MRI image reconstruction method and device, electronic device and storage medium

Abstract

The invention relates to an MRI (Magnetic Resonance Imaging) image reconstruction method and device, an electronic device and a storage medium, and the method comprises the steps: obtaining multi-contrast MRI sample data which comprises a target modal high-resolution image and an auxiliary modal high-resolution image; performing frequency domain conversion on the target modal high-resolution image and cutting off high-frequency components to generate low-resolution frequency domain data; carrying out spatial domain conversion on the image to generate a spatial domain low-resolution image; inputting the low-resolution frequency domain data and the spatial domain low-resolution image into a diffusion model, and performing feature extraction on the two types of low-resolution data in a frequency domain and a spatial domain by using the diffusion model to generate a current time step feature; and carrying out feature fusion on the current time step feature and the feature of the high-resolution image of the auxiliary mode by using a structure attention mechanism to obtain a reconstructed image of the target mode. According to the invention, the balance problem of spectrum fidelity and detail enhancement in a traditional MRI super-division method is solved, and the reconstruction efficiency and the image quality are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of medical image processing, and particularly to an MRI image reconstruction method, apparatus, electronic device, and storage medium. Background Art

[0002] Magnetic Resonance Imaging (MRI), as a core tool for clinical diagnosis, its resolution directly affects the accuracy of lesion detection. However, limited by hardware costs, scanning time, and patient tolerance, low-resolution (LR) MRI images are widespread, making it difficult to identify tiny lesions (such as early tumors, microbleeds).

[0003] Current mainstream technologies are mainly divided into three categories: frequency-domain reconstruction methods, spatial-domain deep learning models, and diffusion models, but all have significant limitations: First, frequency-domain methods (such as k-space reconstruction) directly supplement high-frequency signals in the frequency domain through technologies such as compressed sensing. Although they can maintain spectral physical fidelity, they are prone to losing spatial texture details (such as blood vessel branches, tissue edges), resulting in visually blurred reconstructed images and being difficult to meet the needs of clinical diagnosis. Second, spatial-domain deep learning models (such as U-Net) are based on end-to-end learning in the image domain and can effectively enhance visual details, but lack the modeling of imaging physical constraints, which may lead to anatomical structure distortion (such as organ shape deviation), affecting diagnostic reliability. Third, when directly applying diffusion models to MRI, traditional diffusion models generate images by gradually denoising random noise, which does not match the essence of MRI degradation (loss of low-frequency information), resulting in a large number of iterative steps required to start from pure noise, and long inference time; because MRI degradation is not global noise pollution, blind denoising is prone to introducing artifacts (such as false lesions). In addition, existing dual-domain (frequency domain + spatial domain) fusion mostly adopts static strategies and cannot dynamically exchange cross-domain high-frequency residuals, resulting in error accumulation.

[0004] Currently, no effective solution has been proposed for the balance problem between spectral fidelity and detail enhancement in traditional MRI super-resolution methods. Summary of the Invention

[0005] Embodiments of this application provide an MRI image reconstruction method, apparatus, electronic device, and storage medium to at least solve the balance problem between spectral fidelity and detail enhancement in traditional MRI super-resolution methods.

[0006] In a first aspect, embodiments of this application provide an MRI image reconstruction method, including:

[0007] Obtain multi-contrast MRI sample data, including a high-resolution image of a target modality and a high-resolution image of an auxiliary modality;

[0008] Perform frequency domain conversion on the target modal high-resolution image to obtain initial frequency domain data, truncate the high-frequency components of the initial frequency domain data to generate low-resolution frequency domain data; perform spatial domain conversion on the low-resolution frequency domain data to generate a low-resolution image in the spatial domain;

[0009] Input the low-resolution frequency domain data and the low-resolution image in the spatial domain into a diffusion model, and use the diffusion model to extract features of the low-resolution frequency domain data and the low-resolution image in the spatial domain in the frequency domain and the spatial domain respectively to generate features at the current time step; and use a structural attention mechanism to perform feature fusion on the features at the current time step and the features of the high-resolution image of the auxiliary modality to obtain a reconstructed image of the target modality.

[0010] In some embodiments, the step of truncating the high-frequency components of the initial frequency domain data to generate low-resolution frequency domain data includes:

[0011] Obtain a preset mask that occludes high-frequency components from the edge to the center at a preset step size. Based on the initial frequency domain data, truncate the high-frequency components by shrinking the mask of the frequency center region at the preset step size to generate the low-resolution frequency domain data at different time steps.

[0012] In some embodiments, the step of extracting features of the low-resolution frequency domain data and the low-resolution image in the spatial domain in the frequency domain and the spatial domain respectively to generate features at the current time step; and using a structural attention mechanism to perform feature fusion on the features at the current time step and the features of the high-resolution image of the auxiliary modality to obtain a reconstructed image of the target modality includes:

[0013] The features at the current time step include frequency domain time step features and spatial domain time step features;

[0014] In the frequency domain, perform encoder feature extraction on the low-resolution frequency domain data to generate the frequency domain time step features; use the structural attention mechanism to fuse the frequency domain time step features with the frequency domain features of the high-resolution image of the auxiliary modality to generate high-frequency missing distribution features;

[0015] In the spatial domain, perform encoder feature extraction on the low-resolution image in the spatial domain to generate the spatial domain time step features; use the structural attention mechanism and fuse the spatial domain time step features with the spatial domain features of the high-resolution image of the auxiliary modality to generate local detail features;

[0016] Based on the high-frequency missing distribution features and the local detail features, obtain the reconstructed image of the target modality.

[0017] In some of these embodiments, the structural attention mechanism dynamically fuses the features of the current time step with the features of the high-resolution image of the auxiliary modality through attention weights.

[0018] In some of these embodiments, the method further includes:

[0019] Obtain training MRI sample data, including a training high-resolution image of the target modality and a training high-resolution image of the auxiliary modality;

[0020] Perform frequency domain conversion on the training high-resolution image of the target modality to obtain initial training frequency domain data, truncate the high-frequency components of the initial training frequency domain data to generate low-resolution training frequency domain data; perform spatial domain conversion on the initial training frequency domain data to generate an initial training spatial domain image;

[0021] Input the low-resolution training frequency domain data and the initial training spatial domain image into an initial diffusion model, and use the initial diffusion model to perform feature extraction on the low-resolution training frequency domain data and the initial training spatial domain image in the frequency domain and the spatial domain respectively to generate training features of the current time step; and use a structural attention mechanism to perform feature fusion on the training features of the current time step and the features of the training high-resolution image of the auxiliary modality to reconstruct a training reconstruction image of the target modality;

[0022] The training reconstruction image includes frequency domain reconstruction data and spatial domain reconstruction data;

[0023] Construct a dual-domain data consistency loss function and a reconstruction loss function; wherein, the dual-domain data consistency loss function includes a frequency domain consistency loss function and a spatial domain consistency loss function;

[0024] Based on the dual-domain data consistency loss function and the reconstruction loss function and the dual-domain data consistency loss function and the reconstruction loss function, calculate a frequency domain consistency loss value, a spatial domain consistency loss value, and a reconstruction loss value;

[0025] Weightedly fuse the frequency domain consistency loss value, the spatial domain consistency loss value, and the reconstruction loss value to obtain a total loss function value; based on the total loss function value, iteratively optimize the parameters of the initial diffusion model to generate a trained diffusion model.

