Unsupervised pseudo-CT anti-diffusion model construction method, system, medium and equipment

Abstract

The present invention discloses a method, system, medium and equipment for constructing an unsupervised pseudo CT adversarial diffusion model, and relates to the field of data processing technology. The method comprises: obtaining training samples based on a radiotherapy planning system, a CT device and an MR device, wherein the training samples include patient contour information, CT image data and MR image data; performing tensor conversion and registration processing on the patient contour information, CT image data and MR image data to obtain CT training samples and MR training samples in the form of three-dimensional tensors and after pairing; based on a pre-constructed adversarial diffusion model, performing a preset number of training processes on the CT training samples and the MR training samples, and saving the parameters of the trained adversarial diffusion model to obtain an unsupervised pseudo CT adversarial diffusion model for therapeutic dose prediction. The present invention utilizes artificial intelligence technology to train an adversarial diffusion model, and can more accurately detect and evaluate radiotherapy errors based on the trained model, with greater objectivity and accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method, system, medium and equipment for constructing an unsupervised pseudo CT anti-diffusion model. Background Art

[0002] Computed tomography (CT) has long been established as a fundamental imaging modality for radiotherapy (RT) due to the tissue electron density information provided by grayscale Hounsfield units (HU) and anatomical data. However, a major drawback of CT is the poor soft tissue contrast, which leads to imprecise delineation of tumor targets and organs at risk (OARs), especially in the pelvis, liver, and brain, thus resulting in an increasing demand for dedicated magnetic resonance imaging (MRI) scanners in RT centers worldwide. In clinical practice, CT images are an integral part of the current RT workflow, although the need for planning CT and MR images increases costs for cancer patients and requires additional irradiation. Special attention should be paid to the accuracy of the inherent registration introduced during the planning process, which is caused by the inevitable differences in patient positioning during scanning with different imaging modalities and the changes in the anatomical position of organs between different imaging modalities.

[0003] To address the above limitations, using artificial intelligence (AI) algorithms to estimate the Hu value and generate pseudo CT from MRI may be a potential and effective solution. From a macroscopic point of view, the pseudo CT to MRI is a style transfer problem. In recent years, with the rapid development of promising image-to-image translation schemes in generative frameworks, a large number of studies have focused on using generative adversarial networks (GANs) or their variants to improve the quality of sCT images. Segmentation-based methods are generated by pre-dividing tissue types in MR images and assigning uniform volume density to each segmented structure.

[0004] However, these methods are limited by the need to predetermine tissue types, and they also rely heavily on the accuracy of organ segmentation without accounting for the heterogeneity of each structure. Atlas-based methods are limited by the large amount of computation and the accuracy of deformable image registration, and lack robustness in the presence of large anatomical variations. Sequence-based methods have been shown to increase scanning time, motion artifacts, and patient discomfort. Deep learning-based generative methods, especially GAN-based methods, have attracted great attention in the field of medical image generation. Once the optimal deep learning parameters are estimated, pseudo-CT images can be easily obtained in a few seconds by inputting new MR images into the training model. However, deep learning-based models still require properly aligned and paired datasets for training, and some studies have reported that they cannot preserve details during the conversion process. In addition, these models have high requirements for MRI-CT image registration, which is usually difficult in clinical practice. Summary of the invention

[0005] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a method, system, medium and equipment for constructing an unsupervised pseudo-CT anti-diffusion model, aiming to more accurately detect and evaluate radiotherapy errors by establishing an anti-diffusion model, so as to improve the objectivity and accuracy of radiotherapy treatment dose setting.

[0006] A first aspect of the present invention is to provide a method for constructing an unsupervised pseudo CT anti-diffusion model, the method comprising:

[0007] Acquire training samples based on the radiotherapy planning system, the CT device and the MR device, wherein the training samples include patient contour information, CT image data and MR image data;

[0008] Performing tensor conversion and registration processing on the patient contour information, CT image data and MR image data to obtain paired CT training samples and MR training samples in the form of three-dimensional tensors;

[0009] Based on the pre-constructed adversarial diffusion model, a preset number of training processes are performed on the CT training samples and the MR training samples, and the parameters of the trained adversarial diffusion model are saved to obtain an unsupervised pseudo CT adversarial diffusion model for therapeutic dose prediction.

