A diffusion 3D medical image generation method with decoupling of mask generation and image generation

Abstract

The present invention discloses a diffusion 3D medical image generation method that decouples mask generation from image generation, relates to the technical field of image processing and analysis, and solves the problems of poor generalization ability, low computational efficiency, and high dependence on training data in image conversion in the prior art. The method comprises: obtaining a mask image, processing the mask image using a multi-conditional diffusion probability model to obtain a multi-label mask sequence; screening 3D medical image data according to screening conditions to determine an informed slice set, and inputting the multi-label mask sequence and the informed slice set into a volume diffusion generation model to obtain a 3D composite image; adding random noise to the 3D composite image, and using a diffusion semantic refinement generation model to refine the 3D composite image after adding the random noise to obtain a refined 3D image. The method achieves data decoupling and consistency in the generation model volume, and can generate higher-quality 3D medical images.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing and analysis, and in particular to a diffusion three-dimensional medical image generation method that decouples mask generation and image generation. Background Art

[0002] With advances in medical imaging technologies such as computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET), the volume and complexity of medical image data continues to increase, posing new challenges for data processing and analysis. Traditional image processing methods are often limited by computing power and storage space when processing large-scale three-dimensional data, and it is difficult to ensure processing efficiency and accuracy of the results.

[0003] In existing technologies, the conversion of images from CT to PET involves complex changes in modal characteristics. This method has insufficient generalization capabilities when processing different types or stages of lung cancer, especially for tumors with atypical or rare features. Although the method mentions the extraction of local and global information, how to balance the two to generate images that are both detailed and consistent with overall pathological characteristics remains a challenge. Although existing technologies provide new technical approaches in the field of medical image generation, they may require further optimization and improvement in terms of generalization capabilities, computational efficiency, dependence on training data, and the complexity of clinical applications. Summary of the Invention

[0004] The present invention provides a diffusion 3D medical image generation method that decouples mask generation from image generation, thereby solving the problems of poor generalization ability, low computational efficiency, and high dependence on training data during image conversion in the prior art. It implements a decoupled training process to solve the problem of inconsistent generated model volumes and generate higher-quality 3D medical images.

[0005] The present invention provides a method for generating diffuse three-dimensional medical images by decoupling mask generation and image generation, the method comprising:

[0006] Obtaining a mask image, and processing the mask image using a multi-conditional diffusion probability model to obtain a multi-label mask sequence; processing the mask image using the multi-conditional diffusion probability model to obtain a multi-label mask sequence includes:

[0007] The mask image is cropped and padded to obtain multiple images of uniform size, and the multiple images are represented as two-dimensional sequences respectively;

[0008] Selecting a portion of the sequence from the two-dimensional sequence as a two-dimensional subsequence, and unconditionally generating an initial subsequence of continuous slices from random noise; wherein the length of the two-dimensional subsequence is equal to the length of the initial subsequence;

[0009] Adding the initial subsequence to the two-dimensional subsequence, and performing forward and backward expansion on the two-dimensional subsequence after adding the initial subsequence by an autoregressive method to obtain a multi-label mask sequence;

[0010] The multi-conditional diffusion probability model is expressed as:

[0011]

[0012] in, represents the random variable used to simulate noise in the diffusion model; represents the noise predicted by the multi-conditional diffusion probability model; represents the part of the subsequence to be generated after the noise perturbation after time step t; The part that represents the conditional slice; Represents the relative position condition of the subsequence, z~=z / D; represents the time step; Indicates that the loss function obeys the probability distribution of multiple conditions;

[0013] screening the three-dimensional medical image data according to the screening conditions to determine an informed slice set, and inputting the multi-label mask sequence and the informed slice set into a volume diffusion generation model to obtain a three-dimensional composite image;

[0014] The method of using the diffusion semantic refinement generation model to refine the three-dimensional synthetic image after adding random noise to obtain a refined three-dimensional image includes:

[0015] Reading the three-dimensional composite image after adding random noise according to different dimensions to obtain informed slices of corresponding dimensions; wherein the different dimensions include: coronal slice mode, sagittal slice mode and axial slice mode;

[0016] The informed slices of different dimensions are input into the corresponding diffusion semantic refinement generation model to generate the refined images corresponding to different dimensions;

[0017] Calculate the mean of the refined images corresponding to different dimensions, and obtain the refined three-dimensional image based on the mean;

[0018] adding random noise to the three-dimensional composite image, and using a diffusion semantic refinement generation model to refine the three-dimensional composite image after the random noise is added to obtain a refined three-dimensional image;

[0019] The loss function of the volume diffusion generation model is expressed as:

[0020] ;

[0021] in, Indicates that the loss function obeys the probability distribution of multiple conditions; represents the random variable used to simulate noise in the diffusion model; represents the noise predicted by the volume diffusion generative model; It represents the product of the denoising coefficients accumulated step by step from time step 1 to t, indicating the denoising strength after multiple time steps; represents the time step; represents the original, unnoised image or data point; Indicates that at each iteration, the model determines the conditional slice according to the probability.

[0022] In a possible implementation, the method of forward and backward extending the two-dimensional subsequence after adding the initial subsequence by an autoregressive method to obtain a multi-label mask sequence includes:

[0023] Determining generation parameters of the multi-label mask sequence;

[0024] The multi-conditional diffusion probability model performs forward and backward expansion on the two-dimensional subsequence after adding the initial subsequence according to the generation parameters and the autoregressive method to obtain a multi-label mask sequence.

[0025] In one possible implementation, screening the three-dimensional medical image data according to the screening conditions to determine the informed slice set includes:

[0026] An extraction starting position is determined according to the screening condition, and a plurality of different informed slices are selected from the three-dimensional medical image through a resampling technique and the extraction starting position, and the plurality of different informed slices are combined into the informed slice set.

[0027] In a possible implementation, the informed slices during training of the volume diffusion generation model are extracted from a three-dimensional medical image window.

[0028] One or more technical solutions provided in the present invention have at least the following technical effects or advantages:

[0029] (1) The present invention significantly enhances the quality and fidelity of image generation by adding a multi-conditional diffusion probability model and a diffusion semantic refinement generation model to the volumetric diffusion generation model and decoupling the multi-label mask sequence from the mask image.

[0030] (2) The present invention receives informed slices as input through a volumetric diffusion generative model and combines them with a multi-label mask sequence as conditional information to generate a 3D synthetic image through an iterative denoising process. This process not only improves the image quality but also enhances the controllability of the generation process.

[0031] (3) The present invention allows users to adjust the features of generated images according to specific medical application scenarios by combining multi-label mask sequences and conditional information of informed slices in the volumetric diffusion generation model;

[0032] (4) This invention optimizes computational efficiency and resource consumption while maintaining the quality of generated images. By improving the network architecture and parameter settings, the demand for computing resources is reduced, making the model more suitable for resource-constrained environments.

[0033] (5) The present invention better protects the privacy of patients through the use of informed slices, complies with ethical standards and privacy protection requirements, which is particularly important in the field of medical image analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A flowchart of the steps of a diffusion 3D medical image generation method with decoupling of mask generation and image generation provided by an embodiment of the present invention;

[0035] Figure 2 A schematic diagram of the training process for generating a multi-label mask sequence through a multi-conditional diffusion probability model provided by an embodiment of the present invention;

[0036] Figure 3 A schematic diagram of the inference process for generating a multi-label mask sequence through a multi-conditional diffusion probability model provided by an embodiment of the present invention;

[0037] Figure 4 A schematic diagram of the use of the volume diffusion generation model provided in an embodiment of the present invention;

[0038] Figure 5 A schematic diagram of the reasoning process of the volume diffusion generation model provided in an embodiment of the present invention;

[0039] Figure 6 Schematic diagram of the network architecture of the volume diffusion generation model and semantic refiner provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0041] The present invention provides a method for generating diffuse 3D medical images by decoupling the generation mask and the image. Figure 1 As shown, the method includes the following steps S101 to S103.