[0026] In some of these embodiments, the calculating the frequency domain consistency loss value, the spatial domain consistency loss value, and the reconstruction loss value based on the dual-domain data consistency loss function and the reconstruction loss function and the dual-domain data consistency loss function and the reconstruction loss function includes:

[0027] Based on the frequency-domain consistency loss function, calculate the difference between the frequency-domain reconstructed image features and the frequency-domain features of the corresponding high-resolution image of the training target modality, and generate a frequency-domain consistency loss value;

[0028] Based on the spatial-domain consistency loss function, calculate the difference between the spatial-domain reconstructed image features and the spatial-domain features of the corresponding high-resolution image of the training target modality, and generate a spatial-domain consistency loss value;

[0029] Based on the reconstruction loss function, calculate the difference between the training reconstructed image and the features of the corresponding high-resolution image of the training target modality, and generate a reconstruction loss value.

[0030] In some embodiments, the weighted fusion of the frequency-domain consistency loss value, the spatial-domain consistency loss value, and the reconstruction loss value includes:

[0031] When the proportion of low-frequency components in the training low-resolution frequency-domain data is greater than that of high-frequency components, the weight value of the frequency-domain consistency loss value is greater than the weight value of the spatial-domain consistency loss value;

[0032] When the proportion of low-frequency components in the training low-resolution frequency-domain data is not greater than that of high-frequency components, the weight value of the spatial-domain consistency loss value is greater than the weight value of the frequency-domain consistency loss value;

[0033] Based on the weight value of the spatial-domain consistency loss value, the weight value of the frequency-domain consistency loss value, and the weight value of the reconstruction loss value, perform weighted fusion on the frequency-domain consistency loss value, the spatial-domain consistency loss value, and the reconstruction loss value.

[0034] In a second aspect, an embodiment of the present application provides an MRI image reconstruction device, including:

[0035] An image acquisition module, configured to acquire multi-contrast MRI sample data, including a high-resolution image of the target modality and a high-resolution image of the auxiliary modality;

[0036] A data processing module, configured to perform frequency-domain conversion on the high-resolution image of the target modality to obtain initial frequency-domain data, truncate high-frequency components from the initial frequency-domain data to generate low-resolution frequency-domain data; perform spatial-domain conversion on the low-resolution frequency-domain data to generate a low-resolution image in the spatial domain;

[0037] A reconstruction image module, configured to input the low-resolution frequency-domain data and the spatial-domain low-resolution image into a diffusion model, and use the diffusion model to extract features of the low-resolution frequency-domain data and the spatial-domain low-resolution image in the frequency domain and the spatial domain respectively to generate features of the current time step; and use a structural attention mechanism to perform feature fusion on the features of the current time step and the features of the high-resolution image of the auxiliary modality to obtain a reconstructed image of the target modality.

[0038] In a third aspect, an embodiment of the present application provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the MRI image reconstruction method described in the first aspect above is implemented.

[0039] In a fourth aspect, an embodiment of the present application provides a storage medium, on which a computer program is stored. When the program is executed by a processor, the MRI image reconstruction method described in the first aspect above is implemented.

[0040] Compared with the related art, an MRI image reconstruction method, device, electronic device, and storage medium provided by an embodiment of the present application obtain multi-contrast MRI data (including high-resolution images of the target and auxiliary modalities), transfer the target modality image to the frequency domain and truncate the high frequencies to generate low-resolution frequency-domain data, and then transfer it back to the spatial domain to obtain a low-resolution image; input the two types of low-resolution data into a diffusion model, extract features in the frequency domain and the spatial domain respectively, and combine the structural attention mechanism to fuse the features of the auxiliary modality, and finally reconstruct the high-resolution image of the target modality, solving the balance problem between spectrum fidelity and detail enhancement in traditional MRI super-resolution methods, and significantly improving the reconstruction efficiency and image quality.

[0041] Details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more concise and understandable. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments and descriptions thereof of the present application are used to explain the present application and do not constitute an improper limitation of the present application. In the drawings:

[0043] Figure 1 is a hardware structure block diagram of a terminal of the MRI image reconstruction method according to an embodiment of the present invention;

[0044] Figure 2 is a flowchart of the MRI image reconstruction method according to an embodiment of the present application;

[0045] Figure 3 is a preferred flowchart of the MRI image reconstruction method according to a preferred embodiment of the present application;

[0046] Figure 4 It is a framework diagram of a dual-domain collaborative diffusion model for the MRI image reconstruction method according to the preferred embodiment of the present application;

[0047] Figure 5 It is a structural block diagram of the MRI image reconstruction device according to the embodiment of the present application. Detailed implementation manners

[0048] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments provided in the present application without making creative efforts fall within the scope of protection of the present application. In addition, it can also be understood that although the efforts made in this development process may be complex and lengthy, for those of ordinary skill in the art related to the content disclosed in the present application, some design, manufacturing or production changes made on the basis of the technical content disclosed in the present application are only conventional technical means and should not be understood as the content disclosed in the present application being insufficient.

[0049] Referring to "embodiment" in the present application means that the specific features, structures or characteristics described in connection with the embodiment may be included in at least one embodiment of the present application. The phrase appears in various places in the specification and does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those of ordinary skill in the art explicitly and implicitly understand that the embodiments described in the present application may be combined with other embodiments without conflict.

[0050] Unless otherwise defined, the technical terms or scientific terms involved in this application shall have the ordinary meanings understood by those with ordinary skills in the technical field to which this application belongs. The words such as "a", "an", "one kind", "the" and the like involved in this application do not indicate a limitation in quantity and may represent a singular or plural number. The terms "comprise", "include", "have" and any variations thereof involved in this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may further include unlisted steps or units, or may further include other steps or units inherent to these processes, methods, products or devices. The words such as "connect", "be connected", "couple" and the like involved in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The "plurality" involved in this application means greater than or equal to two. "And / or" describes the association relationship of associated objects and indicates that three relationships may exist. For example, "A and / or B" may represent three situations: A exists alone, A and B exist simultaneously, and B exists alone. The terms "first", "second", "third" and the like involved in this application are only used to distinguish similar objects and do not represent a specific order of the objects.

[0051] The method embodiments provided in this embodiment can be executed on a terminal, a computer or a similar computing device. Taking running on a terminal as an example, Figure 1 is the hardware structure block diagram of the terminal of the MRI image reconstruction method according to the embodiments of the present invention. As Figure 1 shown, the terminal may include one or more ( Figure 1 only one is shown in the figure) processors 102 (the processor 102 may include, but is not limited to, a processing device such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data. Optionally, the above terminal may further include a transmission device 106 for communication functions and an input / output device 108. Those of ordinary skill in the art can understand that Figure 1 the structure shown is only schematic and does not limit the structure of the above terminal. For example, the terminal may further include more or fewer components than those shown in Figure 1 the figure, or have a different configuration from that shown in Figure 1 the figure.

[0052] The memory 104 can be used to store computer programs, such as software programs and modules of application software, such as the computer program corresponding to the MRI image reconstruction method in the embodiments of the present invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, the above-mentioned method is implemented. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some instances, the memory 104 may further include a memory remotely disposed relative to the processor 102, and these remote memories can be connected to the terminal through a network. Examples of the above-mentioned network include but are not limited to the Internet, enterprise intranet, local area network, mobile communication network, and combinations thereof.

[0053] The transmission device 106 is used to receive or send data via a network. Specific examples of the above-mentioned network may include a wireless network provided by a communication provider of the terminal. In one instance, the transmission device 106 includes a network adapter (Network Interface Controller, abbreviated as NIC), which can be connected to other network devices through a base station and thus communicate with the Internet. In one instance, the transmission device 106 can be a radio frequency (Radio Frequency, abbreviated as RF) module, which is used to communicate with the Internet wirelessly.