[0010] According to one aspect of the above technical solution, the steps of performing tensor conversion and registration processing on the patient contour information, CT image data and MR image data to obtain paired CT training samples and MR training samples in the form of three-dimensional tensors include:

[0011] Performing tensor conversion on the patient contour information, CT image data and MR image data;

[0012] The tensor-converted CT image data and the MR image data are linearly normalized to a preset resolution, and the CT image data and the MR image data are rigidly aligned to obtain paired CT training samples and MR training samples in the form of three-dimensional tensors.

[0013] According to one aspect of the above technical solution, the process of performing tensor conversion on the patient contour information, CT image data and MR image data is expressed as:

[0014] ;

[0015] In the formula, represents a single sample of input training sample data, is the number of samples, Represents tensor conversion processing;

[0016] Wherein, the preset resolution is 1mm×1mm×1mm.

[0017] According to one aspect of the above technical solution, the pre-constructed adversarial diffusion model includes a diffusion module based on a source domain conditional mapper, which is used to extract the three-dimensional features of the original MR image and the noise distribution information after gradually adding noise, and a non-diffusion model for regression fitting of the three-dimensional features to the distribution of pseudo CT image information.

[0018] According to one aspect of the above technical solution, based on the pre-constructed counter-diffusion model, the training process is performed on the CT training samples and the MR training samples for a preset number of times, and the parameters of the trained counter-diffusion model are saved to obtain the unsupervised pseudo CT counter-diffusion model for treatment dose prediction, including:

[0019] The diffusion module is used to extract the three-dimensional features and noise distribution information in the original MR image, and the image prediction module is used to estimate the target modality image by combining the three-dimensional features and the noise distribution information;

[0020] Based on the data difference between the pseudo-CT results and the real CT results, the final loss is calculated to obtain the total loss of the model;

[0021] According to the total loss of the model, a back-propagation algorithm is used to update the gradient of the adversarial diffusion model to obtain an unsupervised pseudo-CT adversarial diffusion model for therapeutic dose prediction.

[0022] According to one aspect of the above technical solution, the step of using the diffusion module to extract the three-dimensional features and noise distribution information in the original MR image, combining the three-dimensional features and the noise distribution information, and using the image prediction module to estimate the target modality image includes:

[0023] When estimating the target modality image, two adversarial diffusion processes are used, each paired with a dedicated discriminator;

[0024] When the diffusion process is used in reverse operation, the preset condition generator will produce an estimate of the target image.

[0025] A second aspect of the present invention is to provide an unsupervised pseudo-CT anti-diffusion model construction system, the system comprising:

[0026] A training sample acquisition module, used to acquire training samples based on the radiotherapy planning system, the CT device and the MR device, wherein the training samples include patient contour information, CT image data and MR image data;

[0027] A training sample processing module, used for performing tensor conversion and registration processing on the patient contour information, CT image data and MR image data to obtain paired CT training samples and MR training samples in the form of three-dimensional tensors;

[0028] The model building module performs a preset number of training processes on the CT training samples and the MR training samples based on the pre-built adversarial diffusion model, and saves the parameters of the trained adversarial diffusion model to obtain an unsupervised pseudo CT adversarial diffusion model for treatment dose prediction.

[0029] According to one aspect of the above technical solution, the training sample processing module is specifically used for:

[0030] Performing tensor conversion on the patient contour information, CT image data and MR image data;

[0031] The tensor-converted CT image data and the MR image data are linearly normalized to a preset resolution, and the CT image data and the MR image data are rigidly aligned to obtain paired CT training samples and MR training samples in the form of three-dimensional tensors.

[0032] The third aspect of the present invention is to provide a readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in the above technical solution.

[0033] The fourth aspect of the present invention is to provide an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the above technical solution when executing the computer program.