[0042] S101, obtaining a mask image, and processing the mask image using a multi-conditional diffusion probability model to obtain a multi-label mask sequence.

[0043] Specifically, in step S101, the mask image is processed using a multi-conditional diffusion probability model to obtain a multi-label mask sequence, including:

[0044] The mask image is cropped and padded to obtain multiple images of uniform size, and the multiple images are represented as two-dimensional sequences respectively;

[0045] Selecting a portion of the sequence from the two-dimensional sequence as a two-dimensional subsequence, and unconditionally generating an initial subsequence of continuous slices from random noise; wherein the length of the two-dimensional subsequence is equal to the length of the initial subsequence;

[0046] The initial subsequences are added to the two-dimensional subsequences respectively, and the two-dimensional subsequences after adding the initial subsequences are forward and backward extended by the autoregressive method to obtain a multi-label mask sequence.

[0047] Here, the two-dimensional subsequence after adding the initial subsequence is forward and backward extended by the autoregressive method to obtain a multi-label mask sequence, including:

[0048] (1) Determine the generation parameters of the multi-label mask sequence;

[0049] (2) The multi-conditional diffusion probability model performs forward and backward expansion on the two-dimensional subsequence after adding the initial subsequence according to the generation parameters and autoregressive method to obtain a multi-label mask sequence.

[0050] Specifically, in step S101, the multi-conditional diffusion probability model is expressed as:

[0051] ;

[0052] in, represents the random variable used to simulate noise in the diffusion model; represents the noise predicted by the multi-conditional diffusion probability model; represents the part of the subsequence to be generated after the noise perturbation after time step t; represents the part that represents the conditional slice; Denotes the relative position condition of the subsequence z~=z / D; represents the time step; Indicates that the loss function obeys the probability distribution of multiple conditions.

[0053] For example, a mask generation framework is designed to crop and pad the mask image to a uniform size. , represented as a two-dimensional sequence. An initial subsequence of m consecutive slices is generated unconditionally from random noise. Figure 2 and Figure 3 As shown, after the cropped and padded mask image is represented as a two-dimensional subsequence, the subsequence is extended forward and backward in an autoregressive manner, conditioned on the existing slices, and a multi-conditional diffusion probability model is set to generate a multi-label mask sequence. The generation parameters of the mask sequence include but are not limited to sequence length, conditional probability, etc.

[0054] For a given 3D multi-label mask , select a subsequence of m consecutive slices as , where z is a randomly selected starting index. For each subsequence, we select the first n or last n slices, or no slices, based on the probability To determine the conditional slice The goal of the multi-conditional diffusion probability model is to generate the remaining slices. In order to incorporate the position condition, the relative position of the subsequence is used ,in is the index of the starting slice of the subsequence. The positional conditional embedding is then concatenated with the temporal embedding to aid the generative process. A binary indicator is used for each slice in the subsequence to indicate the presence of the conditional slice. The joint distribution of the inverse diffusion process (RDP) of the conditional slice XC can be written as:

[0055] ;

[0056] During the inference process, the multi-conditional diffusion probability model first In the case of random noise generation A subsequence of slices can then be generated by autoregressively extending in both directions, conditioned on the existing slices, to generate a multi-label mask sequence.

[0057] S102 , screening the three-dimensional medical image data according to the screening conditions, determining an informed slice set, and inputting the multi-label mask sequence and the informed slice set into a volume diffusion generation model to obtain a three-dimensional composite image.

[0058] Specifically, in step S102, the three-dimensional medical image data is screened according to the screening conditions to determine the informed slice set, including: determining an extraction starting position according to the screening conditions, selecting multiple different informed slices in the three-dimensional medical image through resampling technology and the extraction starting position, and forming the multiple different informed slices into an informed slice set.