[0054] This embodiment provides an MRI image reconstruction method. Figure 2 It is a flowchart of the MRI image reconstruction method according to the embodiments of the present application, as Figure 2 shown, and this process includes the following steps:

[0055] Step S201, obtain multi-contrast MRI sample data, including a high-resolution image of a target modality and a high-resolution image of an auxiliary modality;

[0056] Among them, multi-contrast MRI sample data is collected by a medical imaging device (such as an MRI scanner). The multi-contrast MRI sample data includes high-resolution images of the target modality (such as T2) and high-resolution images of the auxiliary modality (such as T1). The high-resolution image of the target modality is a three-dimensional (3D) T2 high-resolution image (HR), which shows tissue water content and highlights pathological features (such as tumors, edema); the high-resolution image of the auxiliary modality is a three-dimensional (3D) T1 high-resolution image (HR), which clearly shows anatomical structures (such as the boundaries between gray matter and white matter in the brain). The collected three-dimensional MRI sample data is sliced and converted into two-dimensional data along the depth axis to adapt to the model input format. Preprocessing operations such as brightness normalization or Z-score standardization are performed on the images to eliminate device differences and noise interference. In this step, by obtaining multi-contrast MRI data, complementary information in different modalities can be fused. MRI images in different modalities provide rich feature information. For example, T1 and T2 modalities may reveal different tissue characteristics. Combining this information can improve the details and accuracy of the reconstructed image. High-quality multi-modal MRI reconstructed images can provide more accurate information for clinical diagnosis, helping doctors make more accurate diagnosis and treatment decisions.

[0057] Step S202: For the high-resolution image of the target modality, perform frequency domain conversion to obtain initial frequency domain data, truncate high-frequency components of the initial frequency domain data to generate low-resolution frequency domain data; perform spatial domain conversion on the low-resolution frequency domain data to generate a low-resolution image in the spatial domain;

[0058] Among them, the target modality high-resolution image (such as T2 HR image) is subjected to frequency-domain conversion (i.e., Fourier transform), which converts it from the spatial domain to the frequency domain to obtain initial frequency-domain data (k-space data). In the frequency domain, a mask that gradually blocks high-frequency components from the edge to the center can be defined to simulate the k-space truncation degradation process. This mask gradually shrinks as the time step increases, thereby gradually truncating high-frequency components to generate low-resolution frequency-domain data. This step simulates the information loss during the low-resolution acquisition process of MRI images. The generated low-resolution frequency-domain data is subjected to spatial-domain conversion (i.e., inverse Fourier transform), which converts it back from the frequency domain to the spatial domain to generate a low-resolution spatial-domain image. This step truncates high-frequency components and uses low-frequency information as the starting point for diffusion, avoiding the interference of invalid noise and being more in line with the essence of MRI image degradation because the degradation of MRI images is mainly due to the loss of low-frequency information rather than global noise pollution; by performing degradation processing in the frequency domain, spectral fidelity can be maintained, which means that during the degradation process, the low-frequency components of the image are better preserved, contributing to the recovery of frequency-domain information in the subsequent reconstruction process; since frequency-domain processing can maintain spectral information while spatial-domain processing focuses more on the recovery of image details, compared with methods that only rely on the frequency domain or the spatial domain, through the conversion processing of the frequency domain and the spatial domain, the loss of spatial texture details can be reduced to a certain extent. The generated low-resolution frequency-domain data and low-resolution spatial-domain image facilitate the subsequent dual-domain collaborative recovery process. These data can be used as inputs to the diffusion model for subsequent feature extraction and reconstruction processes.

[0059] Step S203: Input the low-resolution frequency-domain data and the low-resolution spatial-domain image into the diffusion model. Using the diffusion model, perform feature extraction on the low-resolution frequency-domain data and the low-resolution spatial-domain image in the frequency domain and the spatial domain respectively to generate features at the current time step; and use the structural attention mechanism to perform feature fusion on the features at the current time step and the features of the auxiliary modality high-resolution image to obtain the reconstructed image of the target modality.

[0060] Among them, the low-resolution frequency-domain data and the spatial-domain low-resolution image generated in the above steps are used as inputs and input into the diffusion model. In the diffusion model, feature extraction is respectively performed on the low-resolution frequency-domain data and the spatial-domain low-resolution image. The frequency-domain feature extraction focuses on capturing the spectral characteristics of the image (such as phase distribution, low-frequency contour), while the spatial-domain feature extraction focuses on capturing the spatial structure and details of the image (such as texture features like edges, lesion regions, etc.). Through a specific encoder network structure, feature representations at the current time step are generated respectively in the frequency domain and the spatial domain. Using the structured attention mechanism (SP-former), the features extracted at the current time step in the frequency domain and the spatial domain are fused with the features of the auxiliary modality high-resolution image (such as T1 HR image). The structured attention mechanism can adaptively learn the correlations between different features and perform weighted fusion on the features according to these correlations, thereby generating a richer and more accurate feature representation. Based on the fused feature representation, a reconstructed image of the target modality (such as T2) is generated through a decoder network structure. This reconstructed image has significant improvements in both detail fidelity and reconstruction efficiency and can provide high-quality MRI image support for clinical applications.

[0061] Through the above steps, this application uses the low-resolution image as the starting point of diffusion, and through dynamic collaborative optimization of k-space spectrum fidelity and spatial details, not only preserves the physical fidelity of the anatomical structure but also effectively restores high-frequency details (such as blood vessel textures). At the same time, compared with the defect of anatomical position deviation caused by the lack of frequency-domain constraints in the spatial-domain model of traditional methods, this application uses multi-modal feature fusion (such as combining T1 and T2 modalities) and the structured attention mechanism to enhance visual details in the spatial domain while ensuring the accuracy of the anatomical structure through frequency-domain consistency loss. In addition, aiming at the problems that the traditional diffusion model starts from random noise for reverse denoising, which does not match the essence of MRI degradation, is inefficient, and is prone to introducing artifacts, this application directly uses low-frequency information as the starting point of diffusion, reduces the interference of invalid noise, and avoids a large number of iterative steps for restoring the contour, thereby significantly improving the reconstruction efficiency and image quality, solving the balance problem of spectrum fidelity and detail enhancement in traditional MRI super-resolution methods, and achieving a perfect balance between spectrum fidelity and detail enhancement.

[0062] In some of these embodiments, truncating the high-frequency components of the initial frequency-domain data to generate low-resolution frequency-domain data includes:

[0063] Obtain a preset mask that occludes high-frequency components from the edge to the center at a preset step size. Based on the initial frequency-domain data, by shrinking the mask of the frequency center region at the preset step size, truncate the high-frequency components to generate low-resolution frequency-domain data at different time steps.

[0064] Among them, a mask is defined to block high-frequency components from the edge to the center according to a preset step size, which is used to simulate the process of gradual loss of high-frequency information in k-space (frequency domain). Based on the initial frequency domain data, the high-frequency components are truncated by gradually shrinking the mask in the frequency center region. During this process, the pixel center coordinates of the mask remain unchanged, while the parameter controlling the side length of the k-space center region to be retained (i.e., the low-frequency component) gradually decreases with the increase of the number of steps. In each step, the current mask is used to process the initial frequency domain data to generate the low-resolution frequency domain data of this time step. By gradually shrinking the mask, a series of low-resolution frequency domain data of different time steps can be generated, and these data reflect the process of gradual loss of high-frequency information.