[0034] Compared with the prior art, the unsupervised pseudo-CT anti-diffusion model construction method, system, medium and device shown in the present invention have the following beneficial effects:

[0035] Based on the technical solution shown in the present invention, the CT adversarial diffusion model is trained using artificial intelligence technology, so that the radiotherapy error can be detected and evaluated more accurately based on the trained model, which will have stronger objectivity and accuracy compared with traditional methods and doctors' experience judgment. In addition, the unsupervised adversarial diffusion model uses diffusion and non-diffusion mechanisms for feature extraction and processing, which can enhance the representation and generalization capabilities of the model, so that the prediction results obtained are more accurate and effective. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0037] Figure 1 It is a flowchart of a method for constructing an unsupervised pseudo CT anti-diffusion model in one embodiment of the present invention;

[0038] Figure 2 It is a structural block diagram of a system for building an unsupervised pseudo-CT anti-diffusion model in one embodiment of the present invention. DETAILED DESCRIPTION

[0039] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings. Several embodiments of the present invention are shown in the accompanying drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0040] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be a central element. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be a central element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only.

[0041] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed items.

[0042] Embodiment 1

[0043] See also Figure 1 The first embodiment of the present invention provides a method for constructing an unsupervised pseudo-CT anti-diffusion model. The method shown in this embodiment includes steps S10 to S30:

[0044] Step S10: acquiring training samples based on the radiotherapy planning system, the CT device and the MR device, wherein the training samples include patient contour information, CT image data and MR image data.

[0045] Specifically, patient contour information refers to the external contour information of the tumor patient, which is a common DICOM (Digital Imaging and Communications in Medicine) format in radiotherapy treatment planning systems (such as Pinnacle or Monaco systems). Among them, contour DICOM belongs to radiotherapy files, which records the three-dimensional coordinates of the patient's body, the internal anatomical structure and the distribution information of the external body surface.

[0046] The patient's outer contour is represented in the form of a three-dimensional voxel matrix, where each voxel contains its received Hu value, which represents the response of human tissue density. The MR image data and CT image data are image data of the patient collected using a simulated positioning MR device and a CT device dedicated to radiotherapy.

[0047] Step S20, performing tensor conversion and registration processing on the patient contour information, CT image data and MR image data to obtain paired CT training samples and MR training samples in the form of three-dimensional tensors.

[0048] Specifically, the processing process can be expressed as:

[0049] ;

[0050] In the formula, represents a single sample of input training sample data, is the number of samples, Represents a tensor transformation process.

[0051] Then, all MR images and CT images are linearly normalized to a resolution of 1 mm × 1 mm × 1 mm, and the CT images and MR images are rigidly registered. In this embodiment, the boundary information of the two modes is introduced, and the influence of the image features outside the pelvis on the network performance is minimized. The background intensity of the MR image and the CT image is assigned to 0Hu and -1000Hu respectively by applying the threshold segmentation method to obtain the body mask to obtain the registered training samples in the form of a three-dimensional tensor.

[0052] Step S30, based on the pre-constructed counter-diffusion model, performing a preset number of training processes on the CT training samples and the MR training samples, and saving the parameters of the trained counter-diffusion model to obtain an unsupervised pseudo CT counter-diffusion model for treatment dose prediction.

[0053] In this embodiment, the pre-built adversarial diffusion model is an unsupervised unpaired pseudo-CT neural model, which is used to perform efficient and high-fidelity image conversion between a source domain image and a target image of a given anatomical structure.

[0054] Specifically, the above-mentioned adversarial diffusion model consists of two main modules: a diffusion model based on a source domain conditional mapper for fast and accurate back-diffusion sampling, and a non-diffusion target-source domain paired image prediction module for unsupervised learning.

[0055] By synergizing the strengths of diffusion and predictive modeling, this embodiment creates a breakthrough image translation method with speed, accuracy, and the ability to learn from unpaired data, thereby improving current capabilities in medical imaging and other fields.

[0056] The process of training the pre-built adversarial diffusion model using training samples specifically includes:

[0057] Step S31, using the diffusion module to extract the three-dimensional features and noise distribution information in the original MR image, combining the three-dimensional features and the noise distribution information, and using the image prediction module to estimate the target modality image.

[0058] In this embodiment, two adversarial diffusion processes are used to achieve the estimation of the target modality image, each paired with a dedicated discriminator. In the condition generator We will first generate an estimate of the target image. Specifically, the condition generator Will Processed as a three-channel input, extracting intermediate features ,in Used to represent a sub-block index in the encoder-decoder structure.