[0059] Exemplarily, during training, informed slices are extracted from a 3D medical image window. Reconstructing a 3D medical image from a mask image involves a one-to-many mapping. To alleviate volume inconsistencies and poor generation fidelity and improve generation quality, additional information, namely informed slices, is extracted from the 3D image, which indicates the patient's anatomical appearance, physical position, and other scan patterns.

[0060] During inference, informed slices are initially generated by the model or randomly selected in accessible 3D medical images, and then autoregressively assigned synthetic ones, e.g. Figure 5 In addition to guiding the generation process, informed slice-conditional generation also achieves generation enhancement in the dataset through resampling. Therefore, the input condition in the volume diffusion generation model Including multi-label mask sequences and informed slices , used for 3D medical images slices.

[0061] The selection of informed slices is crucial to the generation process because it carries the patient’s anatomical prior information. Figure 4 As shown, informed slices and 3D mask images from sample 3D medical images are used. Due to the generalization ability, informed slices can be sampled from different 3D medical images by simple random selection to drive conditional generation given the mask volume and introduce information from other patients' 3D medical images. To increase the variation, a simple way to sample informed slices is to extract initial slices at random starting positions from randomly selected 3D medical images. ,Right now It is from The sampled from The samples were taken from represents the number of samples in the dataset, Indicates the The number of slices in a 3D medical image.

[0062] To achieve this, the position information of the slices in the volume is normalized to ensure that the position information of different samples is physically aligned.

[0063] S103 , adding random noise to the 3D composite image, and using a diffusion semantic refinement generation model to refine the 3D composite image after adding the random noise, to obtain a refined 3D image.

[0064] Specifically, in step S103, the diffusion semantic thinning generation model is used to thin the 3D synthetic image after adding random noise to obtain a thinned 3D image, including:

[0065] (1) reading the three-dimensional composite image after adding random noise according to different dimensions to obtain informed slices of the corresponding dimensions; wherein the different dimensions include: coronal slice mode, sagittal slice mode and axial slice mode;

[0066] (2) Inputting informed slices of different dimensions into the corresponding diffusion semantic refinement generation model to generate refined images corresponding to different dimensions;

[0067] (3) Calculate the mean of the refined images corresponding to different dimensions, and obtain the refined three-dimensional image based on the mean.

[0068] Specifically, the loss function of the volume diffusion generation model is expressed as:

[0069] ;

[0070] in, Indicates that the loss function obeys the probability distribution of multiple conditions; represents the random variable used to simulate noise in the diffusion model; represents the noise predicted by the volume diffusion generative model; It represents the product of the denoising coefficients accumulated step by step from time step 1 to t, indicating the denoising strength after multiple time steps; represents the time step; represents the original, unnoised image or data point; Indicates that at each iteration, the model determines the conditional slice according to the probability.

[0071] Specifically, the informed slices during training of the volumetric diffusion generative model are extracted from 3D medical image windows.

[0072] For example, the goal of the volumetric diffusion generative model is to learn the inverse process of denoising noisy data samples from the noise and recovering them to follow the true data distribution. In the inverse process, the denoising network Predictions at time step Data that is damaged noise, and construct a parameterized Gaussian transfer The denoising network is built on top of U-Net and trained with mean squared error:

[0073] ;

[0074] in, Indicates that the loss function obeys the probability distribution of multiple conditions; represents the random variable used to simulate noise in the diffusion model, that is, the noise sampled from the standard normal distribution; represents the noise predicted by the model; It represents the product of the denoising coefficients accumulated step by step from time step 1 to t, indicating the denoising strength after multiple time steps. represents the time step; represents the original, unnoised image or data point; Indicates that at each iteration, the model determines the conditional slice according to the probability.

[0075] 3D synthetic images and set up a diffusion semantic refinement generation model, such as Figure 6 As shown, the multi-label mask sequence in S101 and the informed slice set in S102 are combined as conditional inputs to a volume diffusion generative model to generate a three-dimensional medical image.

