A cross-field intensity magnetic resonance image synthesis method and system

Abstract

The application provides a cross-field strength magnetic resonance image synthesis method and system, and relates to the technical field of medical image processing. The method comprises the following steps: acquiring low-field magnetic resonance volume data to be processed; determining slice position conditions according to neighborhood information and layer position information of a target slice; acquiring local structure conditions and determining low-field image conditions; inputting the three types of conditions into a trained conditional latent space diffusion network to iteratively denoise latent variables corresponding to the target slice, thereby obtaining a synthesized high-field style magnetic resonance image; establishing diffusion trajectory consistency constraints, closed loop consistency constraints and input traceability constraints during the training process, and training the network according to the constraints. Through the above processing, the spatial relationship of the slice and the local anatomical structure can be considered, the abnormal deviation of the denoising path and the structure change without basis can be limited, and the spatial consistency, detail stability and reliability of the cross-field strength magnetic resonance image synthesis are improved.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, and in particular to a method and system for cross-field strong magnetic resonance image synthesis. Background Technology

[0002] Magnetic resonance imaging (MRI) is an important medical imaging technique that uses the differences in magnetic resonance signals among different tissues to obtain images of internal tissues and organs. MRI machines with different field strengths have varying characteristics in terms of equipment cost, application environment, imaging performance, and usage conditions. Low-field or ultra-low-field MRI machines offer advantages such as flexible deployment and wider applicability; however, due to factors such as equipment performance and acquisition conditions, the MRI images they produce differ from those with higher field strengths in terms of signal-to-noise ratio, image contrast, and detail. Therefore, using image processing techniques to convert MRI images of different field strengths to improve the display quality of low-field MRI images has gradually become a research direction in the field of medical image processing.

[0003] In existing technologies, machine learning or deep learning models can be used to learn the feature relationships between magnetic resonance images of different field strengths and generate corresponding high-field magnetic resonance images based on the input low-field magnetic resonance images. However, different devices may differ in scanning parameters, spatial resolution, imaging range, and image quality, and the effective detail information contained in low-field magnetic resonance images is relatively limited, making the image generation process susceptible to the influence of different data distributions and limited input information. Consequently, the generated images may still have deviations in terms of the consistency of spatial structure, the stable representation of fine tissue structures, and the continuity between adjacent images. At the same time, when the generation model enhances the limited information in the input image, the correspondence between some image features in the generated result and the original anatomical information is difficult to fully guarantee, thus affecting the stability and reliability of the cross-field magnetic resonance image synthesis results.

[0004] Therefore, in cross-field strong magnetic resonance image synthesis, the large data differences between images of different field strengths, the unstable spatial consistency and detail representation of the generated images, and the insufficient reliability of the generated results have become problems that urgently need to be solved. Summary of the Invention

[0005] This application provides a method and system for synthesizing cross-field strong magnetic resonance images, aiming to solve the problems of large data differences between images of different field strengths, unstable spatial consistency and detail representation of generated images, and insufficient reliability of generated results in the existing cross-field strong magnetic resonance image synthesis process.

[0006] In a first aspect, this application provides a method for synthesizing cross-field strong magnetic resonance images, the method comprising: Acquire data from the low-field magnetic resonance imaging (MRI) body to be processed; Based on the neighborhood information and layer information of the target slice in the low-field magnetic resonance imaging data to be processed, the slice location conditions are determined. Obtain the local structural conditions associated with the target slice; Based on the target slice, determine the low-field image conditions; The low-field image conditions, the slice location conditions, and the local structure conditions are input into the trained conditional latent space diffusion network, and the latent variables corresponding to the target slice are iteratively denoised to obtain a synthetic high-field pattern magnetic resonance image. The training process of the conditional latent space diffusion network includes: determining a corresponding deterministic reference denoising state based on the actual denoising state, and establishing diffusion trajectory consistency constraints based on the difference between the actual denoising state and the deterministic reference denoising state; back-mapping the high-field patterned magnetic resonance image generated during training to the low-field domain, and establishing closed-loop consistency constraints based on the correspondence between the back-mapping result and the corresponding low-field magnetic resonance image; establishing input traceability constraints based on the anatomical structure correspondence between the generated high-field patterned magnetic resonance image and the corresponding low-field magnetic resonance image; and training the conditional latent space diffusion network based on the diffusion trajectory consistency constraints, the closed-loop consistency constraints, and the input traceability constraints.

[0007] In one possible design, determining the slice location conditions based on the neighborhood information and layer information of the target slice in the low-field magnetic resonance imaging (MRI) data to be processed includes: Centered on the target slice, a slice neighborhood sequence is constructed based on the spatial relationship between the target slice and adjacent slices; Based on the spatial position of the target slice in the low-field magnetic resonance imaging data to be processed, the normalized layer corresponding to the target slice is determined; The neighborhood context of the slice is represented, and the obtained neighborhood context information is fused with the position information corresponding to the normalized layer to obtain the slice position feature; The slice location conditions are determined based on the slice location characteristics.

[0008] In one possible design, the slice location conditions are determined by a pre-trained slice location encoder, the training process of which includes: Determine the slice location features and layer information corresponding to low-field magnetic resonance slices and candidate high-field magnetic resonance slices respectively; Based on the similarity of slice position features and the layer distance between the low-field magnetic resonance slice and each of the candidate high-field magnetic resonance slices, the slice matching information corresponding to each of the candidate high-field magnetic resonance slices is determined. Based on the slice matching information, a target matching slice corresponding to the low-field magnetic resonance slice is determined from each of the candidate high-field magnetic resonance slices; Based on the positional correspondence between the low-field magnetic resonance slice and the target matching slice, the slice position encoder is trained with positional constraints.

[0009] In one possible design, obtaining the local structural conditions associated with the target slice includes: Obtain multi-scale local structural features associated with the target slice and provided by a pre-trained local structure repair network; The local structural conditions are obtained by performing conditional characterization on the multi-scale local structural features. The local structure repair network is obtained by performing local structural perturbation on high-field magnetic resonance slices and then training the perturbed high-field magnetic resonance slices to restore their structure.

[0010] In one possible design, the local structure repair network is obtained by locally perturbing a high-field magnetic resonance image (MRI) slice and then training the perturbed high-field MRI slice to restore its structure, including: A structural disruption operation is performed on a local region in the high-field magnetic resonance slice to obtain a local structural perturbation sample, wherein the structural disruption operation includes at least one of orientation perturbation, region occlusion, and texture replacement; The local structure repair network is used to restore the structure of the local structure perturbation sample to obtain the local structure repair result. Based on the differences in image reconstruction and local structure between the local structure repair results and the corresponding high-field magnetic resonance slices, the local structure repair constraints are determined. The local structure repair network is trained based on the local structure repair constraints to obtain a local structure repair network used to generate the local structure conditions.

[0011] In one possible design, the step of inputting the low-field image conditions, the slice location conditions, and the local structure conditions into a trained conditional latent space diffusion network, and iteratively denoising the latent variables corresponding to the target slice to obtain a synthesized high-field style magnetic resonance image, includes: The low-field image conditions, the slice location conditions, and the local structure conditions are applied to the backbone features, intermediate layer features, and decoding layer features of the conditional latent space diffusion network, respectively, to obtain multi-conditional denoising features. Based on the current diffusion step and the local structure conditions, the intensity of the conditional action corresponding to the local structure conditions is determined, and the local structure of the multi-conditional denoising features is adjusted based on the intensity of the conditional action to obtain the target denoising features; Based on the target denoising features, determine the noise prediction information corresponding to the current diffusion step, and update the latent variables corresponding to the target slice based on the noise prediction information; The conditional denoising and latent variable updates are repeatedly performed on the updated latent variables until iterative denoising is completed, and the synthesized high-field pattern magnetic resonance image is obtained.

[0012] In one possible design, determining a corresponding deterministic reference denoising state based on the actual denoising state, and establishing diffusion trajectory consistency constraints based on the difference between the actual denoising state and the deterministic reference denoising state, includes: Based on the actual denoising state and noise prediction information corresponding to the current diffusion step, determine the clean latent variable estimate corresponding to the current diffusion step; Based on the clean latent variable estimates and diffusion process parameters, a deterministic reference denoising state corresponding to the next diffusion step is determined; Obtain the actual denoising state of the conditional latent space diffusion network for the next diffusion step, and determine the trajectory deviation information based on the difference between the actual denoising state and the deterministic reference denoising state. Based on the trajectory deviation information corresponding to multiple diffusion steps, the diffusion trajectory consistency constraint is established.

[0013] In one possible design, the step of inversely mapping the high-field patterned magnetic resonance images generated during training to the low-field domain, and establishing closed-loop consistency constraints based on the correspondence between the inverse mapping results and the corresponding low-field magnetic resonance images, includes: The high-field pattern magnetic resonance image generated during the training process is reverse-mapped to a low-field region to obtain a reconstructed low-field magnetic resonance image. Based on the difference between the reconstructed low-field magnetic resonance image and the corresponding low-field magnetic resonance image, the closed-loop consistency constraint is established; The step of establishing input traceability constraints based on the anatomical structure correspondence between the generated high-field pattern magnetic resonance image and the corresponding low-field magnetic resonance image includes: The main structural information of the generated high-field pattern magnetic resonance image and the corresponding low-field magnetic resonance image are determined respectively, and the main structural constraint information is determined based on the difference between the main structural information; Based on the edge position relationship between the generated high-field pattern magnetic resonance image and the corresponding low-field magnetic resonance image after spatial correspondence processing, edge structure constraint information is determined. Based on the main structure constraint information and the edge structure constraint information, the input traceability constraint is established.

[0014] In one possible design, the method further includes: Multiple slice neighborhoods are constructed according to the slice order of the low-field magnetic resonance body data to be processed, and high-field pattern magnetic resonance image synthesis is performed on the target slices in each slice neighborhood to obtain at least one synthesis result corresponding to each target slice. When the same target slice corresponds to multiple synthesis results, the multiple synthesis results are fused, and the synthesis results corresponding to adjacent target slices are corrected for low-frequency intensity consistency to obtain synthesized high-field pattern magnetic resonance data. Based on the inverse mapping consistency information, input traceability information, and adjacent slice consistency information corresponding to the synthesized high-field pattern magnetic resonance data, slice quality evaluation information is determined. Based on the slice quality evaluation information, low-confidence synthetic slices are identified, and verification information corresponding to the low-confidence synthetic slices is generated.

