2: Sampling CBCT images from the dataset, initializing paths, and randomly setting the number of interpolation steps; Step 3: For each interpolation step, predicting the synthetic CT image using a wavelet conditional generator, and updating the intermediate state by combining weighted mixing with time-correlated Gaussian noise; Step 4: Calculating three types of losses, and jointly optimizing the wavelet conditional generator, discriminator, energy network, and feature extractor; Step 5: Repeating steps 2 to 4 until the training cycle is complete; Step 6: Fixing the number of interpolation steps, inputting a new CBCT image, and recursively performing interpolation and noise injection through the trained wavelet conditional generator to output the final synthetic CT image. Specifically, through the synergistic design of wavelet neural Schrödinger bridge diffusion, wavelet multi-scale decomposition, and composite loss functions, the method solves the problem of high-fidelity generation of CBCT to sCT from unpaired data, demonstrating significant clinical application value and promising prospects for wider application. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the implementation of a synthetic CT generation method according to Embodiment 1 of the present invention. Figure 2 This is a structural block diagram of a synthetic CT generation system provided in Embodiment 2 of the present invention; Figure 3 This is a structural block diagram of an electronic device provided in Embodiment 3 of the present invention. Detailed Implementation

[0016] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0017] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0019] Example 1 Please see Figure 1 , Figure 1 The flowchart of a synthetic CT generation method provided in Embodiment 1 of the present invention is shown. This synthetic CT generation method is applied to wavelet neural Schrödinger bridge diffusion (WNSBD), where a Schrödinger bridge (SB) framework is used to replace the Gaussian prior of the traditional diffusion model. A random transmission path is directly constructed between the source domain (CBCT, distribution π0) and the target domain (CT, distribution π1). Unpaired domain mapping is achieved through Gaussian interpolation, avoiding structural inconsistency issues. In this embodiment of the invention, an unpaired modal bridge diffusion model based on wavelet neural Schrödinger bridge diffusion is proposed. This model can learn the mapping between arbitrary distributions, thereby promoting the transformation between the unpaired source domain (CBCT, π0) and the target domain (sCT, π1). The core objective is to find the most probable path q* connecting π0 and π1 while minimizing the deviation from the reference diffusion process, which can be expressed as: ; Where T represents the endpoint (equivalent to t=1 in the discrete step); x0: T represents the transition from the initial state x0~π0 (CBCT) to the terminal state x T The state sequence of ~π1 (sCT); Q represents the state sequence subject to endpoint conditions q(x0)=π0, q(x T The set of feasible path distributions with constraints π / 1; p(x0:T) is the reference path distribution, usually a Brownian bridge or Wiener process, serving as a priori information for the random trajectory; KL represents the Koolbek-Leibler divergence; q * This represents the optimal transmission path, and π represents the most efficient random mapping from π0 to π1.

[0020] The synthetic CT generation method specifically includes steps S01 to S06.

[0021] Step S01: Collect unpaired CBCT and CT data, perform preprocessing, and construct the dataset.

[0022] Specifically, unpaired CBCT and CT data (including data from different anatomical sites and institutions) were acquired and resampled (to a uniform 1×1×1mm). 3 CBCT-CT data is used to construct training, validation, and test sets using resolution, rigid registration, deformation registration, and body mask segmentation. The training set does not require pairing, while the validation and test sets must be paired.

[0023] Step S02: Sample CBCT images from the dataset, initialize the path, and randomly set the number of interpolation steps.

[0024] Specifically, the training set samples CBCT source images. Initialization path Randomly set the number of interpolation steps .

[0025] Step S03: For each interpolation step, the synthesized CT image is predicted by the wavelet condition generator, and the intermediate state is updated by combining weighted mixing with time-related Gaussian noise.

[0026] It should be noted that for each interpolation step i=0 to n-1, the synthesized CT image is predicted by a wavelet conditional generator, wherein the wavelet conditional generator is used to provide additional noise embedding to the features of the wavelet residual through the time step, forming a conditional wavelet residual block for diffusion.