[0065] In this embodiment, by gradually shrinking the mask in the frequency center region to truncate high-frequency components, the degradation process of MRI images during low-resolution acquisition can be accurately simulated, thus being more in line with the essence of MRI image degradation. Because the degradation of MRI images is mainly due to the loss of low-frequency information (more significant than the absence of high-frequency information), rather than global noise pollution; by generating a series of low-resolution frequency domain data of different time steps, diverse samples are provided for model training. These data cover different degradation degrees from high resolution to low resolution, which helps the model learn more robust feature representations and improve the quality of the reconstructed images.

[0066] In some of these embodiments, feature extraction is respectively performed on the low-resolution frequency domain data and the low-resolution image in the spatial domain in the frequency domain and the spatial domain to generate the features of the current time step; and using the structural attention mechanism, the features of the current time step are fused with the features of the high-resolution image of the auxiliary modality to obtain the reconstructed image of the target modality, including:

[0067] The features of the current time step include the frequency domain time step features and the spatial domain time step features;

[0068] In the frequency domain, encoder feature extraction is performed on the low-resolution frequency domain data to generate the frequency domain time step features; using the structural attention mechanism, the frequency domain time step features are fused with the frequency domain features of the high-resolution image of the auxiliary modality to generate the high-frequency missing distribution features;

[0069] In the spatial domain, encoder feature extraction is performed on the low-resolution image in the spatial domain to generate the spatial domain time step features; using the structural attention mechanism, and the spatial domain time step features are fused with the spatial domain features of the high-resolution image of the auxiliary modality to generate the local detail features;

[0070] Based on the high-frequency missing distribution features and the local detail features, the reconstructed image of the target modality is obtained.

[0071] Among them, in the frequency domain, encoder features are extracted from the low-resolution frequency-domain data to generate frequency-domain time-step features, so as to capture the low-frequency information and high-frequency missing patterns in the frequency domain. Using the structural attention mechanism, the frequency-domain time-step features are fused with the frequency-domain features of the auxiliary-modal high-resolution image to generate high-frequency missing distribution features. The structural attention mechanism can adaptively learn the correlations between different features and perform weighted fusion on the features according to these correlations, thereby highlighting the regions with high-frequency missing. In the spatial domain, encoder features are extracted from the low-resolution spatial-domain image to generate spatial-domain time-step features, so as to capture the spatial structure and local details of the image. Similarly using the structural attention mechanism, the spatial-domain time-step features are fused with the spatial-domain features of the auxiliary-modal high-resolution image to generate local detail features. Through the structural attention mechanism, the ability to capture and restore local details of the image can be enhanced. Based on the high-frequency missing distribution features and local detail features, a reconstructed image of the target modality is generated through the decoder network structure. These features provide rich information for the decoder, which helps to maintain both spectral fidelity and spatial details during the reconstruction process.

[0072] In this embodiment, by performing feature extraction and fusion in the frequency domain and the spatial domain respectively, the information complementarity of different domains is fully utilized. The frequency-domain features provide the distribution information of high-frequency missing, while the spatial-domain features provide rich local details. The combination of the two can significantly improve the quality of the reconstructed image. The application of the structural attention mechanism in the frequency-domain feature fusion helps to more accurately identify and restore the regions with high-frequency missing, significantly enhancing the clarity and accuracy of high-frequency details (such as blood vessel textures) in the reconstructed image. In the spatial-domain feature fusion, the structural attention mechanism enhances the ability to capture and restore local details, enabling the reconstructed image to better present local details while maintaining the overall structure, improving the visual quality of the image. By combining the high-frequency missing distribution features and local detail features, the decoder can generate a more accurate and clearer reconstructed image. This method achieves a good balance between spectral fidelity and spatial detail enhancement, providing more reliable MRI image support for clinical applications.

[0073] In some of these embodiments, the structural attention mechanism dynamically fuses the features of the current time step with the features of the auxiliary-modal high-resolution image through attention weights.

[0074] Among them, the current time-step features include frequency-domain time-step features and spatial-domain time-step features, which are respectively extracted from low-resolution frequency-domain data and low-resolution spatial-domain images through encoders in the frequency domain and spatial domain; frequency-domain and spatial-domain features are extracted from high-resolution images of the auxiliary modality through corresponding encoders. The structural attention mechanism calculates the similarity or correlation between the current time-step features and the features of the high-resolution image of the auxiliary modality to obtain attention weights. The attention weights reflect the degree of association between different features and determine the contribution of each feature in the fusion process. The specific calculation process may involve operations such as dot product, weighted summation, and normalization of feature vectors to generate an attention weight matrix. Using the calculated attention weights, the current time-step features and the features of the high-resolution image of the auxiliary modality are weighted and fused. In this process, features with higher attention weights occupy a larger proportion in the fusion result, thus realizing the dynamic selection and fusion of features. The fused features contain both the reconstruction information of the current time-step and the beneficial features of the high-resolution image of the auxiliary modality, which helps to improve the quality of the reconstructed image.

[0075] In this embodiment, the structural attention mechanism can adaptively adjust the contribution of each feature according to the correlation between the current time-step features and the features of the high-resolution image of the auxiliary modality by dynamically adjusting the weight distribution in the feature fusion process. This adaptive feature fusion method helps to better utilize the information of the auxiliary modality, improve the accuracy of the reconstructed image, and also makes the model have a certain robustness to the noise and changes of the input data. Even under the input of MRI images with different modalities or different qualities, the model can generate relatively stable reconstruction results.

[0076] In some embodiments, the method further includes:

[0077] Obtain training MRI sample data, including training high-resolution images of the target modality and training high-resolution images of the auxiliary modality;

[0078] For the training high-resolution images of the target modality, perform frequency-domain conversion to obtain training initial frequency-domain data, truncate the high-frequency components of the training initial frequency-domain data to generate training low-resolution frequency-domain data; perform spatial-domain conversion on the training initial frequency-domain data to generate training initial spatial-domain images;

[0079] Input the training low-resolution frequency-domain data and the training initial spatial-domain images into the initial diffusion model. Using the initial diffusion model, extract features from the training low-resolution frequency-domain data and the training initial spatial-domain images in the frequency domain and spatial domain respectively to generate training current time-step features; and use the structural attention mechanism to perform feature fusion on the training current time-step features and the features of the training high-resolution images of the auxiliary modality to reconstruct the training reconstructed images of the target modality;

[0080] The training reconstructed image includes frequency-domain reconstructed data and spatial-domain reconstructed data;

[0081] Construct a dual-domain data consistency loss function and a reconstruction loss function; among them, the dual-domain data consistency loss function includes a frequency-domain consistency loss function and a spatial-domain consistency loss function;

[0082] Based on the dual-domain data consistency loss function and the reconstruction loss function, calculate the frequency-domain consistency loss value, the spatial-domain consistency loss value, and the reconstruction loss value;

[0083] Fuse the frequency-domain consistency loss value, the spatial-domain consistency loss value, and the reconstruction loss value by weighting to obtain the total loss function value; based on the total loss function value, iteratively optimize the parameters of the initial diffusion model to generate a trained diffusion model.

[0084] Among them, obtaining the training MRI sample data, including the training target modality high-resolution image and the training auxiliary modality high-resolution image, provides the necessary input and reference standard for model training. Perform a frequency-domain transformation on the training target modality high-resolution image to obtain the initial training frequency-domain data. Truncate the high-frequency components of the initial training frequency-domain data to generate the training low-resolution frequency-domain data, which simulates the frequency-domain degradation during the low-resolution acquisition process of MRI images. Perform a spatial-domain transformation on the initial training frequency-domain data to generate the initial training spatial-domain image, providing a basis for subsequent spatial-domain feature extraction. Input the training low-resolution frequency-domain data and the initial training spatial-domain image into the initial diffusion model. In the diffusion model, perform frequency-domain and spatial-domain feature extraction on the training low-resolution frequency-domain data and the initial training spatial-domain image respectively to generate the training current time-step features. Use the structured attention mechanism to fuse the features of the training current time-step features and the training auxiliary modality high-resolution image to reconstruct the training reconstructed image of the target modality.