[0015] Secondly, this application provides a cross-field strong magnetic resonance image synthesis system, the system comprising: The data acquisition module is used to acquire the low-field magnetic resonance imager data to be processed. The slice location condition determination module is used to determine the slice location conditions based on the neighborhood information and layer information of the target slice in the low-field magnetic resonance imaging data to be processed. A local structure condition acquisition module is used to acquire local structure conditions associated with the target slice; A low-field image condition determination module is used to determine low-field image conditions based on the target slice; The image synthesis module is used to input the low-field image conditions, the slice location conditions, and the local structure conditions into a trained conditional latent space diffusion network, and to iteratively denoise the latent variables corresponding to the target slice to obtain a synthesized high-field pattern magnetic resonance image. The network training module is used to determine the corresponding deterministic reference denoising state based on the actual denoising state, and to establish diffusion trajectory consistency constraints based on the difference between the actual denoising state and the deterministic reference denoising state; to reverse map the high-field patterned magnetic resonance images generated during training to the low-field domain, and to establish closed-loop consistency constraints based on the correspondence between the reverse mapping results and the corresponding low-field magnetic resonance images; to establish input traceability constraints based on the anatomical structure correspondence between the generated high-field patterned magnetic resonance images and the corresponding low-field magnetic resonance images; and to train the conditional latent space diffusion network based on the diffusion trajectory consistency constraints, the closed-loop consistency constraints, and the input traceability constraints.

[0016] Thirdly, this application provides an electronic device, including: a memory and at least one processor; The memory stores computer-executed instructions; The at least one processor executes computer execution instructions stored in the memory, causing the at least one processor to perform the method described in the first aspect or various possible designs of the first aspect.

[0017] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed, implement the method described in the first aspect or various possible designs of the first aspect.

[0018] This application provides a method and system for synthesizing cross-field magnetic resonance images. The method determines slice location conditions based on the neighborhood and layer information of the target slice in the low-field magnetic resonance image data to be processed. This allows the image synthesis process to combine the spatial location of the target slice in the magnetic resonance image data with its neighborhood relationships for location characterization, thereby reducing the impact of differences in image data of different field intensities on spatial structure consistency. By introducing local structural conditions associated with the target slice and using the low-field image conditions, slice location conditions, and local structural conditions together for latent variable iterative denoising in a conditional latent spatial diffusion network, the generation process retains the input image information while also considering slice location relationships and local structural conditions. The structural features are used to improve the spatial consistency and detail stability of the synthesized image. Furthermore, a diffusion trajectory consistency constraint is established by the difference between the actual denoised state and the deterministic reference denoised state to limit abnormal deviations of the denoising path during the generation process. A closed-loop consistency constraint is established by the reverse mapping of high-field magnetic resonance images to low-field fields. At the same time, an input traceability constraint is established based on the correspondence between the generated image and the corresponding low-field magnetic resonance image. Thus, the conditional latent space diffusion network is constrained from both the generation process and the generation result, reducing structural changes that lack input anatomical basis, thereby improving the stability and reliability of the cross-field strong magnetic resonance image synthesis results. Attached Figure Description

[0019] Figure 1 A schematic flowchart of a cross-field strong magnetic resonance image synthesis method provided in an embodiment of this application; Figure 2 A schematic diagram of the overall process of a cross-field strong magnetic resonance image synthesis method provided in this application embodiment; Figure 3 This is a schematic diagram of a cross-field strong magnetic resonance image synthesis system provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] 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 application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims and drawings of this application are intended to cover non-exclusive inclusion.

[0022] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of the phrase "embodiment" in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0023] In this article, the term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists, A and B can exist simultaneously, and B exists. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0024] Furthermore, the terms "first," "second," etc., in the specification and claims of this application or in the aforementioned drawings are used to distinguish different objects rather than to describe a specific order, and may explicitly or implicitly include one or more of the features.

[0025] In the description of this application, unless otherwise stated, "multiple" and "at least two" mean two or more (including two), and similarly, "multiple groups" and "at least two groups" mean two or more (including two groups).

[0026] In the description of this application, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, "connected" or "linked" can refer not only to a physical connection, but also to an electrical connection or a signal connection. For instance, it can be a direct connection, i.e., a physical connection, or an indirect connection through at least one intermediate component, as long as the circuit is connected. It can also refer to the internal connection between two components. A signal connection can refer not only to a signal connection through a circuit, but also to a signal connection through a medium, such as radio waves. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0027] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, different technical features in this application can be combined with each other.

[0028] Figure 1 This is a schematic flowchart illustrating a cross-field strong magnetic resonance image synthesis method provided in an embodiment of this application. Figure 1 As shown, the method includes S101 to S105, and S101 to S105 are described in detail below.

[0029] The method provided in this application embodiment can be executed by a cross-field strong magnetic resonance image synthesis system. The cross-field strong magnetic resonance image synthesis system can be deployed on an image workstation, server, or other computing device with image processing capabilities. Alternatively, the model training process and the image synthesis process can be deployed on different computing devices according to the computational resource requirements of model training and image synthesis. During the model training phase, a slice location encoder, a local structure repair network, and a conditional latent space diffusion network can be trained using low-field and high-field magnetic resonance data. The conditional latent space diffusion network is then constrained and trained using diffusion trajectory consistency constraints, closed-loop consistency constraints, and input traceability constraints. During the model application phase, the low-field magnetic resonance volume data to be processed can be input into the image synthesis processing flow composed of the trained related networks to generate a corresponding synthesized high-field style magnetic resonance image for the target slice in the low-field magnetic resonance volume data to be processed.

[0030] Figure 2 This is a schematic diagram illustrating the overall flow of a cross-field strong magnetic resonance image synthesis method provided in an embodiment of this application. Figure 2 As shown, the process of synthesizing a cross-field strong magnetic resonance image may include input data acquisition, conditional encoding, latent space diffusion generation, and high-field image restoration.

[0031] During the input data acquisition process, low-field magnetic resonance imaging (MRI) data to be processed can be acquired, and the target slice to be processed, as well as the preceding and following slices adjacent to the target slice, can be determined from the low-field MRI data. The target slice and its adjacent slices are used to provide neighborhood information corresponding to the target slice, and the corresponding layer information is determined based on the spatial position of the target slice in the low-field MRI data. This provides an input basis for the subsequent construction of low-field image conditions, slice position conditions, and local structure conditions.

[0032] In the conditional encoding process, low-field image conditions, slice location conditions, and local structure conditions can be constructed separately. Specifically, an image conditional encoder can be used to feature-encode the low-field magnetic resonance image content in the target slice to obtain the low-field image conditions. A pre-trained slice location encoder can be used to jointly represent the neighborhood information and normalized layers of the target slice, and the slice location conditions can be obtained through conditional mapping. The network can be repaired using pre-trained local structures. Multi-scale local structural features are provided, and the conditions of these features are characterized to obtain local structural conditions. Among them, the low-field image condition is used to characterize the low-field image content of the target slice itself, the slice location condition is used to characterize the spatial layer and neighborhood context of the target slice, and the local structure condition is used to provide prior knowledge of the local anatomical structure.

[0033] During the latent space diffusion generation process, the low-field image conditions are... Slice location conditions and local structural conditions Input the trained conditional latent space diffusion network U θ The Conditional Latent Space Diffusion Network (CLDN) processes noisy latent variables corresponding to the target slice and performs conditional denoising and latent variable updates using three types of conditional information in multiple diffusion steps. This gradually reduces the noise component in the latent variables and causes them to evolve towards a high-field pattern latent space state corresponding to the target slice. Low-field image conditions, slice location conditions, and local structure conditions can be applied to different feature levels of the CLDN, allowing image content information, slice spatial relationships, and prior local anatomical structures to collaboratively participate in the generation of high-field pattern magnetic resonance images during iterative denoising. The specific mechanisms by which these three types of conditions function in the CLDN will be further explained in subsequent embodiments.

[0034] In the high-field image restoration process, a latent space decoding module can be used to decode the latent variables after iterative denoising, restoring the latent space representation to the image space, thus obtaining a synthetic high-field pattern magnetic resonance image corresponding to the target slice. Therefore, the generated result can, while preserving the main anatomical information of the input low-field magnetic resonance image, combine the spatial layer relationship of the target slice and prior local structures to form a corresponding high-field pattern image representation.

[0035] pass Figure 2 The processing flow shown converts the image information of the target slice itself, the neighborhood and layer information of the target slice, and the prior local structure in the low-field magnetic resonance imaging data into low-field image conditions, slice location conditions, and local structure conditions, respectively. These conditions are then used together for latent variable iterative denoising in a conditional latent spatial diffusion network. This allows the image generation process to simultaneously consider the input image content, slice spatial location, and local anatomical structure, thus providing a foundation for improving the spatial consistency, local structural stability, and reliability of cross-field high magnetic resonance imaging synthesis results.

[0036] S101. Obtain the low-field magnetic resonance image data to be processed.

[0037] The low-field magnetic resonance imaging (MRI) data to be processed can be three-dimensional MRI data obtained by scanning a target object using a low-field MRI or ultra-low-field MRI device, which includes multiple low-field MRI slices arranged in spatial order. The target object can be the human head or other target organs or tissue regions to be subjected to MRI imaging, and the target slice is the slice in the low-field MRI data to be processed that is currently used to generate a high-field MRI image.

[0038] In some embodiments, before further processing the low-field magnetic resonance imaging (MRI) data, at least one of the following preprocessing techniques can be performed: bias field correction, spatial orientation unification, voxel resampling, target region cropping, and intensity normalization. Bias field correction reduces low-frequency intensity deviations caused by magnetic field inhomogeneity; spatial orientation unification ensures consistent spatial coordinates across different scans; voxel resampling converts the data to be processed to a preset voxel scale; target region cropping reduces background information irrelevant to the target anatomical region; and intensity normalization can employ image intensity quantile-based normalization to reduce differences in signal intensity range between different scans.

