An MRI image reconstruction method and system based on a shared latent space diffusion model and physical consistency constraints, a terminal and a storage medium

Abstract

This invention discloses an MRI image reconstruction method, system, terminal, and storage medium based on a shared latent space diffusion model and physical consistency constraints. The method includes: acquiring a low-field MRI image and preprocessing it, then inputting it into a low-field encoder to extract latent variables; inputting the latent variables into a high-field decoder to generate an initial high-field baseline image, and inputting it into a structure-aware diffusion enhancement module to obtain a structure-enhanced image; correcting the current reconstruction result based on the low-field MRI image to obtain a corrected image; using a detail diffusion model to obtain a detail-enhanced image based on the structure-enhanced image and the corrected image; and using a residual diffusion model to learn from the detail-enhanced image and compensate for remaining high-frequency residual information, outputting the final high-field MRI reconstruction result. This invention achieves low-field to high-field MRI reconstruction by encoding an initial high-field image, combining structure diffusion and high-frequency residual diffusion to recover details, and utilizing shape-gated iterative optimization.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance imaging technology, and in particular to an MRI image reconstruction method, system, terminal, and computer-readable storage medium based on a shared latent space diffusion model and physical consistency constraints. Background Technology

[0002] In recent years, low-field magnetic resonance imaging (MRI) has gained increasing attention in primary healthcare, bedside imaging, emergency care, and resource-constrained areas due to its advantages such as low equipment cost, low power consumption, portability, and flexible deployment. For example, 0.3T and 0.064T low-field MRI devices can perform rapid scans in environments where traditional high-field MRI is difficult to deploy, making them particularly suitable for mobile healthcare and real-time clinical support scenarios.

[0003] However, compared to the 1.5T or 3T high-field MRI systems commonly used in clinical practice, low-field MRI, due to its lower magnetic field strength, results in a significantly weakened magnetic resonance signal. This manifests as a low signal-to-noise ratio (SNR), insufficient spatial resolution, blurred edges, loss of high-frequency textures, and noticeable artifacts. In brain, knee joint, and soft tissue imaging, this decline in image quality directly affects the discernibility of tissue boundaries, the visibility of subtle anatomical structures, and the clarity of overall image interpretation. For example, problems such as blurred boundaries between white and gray matter, loss of small blood vessel structures, and unclear cartilage edges in low-field MRI reduce the visibility of key structural features, thereby affecting the accuracy of image-based post-processing analysis or manual interpretation. Therefore, how to restore images with near-high-field MRI quality while maintaining the low-cost advantage of low-field equipment has become an important research direction in the field of medical image reconstruction.

[0004] Existing methods mainly include traditional super-resolution methods, convolutional neural network reconstruction methods, generative adversarial network (GAN) methods, and diffusion model methods developed in recent years. However, these methods still have significant shortcomings.

[0005] First, traditional convolutional network methods typically rely on a large number of strictly paired low-field / high-field images as supervised data. However, in real clinical scenarios, it is difficult to obtain fully registered low-field and high-field images of the same patient in the same pose and at the same level, thus severely limiting supervised training.

[0006] Second, although unpaired image translation methods such as CycleGAN can learn the mapping relationship from low field to high field in the absence of paired data, these methods are essentially single-step generation models that focus more on overall style transfer and are difficult to recover real high-frequency texture details. The generated results are prone to structural drift, edge blurring and pseudo-texture.

[0007] Third, although existing diffusion models have achieved high-quality generation results in the field of natural image generation, when directly applied to low-field MRI enhancement, they often have the following problems: (1) lack of constraints from the physical model of magnetic resonance imaging; (2) inability to guarantee the consistency of k-space data; (3) insufficient recovery of high-frequency details; (4) mismatch between the generated results and the real sampling physical process; and (5) weak generalization ability on real low-field datasets.

[0008] Fourth, most current medical image translation models only optimize within the image domain, without incorporating physical information such as Fourier transforms, sampling masks, and coil sensitivity from the original MRI sampling process. This often results in generated images that, while visually appealing, do not accurately reflect the actual acquired data. This inconsistency reduces the model's reliability and the fidelity of image reconstruction in real-world applications.

[0009] Fifth, most current low-field enhancement methods only focus on improving overall contrast, without specifically addressing the crucial issue of "high-frequency texture restoration." In fact, one of the core differences between low-field and high-field MRI is the loss of high-frequency information; for example, fine structures such as small cerebral sulci, cartilage boundaries, and subtle tissue textures cannot be accurately restored. Therefore, simply increasing image brightness or overall sharpness cannot truly achieve the image quality level corresponding to high-field MRI, and is insufficient to meet the needs of subsequent image analysis or auxiliary observation.