[0085] The training reconstructed image includes frequency-domain reconstructed data and spatial-domain reconstructed data, corresponding to the reconstruction results in the frequency domain and the spatial domain respectively. Construct a dual-domain data consistency loss function and a reconstruction loss function. Among them, the dual-domain data consistency loss function includes a frequency-domain consistency loss function and a spatial-domain consistency loss function, which are respectively used to measure the consistency between the frequency-domain and spatial-domain reconstructed data and the original high-resolution data. The reconstruction loss function is used to measure the difference between the reconstructed image and the true high-resolution image. Based on the dual-domain data consistency loss function and the reconstruction loss function, calculate the frequency-domain consistency loss value, the spatial-domain consistency loss value, and the reconstruction loss value. Fuse these loss values by weighting to obtain the total loss function value. By adjusting the weights, the influence of different loss terms on model training can be balanced. Based on the total loss function value, use an optimization algorithm (such as gradient descent) to iteratively optimize the parameters of the initial diffusion model to generate a trained diffusion model.

[0086] In this embodiment, by combining and using the dual-domain data consistency loss function and the reconstruction loss function, the model can simultaneously focus on the consistency in the frequency domain and the spatial domain, as well as the difference between the reconstructed image and the real image during the training process, which helps to improve the quality of the reconstructed image and make it closer to the real high-resolution image; by weighted fusion of different loss terms, the model can better balance the performance requirements in different aspects, thereby enhancing the robustness to the noise and changes of the input data. Even when facing MRI image inputs of different qualities, the model can generate relatively stable reconstruction results; the process of iteratively optimizing the model parameters enables the model to gradually learn more accurate feature representations and reconstruction strategies. By continuously adjusting the parameters to minimize the total loss function value, the model can gradually approach the optimal solution, improving the accuracy and clarity of the reconstructed image. The trained diffusion model can provide high-quality MRI image reconstruction results for clinical use, which helps doctors make more accurate diagnosis and treatment decisions. By improving the quality and robustness of the reconstructed image, this model has higher practical value and reliability in clinical applications.

[0087] In some of these embodiments, based on the dual-domain data consistency loss function and the reconstruction loss function, the frequency-domain consistency loss value, the spatial-domain consistency loss value, and the reconstruction loss value are calculated, including:

[0088] Based on the frequency-domain consistency loss function, calculate the difference between the frequency-domain reconstructed image features and the frequency-domain features of the corresponding training target modality high-resolution image to generate the frequency-domain consistency loss value;

[0089] Based on the spatial-domain consistency loss function, calculate the difference between the spatial-domain reconstructed image features and the spatial-domain features of the corresponding training target modality high-resolution image to generate the spatial-domain consistency loss value;

[0090] Based on the reconstruction loss function, calculate the difference between the training reconstructed image and the features of the corresponding training target modality high-resolution image to generate the reconstruction loss value.

[0091] Among them, the difference between the frequency-domain reconstructed image features and the frequency-domain features of the corresponding training target modality high-resolution image is calculated, and a frequency-domain consistency loss function (such as mean square error loss function) is used to measure this difference, generating a frequency-domain consistency loss value, ensuring that the reconstructed frequency-domain image is consistent with the real high-resolution image in spectral characteristics. The difference between the spatial-domain reconstructed image features and the spatial-domain features of the corresponding training target modality high-resolution image is calculated, and a spatial-domain consistency loss function (such as structural similarity loss function) is used to measure this difference, generating a spatial-domain consistency loss value, ensuring that the reconstructed spatial-domain image is consistent with the real high-resolution image in spatial structure and details. The difference between the training reconstructed image (including frequency-domain and spatial-domain reconstruction data) and the overall features of the corresponding training target modality high-resolution image is calculated, and a reconstruction loss function (such as L1 loss function or L2 loss function) is used to measure this difference, generating a reconstruction loss value, ensuring that the reconstructed image is consistent with the real high-resolution image in overall visual quality.

[0092] The calculation of the frequency-domain consistency loss value in this embodiment ensures the accuracy and consistency of the reconstructed image in the frequency domain, helps to retain and restore the high-frequency details of the image, such as blood vessel textures, etc., thus improving the quality of frequency-domain reconstruction; the calculation of the spatial-domain consistency loss value ensures the structural similarity and detail expressiveness of the reconstructed image in the spatial domain, helps to restore the local details and overall structure of the image, making the reconstructed image clearer and more accurate visually; the calculation of the reconstruction loss value measures the difference between the reconstructed image and the real image as a whole, ensures the quality of the reconstructed image globally and locally, helps to balance the reconstruction effects in the frequency domain and the spatial domain, generates a more natural and real-like high-resolution image; by calculating the loss values in the frequency domain and the spatial domain respectively, the model can better utilize the information of the auxiliary modality to guide the reconstruction process. This multi-modal fusion method helps to improve the model's processing ability for complex MRI data, generates more reliable reconstruction results, and the clear calculation of the loss value provides a clear optimization direction for model training. By continuously adjusting the model parameters to minimize these loss values, the model can converge to the optimal solution faster and improve the training efficiency.

[0093] In some of these embodiments, the frequency-domain consistency loss value, the spatial-domain consistency loss value, and the reconstruction loss value are weighted and fused, including:

[0094] When the proportion of low-frequency components in the training low-resolution frequency-domain data is greater than that of high-frequency components, the weight value of the frequency-domain consistency loss value is greater than the weight value of the spatial-domain consistency loss value;

[0095] When the proportion of low-frequency components in the training low-resolution frequency-domain data is not greater than that of high-frequency components, the weight value of the spatial-domain consistency loss value is greater than the weight value of the frequency-domain consistency loss value;

[0096] Based on the weight value of the spatial domain consistency loss value, the weight value of the frequency domain consistency loss value, and the weight value of the reconstruction loss value, the frequency domain consistency loss value, the spatial domain consistency loss value, and the reconstruction loss value are weighted and fused.

[0097] Among them, the training low-resolution frequency domain data is analyzed to determine the proportion of the low-frequency components and the high-frequency components therein. This analysis can be achieved through frequency domain energy distribution calculation or frequency domain feature statistics. When the proportion of the low-frequency components in the training low-resolution frequency domain data is greater than that of the high-frequency components, the weight value of the frequency domain consistency loss value is increased so that it occupies a larger proportion in the total loss function, which helps the model to pay more attention to the accuracy of frequency domain reconstruction when the low-frequency information dominates. On the contrary, when the proportion of the low-frequency components is not greater than that of the high-frequency components, the weight value of the spatial domain consistency loss value is increased so that it occupies a larger proportion in the total loss function, which helps the model to pay more attention to the restoration of spatial domain details and structures when the high-frequency information is relatively rich. Based on the adjusted weight of the spatial domain consistency loss value, the weight of the frequency domain consistency loss value, and the preset weight of the reconstruction loss value, the frequency domain consistency loss value, the spatial domain consistency loss value, and the reconstruction loss value are weighted and fused, and the total loss function value is generated by weighted summation to guide the optimization of the model parameters. The phased fusion of the attention mechanism is consistent with the spectrum energy transfer logic. In the early iteration, the structural attention mechanism (SP-former) preferentially fuses the global low-frequency features of the auxiliary modality (such as T1); in the later stage, the local high-frequency features of the target modality (such as T2) are enhanced.