[0039] During model training, high-field magnetic resonance (MR) data corresponding to low-field MR data can also be acquired. For low-field and high-field MR data with a pairing relationship, rigid registration can be performed first, followed by spatial registration that allows for small-range deformation, in order to reduce differences in scanning positions. For data without a strict pairing relationship, the spatial orientation, imaging field of view, and voxel scale of the two can be unified without forcibly establishing a pixel-by-pixel correspondence, thereby avoiding directly using the objectively existing scanning layer deviation between different devices as image generation error.

[0040] S102. Determine the slice location conditions based on the neighborhood information and layer information of the target slice in the low-field magnetic resonance imaging data to be processed.

[0041] Specifically, a slice neighborhood can be formed by taking the target slice as the center and combining it with at least one adjacent slice before and after the target slice. The slice neighborhood can represent the local spatial variation relationship between the target slice and its adjacent slices. At the same time, the corresponding layer information can be determined according to the spatial position of the target slice in the low-field magnetic resonance data to be processed, so that the layer information can represent the spatial position of the target slice relative to the entire magnetic resonance data.

[0042] For example, a three-slice neighborhood can be formed using the target slice and its preceding and following adjacent slices, or the number of adjacent slices can be further increased to form a five-slice neighborhood. For data with consistent slice thickness and continuous slices, the normalized slice level can be determined based on the slice index; for data with missing slices, inconsistent slice thickness, or differences in scan start positions, the normalized slice level can be determined based on the physical spatial coordinates corresponding to the slices, so that the slice information in different magnetic resonance imaging data has a more stable spatial meaning.

[0043] Furthermore, a slice position encoder can be used to perform contextual representation of the image content in the slice's neighborhood, and the obtained neighborhood context information can be fused with the positional representation corresponding to the layer information to obtain the slice position conditions corresponding to the target slice. Thus, the slice position conditions can not only reflect the image representation of the target slice itself, but also carry the spatial relationship between the target slice and adjacent slices, as well as the relative layer information of the target slice in the magnetic resonance imaging data.

[0044] S103. Obtain the local structural conditions associated with the target slice.

[0045] Local structural conditions are used to characterize the local anatomical features associated with the target slice, providing local structural priors to the subsequent image synthesis process, so that the conditional latent space diffusion network can be constrained by local anatomical information while performing high-field pattern transformation.

[0046] Specifically, a pre-trained local structure repair network can be used to provide multi-scale local structural features, and these features can be conditionally represented to form local structural conditions suitable for inputting a conditional latent space diffusion network. The multi-scale local structural features can represent information such as tissue contours, local edges, fine structures, and texture continuity at different spatial scales, thus enabling the local structural conditions to simultaneously contain relatively coarse-grained structural relationships and relatively fine-grained local anatomical information.

[0047] Local structure restoration networks can be pre-trained by artificially perturbing local structures in high-field magnetic resonance imaging (MRI) slices and then restoring the original local structures from the perturbed image. This training process allows the local structure restoration network to learn relatively stable local anatomical relationships within high-field MRI images, rather than simply learning the overall image intensity or texture distribution.

[0048] S104. Determine the low-field image conditions based on the target slice.

[0049] Specifically, the preprocessed target slice can be input into an image conditional encoder. The image conditional encoder performs feature encoding on the image content of the target slice to obtain low-field image conditions. Low-field image conditions are used to characterize the actual low-field magnetic resonance image information existing in the current target slice, including the main anatomical morphology of the target object, tissue signal distribution, and structural information that can be observed in low-field magnetic resonance images.

[0050] Image conditional encoders can be implemented using conditional coding structures configured in convolutional neural networks, encoder networks, or conditional latent space diffusion networks. The low-field image conditions output can be single-scale conditional features or multi-scale conditional features corresponding to different processing scales of the conditional latent space diffusion network. By preserving the image conditions of the target slice itself, subsequent high-field pattern image generation can always be based on the actual anatomical information present in the low-field MRI image to be processed, rather than solely on the statistical distribution of the high-field MRI image.

[0051] It should be noted that low-field image conditions, slice location conditions, and local structure conditions are used to provide different types of conditional information. Among them, low-field image conditions mainly provide image content information of the target slice itself, slice location conditions mainly provide the layer and neighborhood spatial relationship of the target slice, and local structure conditions mainly provide prior information of local anatomical structures. Together, these three constitute the generation conditions of the subsequent conditional latent space diffusion network.

[0052] S105. Input the low-field image conditions, slice location conditions, and local structure conditions into the trained conditional latent space diffusion network, and perform iterative denoising on the latent variables corresponding to the target slice to obtain a synthesized high-field pattern magnetic resonance image.

[0053] Specifically, the conditional latent space diffusion network can include an image latent space coding module, a diffusion denoising network, and a latent space decoding module. By mapping the magnetic resonance image to a latent space with a dimension lower than the original image space through the image latent space coding module, and performing diffusion noise prediction and iterative denoising within the latent space, the computational resources required to perform diffusion operations directly in the original image space can be reduced.

[0054] During model training, clean latent variables can be encoded from real high-field magnetic resonance images. Gaussian noise is then added to these clean latent variables according to a pre-defined diffusion process, resulting in noisy latent variables for different diffusion steps. The conditional latent space diffusion network uses these noisy latent variables as the denoising target and combines low-field image conditions, slice location conditions, and local structure conditions to predict noise information in the corresponding diffusion steps. This allows the network to learn the generation process of gradually recovering high-field magnetic resonance image latent variables from noisy latent variables.

[0055] In the application of the model, latent variables to be denoised can be established for the target slice. Under the combined effects of low-field image conditions, slice location conditions, and local structure conditions, the trained conditional latent space diffusion network repeatedly performs noise prediction and latent variable updates, causing the latent variables to gradually evolve from a high-noise state to a high-field pattern latent space state corresponding to the target slice. After completing a preset number of iterations of denoising, the final latent variables are image decoded to obtain the synthetic high-field pattern magnetic resonance image corresponding to the target slice.

[0056] Among them, low-field image conditions, slice location conditions, and local structure conditions can be applied to different processing levels of the conditional latent space diffusion network so that different types of conditional information play different roles in the diffusion denoising process; the degree of influence of local structure conditions can also be adjusted according to the current diffusion stage, so that the early stage of generation focuses more on the overall structure restoration, while the later stage of generation further strengthens the correction of local details.

[0057] The training process of the Conditional Latent Space Diffusion Network includes: determining the corresponding deterministic reference denoising state based on the actual denoising state, and establishing diffusion trajectory consistency constraints based on the difference between the actual denoising state and the deterministic reference denoising state; back-mapping the high-field patterned magnetic resonance images generated during training to the low-field domain, and establishing closed-loop consistency constraints based on the correspondence between the back-mapping results and the corresponding low-field magnetic resonance images; establishing input traceability constraints based on the anatomical structure correspondence between the generated high-field patterned magnetic resonance images and the corresponding low-field magnetic resonance images; and training the Conditional Latent Space Diffusion Network based on the diffusion trajectory consistency constraints, closed-loop consistency constraints, and input traceability constraints.

[0058] Specifically, during denoising training of the Conditional Latent Space Diffusion Network, the actual denoising state corresponding to different diffusion steps can be recorded. For the current diffusion step, based on the current actual denoising state and the noise prediction result output by the Conditional Latent Space Diffusion Network, the corresponding clean latent variable can be estimated, and the deterministic reference denoising state corresponding to the next diffusion step can be further determined according to the state change relationship defined in the diffusion process. The deterministic reference denoising state does not introduce additional random sampling noise during the determination process, and therefore can serve as a stable reference state corresponding to the actual denoising process. By comparing the actual denoising state with the corresponding deterministic reference denoising state, the degree of deviation of the actual generation path from the reference generation path during the diffusion process can be obtained, and diffusion trajectory consistency constraints can be established accordingly. By restricting the generation process at the intermediate denoising state level, it is possible to avoid large shifts in the intermediate diffusion state and their gradual accumulation into local artifacts when training only based on the final generated image. The specific determination method of the deterministic reference denoising state and diffusion trajectory consistency constraints will be further explained in subsequent embodiments.

[0059] For closed-loop consistency constraints, a reverse mapping network can be set up to convert high-field patterned MRI images to low-field images. After the conditional latent space diffusion network generates the high-field patterned MRI image, this high-field patterned MRI image is input into the reverse mapping network to obtain the reconstructed low-field MRI image. The reconstructed low-field MRI image is then compared with the corresponding low-field MRI image used as the basis for generation. If image structures that are difficult to be supported by the original low-field MRI image are generated during the forward generation process, such structures are usually difficult to stably recover the corresponding low-field MRI image during the reverse mapping process, thus increasing the difference between the reverse mapping result and the corresponding low-field MRI image. Establishing a closed-loop consistency constraint based on this difference can ensure that the high-field patterned MRI image generated by the conditional latent space diffusion network maintains a reversible image content relationship with its corresponding low-field MRI image.

[0060] For input traceability constraints, the structural correspondence between the generated high-field pattern MRI image and the corresponding low-field MRI image can be evaluated from two levels: the main anatomical structure and the observable edge positions. Specifically, the main structural information of the two images can be obtained through low-frequency structure extraction to ensure consistency in organ contours and major tissue structures. Furthermore, image gradients or other information characterizing edge positions can be acquired, and spatial correspondence processing can be performed considering the registration relationship and resolution differences between the low-field and high-field pattern MRI images to ensure that observable anatomical edges in the input low-field MRI image maintain reasonable spatial positions in the synthesized image. Thus, while allowing the high-field pattern MRI image to produce reasonable high-frequency textures and contrast enhancements, unfounded changes to the main anatomical structure and existing edge positions can be restricted. The specific construction methods of closed-loop consistency constraints and input traceability constraints will be further explained in subsequent embodiments.