[0059] Then for each time step Computing a learnable temporal embedding , As a channel-specific bias term added to the feature map of each sub-block index, it is expressed as Then, the target image is synthesized using the denoising distribution unique to each image modality. as well as and The discriminator as input Differentiate estimated denoised distribution and the true denoised distribution Samples of . Discriminator deal with or As dual channel input. Time embedding Also used as a bias term in the discriminator When the denoising distribution is parameterized from the network When sampling in the condition generator The distribution is performed as follows:

[0060] ;

[0061] In the formula, the condition generator Predicted Yes t / k steps, and then use the denoised distribution of each modality to synthesize the target image .

[0062] Furthermore, the adversarial diffusion model further includes a diffusion module for estimating and synthesizing a target domain image from a source domain image output by a non-diffusion module.

[0063] To achieve this, two adversarial diffusion processes are used, each with its own discriminator. In each step of the reverse process, the conditional generator First, a deterministic estimation is performed on the denoised target image, and then the target image is synthesized using the denoising distribution unique to each image modality. Based on the obtained pseudo CT results and the real CT, the loss is calculated to obtain the model loss, and based on the model loss, the gradient of the generative network model is updated using the back propagation algorithm.

[0064] Among them, the anti-diffusion model proposed in this embodiment is arrive The change is fast with T / k steps. , each forward step adds more noise to compensate for the larger k, breaking the normality assumption of the backward transition probabilities.

[0065] To improve accuracy, the back-diffusion step is performed by using a conditional generator and the discriminator The new adversarial projector operates on the conditional generator First generate the given and The target image is approximately Then synthesize denoised image samples from the denoising distribution , while the discriminator Used to distinguish between real and synthetic samples of denoised images.

[0066] However, the traditional diffusion model usually requires a relatively large time step, because this number requires enough clients to satisfy the assumption that Gaussian noise gradually increases, which greatly limits the efficiency of image generation. To solve this limitation, this embodiment proposes a fast and effective forward diffusion process. In the adversarial diffusion model shown in this embodiment, the value of k is significantly greater than 1. represents the noise variance. Considering that k is much larger than 1, the large-step backward process lacks a closed-form expression. To solve this problem, this embodiment introduces a source domain conditional mapper to capture the complex transition probability in the large-step conditional diffusion model Here we borrow from the adversarial projector The idea is characterized by a cycle-consistent structure and the diffusion process is integrated. In the adversarial diffusion model proposed in this embodiment, the condition generator Source Mode Input and the corresponding newly introduced boundary mask As input, extract intermediate features And perform step-by-step denoising.

[0067] In addition, based on time compensation Computational Learnable Embeddings are merged into these extracted intermediate features Then, a discriminant factor is introduced to distinguish The generated estimated denoised distribution and actual denoised samples. The temporal embedding is added as a bias term to This uses a non-saturated adversarial loss and is used to supervise the learning of , as shown below:

[0068] ;

[0069] ;

[0070] The inverse diffusion process is iterated T / k times to obtain the final denoising result. In order to maintain randomness, the generator distribution is operationalized as shown in the above formula. This implementation allows for more detailed and effective training of the model, thereby facilitating the generation of high-quality denoised outputs while maintaining robustness to adversarial noise.

[0071] In the development process of model training and translation, a hybrid diffusion model was selected in this embodiment, which combines boundary structure information and adversarial projection in the reverse diffusion direction, and the diffusion step size is large, which greatly improves the efficiency of pseudo CT generation while ensuring image quality. In addition, in order to facilitate the training of unpaired MR image data to CT image data sets, a cycle-consistent architecture with coupled diffusion and non-diffusion modules is designed in this embodiment, which can accurately convert between the two modes. The model proposed in this embodiment can sample images in less than 8 steps, thereby greatly speeding up the inference speed, while the original adversarial diffusion model requires more than 1000 steps.

[0072] In mathematical terms, the training process is done by matching the inverse denoising model and To express it, the final adversarial loss is used to minimize the divergence of each denoising step as follows:

[0073] ;

[0074] In this embodiment, a recursive method is used to perform denoising until the final pseudo CT image is obtained. , to express a synthetic image that anatomically matches the input MR image.