[0098] In this embodiment, by dynamically adjusting the loss weights, the model can automatically balance the reconstruction priorities of the frequency domain and the spatial domain according to the characteristics of the input data. When the low-frequency information dominates, the accuracy of frequency domain reconstruction is enhanced; when the high-frequency information is rich, the ability to restore spatial domain details is enhanced. This dynamic adjustment strategy is particularly applicable to the multi-modal MRI reconstruction task, which can make full use of the complementary information between different modalities and improve the overall reconstruction performance. By combining the advantages of the frequency domain and the spatial domain, the model can generate more comprehensive and reliable reconstruction results.

[0099] The embodiments of the present application will be described and illustrated below through preferred embodiments.

[0100] The preferred embodiments of the MRI image reconstruction method of the present application include the following steps:

[0101] Step S301, data acquisition and preprocessing of multi-contrast MRI samples;

[0102] Step S302, constructing a low-frequency truncation degradation process;

[0103] Step S303, constructing a dual-domain collaborative recovery process;

[0104] Step S304, construct the dual-domain data consistency loss and the reconstruction loss;

[0105] Step S305, train and infer the MRI super-resolution framework of the multi-contrast dual-domain collaborative diffusion model guided by the low-frequency prior.

[0106] For step S301, the target modality (T2) and the auxiliary modality (T1) in the multi-contrast MRI data are sliced into 2D data along the depth axis, obtaining , , and perform normalization preprocessing;

[0107] Among them, is a three-dimensional (3D) T2-weighted high-resolution (HR) magnetic resonance image; 3D represents three-dimensional data; T2 represents the target modality (T2-weighted imaging); H, W, D represent the height, width, and depth of the image. is a three-dimensional (3D) T1-weighted high-resolution (HR) magnetic resonance image; T1 represents the auxiliary modality (T1-weighted imaging). represents the two-dimensional (2D) T2-weighted image after slicing along the depth axis; represents the two-dimensional (2D) T1-weighted image after slicing along the depth axis, and d represents the feature dimension (scaling factor in the attention mechanism).

[0108] For step S302, the traditional diffusion model uses Gaussian noise as the degradation operator. Since it requires the low-frequency as the prior, Figure 3 is the preferred flowchart of the MRI image reconstruction method according to the preferred embodiment of the present application, as Figure 3 shown: Perform Fourier transform on the high-resolution image of the target modality (T2 HR), i.e., , to obtain the k-space data (i.e., the initial frequency-domain data). Among them, is the high-resolution image of the target modality, which is a high-resolution image belonging to the spatial domain (image domain); s represents the spatial domain (spatial domain); 0 represents the initial time step, and n represents the size of the image (e.g., n×n represents a square image). represents converting the spatial-domain image to the frequency-domain (k-space) data through the Fourier transform F, and k represents the frequency domain (k-space).

[0109] By defining a mask that gradually occludes the high-frequency components from the edge to the center, simulate k-space truncation degradation to generate the k-space data at time step t (the degraded k-space low-frequency data, i.e., the low-resolution frequency-domain data).

[0110] The degradation process of k-space is as follows: ;

[0111] In the above formula, represents the low-resolution frequency-domain data, represents element-wise multiplication, is the degradation mask at the t-th step, which is used to truncate the high-frequency components of k-space. Its formula is as follows:

[0112] ;

[0113] In the above formula, β t represents the mask parameter (dynamically adjusted with the number of steps); (u, v) represents the pixel coordinates of the degradation mask in k-space, and c represents the pixel center coordinates of the degradation mask. controls the side length of the central region of k-space (low-frequency components) retained at the t-th step, which gradually decreases as the number of steps t increases, and is used to simulate the gradual loss of high-frequency information.

[0114] Then, perform the inverse Fourier transform on back to the image domain , that is, the low-resolution image in the spatial domain. F represents the Fourier transform, represents the inverse transform. This process directly uses the low-frequency information as the diffusion starting point to avoid the degradation of invalid noise. Specifically, for the degradation process in the image domain, it is as follows:

[0115] ;

[0116] In the above formula, represents element-wise multiplication, is the degradation mask at the t-th step, represents the inverse Fourier transform, represents the low-resolution image in the spatial domain.

[0117] For step S303, Figure 4 is the framework diagram of the dual-domain collaborative diffusion model of the MRI image reconstruction method according to the preferred embodiment of the present application, as shown in Figure 4 : Construct a dual-domain collaborative recovery process, and introduce the auxiliary modality high-resolution image (T1 HR) during the recovery process. For the k-space part, and After extracting features through the encoder, use the structural prior attention mechanism (SP-former) for feature fusion, and then restore it to through the decoder; for the spatial domain there is no need to perform the Fourier transform, After extracting features through the encoder, use SP-former for feature fusion, and then use the encoder to restore to 。

[0118] The specific process of dual-domain collaborative restoration is as follows;

[0119] Construct a restoration operator to reverse the degradation process D, and this operator can be implemented by a neural network with parameters as follows:

[0120] ;

[0121] In the above formula, the restoration operator (neural network) with parameter θ is used to reverse the degradation process; x t represents the input data at the t-th step of the current time step. represents the reconstruction loss, which constrains the difference between the restoration result and the real HR image.

[0122] Construct a structural attention mechanism (SP-former) to fuse the auxiliary modality features with the features of the current time step, see Figure 4 as follows:

[0123] ;

[0124] Among them, , , ;

[0125] In the above formula, , is a learnable attention weight (query, key, value), represents the downsampled features of the current time step, represents the downsampled features of the auxiliary modality (T1). In Figure 4 , MLP (Multi-Layer Perceptron) is a feedforward neural network structure in deep learning; Scale is the scale (resolution); the Sigmoid function is a commonly used activation function in neural networks, with an S-shaped curve, and its core role is to map any real number to the interval (0, 1), and the output can be interpreted as a probability value.

[0126] For step S304, construct the dual-domain data consistency loss and the reconstruction loss, introduce the spectral consistency loss to ensure cross-domain physical fidelity, jointly optimize the visual fidelity with the spatial domain consistency loss, and fuse them with the reconstruction loss of the diffusion model with weights; specifically as follows:

[0127] Based on the restoration operator The output and physical imaging constraints define the dual-domain data consistency loss:

[0128] For the frequency domain:

[0129] ;

[0130] For the spatial domain:

[0131] ;

[0132] where , , which constrains the dual-domain to maintain frequency-domain consistency and visual consistency during the iterative recovery process, reducing the cumulative error.

[0133] In the above formula, represents the frequency-domain consistency loss, which constrains the reconstructed spectrum to align with the true spectrum; represents the spatial-domain consistency loss, which constrains the reconstructed image to align with the true image. , represents the loss weight that adjusts the frequency-domain fidelity and the spatial-domain visual quality.

[0134] Construct the reconstruction loss:

[0135] For the frequency domain:

[0136] ;

[0137] For the spatial domain:

[0138] ;

[0139] The total loss function is defined as follows:

[0140] For the frequency domain:

[0141] ;

[0142] For the spatial domain:

[0143] ;

[0144] In the above formula, represents the total frequency-domain loss function, which combines the reconstruction loss and the consistency loss; represents the total spatial-domain loss function, which combines the reconstruction loss and the consistency loss. , is the weight adjustment coefficient, which is used to adjust the weights of the reconstruction loss and the dual-domain data consistency loss.