[0061] In some embodiments, the training objective of the conditional latent space diffusion network may further include a diffusion noise prediction loss used to constrain the noise prediction results, and combined with the position constraint training objective corresponding to the slice position encoder and the structure repair training objective corresponding to the local structure repair network, to perform phased training or joint training on each network. Preferably, the slice position encoder and the local structure repair network can be trained first, and then the trained slice position encoder and the local structure repair network can be connected to the conditional latent space diffusion network with parameter freezing or a low learning rate; after the conditional latent space diffusion network reaches a preset training stable state, the network parameters related to the conditional information are jointly optimized so that the low-field image conditions, slice position conditions, and local structure conditions can work together to participate in the diffusion denoising process. During training, for adjacent slices belonging to the same continuous magnetic resonance imaging (MRI) data, consistent random intensity transformation parameters can be used to avoid unreal intensity fluctuations between adjacent slices caused by the data augmentation process itself.

[0062] This application's embodiments determine slice location conditions based on the target slice's neighborhood and layer information, enabling cross-field image synthesis to consider both the spatial location of the target slice in the magnetic resonance imaging (MRI) data and the relationship between adjacent slices. By combining low-field image conditions, slice location conditions, and local structural conditions for latent space diffusion denoising, the generation process is further constrained by slice spatial relationships and local anatomical structure information while maintaining the main information of the input image, thereby improving the spatial consistency and detail stability of the synthesized image. Simultaneously, a diffusion trajectory consistency constraint is established by the difference between the actual denoising state and the deterministic reference denoising state, limiting abnormal deviations in the intermediate generation path. A closed-loop consistency constraint is established through the reverse mapping from high-field MRI images to the low-field domain, and an input traceability constraint is established based on the anatomical structure correspondence between the generated image and the corresponding low-field MRI image. These measures constrain the conditional latent space diffusion network from both the generation process and the generation result levels, reducing structural changes lacking input anatomical basis, and thus improving the stability and reliability of the cross-field MRI image synthesis results.

[0063] In one possible embodiment, the method steps shown in S102 are implemented by S1021 to S1024, which are described in detail below.

[0064] S1021. Using the target slice as the center, construct a slice neighborhood sequence based on the spatial relationship between the target slice and adjacent slices.

[0065] Specifically, the low-field magnetic resonance imaging (MRI) data to be processed includes multiple low-field MR slices arranged sequentially along the slice direction. For the target slice to be synthesized, adjacent slices can be selected from the front and rear sides of the target slice according to the slice arrangement direction, and combined according to the corresponding spatial order to obtain a slice neighborhood sequence. The slice neighborhood sequence is used to provide local three-dimensional context near the target slice, so that subsequent location representation can simultaneously consider the image information of the target slice itself and the structural change relationship between adjacent slices.

[0066] In one specific implementation, a three-slice neighborhood sequence can be formed by the target slice and one adjacent slice before and after it, or a five-slice neighborhood sequence can be formed by the target slice and two adjacent slices before and after it. The number of slices included in the slice neighborhood sequence can be set according to the slice thickness of the magnetic resonance imaging data, the degree of local structural changes, and processing resources.

[0067] S1022. Determine the normalized layer corresponding to the target slice based on its spatial location in the low-field magnetic resonance imaging data to be processed.

[0068] Specifically, the slice index of the target slice in the low-field magnetic resonance imaging (MRI) data to be processed can be converted into a normalized slice level. When the slices are continuous and of uniform thickness, and the i-th slice is used as the target slice, the normalized slice level is determined according to the formula... Determined; among them, This represents the normalized layer position corresponding to the i-th target slice, where i represents the slice index corresponding to the target slice, i=0,…,N 1, N represents the number of slices contained in the low-field magnetic resonance volume data to be processed. The normalized slice level is used to characterize the relative position of the target slice from the start end to the end end of the volume data.

[0069] For samples with missing slices or inconsistent slice thickness, normalized layers can be calculated using the physical spatial coordinates of each slice instead of simple slice indices, so that the normalized layers in different scan data continue to reflect the actual spatial location of the target slice.

[0070] S1023. Perform neighborhood context representation on the slice neighborhood sequence, and fuse the obtained neighborhood context information with the position information corresponding to the normalized layer to obtain the slice position features.

[0071] Specifically, the slice neighborhood sequence can be input into the slice location encoder. The slice location encoder receives the target slice and its neighboring slices. In one specific implementation, a three-layer shared convolution is used to extract the local texture of each slice, and then the context is aggregated by a one-dimensional convolution along the slice dimension to obtain the neighborhood context information. The shared convolution enables different slices to be represented in a unified feature space, and the one-dimensional convolution along the slice dimension is used to extract the contextual relationships between adjacent slices that change with spatial position.

[0072] For the normalized layer determined in S1022, sinusoidal position encoding can be used to obtain the corresponding position information, which is then concatenated and fused with the neighborhood context information. In one specific implementation, the fused output is a 128-dimensional unitized position feature, thereby obtaining the slice position feature, which simultaneously contains the slice image content, the relative layer, and the neighborhood context.

[0073] S1024. Determine the slice location conditions based on the slice location characteristics.

[0074] Specifically, the slice location features output by the trained slice location encoder can be used as the location representation of the target slice, and the slice location conditions can be obtained through a conditional mapping adapted to the conditional latent space diffusion network. In one specific implementation, the slice location features can be linearly projected to form the slice location conditions for the conditional latent space diffusion network. The trained slice location features can simultaneously encode slice identity, relative order, and neighborhood context, thereby providing location constraints for cross-field strong magnetic resonance image synthesis.

[0075] In this embodiment, a local three-dimensional context is formed by the target slice and its adjacent slices, and the slice position features are formed by combining the normalized layer of the target slice. This makes the slice position features not only reflect the image performance of the target slice itself, but also its relative position and neighborhood relationship in the entire magnetic resonance imaging data. This reduces the possibility of layer confusion when image synthesis is performed based solely on the appearance of a single slice, and provides a position condition basis for improving the spatial consistency of the synthesis results.

[0076] In one possible embodiment, the slice position encoder used in S1024 is trained by the method shown in S10241 to S10244, which are described in detail below.

[0077] S10241. Determine the slice location features and layer information corresponding to the low-field magnetic resonance slice and the candidate high-field magnetic resonance slice, respectively.

[0078] Specifically, during the training of the slice position encoder, corresponding slice neighborhood sequences can be constructed for low-field MRI slices and candidate high-field MRI slices respectively, and the slice position features can be obtained using the same slice position encoding method, while determining the corresponding normalized layer. For paired data, the candidate range can be determined based on rigid registration and small deformation registration; for unpaired data, candidate high-field MRI slices can be determined based on the uniformity of spatial orientation, field of view, and voxel scale, without requiring strict pixel-by-pixel correspondence.

[0079] In one specific implementation, the positional characteristics of the low-field magnetic resonance slice are denoted as u. i The positional characteristics of the high-field magnetic resonance slice are denoted as v. j The normalized layers corresponding to the low-field magnetic resonance slices and the candidate high-field magnetic resonance slices are denoted as r, respectively. i and r j .

[0080] S10242. Based on the similarity of slice position features and the layer distance relationship between the low-field magnetic resonance slices and each candidate high-field magnetic resonance slice, determine the slice matching information corresponding to each candidate high-field magnetic resonance slice.

[0081] Specifically, to avoid mistaking adjacent but different slices for corresponding slices based solely on the similarity of image appearance, both the cosine similarity between slice position features and the distance between normalized layers can be considered simultaneously.

[0082] The normalized matching weight between the i-th low-field MRI slice and the j-th candidate high-field MRI slice can be calculated using the formula... It is confirmed that, among them, Indicates the lower field Zhang slice and Gaochangdi Normalized matching weights for slices; sim(·) represents cosine similarity; and These are the low-field and high-field location characteristics, respectively; and For normalized layers; Characteristic temperature coefficient; The allowable stratigraphic deviation scale is used; the denominator is normalized to the candidate high-field slice set. This formula gives lower weights to candidates that are similar in appearance but far apart in stratigraphic position.

[0083] Preferably, It can be taken as 0.05 to 0.2. It can be between 0.03 and 0.15.

[0084] S10243. Based on the slice matching information, determine the target matching slice corresponding to the low-field magnetic resonance slice from each candidate high-field magnetic resonance slice.

[0085] Specifically, the normalized matching weights corresponding to each candidate high-field magnetic resonance (MR) slice can be compared within the candidate window, and the candidate MR slice with the largest matching weight is selected as the target matching slice for the current low-field MR slice, i.e., as the positive sample for position matching. Thus, the determination of the target matching slice is constrained by both the slice position feature similarity relationship and the slice distance relationship.

[0086] S10244. Based on the positional correspondence between the low-field magnetic resonance slice and the target matching slice, perform positional constraint training on the slice position encoder.

[0087] Specifically, after determining the target matching slice, a position-constrained contrast loss can be constructed based on the normalized matching weights between the low-field magnetic resonance slice and the selected positive sample. The position-constrained contrast loss is determined according to the following formula: ;in, This indicates the contrast loss at slice locations; This represents the weight of the selected positive samples. By minimizing the slice position contrast loss, the consistency of positional representations between low-field and high-field magnetic resonance slices at corresponding layers can be improved, while suppressing the matching degree of non-corresponding layer candidates. The positional features after training simultaneously encode slice identity, relative order, and neighborhood context, and serve as slice position conditions for the conditional latent space diffusion network.

[0088] In this embodiment, a layer distance kernel is further introduced based on the similarity relationship of slice position features, and positive sample selection and position constraint training are completed accordingly. This enables the slice position encoder to learn a more stable layer correspondence between magnetic resonance slices with different field strengths, thereby reducing the possibility of incorrect layer matching due to similar appearance.

[0089] In one possible embodiment, the method steps shown in S103 are implemented by S1031 to S1032, which are described in detail below.

[0090] S1031. Obtain multi-scale local structural features associated with the target slice and provided by a pre-trained local structural repair network.