[0027] In this embodiment of the invention, the input CBCT image is decomposed into low-frequency and high-frequency components using discrete wavelet transform to separate anatomical structures from artifacts. The expression is as follows: ; Among them, A t This indicates a low-frequency component, primarily capturing global structural information, D t Indicates high-frequency components. D t,h D t,v D t,d These represent the local details of the high-frequency components in the horizontal, vertical, and diagonal directions, respectively, and WaveletDecompose is the wavelet decomposition function; The time step is converted into a time embedding vector using sinusoidal embedding, expressed as follows: ; e t Let t be the time embedding vector, and t be the time step parameter of the diffusion process. These are preset angular frequency parameters; These embeddings modulate wavelet features through affine transformation. Then, the domain information is converted into domain embedding vectors through linear transformation. The time embedding vector and the domain embedding vector are then integrated into the wavelet decomposition features through affine transformation to obtain the integrated result, expressed as: ; e t e is the temporal embedding vector. z For field embedding vectors, and Let be a mapping function that can be implemented through a linear transformation. Preserve tumor / organ morphology. Suppress noise without blurring edges; The integrated result is then subjected to inverse wavelet transform to reconstruct the feature map, i.e., synthesize the CT image, as expressed by: ; For synthesizing CT images, InvWavelet is the inverse wavelet transform function; Based on the synthesized CT image, the next intermediate state is calculated by combining the interpolation ratio and Gaussian noise for updating. The expression for calculating the next intermediate state is: ; ; in, Let s be the intermediate state corresponding to step i+1, s be the normalized interpolation ratio, and t be the intermediate state. i Let t be the time step parameter for the i-th step. i+1 For the (i+1)th step time step parameter, This represents the current intermediate state corresponding to step i. The image is a synthetic CT image predicted by a wavelet conditional generator, where ε represents Gaussian noise, introducing randomness into the diffusion process to ensure path diversity. Representing a standard multivariate normal distribution with a mean of 0 and a covariance matrix of identity matrix I, it can be understood that in the above process, at each step i, the wavelet conditional generator predicts the terminal image. Then, the next intermediate image is updated by random interpolation.

[0028] More specifically, the wavelet condition generator includes a convolutional layer, a wavelet residual block group, and an upsampling network connected in sequence. The convolutional layer includes a first sub-convolutional layer, a second sub-convolutional layer, and a third sub-convolutional layer. The first sub-convolutional layer and the third sub-convolutional layer each include a 7×7 convolutional unit, a batch normalization unit, and a corrected linear unit. The second sub-convolutional layer includes a 3×3 convolutional unit, a batch normalization unit, and a corrected linear unit. The wavelet residual block group comprises three cascaded wavelet residual blocks. In each wavelet residual block, the input is the feature map of the current level, the time step of the diffusion process, and the domain condition. The feature map of the current level is processed by wavelet convolution + instance normalization + activation. The time step of the diffusion process and the domain condition are encoded into vectors of the same dimension by linear layers. The encoded feature map of the current level, the time step of the diffusion process, and the domain condition are added together. During the addition process, affine modulation is performed through scaling factors and offset factors to guide the features with time / domain information. Finally, the modulated features are processed again by wavelet convolution + instance normalization + activation, and finally added to the original input feature map of the current level to output the updated features.

[0029] Step S04: Calculate the three types of losses respectively, and jointly optimize the wavelet conditional generator, discriminator, energy network and feature extractor.

[0030] It should be noted that the three types of loss include adversarial loss, Schrödinger bridge loss, and regularization loss. The expression for the adversarial loss is as follows: ; To counteract the loss, CE is the cross-entropy loss. Let x1 be the probability that the discriminator considers x1 to be a real sample. The discriminator considers It is the probability of being a real sample; The expression for the Schrödinger bridge loss is: ; The Schrödinger bridge loss is represented by t, which is the time step parameter of the diffusion process, ranging from [0,1], and E represents the energy network. For the input sample pairs of the energy network, x t This is an intermediate state in the diffusion process. The synthesized CT image is predicted by the wavelet conditional generator. For negative samples, The contrast energy term for negative samples. This is the L2 loss term; The expression for the regularization loss is: ; The loss is the regularization loss, and L is the total number of layers in the feature extractor. To estimate the loss for information noise contrast, Let q be the l-th layer feature extracted by the feature extractor F from the source image, and let q be the query. Let k be the l-th layer feature extracted by the feature extractor F from the generated image; ; For the total loss, , , User-defined balancing weights.

[0031] Specifically, the wavelet conditional generator and feature extractor are optimized using total loss, the discriminator is updated using adversarial loss, and the energy network is optimized using Schrödinger bridge loss.

[0032] Step S05: Repeat steps S02 to S04 until the training cycle is completed.

[0033] Step S06: Fix the number of interpolation steps, input a new CBCT image, recursively perform interpolation and noise injection through the trained wavelet condition generator, and output the final synthesized CT image.