[0076] Although the volumetric diffusion generative model achieves high cross-slice consistency and spatial continuity, problems such as blocking, blurring, and suboptimal texture generation still exist. Therefore, a diffusion semantic refinement generative model is used to refine the previously generated images.

[0077] For three different views, specifically axial view, coronal view and sagittal view, a two-dimensional mask and an image with added noise are used as input to generate an image aligned with the input mask.

[0078] During inference, a small amount of noise is applied to the generated image to preserve the overall anatomical structure and spatial continuity. Afterwards, the final refined 3D image is the average result from the three views using a pre-trained diffusion semantic refinement generative model.

[0079] Compared to existing technologies, this invention significantly improves the quality of 3D medical image generation by introducing a training process that decouples mask generation from image generation and fine-tuning the generated structure through a diffusion semantic refinement generative model (SDM). This addresses the shortcomings of volumetric diffusion generative models in capturing medical image detail. By using informed slices, this invention enhances the controllability of the generation process, allowing users to adjust the characteristics of the generated image according to specific needs.

[0080] While maintaining computational efficiency, the present invention reduces the need for computing resources, making the model more suitable for resource-constrained environments. Through an improved generation framework, the present invention alleviates the problem of volume inconsistency and improves the fidelity and practicality of 3D medical images.

[0081] In a specific use embodiment provided by the present invention, the method includes: step 1, generating a multi-label mask sequence; step 2, determining an informed slice set and generating a three-dimensional composite image; step 3, using a diffusion semantic refinement generation model to obtain a refined three-dimensional image.

[0082] In step 1, the mask image is cropped and padded to a uniform size and represented as a two-dimensional sequence. An initial subsequence of m consecutive slices is unconditionally generated from random noise. This initial subsequence is then extended forward and backward in an autoregressive manner, conditioned on the existing slices, and a multi-conditional volumetric diffusion generative model is set to generate a multi-label mask sequence. The generation parameters of the multi-label mask sequence include but are not limited to sequence length, conditional probability, etc.

[0083] During inference, the multi-conditional diffusion probability model first generates a subsequence of 6 slices from random noise given a random position z. It can then generate a multi-label mask sequence by autoregressively expanding in two directions, conditioned on the existing slices.

[0084] Specifically, in the first step, a starting index z is randomly selected from the 3D multi-label mask dataset.

[0085] In the second step, random noise is used to unconditionally generate an initial subsequence of m consecutive slices.

[0086] In the third step, based on the existing slices, the subsequence is expanded in a backward and forward autoregressive manner, using a multi-conditional diffusion probability model.

[0087] The fourth step is to use the probability p C Select conditional slice X C , which may be the first n or last n slices of the subsequence, or no slice.

[0088] In the fifth step, the relative position z˜ of the subsequence is used as the conditional,embedding and combined with the temporal embedding, while a binary indicator is used to mark the conditional slices.

[0089] Step 6: Train the multi-conditional diffusion probability model to minimize the loss function and generate the remaining slices X P .

[0090] Step 7: In the inference phase, starting from a random position z, a subsequence of m slices is generated.

[0091] In the eighth step, we autoregressively extend the subsequence to the entire mask sequence, conditioned on the existing slices.

[0092] In the ninth step, the mask sequence is generated using the detailed process and network structure in the supplementary material.

[0093] Differences from existing technologies: Traditional methods do not use autoregressive methods and multi-conditional diffusion probability models to generate mask sequences. This invention improves the quality and accuracy of mask sequence generation through a volumetric diffusion generation model and relative position embedding.

[0094] The results achieved are: The generated mask sequences have higher quality and diversity, improving the realism and practicality of 3D medical images.

[0095] In the second step, informed slices are extracted from image windows during training; during inference, they are generated by existing models or randomly selected from accessible images, and then autoregressively integrated into the synthesis process.