[0075] To synthesize the target CT image, the inverse diffusion fusion step requires guidance from source MR images with the same anatomical structure. Therefore, the adversarial diffusion model uses non-diffusion and diffusion processes to estimate the source domain image corresponding to each target image in the training set. Initially, the conditional generator Create a source domain image based on the target domain image as follows:

[0076] .

[0077] At the same time, the discriminator A non-saturated adversarial loss is used to distinguish the estimated source domain image from the actual source domain image as follows:

[0078] ;

[0079] In the formula, Represents the true conditional distribution of the source given a target image.

[0080] Subsequently, the target modality image is estimated using the diffusion module, guided by the source modality image generated by the non-diffusion module.

[0081] To achieve this, two adversarial diffusion processes are used, each paired with a dedicated discriminator. In the condition generator First an estimate of the target image is generated.

[0082] In some embodiments, the L1 loss is calculated based on the obtained predicted synthetic CT result and the real CT to obtain the model loss; based on the model loss, the gradient of the adversarial diffusion model is updated using the forward adversarial noise and backward denoising propagation algorithms. To promote unsupervised learning, the production model uses the cycle consistency loss to compare the real target image with the reconstructed image. In the diffusion module, the reconstructed image is called the synthetic target image.

[0083] In the non-diffusion module, the estimated source domain image is mapped to the target domain through the generator. The diffusion module and the non-diffusion module are trained jointly without pre-training procedure.

[0084] Based on the above method, the total generation loss and identification loss It is expressed as follows:

[0085] ;

[0086] ;

[0087] In the formula, and are the weights of the adversarial loss terms.

[0088] Based on the above approach, radiation therapy enables the superior soft tissue contrast of MR to be used for delineation without introducing the inherent alignment errors between MR and CT; this can improve the geometric accuracy of treatment, reduce unnecessary radiation exposure to patients, and improve the efficiency of clinical workflow.

[0089] Mathematically, the CT image synthesis problem can be expressed as:

[0090] ;

[0091] In the formula, is the source modality image.

[0092] Based on the above method, the generated diffusion model adopts the forward diffusion process , gradually moving towards the original data in T steps Add Gaussian white noise with constant variance as follows:

[0093] ;

[0094] ;

[0095] In the formula, is a hyperparameter of the noise variance, is the unit covariance matrix. The forward modeling process is constructed in hour, is almost indistinguishable from Gaussian noise. The Gaussian parameterization of the forward modeling process also allows for a given of The closed form of the simulation is performed on the posterior distribution of .

[0096] Step S32, based on the data difference between the pseudo CT result and the real CT result, the final loss is calculated to obtain the total loss of the model.

[0097] The true value refers to the real simulated positioning CT image corresponding to the training sample obtained through actual scanning. The total loss is the generator mean absolute error (MAE) loss plus the cycle consistency loss. By calculating the loss, the deviation between the predicted result and the true value can be expressed. The smaller the deviation, the better the performance of the model.

[0098] Step S33: According to the total loss of the model, a back propagation algorithm is used to update the gradient of the anti-diffusion model to obtain an unsupervised pseudo-CT anti-diffusion model for treatment dose prediction.

[0099] Specifically, the traditional diffusion model gradually maps pure noise samples to the actual image over T time steps. In the forward process, A small amount of Gaussian noise is gradually added to T. When T becomes large enough, this will lead to isotropic Gaussian distributed samples. In the reverse process, the neural network generates estimates for each step, converting the random noise back into actual image samples. Both the forward and reverse processes can be defined as parameterized Markov chains.

[0100] Due to the use of this diffusion generation process, there are three main issues that need to be considered when constructing a diffusion-based CT generation method for practical application in the field of radiotherapy.

[0101] First, traditional diffusion models usually require relatively many time steps T to ensure that these steps are small enough to satisfy the Gaussian assumption inherent in the traditional diffusion process. However, this greatly limits the efficiency of CT image generation.

[0102] Second, the forward diffusion process of the source domain image leads to the loss of structural details, which cannot be fully restored in the inverse denoising process. However, the geometric and structural integrity of CT is particularly important for the delineation of target volumes in RT, as this helps in planning the precise radiation dose to tumor and normal tissues. Impaired anatomical accuracy may lead to serious consequences, including potential harm to the patient or inadequate treatment of the disease.