[0145] For step S305, train and infer the MRI super-resolution framework of the multi-contrast dual-domain collaborative diffusion model guided by low-frequency priors: For the training process, after degrading the standard T2 HR through step S302, restore it through step S303, that is, generate low-resolution data through the degradation process of step S302 for T2HR, and then reconstruct the image through the dual-domain collaborative restoration process (combining T1HR) of step S303, and optimize the model parameters using the total loss function defined in step S304; for the inference process, directly use T2 LR to restore T2 HR through step S303, that is, directly input T2LR, and generate a high-resolution T2 image through the dual-domain collaborative restoration process (combining T1HR) of step S303, without the degradation step.

[0146] This embodiment also provides an MRI image reconstruction device, which is used to implement the above embodiments and preferred implementation manners, and those that have been described will not be repeated. As used hereinafter, terms such as "module", "unit", "sub-unit", etc. can be a combination of software and / or hardware that can achieve a predetermined function. Although the devices described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware is also possible and contemplated.

[0147] Figure 5 is a structural block diagram of an MRI image reconstruction device according to an embodiment of the present application, as Figure 5 shown, the device includes:

[0148] An image acquisition module 10, configured to acquire multi-contrast MRI sample data, including a high-resolution image of a target modality and a high-resolution image of an auxiliary modality;

[0149] A data processing module 20, configured to perform frequency-domain conversion on the high-resolution image of the target modality to obtain initial frequency-domain data, truncate high-frequency components of the initial frequency-domain data to generate low-resolution frequency-domain data; perform spatial-domain conversion on the low-resolution frequency-domain data to generate a low-resolution image in the spatial domain;

[0150] A reconstructed image module 30, configured to input the low-resolution frequency-domain data and the low-resolution image in the spatial domain into a diffusion model, and use the diffusion model to perform feature extraction on the low-resolution frequency-domain data and the low-resolution image in the spatial domain in the frequency domain and the spatial domain respectively to generate features at the current time step; and use a structural attention mechanism to perform feature fusion on the features at the current time step and the features of the high-resolution image of the auxiliary modality to obtain a reconstructed image of the target modality.

[0151] The above data processing module 20 is further configured to obtain a preset mask that occludes high-frequency components from the edge to the center at a preset step length, and based on the initial frequency-domain data, truncate the high-frequency components by gradually shrinking the mask of the frequency center region to generate low-resolution frequency-domain data at different time steps.

[0152] The above-mentioned reconstruction image module 30 is also used for the current time step features including frequency domain time step features and spatial domain time step features;

[0153] In the frequency domain, encoder feature extraction is performed on the low-resolution frequency domain data to generate frequency domain time step features; using the structural attention mechanism, the frequency domain time step features are fused with the frequency domain features of the auxiliary modality high-resolution image to generate high-frequency missing distribution features;

[0154] In the spatial domain, encoder feature extraction is performed on the spatial domain low-resolution image to generate spatial domain time step features; using the structural attention mechanism, and the spatial domain time step features are fused with the spatial domain features of the auxiliary modality high-resolution image to generate local detail features;

[0155] Based on the high-frequency missing distribution features and local detail features, a reconstructed image of the target modality is obtained.

[0156] The above-mentioned reconstruction image module 30 is also used for the structural attention mechanism to dynamically fuse the current time step features with the features of the auxiliary modality high-resolution image through attention weights.

[0157] The above-mentioned MRI image reconstruction device further includes a training module; the training module is used to obtain training MRI sample data, including training target modality high-resolution images and training auxiliary modality high-resolution images;

[0158] For the training target modality high-resolution image, frequency domain conversion is performed to obtain training initial frequency domain data, the high-frequency components of the training initial frequency domain data are truncated to generate training low-resolution frequency domain data; spatial domain conversion is performed on the training initial frequency domain data to generate a training spatial domain initial image;

[0159] The training low-resolution frequency domain data and the training spatial domain initial image are input into the initial diffusion model, and the initial diffusion model is used to perform feature extraction on the training low-resolution frequency domain data and the training spatial domain initial image in the frequency domain and the spatial domain respectively to generate training current time step features; and using the structural attention mechanism, feature fusion is performed on the training current time step features and the features of the training auxiliary modality high-resolution image to reconstruct the training reconstructed image of the target modality;

[0160] The training reconstructed image includes frequency domain reconstruction data and spatial domain reconstruction data;

[0161] A dual-domain data consistency loss function and a reconstruction loss function are constructed; among them, the dual-domain data consistency loss function includes a frequency domain consistency loss function and a spatial domain consistency loss function;

[0162] Calculate the frequency-domain consistency loss value, spatial-domain consistency loss value, and reconstruction loss value based on the dual-domain data consistency loss function and the reconstruction loss function, as well as the dual-domain data consistency loss function and the reconstruction loss function;

[0163] Weightedly fuse the frequency-domain consistency loss value, spatial-domain consistency loss value, and reconstruction loss value to obtain the total loss function value; based on the total loss function value, iteratively optimize the parameters of the initial diffusion model to generate a trained diffusion model.

[0164] The above training module is also used to calculate the difference between the frequency-domain reconstructed image features and the frequency-domain features of the corresponding training target modality high-resolution image based on the frequency-domain consistency loss function, and generate the frequency-domain consistency loss value;

[0165] Calculate the difference between the spatial-domain reconstructed image features and the spatial-domain features of the corresponding training target modality high-resolution image based on the spatial-domain consistency loss function, and generate the spatial-domain consistency loss value;

[0166] Calculate the difference between the training reconstructed image and the features of the corresponding training target modality high-resolution image based on the reconstruction loss function, and generate the reconstruction loss value.

[0167] The above training module is also used to make the weight value of the frequency-domain consistency loss value greater than the weight value of the spatial-domain consistency loss value when the proportion of low-frequency components in the training low-resolution frequency-domain data is greater than that of high-frequency components;

[0168] When the proportion of low-frequency components in the training low-resolution frequency-domain data is not greater than that of high-frequency components, the weight value of the spatial-domain consistency loss value is greater than the weight value of the frequency-domain consistency loss value;

[0169] Based on the weight value of the spatial-domain consistency loss value, the weight value of the frequency-domain consistency loss value, and the weight value of the reconstruction loss value, weightedly fuse the frequency-domain consistency loss value, spatial-domain consistency loss value, and reconstruction loss value.

[0170] This embodiment also provides an electronic device, including a memory and a processor. A computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0171] Optionally, the above electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the above processor, and the input / output device is connected to the above processor.

[0172] Optionally, in this embodiment, the above processor may be configured to execute the following steps through a computer program:

[0173] S1, Obtain multi-contrast MRI sample data, including a target modality high-resolution image and an auxiliary modality high-resolution image;

[0174] S2. For the high-resolution target modality image, perform frequency domain conversion to obtain initial frequency domain data, truncate the high-frequency components of the initial frequency domain data to generate low-resolution frequency domain data; perform spatial domain conversion on the low-resolution frequency domain data to generate a low-resolution spatial domain image.

[0175] S3. Input the low-resolution frequency domain data and the low-resolution spatial domain image into a diffusion model. Using the diffusion model, perform feature extraction on the low-resolution frequency domain data and the low-resolution spatial domain image in the frequency domain and the spatial domain respectively to generate features at the current time step; and use a structural attention mechanism to perform feature fusion on the features at the current time step and the features of the high-resolution auxiliary modality image to obtain a reconstructed image of the target modality.

[0176] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation manners, and will not be elaborated here.