[0034] In summary, the synthetic CT generation method in the above embodiments of the present invention comprises the following steps: Step 1, acquiring unpaired CBCT and CT data, preprocessing them, and constructing a dataset; Step 2, sampling CBCT images from the dataset, initializing the path, and randomly setting the number of interpolation steps; Step 3, for each interpolation step, predicting the synthetic CT image using a wavelet conditional generator, and updating the intermediate state by combining weighted mixing with time-correlated Gaussian noise; Step 4, calculating three types of losses respectively, and jointly optimizing the wavelet conditional generator, discriminator, energy network, and feature extractor; Step 5, repeating steps 2 to 4 until the training cycle is completed; Step 6, fixing the number of interpolation steps, inputting a new CBCT image, recursively performing interpolation and noise injection through the trained wavelet conditional generator, and outputting the final synthetic CT image. Specifically, through the synergistic design of wavelet neural Schrödinger bridge diffusion, wavelet multi-scale decomposition, and composite loss function, the method solves the problem of high-fidelity generation of CBCT to sCT under unpaired data, and has significant clinical application value and promotion prospects.

[0035] Example 2 Please see Figure 2 , Figure 2This is a structural block diagram of a synthetic CT generation system provided in Embodiment 2 of the present invention. The synthetic CT generation system 200 includes: a construction module 21, an adoption module 22, a prediction module 23, an optimization module 24, a training module 25, and an input module 26, wherein: Module 21 is used to acquire unpaired CBCT and CT data, and after preprocessing, to construct the dataset. Module 22 is used to sample CBCT images from the dataset, initialize the path, and randomly set the number of interpolation steps; The prediction module 23 is used to predict the synthesized CT image for each interpolation step by using a wavelet conditional generator, and to update the intermediate state by combining weighted mixing with time-related Gaussian noise. The wavelet conditional generator is used to provide additional noise embedding to the features of the wavelet residuals through the time step to form a conditional wavelet residual block for diffusion. In addition, the wavelet conditional generator includes a convolutional layer, a group of wavelet residual blocks, and an upsampling network connected in sequence. The convolutional layer includes a first sub-convolutional layer, a second sub-convolutional layer, and a third sub-convolutional layer. The first sub-convolutional layer and the third sub-convolutional layer each include a 7×7 convolutional unit, a batch normalization unit, and a corrected linear unit. The second sub-convolutional layer includes a 3×3 convolutional unit, a batch normalization unit, and a corrected linear unit. The wavelet residual block group consists of three cascaded wavelet residual blocks. In each wavelet residual block, the input is the feature map of the current level, the time step of the diffusion process, and the domain condition. The feature map of the current level is processed by wavelet convolution + instance normalization + activation. The time step of the diffusion process and the domain condition are encoded into vectors of the same dimension by linear layers. The encoded feature map of the current level, the time step of the diffusion process, and the domain condition are added together. During the addition process, affine modulation is performed through scaling factors and offset factors to guide the features with time / domain information. Finally, the modulated features are processed again by wavelet convolution + instance normalization + activation, and finally added to the original input feature map of the current level to output the updated features. Optimization module 24 is used to calculate three types of losses separately and jointly optimize the wavelet conditional generator, discriminator, energy network, and feature extractor. The three types of losses include adversarial loss, Schrödinger bridge loss, and regularization loss. The expression for the adversarial loss is: ; To counteract the loss, CE is the cross-entropy loss. Let x1 be the probability that the discriminator considers x1 to be a real sample. The discriminator considers It is the probability of being a real sample; The expression for the Schrödinger bridge loss is: ; The Schrödinger bridge loss is represented by t, which is the time step parameter of the diffusion process, ranging from [0,1], and E represents the energy network. For the input sample pairs of the energy network, x t This is an intermediate state in the diffusion process. The synthesized CT image is predicted by the wavelet conditional generator. For negative samples, The contrast energy term for negative samples. This is the L2 loss term; The expression for the regularization loss is: ; The loss is the regularization loss, and L is the total number of layers in the feature extractor. To estimate the loss for information noise contrast, Let q be the l-th layer feature extracted by the feature extractor F from the source image, and let q be the query. Let k be the l-th layer feature extracted by the feature extractor F from the generated image; ; For the total loss, , , User-defined balancing weights; The wavelet conditional generator and feature extractor are optimized using total loss, while the discriminator is updated using adversarial loss, and the energy network is optimized using Schrödinger bridge loss. Training module 25 is used to repeatedly use the module to the prediction module until the training cycle is completed; Input module 26 is used to fix the number of interpolation steps, input a new CBCT image, recursively perform interpolation and noise injection through a trained wavelet condition generator, and output the final synthesized CT image.