[0091] Specifically, the local structure restoration network is used to learn relatively stable local anatomical relationships in high-field magnetic resonance images. Before applying the model, the local structure restoration network is trained using high-field magnetic resonance slices with local structural perturbations, enabling it to recover the directional relationships, boundary continuity, and local structural information in the original high-field magnetic resonance image based on the perturbed local regions.

[0092] In one specific implementation, the local structure repair network can adopt a four-level U-Net structure, with the encoder channel count preferably being 32, 64, 128, or 256. The decoder utilizes skip connections to recover edges. After pre-training, the features output at each scale of the local structure repair network encoder are retained as multi-scale local structure features that can be accessed by the subsequent conditional latent space diffusion network. Features at different scales reflect information about the main structure, local edges, and fine structures within different spatial ranges.

[0093] In subsequent image synthesis, multi-scale local structural features associated with the current target slice are acquired to provide local structural priors to the conditional latent space diffusion network. The structural perturbation method, structural restoration method, and training constraints of the local structure repair network are further explained in the following embodiments.

[0094] S1032. Perform conditional characterization on multi-scale local structural features to obtain local structural conditions.

[0095] Specifically, according to the input requirements of conditional information at different processing scales of the conditional latent space diffusion network, the multi-scale local structural features output by the encoder at each scale of the local structure repair network can be organized and conditionally mapped to obtain local structural conditions. These local structural conditions preserve the prior knowledge of multi-scale local anatomy and participate in the local structure recovery of the high-resolution decoding layer during subsequent diffusion denoising.

[0096] In this embodiment, the multi-scale structural priors learned by the local structure repair network through the local structure restoration task are converted into local structural conditions that can participate in the conditional latent space diffusion. This allows the image synthesis process to be further constrained by local anatomical structural relationships in addition to low-field image information and slice location information, thereby improving the stable performance of fine structures and local edges.

[0097] In one possible embodiment, the training process of the local structure repair network in S1031 is implemented through S10311 to S10314, which are described in detail below.

[0098] S10311. Perform a structural destruction operation on a local region in a high-field magnetic resonance slice to obtain a local structural perturbation sample, wherein the structural destruction operation includes at least one of orientation perturbation, region occlusion, and texture replacement.

[0099] Specifically, the high-field magnetic resonance slice can be divided into non-overlapping or partially overlapping local blocks. The preferred size of the local blocks is 16×16, 24×24, or 32×32 pixels, and structural destruction operations are performed on some of these local blocks randomly. Orientation perturbation can be achieved by rotating the local blocks by 90, 180, or 270 degrees (R(·)) to disrupt the original orientation and boundary continuity. Region occlusion or texture replacement is used to simulate local structural defects, where texture replacement can use noise, local mean, or low-frequency textures from adjacent blocks. The local perturbation ratio is preferably 20% to 60%.

[0100] In one specific implementation, the local structural perturbation samples are generated according to the following formula: ;in, This represents the perturbed image, i.e., a sample with local structural perturbation; M is a binary or soft mask, where the rotated image patch is preserved at M=1, and the patch at M=0 is replaced with noise, local mean, or low-frequency texture ξ from neighboring blocks; ⊙ represents element-wise multiplication. Rotation is used to disrupt orientation and boundary continuity, while occlusion or texture replacement is used to simulate missing local structures.

[0101] S10312. Use a local structure repair network to restore the structure of the locally perturbed samples and obtain the local structure repair results.

[0102] Specifically, samples with local structural perturbations can be input into a local structure repair network, which then reconstructs the original high-field magnetic resonance slices before the perturbation. In one specific implementation, the local structure repair network... A four-level U-Net structure is adopted, with the encoder channel count preferably being 32, 64, 128, and 256. Skip connections are used at the decoder to recover local edges. Since the local orientation, boundary continuity, or local texture of the input sample has been disrupted, the network needs to utilize the surrounding context and its learned structural rules to complete the recovery, thereby obtaining the local structure repair result.

[0103] S10313. Based on the differences in image reconstruction and local structure between the local structure repair results and the corresponding high-field magnetic resonance slices, determine the local structure repair constraints.

[0104] Specifically, the differences between the local structure restoration results and the original high-field magnetic resonance slices can be compared from three aspects: pixel reconstruction, structural similarity, and image gradient. A local structure restoration loss is then established, which is determined according to the following formula: ;in, Indicates localized structural repair losses; The results represent the local structure repair; the first term is the pixel L1 loss, used to constrain the differences in overall image reconstruction; SSIM represents structural similarity, used to constrain the degree of local structural similarity. Represents the image gradient; and The weights are non-negative. The gradient term enables the local structure repair network to not only restore the overall brightness, but also the orientation and position of local edges, thus forming training constraints directly related to the local anatomical structure.

[0105] S10314. Train the local structure repair network based on local structure repair constraints to obtain the local structure repair network used to generate local structure conditions.

[0106] Specifically, the parameters of the local structure repair network are iteratively updated with the goal of reducing the loss of local structure repair, allowing the network to gradually learn to recover the original local anatomical structure from samples with directional perturbations, region occlusion, or texture replacement. After pre-training, the parameters are retained... The encoder's features at various scales serve as the multi-scale local structural features of the subsequent conditional latent space diffusion network.

[0107] During the overall model training process, the pre-training of the slice location encoder and the local structure repair network can be completed first, and then the conditional latent space diffusion network can be connected with parameter freezing or a low learning rate. After the diffusion network is trained and stabilized, the high-level features of the conditional projection layer and the local structure repair network can be further unfrozen to enable joint optimization of the three types of conditions.

[0108] In this embodiment, training samples with clear local anomalies are formed by artificially disrupting the structure. At the same time, the local structure restoration process is constrained by pixel reconstruction, structural similarity and gradient difference. This enables the local structure repair network to learn the relatively stable local anatomical structure rules in high-field magnetic resonance images and use its encoder multi-scale features for subsequent local structure condition construction, thereby improving the local structure preservation ability in the image synthesis process.

[0109] In one possible embodiment, the method steps shown in S105 are implemented by S1051 to S1054, which are described in detail below.

[0110] S1051. Apply the low-field image conditions, slice location conditions, and local structure conditions to the backbone features, intermediate layer features, and decoding layer features of the conditional latent space diffusion network, respectively, to obtain multi-conditional denoising features.

[0111] Specifically, the conditional latent space diffusion network can adopt a latent space diffusion structure. During model training, the real high-field magnetic resonance image is converted into a clean latent variable x0 using an autoencoder, and Gaussian noise is added at the t-th diffusion step. The noisy latent variable is calculated according to the formula... Determined; among them, Let t be the noisy latent variable at step t; The cumulative signal retention coefficient; Standard Gaussian noise. The low-field image is obtained through an image conditional encoder to obtain the low-field image condition. The slice location features are obtained through linear projection to obtain the slice location conditions. The multi-scale local structural features output by the local structural repair network form local structural conditions. .

[0112] The noise prediction relationship of a diffusion U-Net with three types of conditions as input is expressed by the following formula: ;in, Predict noise for the network; For the diffusion of parameter θ in U-Net; represents the noisy latent variable corresponding to the current diffusion step; t represents the current diffusion step.

[0113] In one specific implementation, low-field image conditions At each scale, zero-initialized convolutions are applied to the backbone features of the diffused U-Net; slice location conditions. Intermediate layer features are modulated through cross-attention; local structural conditions. High-resolution decoding layer features are injected through gated skip connections. Thus, three types of conditions act on different processing levels, forming multi-condition denoising features.

[0114] S1052. Based on the current diffusion step and local structure conditions, determine the conditional action intensity corresponding to the local structure conditions, and adjust the local structure of the multi-condition denoising features based on the conditional action intensity to obtain the target denoising features.

[0115] Specifically, local structural conditions participate in the high-resolution decoding process through gated skip connections. Their gating coefficients are jointly determined by the current diffusion step and the local structural features, serving as the conditional strength of the local structural conditions. By dynamically adjusting this conditional strength, early denoising can primarily recover the global structure, while later denoising stages can enhance the influence of local structural conditions on fine structures and edge details, thus obtaining the target denoised features after local structural adjustment.

[0116] S1053. Determine the noise prediction information corresponding to the current diffusion step based on the target denoising features, and update the latent variables corresponding to the target slice based on the noise prediction information.

[0117] Specifically, the diffusion U-Net can be used to output the network prediction noise corresponding to the current diffusion step based on the target denoising features. The latent variables corresponding to the target slice are updated according to a preset reverse denoising relationship, based on the diffusion process parameters of the current diffusion step. This results in the updated latent variables having a lower noise level and retaining more latent spatial information related to the target high-field pattern magnetic resonance image compared to the current latent variables. During model training, the network prediction noise is also used to construct the diffusion noise prediction loss.

[0118] S1054. Repeat the conditional denoising and latent variable update for the updated latent variables until iterative denoising is completed, and a synthetic high-field pattern magnetic resonance image is obtained.

[0119] Specifically, the updated latent variables are used as input for the next diffusion step, and the three types of conditional injection, local structure dynamic adjustment, noise prediction, and latent variable update are repeatedly executed until the preset number of diffusion sampling steps is completed. The number of diffusion sampling steps is preferably between 10 and 50. After iterative denoising, the final latent variables are transformed into the image space through the latent space decoding module to obtain the synthetic high-field pattern magnetic resonance image corresponding to the target slice. During training, the same random intensity transformation parameters can be used for consecutive slices to prevent the data augmentation itself from introducing unrealistic inter-slice intensity flicker.

[0120] In this embodiment, low-field image conditions, slice position conditions, and local structure conditions are injected into different processing levels of the conditional latent space diffusion network, and the intensity of the local structure conditions is dynamically adjusted according to the diffusion stage. This allows the network to prioritize the restoration of the overall structure in the early denoising stage and focus on correcting local details in the later denoising stage, thereby taking into account the overall anatomical structure, slice position relationship, and local structural stability during cross-field intensity pattern conversion.

[0121] In one possible embodiment, the steps of determining the corresponding deterministic reference denoising state based on the actual denoising state and establishing diffusion trajectory consistency constraints based on the difference between the actual denoising state and the deterministic reference denoising state are implemented by S1061 to S1064, and S1061 to S1064 are described in detail below.