[0103] Finally, the conditional diffusion model, similar to GANs, requires well-aligned paired data and matching anatomy for training, which is challenging due to anatomical motion between scans, as well as additional imaging costs and radiation dose.

[0104] Therefore, compared with the prior art, the unsupervised pseudo-CT anti-diffusion model construction method shown in this embodiment has the following beneficial effects:

[0105] Based on the solution of the above embodiment, artificial intelligence technology is used to train the CT adversarial diffusion model, so that the trained model can be used to more accurately detect and evaluate radiotherapy errors, which will have stronger objectivity and accuracy compared with traditional methods and doctors' experience judgment. In addition, the unsupervised adversarial diffusion model uses diffusion and non-diffusion mechanisms for feature extraction and processing, which can enhance the representation and generalization capabilities of the model, so that the prediction results obtained are more accurate and effective.

[0106] Embodiment 2

[0107] See also Figure 2 The second embodiment of the present invention provides an unsupervised pseudo-CT anti-diffusion model construction system, the system comprising: a training sample acquisition module 10, a training sample processing module 20 and a model construction module 30.

[0108] A training sample acquisition module 10 is used to acquire training samples based on a radiotherapy planning system, a CT device and an MR device, wherein the training samples include patient contour information, CT image data and MR image data;

[0109] The training sample processing module 20 is used to perform tensor conversion and registration processing on the patient contour information, CT image data and MR image data to obtain paired CT training samples and MR training samples in the form of three-dimensional tensors;

[0110] The model building module 30 performs a preset number of training processes on the CT training samples and the MR training samples based on the pre-built counter-diffusion model, and saves the parameters of the trained counter-diffusion model to obtain an unsupervised pseudo-CT counter-diffusion model for treatment dose prediction.

[0111] The training sample processing module 20 is specifically used for:

[0112] Performing tensor conversion on the patient contour information, CT image data and MR image data;

[0113] The tensor-converted CT image data and the MR image data are linearly normalized to a preset resolution, and the CT image data and the MR image data are rigidly aligned to obtain paired CT training samples and MR training samples in the form of three-dimensional tensors.

[0114] Compared with the prior art, the unsupervised pseudo-CT anti-diffusion model construction system shown in this embodiment has the following beneficial effects:

[0115] Based on the solution of the above embodiment, artificial intelligence technology is used to train the CT adversarial diffusion model, so that the trained model can be used to more accurately detect and evaluate radiotherapy errors, which will have stronger objectivity and accuracy compared with traditional methods and doctors' experience judgment. In addition, the unsupervised adversarial diffusion model uses diffusion and non-diffusion mechanisms for feature extraction and processing, which can enhance the representation and generalization capabilities of the model, so that the prediction results obtained are more accurate and effective.

[0116] Embodiment 3

[0117] A third embodiment of the present invention provides a readable storage medium having a computer program stored thereon, which implements the steps of the method described in the above embodiment when executed by a processor.

[0118] Embodiment 4

[0119] A fourth embodiment of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the method described in the above embodiment when executing the computer program.