[0096] In subsequent operations, different informed slices are selected through resampling technology. In the volume diffusion generation model, the conditions of the volume diffusion generation model are By multi-label mask sequence and informed slices Composition, working together on The volume of The denoising network in the volumetric diffusion generative model predicts the noise in the data at each time step and constructs a Gaussian transfer model based on this noise. The volumetric diffusion generative model is built on a U-Net architecture and is trained by minimizing mean squared error.

[0097] Specifically, the steps for selecting and generating informed slices are as follows:

[0098] In the first step, during the training phase, informed slices are randomly selected from the 3D medical image window or generated based on the model. These slices contain information such as the patient's anatomical appearance, physical position, and scanning mode.

[0099] Second, during the inference phase, the initial informed slices can be generated by the model or randomly selected from the accessible volume and synthesized through an autoregressive process.

[0100] In the third step, due to memory limitations, the entire volume cannot be fed into the model, so windows are randomly sampled from the original volume to introduce variations in different conditions and starting positions for training and sampling.

[0101] The fourth step is to provide the Volume window , including n slices starting from the jth slice.

[0102] Step 5: Randomly select informed slices in the training unit and compare it with the multi-label mask sequence The volume diffusion generative model is injected together for denoising.

[0103] In the sixth step, during the autoregressive volume sampling process, the positional relationship of the informed slices at the window endpoints is ensured and processed by diffusion in-painting.

[0104] In the seventh step, the conditional information, including informed slices and mask volumes, is integrated in the latent space through connections.

[0105] In the eighth step, the conditional latent features are combined with the noisy target latent features through a simple connection method to achieve control injection.

[0106] In the ninth step, we train a position ID guided slice generation model, using the normalized position ID p and embedding the sinusoidal embedding as a cross attention condition.

[0107] In the tenth step, a cascaded volume diffusion generative model is constructed, which first samples the informed slice conditions and then samples the volume during inference to achieve generative enhancement within the dataset.

[0108] Difference from existing technologies: Existing technologies do not utilize the role of informed slices in the generation process. This invention, through the use of informed slices, enhances the controllability of generated images and meets the needs of medical application scenarios.

[0109] Effect achieved: The generated images are more in line with the needs of medical application scenarios, improving the quality and practicality of the images.

[0110] In step three, based on the volumetric diffusion generation, the Diffusion Semantic Refinement Model (SDM) further refines the image using a diffusion probability model. During the refinement process, a limited number of noise steps are applied to the image to preserve anatomical structure and spatial continuity, and then the image is optimized using a pre-trained model. By integrating the refinement results from three different views, a high-quality 3D image is obtained.

[0111] Specifically, the diffusion semantic refinement generation model is used to refine the 3D synthetic image after adding random noise to obtain a refined 3D image, including:

[0112] The first step is to use convolutional networks to generate preliminary 3D medical images with high cross-slice consistency and spatial continuity.

[0113] Step 2: For each different view, such as axial view, coronal view and sagittal view, a diffuse semantic refinement generative model (SDM) is trained to accept multi-label mask sequences and informed slices as input.

[0114] Step 3: During the inference phase, a small amount of noise is applied to the initially generated image to preserve the overall anatomical structure and spatial continuity of the image.

[0115] Step 4: Use the pre-trained SDM to refine the slightly noisy image to improve image quality and reduce blur and texture issues.

[0116] Step 5: Calculate the average of the refined image results obtained from three different views to obtain the final 3D refined image.

[0117] Difference from existing techniques: Existing techniques do not use semantic diffusion refiners to improve the quality of the initially generated image. The new method reduces blur and texture issues through the refinement process, improving the image's detail.

[0118] The effect achieved: The refined 3D image has higher quality, reduces blur and texture problems, and improves image clarity and realism.

[0119] The various embodiments in this specification are described in a progressive manner. References to the same or similar parts between the various embodiments are sufficient. Each embodiment focuses on the differences from other embodiments. All or part of the present invention can be used in a variety of general or specialized computer system environments or configurations. For example, personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments that include any of the above systems or devices.