[0177] In addition, in combination with the MRI image reconstruction method in the above embodiments, an embodiment of the present application can be implemented by providing a storage medium. A computer program is stored on the storage medium; when the computer program is executed by a processor, any one of the MRI image reconstruction methods in the above embodiments is implemented.

[0178] Those skilled in the art should understand that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0179] The above embodiments only represent several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the scope of the invention patent. It should be pointed out that for those of ordinary skill in the art, without departing from the concept of the present application, several deformations and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. An MRI image reconstruction method, characterized in that, Including: Obtain multi-contrast MRI sample data, including a target modality high-resolution image and an auxiliary modality high-resolution image; Perform frequency domain conversion on the target modality high-resolution image to obtain initial frequency domain data, truncate high-frequency components of the initial frequency domain data, and generate low-resolution frequency domain data; Perform spatial domain conversion on the low-resolution frequency domain data to generate a spatial domain low-resolution image; Input the low-resolution frequency domain data and the spatial domain low-resolution image into a diffusion model, and use the diffusion model to extract features of the low-resolution frequency domain data and the spatial domain low-resolution image in the frequency domain and the spatial domain respectively to generate features at the current time step; and use a structured attention mechanism to fuse the features of the current time step and the features of the auxiliary modality high-resolution image to obtain a reconstructed image of the target modality.

2. The MRI image reconstruction method according to claim 1, wherein The truncating high-frequency components of the initial frequency domain data to generate low-resolution frequency domain data includes: Obtain a preset mask that occludes high-frequency components from the edge to the center at a preset step size, and based on the initial frequency domain data, truncate the high-frequency components by shrinking the mask of the frequency center region at the preset step size to generate the low-resolution frequency domain data at different time steps.

3. The MRI image reconstruction method according to claim 1, wherein The extracting features of the low-resolution frequency domain data and the spatial domain low-resolution image in the frequency domain and the spatial domain respectively to generate features at the current time step; And using a structured attention mechanism to fuse the features of the current time step and the features of the auxiliary modality high-resolution image to obtain a reconstructed image of the target modality includes: The features at the current time step include frequency domain time step features and spatial domain time step features; In the frequency domain, perform encoder feature extraction on the low-resolution frequency domain data to generate the frequency domain time step features; use the structured attention mechanism to fuse the frequency domain time step features and the frequency domain features of the auxiliary modality high-resolution image to generate a high-frequency missing distribution feature; In the spatial domain, perform encoder feature extraction on the spatial domain low-resolution image to generate the spatial domain time step features; use the structured attention mechanism and fuse the spatial domain time step features and the spatial domain features of the auxiliary modality high-resolution image to generate a local detail feature; Based on the high-frequency missing distribution feature and the local detail feature, obtain the reconstructed image of the target modality.

4. The MRI image reconstruction method according to claim 1, wherein, The structured attention mechanism dynamically fuses the features of the current time step and the features of the auxiliary modality high-resolution image through attention weights.

5. The MRI image reconstruction method according to claim 1, characterized in that The method further includes: Obtain training MRI sample data, including a training target modality high-resolution image and a training auxiliary modality high-resolution image; Perform frequency domain conversion on the training target modality high-resolution image to obtain training initial frequency domain data, truncate high-frequency components of the training initial frequency domain data to generate training low-resolution frequency domain data; perform spatial domain conversion on the training initial frequency domain data to generate a training spatial domain initial image; Input the training low-resolution frequency-domain data and the initial training spatial-domain image into the initial diffusion model. Using the initial diffusion model, perform feature extraction on the training low-resolution frequency-domain data and the initial training spatial-domain image in the frequency domain and the spatial domain respectively to generate the feature of the current training time step; and use the structural attention mechanism to perform feature fusion on the feature of the current training time step and the feature of the training auxiliary-modal high-resolution image to reconstruct the training reconstruction image of the target modality; The training reconstruction image includes frequency-domain reconstruction data and spatial-domain reconstruction data; Construct a dual-domain data consistency loss function and a reconstruction loss function; among them, the dual-domain data consistency loss function includes a frequency-domain consistency loss function and a spatial-domain consistency loss function; Based on the dual-domain data consistency loss function, the reconstruction loss function, and the dual-domain data consistency loss function and the reconstruction loss function, calculate the frequency-domain consistency loss value, the spatial-domain consistency loss value, and the reconstruction loss value; Fuse the frequency-domain consistency loss value, the spatial-domain consistency loss value, and the reconstruction loss value with weights to obtain the total loss function value; based on the total loss function value, iteratively optimize the parameters of the initial diffusion model to generate a trained diffusion model.

6. The MRI image reconstruction method according to claim 5, wherein, The calculating the frequency-domain consistency loss value, the spatial-domain consistency loss value, and the reconstruction loss value based on the dual-domain data consistency loss function, the reconstruction loss function, and the dual-domain data consistency loss function and the reconstruction loss function includes: Based on the frequency-domain consistency loss function, calculate the difference between the frequency-domain feature of the frequency-domain reconstruction image and the frequency-domain feature of the corresponding training target-modal high-resolution image to generate a frequency-domain consistency loss value; Based on the spatial-domain consistency loss function, calculate the difference between the spatial-domain feature of the spatial-domain reconstruction image and the spatial-domain feature of the corresponding training target-modal high-resolution image to generate a spatial-domain consistency loss value; Based on the reconstruction loss function, calculate the difference between the feature of the training reconstruction image and the feature of the corresponding training target-modal high-resolution image to generate a reconstruction loss value.

7. The MRI image reconstruction method according to claim 5, wherein The fusing the frequency-domain consistency loss value, the spatial-domain consistency loss value, and the reconstruction loss value with weights includes: When the proportion of low-frequency components in the training low-resolution frequency-domain data is greater than that of high-frequency components, the weight value of the frequency-domain consistency loss value is greater than the weight value of the spatial-domain consistency loss value; When the proportion of low-frequency components in the training low-resolution frequency-domain data is not greater than that of high-frequency components, the weight value of the spatial-domain consistency loss value is greater than the weight value of the frequency-domain consistency loss value; Based on the weight value of the spatial-domain consistency loss value, the weight value of the frequency-domain consistency loss value, and the weight value of the reconstruction loss value, fuse the frequency-domain consistency loss value, the spatial-domain consistency loss value, and the reconstruction loss value with weights.

8. An MRI image reconstruction device, characterized in that, including: An image acquisition module for acquiring multi-contrast MRI sample data, including a target-modal high-resolution image and an auxiliary-modal high-resolution image; A data processing module for performing frequency-domain conversion on the target-modal high-resolution image to obtain initial frequency-domain data, truncating high-frequency components of the initial frequency-domain data to generate low-resolution frequency-domain data; Perform spatial domain conversion on the low-resolution frequency domain data to generate a low-resolution image in the spatial domain; A reconstruction image module for inputting the low-resolution frequency domain data and the low-resolution image in the spatial domain into a diffusion model, using the diffusion model to perform feature extraction on the low-resolution frequency domain data and the low-resolution image in the spatial domain and the frequency domain respectively to generate features at the current time step; and using a structural attention mechanism to perform feature fusion on the features at the current time step and the features of the high-resolution image in the auxiliary modality to obtain a reconstructed image of the target modality.

9. An electronic device, comprising a memory and a processor, characterized in that, A computer program is stored in the memory, and the processor is configured to run the computer program to execute the MRI image reconstruction method according to any one of claims 1 to 7.

10. A storage medium, characterized in that, A computer program is stored in the storage medium, wherein the computer program is configured to execute the MRI image reconstruction method according to any one of claims 1 to 7 when running.

Images