[0036] Furthermore, in some other embodiments of the present invention, the prediction module 23 includes: The decomposition unit is used to decompose the input CBCT image into low-frequency and high-frequency components through discrete wavelet transform, so as to separate anatomical structures from artifacts. An integration unit is used to convert the time step into a time embedding vector through sinusoidal embedding, convert the domain information into a domain embedding vector through linear transformation, and integrate the time embedding vector and the domain embedding vector into the wavelet decomposition features through affine transformation to obtain the integration result. The reconstruction unit is used to reconstruct the feature map, i.e., synthesize the CT image, from the integrated result through inverse wavelet transform. The calculation unit is used to calculate the next round of intermediate states based on the synthesized CT image, combining the interpolation ratio and Gaussian noise, for updating. The expression for calculating the next round of intermediate states is: ; ; in, Let s be the intermediate state corresponding to step i+1, s be the normalized interpolation ratio, and t be the intermediate state. i Let t be the time step parameter for the i-th step. i+1 For the (i+1)th step time step parameter, This represents the current intermediate state corresponding to step i. The image is a synthetic CT image predicted by a wavelet conditional generator, where ε represents Gaussian noise, introducing randomness into the diffusion process to ensure path diversity. Let I represent a standard multivariate normal distribution with a mean of 0 and a covariance matrix that is the identity matrix.

[0037] Example 3 In another aspect, the present invention also proposes an electronic device, please refer to [link to relevant documentation]. Figure 3 The image shows an electronic device according to Embodiment 3 of the present invention, including a memory 20, a processor 10, and a computer program 30 stored in the memory and executable on the processor. When the processor 10 executes the computer program 30, it implements the synthetic CT generation method as described above.

[0038] In some embodiments, the processor 10 may be a central processing unit (CPU), controller, microcontroller, microprocessor or other data processing chip, used to run program code stored in memory 20 or process data, such as executing access restriction programs.

[0039] The memory 20 includes at least one type of readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 20 can be an internal storage unit of an electronic device, such as the hard disk of the electronic device. In other embodiments, the memory 20 can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc. Furthermore, the memory 20 can include both internal and external storage units of the electronic device. The memory 20 can be used not only to store application software and various types of data of the electronic device, but also to temporarily store data that has been output or will be output.

[0040] It should be pointed out that, Figure 3 The structure shown does not constitute a limitation on the electronic device. In other embodiments, the electronic device may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0041] This invention also proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the synthetic CT generation method described above.

[0042] Those skilled in the art will understand that the logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0043] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0044] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0045] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0046] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. A synthetic CT generation method, characterized by, The method is applied to wavelet neural Schrodinger bridge diffusion, and the method comprises: Step one, collecting unpaired CBCT and CT data, and constructing a data set after preprocessing; Step two, sampling CBCT images from the data set and initializing a path, and randomly setting the number of interpolation steps; Step three, for each interpolation step, predicting a synthesized CT image by a wavelet conditional generator, combining weighted mixing and time-dependent Gaussian noise, and updating the intermediate state; Step four, respectively calculating three types of loss, and jointly optimizing the wavelet conditional generator, the discriminator, the energy network and the feature extractor; Step five, repeating steps two to four until the training period is completed; Step six, fixing the number of interpolation steps, inputting a new CBCT image, recursively executing interpolation and noise injection by the trained wavelet conditional generator, and outputting the final synthesized CT image.

2. The synthetic CT generation method of claim 1, wherein, The wavelet conditional generator is used to provide additional noise embedding to the features of the wavelet residual through a time step, and to form a conditional wavelet residual block for diffusion.

3. The synthetic CT generation method of claim 2, wherein, The step of predicting a synthesized CT image by a wavelet conditional generator for each interpolation step, combining weighted mixing and time-dependent Gaussian noise, and updating the intermediate state comprises: Decomposing the input CBCT image into a low-frequency component and a high-frequency component through a discrete wavelet transform to separate the anatomical structure and the artifact; Converting the time step into a time embedding vector through a sinusoidal embedding, converting the domain information into a domain embedding vector through a linear transformation, integrating the time embedding vector and the domain embedding vector into the wavelet decomposition features through an affine transformation to obtain an integration result; Reconstructing the feature map, i.e., the synthesized CT image, by inverse wavelet transform of the integration result; According to the synthesized CT image, combining the interpolation ratio and the Gaussian noise to calculate the next round of intermediate state for updating.