[0122] S1061. Based on the actual denoising state and noise prediction information corresponding to the current diffusion step, determine the clean latent variable estimate corresponding to the current diffusion step.

[0123] Specifically, during the training of the conditional latent space diffusion network, the actual denoising state of the actual denoising chain at the current diffusion step t can be recorded. And obtain the network predicted noise of the conditional latent space diffusion network based on the actual denoising state. .according to , And the clean latent variable estimation of the diffusion process parameters corresponding to the current diffusion step. Clean latent variable estimation is used to characterize the noiseless latent space state corresponding to the current noise prediction result, and serves as the basis for calculating the deterministic reference state for the next diffusion step.

[0124] S1062. Based on the clean latent variable estimation and diffusion process parameters, determine the deterministic reference denoising state corresponding to the next diffusion step.

[0125] Specifically, after obtaining the clean latent variable estimates, the diffusion equation can be used to directly calculate the diffusion from the current diffusion step. To the next diffusion step The deterministic reference denoising state is determined by the following formula: ;in, This represents the deterministic reference denoised state calculated directly from the diffusion equation; Indicates the first Step-by-step cumulative signal retention coefficient; Indicates by and Estimation of the recovered clean latent variables; This represents the network prediction noise corresponding to the current diffusion step. This deterministic reference denoising state does not sample additional random noise, thus providing a stable path benchmark for the actual generated chain.

[0126] S1063. Obtain the actual denoising state of the conditional latent space diffusion network for the next diffusion step, and determine the trajectory deviation information based on the difference between the actual denoising state and the deterministic reference denoising state.

[0127] Specifically, after the conditional latent space diffusion network is updated from the current diffusion step to the next diffusion step, the actual denoising chain is recorded at the th step. The actual denoising state obtained in the step and compare it with the deterministic reference denoising state determined in S1062. A comparison is made. The difference between the two represents the degree of deviation of the actual generated chain from the deterministic reference path, and can be used as trajectory deviation information.

[0128] S1064. Based on the trajectory deviation information corresponding to multiple diffusion steps, establish diffusion trajectory consistency constraints.

[0129] Specifically, the difference between the actual denoised state and the corresponding deterministic reference denoised state can be calculated for each of the multiple diffusion steps, and the L2 squared norm can be used as a metric to establish the trajectory consistency loss. The trajectory consistency loss is calculated according to the formula... Determined; among them, This represents the trajectory consistency loss; Indicates the first Step weights can be set to have larger weights in the later steps closer to the final output; T represents the total number of diffusion steps; the L2 squared norm is used to limit the deviation of the actual denoising state from the corresponding deterministic reference denoising state. This loss directly constrains the intermediate generation path, rather than only judging the result on the final image.

[0130] In this embodiment, a deterministic reference denoising state without additional random noise sampling is constructed using the current actual denoising state and predicted noise. The deviation between the actual state and the reference state in multiple diffusion steps is directly constrained, so that the conditional latent space diffusion network can suppress abnormal shifts in intermediate denoising paths during the training phase and reduce the possibility of path deviations accumulating to form local artifacts in subsequent denoising processes.

[0131] In one possible embodiment, the process of establishing the closed-loop consistency constraint and input traceability constraint shown in S106 is implemented through S1065 to S1069, which are described in detail below.

[0132] S1065. Perform low-field inverse mapping on the high-field pattern magnetic resonance image generated during the training process to obtain the reconstructed low-field magnetic resonance image.

[0133] Specifically, a high-field to low-field inverse mapping network can be set up. The high-field patterned magnetic resonance image generated by the conditional latent space diffusion network is converted back to the low-field domain to obtain a reconstructed low-field magnetic resonance image. In one specific implementation, the forward network from low field to high field... With reverse mapping network It shares some image encoders but uses independent normalization parameters and output heads to adapt to low-field and high-field image distributions respectively.

[0134] S1066. Based on the difference between the reconstructed low-field magnetic resonance image and the corresponding low-field magnetic resonance image, establish a closed-loop consistency constraint.

[0135] Specifically, a low-field-high-field-low-field closed loop can be formed through forward and reverse mapping, and a reverse loop in the high-field-low-field-high-field direction can be further formed using real high-field magnetic resonance images. The closed-loop consistency constraint is represented by a cyclic consistency loss in one specific implementation, as shown in the following formula: ;in, This represents the loss of cycle consistency. This represents a low-field magnetic resonance image; Represents a true high-field magnetic resonance image; This represents a positive mapping from a low field to a high field; This represents the inverse mapping from high field to low field; This represents the reverse loop weight.

[0136] The first term directly corresponds to the closed-loop difference between the generated high-field pattern magnetic resonance image after inverse mapping in the low-field domain and the corresponding low-field magnetic resonance image. The second term is used to provide auxiliary constraints on the inverse loop in the high-field-low-field-high-field direction. If the forward network generates a structure that does not exist in the input, this structure is usually difficult to stably recover the original low-field magnetic resonance image after inverse mapping, and therefore will be limited by the cycle consistency loss.

[0137] S1067. Determine the main structural information of the generated high-field pattern magnetic resonance image and the corresponding low-field magnetic resonance image respectively, and determine the main structural constraint information based on the differences between the main structural information.

[0138] Specifically, to distinguish between permissible high-field pattern enhancement and permissible anatomical structural alterations, a fixed low-pass filter or a structure extraction operator that is downsampled and then upsampled can be used. Low-frequency structural information of the main body is extracted from the generated high-field pattern magnetic resonance images and their corresponding low-field magnetic resonance images. The difference between the low-frequency structures of the two images is used as structural constraint information to preserve organ contours and main anatomical structures.

[0139] S1068. Determine the edge structure constraint information based on the edge position relationship between the generated high-field pattern magnetic resonance image and the corresponding low-field magnetic resonance image after spatial correspondence processing.

[0140] Specifically, image gradients can be used to characterize the locations of observable edges. Considering the potential registration relationships and resolution differences between low-field and high-field magnetic resonance images, spatial mapping operators can be utilized. Spatial mapping is performed on the gradient of the input low-field magnetic resonance image based on the registration field and resolution difference. The difference between the generated image gradient and the mapped input gradient is then compared to obtain edge structure constraint information, which constrains the position of observable edges without requiring the low-field image and the high-field pattern image to have exactly the same high-frequency texture.

[0141] S1069. Based on the main structure constraint information and the edge structure constraint information, establish input traceability constraints.

[0142] Specifically, the low-frequency main structure differences and the gradient position differences after spatial correspondence processing can be used together to construct the input traceability loss, which is based on the formula... It is confirmed that, among them, Indicates input traceability loss; This represents a synthesized high-field pattern magnetic resonance image; Represents the gradient; This represents the gradient weights. The first term corresponds to the main structure constraint information, used to maintain the organ contour and main structure; the second term corresponds to the edge structure constraint information, used to constrain the position of observable edges in the input image.

[0143] Furthermore, in one specific implementation, the diffusion noise prediction loss, slice position comparison loss, local structure repair loss, cycle consistency loss, trajectory consistency loss, and input traceability loss can be combined to form an overall training objective, which is based on the formula... It is confirmed that, among them, Predicting loss for diffused noise; For slice location comparison loss; For local structural repair losses; This is a loss due to circular consistency. For trajectory consistency loss; For input traceable loss; , , , and The weights are non-negative. Ideally, the slice location encoder and the local structure repair network should be trained first, and then connected to the conditional latent space diffusion network in a frozen or low-learning-rate manner. Once the diffusion network is stable, the high-level features of the conditional projection layer and the local structure repair network are unfrozen, allowing for joint optimization of the three types of conditions.

[0144] In this embodiment, the low-field-high-field-low-field reverse mapping closed loop is used to restrict the generated result from deviating from the input image content. At the same time, the input traceability constraint is established by the low-frequency main structure and the spatially corresponding edge position. This allows the generated high-field style magnetic resonance image to maintain the main anatomical structure and observable edge position that can be traced from the input low-field magnetic resonance image while allowing reasonable improvement in contrast and high-frequency texture. This improves the reliability of the cross-field strong magnetic resonance image synthesis result.

[0145] In one possible embodiment, after the method completes the high-field pattern magnetic resonance image synthesis of the target slice, it can also realize three-dimensional inference, post-processing and quality gating through S1071 to S1074. S1071 to S1074 are described in detail below.

[0146] S1071. Construct multiple slice neighborhoods according to the slice order of the low-field magnetic resonance body data to be processed, and perform high-field pattern magnetic resonance image synthesis on the target slices in each slice neighborhood to obtain at least one synthesis result corresponding to each target slice.

[0147] Specifically, during the model inference phase, a sliding slice neighborhood can be constructed according to the slice order of the low-field magnetic resonance imaging (MRI) data to be processed. The position encoder outputs the slice position conditions corresponding to each target slice, the local structure repair network provides multi-scale local structure priors associated with the target slice, and the conditional latent space diffusion network combines the low-field image conditions, slice position conditions, and local structure conditions to generate the corresponding high-field pattern MRI slices. Due to the overlap between adjacent sliding windows, the same target slice may generate multiple prediction results from multiple sliding windows.

[0148] S1072. When the same target slice corresponds to multiple synthesis results, the multiple synthesis results are fused, and the low-frequency intensity consistency correction is performed on the synthesis results corresponding to adjacent target slices to obtain the synthesized high-field pattern magnetic resonance data.

[0149] Specifically, when adjacent sliding windows produce multiple composite results for the same target slice, the fusion weight can be determined based on the position of the target slice relative to the center of each sliding window, and a weighted average can be performed according to the center distance, so that the prediction result located at the center of the window has a relatively high fusion contribution.

[0150] After fusing the target slices, a synthetic slice sequence is formed according to the original slice order, and a lightweight one-dimensional smoothing is performed along the slice dimensions to correct the low-frequency intensity consistency of the synthetic results corresponding to adjacent target slices. One-dimensional smoothing only corrects the low-frequency intensity variations between slices without moving the already generated anatomical edges, thereby reducing inter-slice brightness flicker while maintaining the position of local anatomical structures, resulting in synthetic high-field pattern magnetic resonance imaging (MRI) data.