[0120] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, it should be understood by those skilled in the art that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the present invention.

Claims

1. A diffusion 3D medical image generation method with decoupling of mask generation and image generation, characterized in that: include: Obtaining a mask image, and processing the mask image using a multi-conditional diffusion probability model to obtain a multi-label mask sequence; The method of processing the mask image using a multi-conditional diffusion probability model to obtain a multi-label mask sequence includes: The mask image is cropped and padded to obtain multiple images of uniform size, and the multiple images are represented as two-dimensional sequences respectively; Selecting a portion of the sequence from the two-dimensional sequence as a two-dimensional subsequence, and unconditionally generating an initial subsequence of continuous slices from random noise; wherein the length of the two-dimensional subsequence is equal to the length of the initial subsequence; Adding the initial subsequence to the two-dimensional subsequence, and performing forward and backward expansion on the two-dimensional subsequence after adding the initial subsequence by an autoregressive method to obtain a multi-label mask sequence; The multi-conditional diffusion probability model is expressed as: in, represents the random variable used to simulate noise in the diffusion model; represents the noise predicted by the multi-conditional diffusion probability model; represents the part of the subsequence to be generated after the noise perturbation after time step t; The part that represents the conditional slice; Represents the relative position condition of the subsequence, z~=z / D; represents the time step; Indicates that the loss function obeys the probability distribution of multiple conditions; screening the three-dimensional medical image data according to the screening conditions to determine an informed slice set, and inputting the multi-label mask sequence and the informed slice set into a volume diffusion generation model to obtain a three-dimensional composite image; The diffusion semantic refinement generation model is used to refine the 3D synthetic image after adding random noise to obtain a refined 3D image, including: Reading the three-dimensional composite image after adding random noise according to different dimensions to obtain informed slices of corresponding dimensions; wherein the different dimensions include: coronal slice mode, sagittal slice mode and axial slice mode; The informed slices of different dimensions are input into the corresponding diffusion semantic refinement generation model to generate the refined images corresponding to different dimensions; Calculate the mean of the refined images corresponding to different dimensions, and obtain the refined three-dimensional image based on the mean; adding random noise to the three-dimensional composite image, and using a diffusion semantic refinement generation model to refine the three-dimensional composite image after the random noise is added to obtain a refined three-dimensional image; The loss function of the volume diffusion generation model is expressed as: ； in, Indicates that the loss function obeys the probability distribution of multiple conditions; represents the random variable used to simulate noise in the diffusion model; represents the noise predicted by the volume diffusion generative model; It represents the product of the denoising coefficients accumulated step by step from time step 1 to t, indicating the denoising strength after multiple time steps; represents the time step; represents the original, unnoised image or data point; Indicates that at each iteration, the model determines the conditional slice according to the probability.

2. The method for generating diffuse 3D medical images with decoupling of mask generation and image generation according to claim 1, characterized in that: The method of forward and backward expansion of the two-dimensional subsequence after adding the initial subsequence is performed by the autoregressive method to obtain a multi-label mask sequence, including: Determining generation parameters of the multi-label mask sequence; The multi-conditional diffusion probability model performs forward and backward expansion on the two-dimensional subsequence after adding the initial subsequence according to the generation parameters and the autoregressive method to obtain a multi-label mask sequence.

3. The method for generating diffuse 3D medical images with decoupling of mask generation and image generation according to claim 1, wherein: The screening of the three-dimensional medical image data according to the screening conditions to determine the informed slice set includes: An extraction starting position is determined according to the screening condition, and a plurality of different informed slices are selected from the three-dimensional medical image through a resampling technique and the extraction starting position, and the plurality of different informed slices are combined into the informed slice set.

4. The method for generating diffuse 3D medical images with decoupling of mask generation and image generation according to claim 1, characterized in that: The informed slices during training of the volume diffusion generation model are extracted from a three-dimensional medical image window.

Images