[0120] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0121] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for those of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A method for constructing an unsupervised pseudo-CT anti-diffusion model, characterized in that: The method comprises: Acquire training samples based on the radiotherapy planning system, the CT device and the MR device, wherein the training samples include patient contour information, CT image data and MR image data; Performing tensor conversion and registration processing on the patient contour information, CT image data and MR image data to obtain paired CT training samples and MR training samples in the form of three-dimensional tensors; Based on the pre-constructed adversarial diffusion model, a preset number of training processes are performed on the CT training samples and the MR training samples, and the parameters of the trained adversarial diffusion model are saved to obtain an unsupervised pseudo CT adversarial diffusion model for therapeutic dose prediction; The step of performing tensor conversion and registration processing on the patient contour information, CT image data and MR image data to obtain paired CT training samples and MR training samples in the form of three-dimensional tensors includes: Performing tensor conversion on the patient contour information, CT image data and MR image data; Linearly normalizing the tensor-converted CT image data and the MR image data to a preset resolution, and performing rigid registration on the CT image data and the MR image data to obtain paired CT training samples and MR training samples in the form of three-dimensional tensors; Among them, the pre-built adversarial diffusion model includes a diffusion module based on a source domain conditional mapper, which is used to extract the three-dimensional features of the original MR image and the noise distribution information after gradually adding noise, and a non-diffusion model for regression fitting of the three-dimensional features to the distribution of pseudo-CT image information; The step of performing a preset number of training processes on the CT training samples and the MR training samples based on a pre-constructed counter-diffusion model, and saving the parameters of the trained counter-diffusion model to obtain an unsupervised pseudo CT counter-diffusion model for therapeutic dose prediction includes: The diffusion module is used to extract the three-dimensional features and noise distribution information in the original MR image, and the image prediction module is used to estimate the target modality image by combining the three-dimensional features and the noise distribution information; Based on the data difference between the pseudo-CT results and the real CT results, the final loss is calculated to obtain the total loss of the model; According to the total loss of the model, a back propagation algorithm is used to update the gradient of the adversarial diffusion model to obtain an unsupervised pseudo-CT adversarial diffusion model for therapeutic dose prediction; The step of using the diffusion module to extract the three-dimensional features and noise distribution information in the original MR image, combining the three-dimensional features and the noise distribution information, and using the image prediction module to estimate the target modality image includes: When estimating the target modality image, two adversarial diffusion processes are used, each paired with a dedicated discriminator; When the diffusion process is used to perform the reverse operation, the preset condition generator will produce an estimate of the target image; In order to achieve the estimation of the target modality image, two adversarial diffusion processes are used, each of which is paired with a dedicated discriminator. In the condition generator Will Processed as a three-channel input, extracting intermediate features ,in Used to represent the sub-block index in the encoder-decoder structure and for each time step Computing a learnable temporal embedding , As a channel-specific bias term added to the feature map of each sub-block index, it is expressed as , and synthesize the target image using the denoising distribution unique to each image modality. as well as and The discriminator as input Differentiate estimated denoised distribution and the true denoised distribution The sample of the discriminator deal with or As a dual-channel input, time embedding Also used as a bias term in the discriminator The feature map in the range, when denoising the distribution from the network parameterization When sampling, the condition generator The distribution is performed as follows: ； In the formula, the condition generator Predicted Yes t / k steps, and use the denoised distribution of each modality to synthesize the target image ; Among them, the condition generator Source Mode Input and the corresponding newly introduced boundary mask As input, extract intermediate features And perform step-by-step denoising based on the time step Computational learnable temporal embeddings are merged into these extracted intermediate features and introduce a discriminant factor to distinguish the conditional generator The generated estimated denoised distribution and actual denoised samples embed the time As a bias term added to The feature map is used to supervise the learning , as shown below: ； ； The inverse diffusion process is iterated T / k times to obtain the final denoising result.

2. The unsupervised pseudo-CT anti-diffusion model construction method according to claim 1 is characterized in that: The process of performing tensor conversion on the patient contour information, CT image data and MR image data is expressed as: ； In the formula, represents a single sample of input training sample data, is the number of samples, Represents tensor conversion processing; Wherein, the preset resolution is 1mm×1mm×1mm.

3. An unsupervised pseudo-CT anti-diffusion model construction system, characterized in that: The method applied to any one of claims 1-2, wherein the system comprises: A training sample acquisition module, used to acquire training samples based on the radiotherapy planning system, the CT device and the MR device, wherein the training samples include patient contour information, CT image data and MR image data; A training sample processing module, used for performing tensor conversion and registration processing on the patient contour information, CT image data and MR image data to obtain paired CT training samples and MR training samples in the form of three-dimensional tensors; The model building module performs a preset number of training processes on the CT training samples and the MR training samples based on the pre-built adversarial diffusion model, and saves the parameters of the trained adversarial diffusion model to obtain an unsupervised pseudo CT adversarial diffusion model for treatment dose prediction.

4. The unsupervised pseudo-CT anti-diffusion model construction system according to claim 3 is characterized in that: The training sample processing module is specifically used for: Performing tensor conversion on the patient contour information, CT image data and MR image data; The tensor-converted CT image data and the MR image data are linearly normalized to a preset resolution, and the CT image data and the MR image data are rigidly aligned to obtain paired CT training samples and MR training samples in the form of three-dimensional tensors.

5. A readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 2 are implemented.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 2 are implemented.

Images