[0151] S1073. Determine slice quality evaluation information based on the inverse mapping consistency information, input traceability information, and adjacent slice consistency information corresponding to the synthesized high-field pattern magnetic resonance data.

[0152] Specifically, for each target synthetic slice in the synthetic high-field pattern magnetic resonance image data, the inverse cyclic residual, the input traceable residual, and the feature difference between adjacent slices can be calculated separately.

[0153] Specifically, the reverse cyclic residual is used to characterize the degree of closed-loop consistency between the synthesized slice and the corresponding low-field magnetic resonance slice after low-field reverse mapping; the input traceable residual is used to characterize the traceability of the main structure and edge positions between the synthesized slice and the corresponding low-field magnetic resonance slice; and the adjacent slice feature difference is used to characterize the continuity between the current synthesized slice and the adjacent synthesized slices before and after it. Slice quality evaluation information is formed based on the above information.

[0154] In one possible implementation, diffusion trajectory consistency information corresponding to the target synthesized slice can also be obtained. Specifically, the trajectory consistency residual corresponding to the target synthesized slice can be determined based on the differences between the actual denoising state and the deterministic reference denoising state corresponding to multiple diffusion steps during the generation of the target synthesized slice. The trajectory consistency residual is used to characterize the degree of deviation of the actual denoising path of the conditional latent space diffusion network from the deterministic reference denoising path during the generation of the target synthesized slice. Further, the diffusion trajectory consistency information can be used together with at least one of the following: the reverse mapping consistency information, the input traceability information, and the adjacent slice consistency information, to determine the slice quality evaluation information. When the trajectory consistency residual is large, it indicates that there is a significant abnormal deviation in the intermediate denoising path of the corresponding target synthesized slice. Therefore, the credibility of the target synthesized slice can be comprehensively evaluated by combining it with other quality evaluation information corresponding to the final synthesis result, thereby improving the reliability of low-credibility synthesized slice identification.

[0155] S1074. Determine low-confidence synthetic slices based on slice quality evaluation information, and generate verification information corresponding to the low-confidence synthetic slices.

[0156] Specifically, thresholds can be set for the inverse cyclic residual, the input traceable residual, and the feature difference between adjacent slices. When any indicator exceeds the corresponding threshold, the corresponding synthesized slice is marked as a low-confidence synthesized slice. For low-confidence synthesized slices, the corresponding original low-field magnetic resonance slice, the synthesized high-field pattern magnetic resonance slice, and the difference map between the two can be output as verification information for manual verification. Low-confidence synthesized slices are not used to directly replace the original low-field magnetic resonance image.

[0157] In one specific implementation, after completing the synthesis results and quality gating, the original spatial orientation, voxel spacing, patient anonymity identifier, and sequence description can be written back to form a new DICOM sequence, so as to maintain the spatial metadata correspondence between the generated volume data and the original magnetic resonance data.

[0158] In this embodiment, sliding slice neighborhood inference is performed according to the volume data sequence, and multiple prediction results for the same target slice are fused and inter-layer low-frequency intensity consistency correction is performed. This can improve the inter-layer brightness flicker and structural discontinuity caused by independent inference for each slice. Furthermore, quality evaluation is performed by combining reverse mapping consistency, input traceability and adjacent slice consistency, and verification information is output for low-confidence synthesized slices. This can effectively screen abnormal synthesized slices while forming three-dimensional synthesis results, thereby improving the continuity and reliability of synthesized high-field pattern magnetic resonance volume data.

[0159] In one possible embodiment, Figure 3 This is a schematic diagram of a cross-field strong magnetic resonance image synthesis system provided in an embodiment of this application. Figure 3 As shown, the cross-field strong magnetic resonance image synthesis system 300 includes a data acquisition module 301, a slice position condition determination module 302, a local structure condition acquisition module 303, a low-field image condition determination module 304, an image synthesis module 305, and a network training module 306.

[0160] The data acquisition module 301 is used to acquire low-field magnetic resonance body data to be processed.

[0161] Specifically, the data acquisition module can acquire low-field magnetic resonance imaging (MRI) data to be processed from a low-field MRI device, a medical image storage system, or a preset data storage space. The low-field MRI data to be processed includes multiple low-field MRI slices arranged in spatial order, and at least one of the following preprocessing steps can be performed after input: bias field correction, spatial orientation unification, voxel resampling, target region cropping, and intensity normalization, to form low-field MRI data for subsequent image synthesis.

[0162] The slice location condition determination module 302 is used to determine the slice location conditions based on the neighborhood information and layer information of the target slice in the low-field magnetic resonance imaging data to be processed.

[0163] Specifically, the slice location condition determination module 302 can construct a slice neighborhood sequence centered on the target slice, determine the corresponding normalized layer based on the spatial location of the target slice in the low-field magnetic resonance imaging data to be processed, and then jointly represent the neighborhood context information and the location information corresponding to the normalized layer through a pre-trained slice location encoder to obtain slice location conditions used to describe the spatial identity, relative order and neighborhood relationship of the target slice.

[0164] The local structure condition acquisition module 303 is used to acquire the local structure conditions associated with the target slice.

[0165] Specifically, the local structure condition acquisition module 303 can acquire multi-scale local structure features provided by a pre-trained local structure repair network, and perform conditional characterization on the multi-scale local structure features to obtain local structure conditions. The local structure repair network learns stable local anatomical relationships in high-field magnetic resonance images by perturbing the local structure of high-field magnetic resonance slices and performing structure restoration training, thereby providing local structure priors for image synthesis.

[0166] The low-field image condition determination module 304 is used to determine the low-field image conditions based on the target slice.

[0167] Specifically, the low-field image condition determination module 304 can perform feature encoding on the target slice to obtain low-field image conditions that characterize the main anatomical structure, tissue signal distribution, and low-field magnetic resonance image content of the target slice, so that the subsequent generation process of high-field style magnetic resonance image can be based on the image information that actually exists in the target slice.

[0168] The image synthesis module 305 is used to input the low-field image conditions, the slice location conditions, and the local structure conditions into a trained conditional latent space diffusion network, and to iteratively denoise the latent variables corresponding to the target slice to obtain a synthesized high-field pattern magnetic resonance image.

[0169] Specifically, the image synthesis module 305 can apply low-field image conditions, slice location conditions, and local structure conditions to different feature levels of the conditional latent space diffusion network, and repeatedly perform conditional denoising and latent variable updates in multiple diffusion steps. As the iterative denoising process progresses, the latent variables corresponding to the target slice gradually recover from a noise state to a high-field latent space representation, and after the iteration is completed, a synthesized high-field magnetic resonance image is obtained through latent space decoding.

[0170] The network training module 306 is used to determine a corresponding deterministic reference denoising state based on the actual denoising state, and to establish a diffusion trajectory consistency constraint based on the difference between the actual denoising state and the deterministic reference denoising state; to reverse map the high-field pattern magnetic resonance image generated during training to the low-field domain, and to establish a closed-loop consistency constraint based on the correspondence between the reverse mapping result and the corresponding low-field magnetic resonance image; to establish an input traceability constraint based on the anatomical structure correspondence between the generated high-field pattern magnetic resonance image and the corresponding low-field magnetic resonance image; and to train the conditional latent space diffusion network based on the diffusion trajectory consistency constraint, the closed-loop consistency constraint, and the input traceability constraint.

[0171] Specifically, the network training module 306 can record the actual denoising state corresponding to multiple diffusion steps during the training of the conditional latent space diffusion network, and use the latent variables and noise prediction information corresponding to the current diffusion step to determine a deterministic reference denoising state that does not introduce additional random noise, thereby limiting abnormal deviations in the actual denoising path. Simultaneously, it can constrain the closed-loop relationship between the generated image and the corresponding low-field magnetic resonance image through a high-field to low-field inverse mapping process, and constrain the anatomical structure correspondence between the generated image and the input low-field magnetic resonance image from both the main structure and edge position aspects. After training the conditional latent space diffusion network based on the above constraints, a trained conditional latent space diffusion network can be obtained for use by the image synthesis module.

[0172] In actual deployment, the above modules can be implemented by different software functional units within the same computing device, or they can be deployed separately according to the computing resource requirements of model training and model application. For example, the network training module can be deployed on a training server, while the data acquisition module, slice location condition determination module, local structure condition acquisition module, low-field image condition determination module, and image synthesis module can be deployed on an image workstation or inference server. Each module completes the transfer of low-field magnetic resonance body data, conditional features, and model parameters through a data interface.

[0173] In one possible implementation, the slice location encoder and the local structure repair network can be trained beforehand and then integrated into the conditional latent space diffusion network as conditional processing modules with fixed or fine-tunable parameters after training. When using the fixed-parameter approach, the corresponding network parameters can remain unchanged during the training process of the conditional latent space diffusion network. When using the fine-tunable approach, the corresponding network parameters can be jointly optimized with a low learning rate, and the network parameters related to the conditional information can be further adjusted after the conditional latent space diffusion network reaches a preset stable state, so that the slice location conditions and local structure conditions can adapt to the cross-field strong magnetic resonance image synthesis task.

[0174] During model application, back-mapping correlation processing and diffusion trajectory consistency correlation processing can be primarily used for quality assessment of the synthesis results, rather than serving as a necessary inference path for obtaining synthesized high-field pattern magnetic resonance images. Specifically, at least one of back-mapping correlation processing and diffusion trajectory consistency correlation processing can be enabled based on the processing resources of the computing device, inference efficiency requirements, and quality assessment needs. When the corresponding quality assessment is not required, the corresponding processing branch can be disabled to reduce computational overhead during model application.

[0175] The aforementioned cross-field high magnetic resonance image synthesis system utilizes the combined constraints of slice location conditions, local structural conditions, and low-field image conditions on the latent space diffusion network image generation process. This enables the synthesized high-field magnetic resonance images to take into account the original image content, spatial layering, and local anatomical structures of the target slice. Simultaneously, it limits deviations in the intermediate denoising path through diffusion trajectory consistency constraints, and restricts deviations from the anatomical basis in the input low-field magnetic resonance image through closed-loop consistency constraints and input traceability constraints. This improves the spatial consistency, detail stability, and reliability of the cross-field high magnetic resonance image synthesis results.