4. The synthetic CT generation method of claim 3, wherein, The wavelet conditional generator comprises a convolution layer, a wavelet residual block group and an up-sampling network connected in sequence, the convolution layer comprises a first sub-convolution layer, a second sub-convolution layer and a third sub-convolution layer, the first sub-convolution layer and the third sub-convolution layer each comprise a 7x7 convolution unit, a batch normalization unit and a rectified linear unit, and the second sub-convolution layer comprises a 3x3 convolution unit, a batch normalization unit and a rectified linear unit; The wavelet residual block group comprises three wavelet residual blocks connected in series, in each wavelet residual block, the feature map of the current level, the time step of the diffusion process and the domain condition are input, the feature map of the current level is processed through wavelet convolution + instance normalization + activation, the time step of the diffusion process and the domain condition are respectively encoded into vectors of the same dimension through a linear layer, the encoded feature map of the current level, the time step of the diffusion process and the domain condition are added, in the adding process, affine modulation is performed through a scaling factor and an offset factor to realize the guidance of time / domain information to the features, and finally the modulated features are again processed through wavelet convolution + instance normalization + activation, and finally added to the original input feature map of the current level to output the updated features.

5. The synthetic CT generation method of claim 4, wherein, In the step of calculating the next round of intermediate state according to the synthetic CT image, combining the interpolation ratio and the Gaussian noise to update, the expression of the next round of intermediate state is: ； ； wherein, is the intermediate state corresponding to the i+1th step, s is the normalized interpolation ratio, t i is the time step parameter of the i+1th step, t i+1 is the time step parameter of the i+1th step, is the current intermediate state corresponding to the i+1th step, is the synthesized CT image predicted by the wavelet condition generator, ε represents Gaussian noise, is introduced to introduce randomness to the diffusion process to ensure path diversity, represents the standard multivariate normal distribution, the mean is 0, and the covariance matrix is the unit matrix I.

6. The synthetic CT generation method of claim 5, wherein, In the step of calculating three types of losses respectively, and jointly optimizing the wavelet condition generator, the discriminator, the energy network and the feature extractor, the three types of losses include the adversarial loss, the Schrodinger bridge loss and the regularization loss, and the expression of the adversarial loss is: ； To combat loss, CE is cross-entropy loss, is the probability that the discriminator thinks x1 is a real sample, is the probability that the discriminator thinks is a real sample; The expression of the Schrodinger bridge loss is: ； is the Schrödinger bridge loss, t is the time step parameter of the diffusion process, taking values in the range [0, 1], E is the energy network, is the input sample pair of the energy network, x t is the intermediate state in the diffusion process, is the synthetic CT image predicted by the wavelet condition generator, is the negative sample, is the contrastive energy term for the negative sample, is the L2 loss term; The expression of the regularization loss is: ； L is the total number of layers of the feature extractor, L is the information-noise contrast estimation loss, F is the l-th layer feature extracted from the source image by the feature extractor, q is the query, F is the l-th layer feature extracted from the generated image by the feature extractor, k is the key; ； is the total loss, , , is a user-defined balancing weight.

7. The synthetic CT generation method of claim 6, wherein, In the step of calculating three types of losses respectively, and jointly optimizing the wavelet condition generator, the discriminator, the energy network and the feature extractor, the total loss is used to optimize the wavelet condition generator and the feature extractor, the adversarial loss is used to update the discriminator, and the Schrodinger bridge loss is used to optimize the energy network.

8. A synthetic CT generation system characterized by, The system is used for implementing the synthetic CT generation method in any one of claims 1-7. The system comprises: A construction module, which is used for collecting unpaired CBCT and CT data, and constructing a data set after preprocessing; An adoption module, which is used for sampling CBCT images from the data set, and initializing a path and randomly setting the number of interpolation steps; A prediction module, which is used for predicting a synthetic CT image by the wavelet condition generator for each interpolation step, and updating an intermediate state by combining weighted mixing and time-dependent Gaussian noise; An optimization module, which is used for calculating three types of losses respectively, and jointly optimizing the wavelet condition generator, the discriminator, the energy network and the feature extractor; A training module, which is used for repeating the adoption module to the prediction module until a training period is completed; 9. A computer readable storage medium having stored thereon a computer program, characterized in that, An input module, which is used for fixing the number of interpolation steps, inputting a new CBCT image, recursively performing interpolation and noise injection by the trained wavelet condition generator, and outputting a final synthetic CT image.

10. An electronic device, comprising: The program is executed by the processor to implement the synthetic CT generation method in any one of claims 1-7. The computer program is stored in the memory and run on the processor, and the processor implements the synthetic CT generation method in any one of claims 1-7 when executing the program. The computer program is stored in the memory and run on the processor, and the processor implements the synthetic CT generation method in any one of claims 1-7 when executing the program.

Images