[0176] In one possible embodiment, this application also provides an electronic device, the electronic device including a processor and a memory communicatively connected to the processor. The memory stores a computer program executable by the processor, which, when executed by the processor, causes the electronic device to perform the cross-field strong magnetic resonance image synthesis method described in any of the above embodiments.

[0177] The electronic device can be an image workstation, a model training server, a model inference server, or other computing devices capable of processing magnetic resonance image data. The processor can call computer programs stored in the memory to perform slice location condition determination, local structure condition acquisition, low-field image condition determination, conditional latent space diffusion denoising, and corresponding network training or quality evaluation processing on the low-field magnetic resonance body data to be processed.

[0178] In one possible embodiment, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the cross-field strong magnetic resonance image synthesis method described in any of the above embodiments. The computer-readable storage medium can be a non-transitory storage medium capable of storing a computer program and being read by a processor.

[0179] The above are merely specific embodiments of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to these embodiments, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for synthesizing cross-field strong magnetic resonance images, characterized in that, The method includes: Acquire data from the low-field magnetic resonance imaging (MRI) body to be processed; Based on the neighborhood information and layer information of the target slice in the low-field magnetic resonance imaging data to be processed, the slice location conditions are determined. Obtain the local structural conditions associated with the target slice; Based on the target slice, determine the low-field image conditions; The low-field image conditions, the slice location conditions, and the local structure conditions are input into the trained conditional latent space diffusion network, and the latent variables corresponding to the target slice are iteratively denoised to obtain a synthetic high-field pattern magnetic resonance image. The training process of the conditional latent space diffusion network includes: determining a corresponding deterministic reference denoising state based on the actual denoising state, and establishing diffusion trajectory consistency constraints based on the difference between the actual denoising state and the deterministic reference denoising state; back-mapping the high-field patterned magnetic resonance image generated during training to the low-field domain, and establishing closed-loop consistency constraints based on the correspondence between the back-mapping result and the corresponding low-field magnetic resonance image; establishing input traceability constraints based on the anatomical structure correspondence between the generated high-field patterned magnetic resonance image and the corresponding low-field magnetic resonance image; and training the conditional latent space diffusion network based on the diffusion trajectory consistency constraints, the closed-loop consistency constraints, and the input traceability constraints.

2. The method according to claim 1, characterized in that, The step of determining the slice location conditions based on the neighborhood information and layer information of the target slice in the low-field magnetic resonance imaging data to be processed includes: Centered on the target slice, a slice neighborhood sequence is constructed based on the spatial relationship between the target slice and adjacent slices; Based on the spatial position of the target slice in the low-field magnetic resonance imaging data to be processed, the normalized layer corresponding to the target slice is determined; The neighborhood context of the slice is represented, and the obtained neighborhood context information is fused with the position information corresponding to the normalized layer to obtain the slice position feature; The slice location conditions are determined based on the slice location characteristics.

3. The method according to claim 2, characterized in that, The slice location conditions are determined by a pre-trained slice location encoder, the training process of which includes: Determine the slice location features and layer information corresponding to low-field magnetic resonance slices and candidate high-field magnetic resonance slices respectively; Based on the similarity of slice position features and the layer distance between the low-field magnetic resonance slice and each of the candidate high-field magnetic resonance slices, the slice matching information corresponding to each of the candidate high-field magnetic resonance slices is determined. Based on the slice matching information, a target matching slice corresponding to the low-field magnetic resonance slice is determined from each of the candidate high-field magnetic resonance slices; Based on the positional correspondence between the low-field magnetic resonance slice and the target matching slice, the slice position encoder is trained with positional constraints.

4. The method according to claim 1, characterized in that, The step of obtaining the local structural conditions associated with the target slice includes: Obtain multi-scale local structural features associated with the target slice and provided by a pre-trained local structure repair network; The local structural conditions are obtained by performing conditional characterization on the multi-scale local structural features. The local structure repair network is obtained by performing local structural perturbation on high-field magnetic resonance slices and then training the perturbed high-field magnetic resonance slices to restore their structure.

5. The method according to claim 4, characterized in that, The local structure repair network is obtained by performing local structural perturbation on high-field magnetic resonance slices and then training the perturbed high-field magnetic resonance slices to restore their structure, including: A structural disruption operation is performed on a local region in the high-field magnetic resonance slice to obtain a local structural perturbation sample, wherein the structural disruption operation includes at least one of orientation perturbation, region occlusion, and texture replacement; The local structure repair network is used to restore the structure of the local structure perturbation sample to obtain the local structure repair result. Based on the differences in image reconstruction and local structure between the local structure repair results and the corresponding high-field magnetic resonance slices, the local structure repair constraints are determined. The local structure repair network is trained based on the local structure repair constraints to obtain a local structure repair network used to generate the local structure conditions.

6. The method according to claim 1, characterized in that, The step of inputting the low-field image conditions, the slice location conditions, and the local structure conditions into a trained conditional latent space diffusion network, and iteratively denoising the latent variables corresponding to the target slice to obtain a synthesized high-field pattern magnetic resonance image includes: The low-field image conditions, the slice location conditions, and the local structure conditions are applied to the backbone features, intermediate layer features, and decoding layer features of the conditional latent space diffusion network, respectively, to obtain multi-conditional denoising features. Based on the current diffusion step and the local structure conditions, the intensity of the conditional action corresponding to the local structure conditions is determined, and the local structure of the multi-conditional denoising features is adjusted based on the intensity of the conditional action to obtain the target denoising features; Based on the target denoising features, determine the noise prediction information corresponding to the current diffusion step, and update the latent variables corresponding to the target slice based on the noise prediction information; The conditional denoising and latent variable updates are repeatedly performed on the updated latent variables until iterative denoising is completed, and the synthesized high-field pattern magnetic resonance image is obtained.

7. The method according to claim 1, characterized in that, The step of determining a corresponding deterministic reference denoising state based on the actual denoising state, and establishing a diffusion trajectory consistency constraint based on the difference between the actual denoising state and the deterministic reference denoising state, includes: Based on the actual denoising state and noise prediction information corresponding to the current diffusion step, determine the clean latent variable estimate corresponding to the current diffusion step; Based on the clean latent variable estimates and diffusion process parameters, a deterministic reference denoising state corresponding to the next diffusion step is determined; Obtain the actual denoising state of the conditional latent space diffusion network for the next diffusion step, and determine the trajectory deviation information based on the difference between the actual denoising state and the deterministic reference denoising state. Based on the trajectory deviation information corresponding to multiple diffusion steps, the diffusion trajectory consistency constraint is established.

8. The method according to claim 1, characterized in that, The step of inversely mapping the high-field pattern magnetic resonance images generated during training to the low-field domain, and establishing closed-loop consistency constraints based on the correspondence between the inverse mapping results and the corresponding low-field magnetic resonance images, includes: The high-field pattern magnetic resonance image generated during the training process is reverse-mapped to a low-field region to obtain a reconstructed low-field magnetic resonance image. Based on the difference between the reconstructed low-field magnetic resonance image and the corresponding low-field magnetic resonance image, the closed-loop consistency constraint is established; The step of establishing input traceability constraints based on the anatomical structure correspondence between the generated high-field pattern magnetic resonance image and the corresponding low-field magnetic resonance image includes: The main structural information of the generated high-field pattern magnetic resonance image and the corresponding low-field magnetic resonance image are determined respectively, and the main structural constraint information is determined based on the difference between the main structural information; Based on the edge position relationship between the generated high-field pattern magnetic resonance image and the corresponding low-field magnetic resonance image after spatial correspondence processing, edge structure constraint information is determined. Based on the main structure constraint information and the edge structure constraint information, the input traceability constraint is established.

9. The method according to claim 1, characterized in that, The method further includes: Multiple slice neighborhoods are constructed according to the slice order of the low-field magnetic resonance body data to be processed, and high-field pattern magnetic resonance image synthesis is performed on the target slices in each slice neighborhood to obtain at least one synthesis result corresponding to each target slice. When the same target slice corresponds to multiple synthesis results, the multiple synthesis results are fused, and the synthesis results corresponding to adjacent target slices are corrected for low-frequency intensity consistency to obtain synthesized high-field pattern magnetic resonance data. Based on the inverse mapping consistency information, input traceability information, and adjacent slice consistency information corresponding to the synthesized high-field pattern magnetic resonance data, slice quality evaluation information is determined. Based on the slice quality evaluation information, low-confidence synthetic slices are identified, and verification information corresponding to the low-confidence synthetic slices is generated.

10. A cross-field strong magnetic resonance image synthesis system, characterized in that, The system includes: The data acquisition module is used to acquire the low-field magnetic resonance imager data to be processed. The slice location condition determination module is used to determine the slice location conditions based on the neighborhood information and layer information of the target slice in the low-field magnetic resonance imaging data to be processed. A local structure condition acquisition module is used to acquire local structure conditions associated with the target slice; A low-field image condition determination module is used to determine low-field image conditions based on the target slice; The image synthesis module is used to input the low-field image conditions, the slice location conditions, and the local structure conditions into a trained conditional latent space diffusion network, and to iteratively denoise the latent variables corresponding to the target slice to obtain a synthesized high-field pattern magnetic resonance image. The network training module is used to determine the corresponding deterministic reference denoising state based on the actual denoising state, and to establish diffusion trajectory consistency constraints based on the difference between the actual denoising state and the deterministic reference denoising state; to reverse map the high-field patterned magnetic resonance images generated during training to the low-field domain, and to establish closed-loop consistency constraints based on the correspondence between the reverse mapping results and the corresponding low-field magnetic resonance images; to establish input traceability constraints based on the anatomical structure correspondence between the generated high-field patterned magnetic resonance images and the corresponding low-field magnetic resonance images; and to train the conditional latent space diffusion network based on the diffusion trajectory consistency constraints, the closed-loop consistency constraints, and the input traceability constraints.

Images