[0010] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0011] The main objective of this invention is to provide an MRI image reconstruction method, system, terminal, and computer-readable storage medium based on a shared latent space diffusion model and physical consistency constraints, aiming to solve the problems of high data dependence, lack of physical constraints, insufficient high-frequency detail recovery, and poor model generalization in existing low-field magnetic resonance imaging (MRI) enhancement techniques.

[0012] To achieve the above objectives, this invention provides an MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints. The MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints includes the following steps: Acquire low-field MRI images, preprocess the low-field MRI images to obtain preprocessed low-field MRI images, input the preprocessed low-field MRI images into a low-field encoder, and extract cross-modal shared latent variables. The cross-modal shared latent variables are input into the high-field decoder for decoding to generate an initial high-field baseline image. The initial high-field baseline image is then input into the structure-aware diffusion enhancement module for structure recovery to obtain a structure-enhanced image, which is then input into the detail diffusion model. A latent space data consistency proximal optimization module is introduced to correct the current reconstruction result based on the low-field MRI image, obtain a corrected image, and input the corrected image into the detail diffusion model; Based on the detail diffusion model, detail restoration processing is performed on the structure-enhanced image and the corrected image to restore the local texture and high-frequency structure in high-field MRI, thereby obtaining a detail-enhanced image, which is then input into the residual diffusion model. The detailed enhancement image is learned based on the residual diffusion model, and the remaining high-frequency residual information is compensated to output the final high-field MRI reconstruction result.

[0013] Optionally, the MRI image reconstruction method based on the shared latent space diffusion model and physical consistency constraints, wherein the steps of acquiring a low-field MRI image, preprocessing the low-field MRI image to obtain a preprocessed low-field MRI image, and inputting the preprocessed low-field MRI image into a low-field encoder to extract cross-modal shared latent variables, specifically include: A low-field MRI image is acquired, and the low-field MRI image is preprocessed to obtain a preprocessed low-field MRI image. The preprocessing includes intensity normalization, center cropping, size unification, Fourier transform standardization, multi-coil reconstruction, and noise standardization. The preprocessed low-field MRI image is input into a low-field encoder for encoding to obtain shared latent variables: ; in, To share latent variables across modalities, The image is a preprocessed low-field MRI image. It is a low-field encoder.

[0014] Optionally, the MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints, wherein the step of inputting the cross-modal shared latent variables into a high-field decoder for decoding to generate an initial high-field baseline image, inputting the initial high-field baseline image into a structure-aware diffusion enhancement module for structure recovery to obtain a structure-enhanced image, and inputting it into a detail diffusion model, specifically includes: Sharing latent variables across modalities The input is fed into the high-field decoder for decoding, generating the initial high-field baseline image: ; in, For the initial high-field baseline image, For high-field decoders; Initial high-field baseline image The image is input into the structure-aware diffusion enhancement module for structure reconstruction processing, resulting in an enhanced structure image. ; Enhanced image Input into the detail diffusion model.

[0015] Optionally, in the MRI image reconstruction method based on the shared latent space diffusion model and physical consistency constraints, the structure-aware diffusion enhancement module adopts a conditional diffusion model, and the forward diffusion process is as follows: ; in, Represents a true high-field MRI image. Indicates the first The noisy image corresponding to each diffusion time step This represents the conditional probability distribution during the forward diffusion process. Indicates a Gaussian distribution. Indicates the first Signal retention coefficients for each diffusion time step Indicates the first The noise intensity at each diffusion time step Represents the identity matrix; The reverse denoising process is as follows: ; in, Indicates the first The noisy image corresponding to each diffusion time step Indicates the current noisy image and initial high-field baseline image As a condition, through learnable parameters The denoising network predicts the noisy image from the previous time step. The conditional probability distribution.

[0016] Optionally, the MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints, wherein the introduction of a latent space data consistency proximal optimization module to correct the current reconstruction result based on the low-field MRI image to obtain a corrected image, and inputting the corrected image into the detail diffusion model, specifically includes: Introducing k-space data consistency constraints, the latent space data consistency proximal optimization module adopts the MRI forward model, which is represented as follows: ; in, For sampling mask, For Fourier transform, For coil sensitivity, MRI images in the image domain; To observe k-space data; Data consistency loss for: ; in: To generate the result; The current reconstruction result is corrected based on the low-field MRI image to obtain a corrected image, which is then input into the detail diffusion model.

[0017] Optionally, the MRI image reconstruction method based on the shared latent space diffusion model and physical consistency constraints further includes: A manifold geometry-gated optimization mechanism is introduced, which decomposes the data consistency correction direction into the manifold tangential direction and the manifold normal direction. The expression of the manifold geometry-gated optimization mechanism is as follows: ; in, Indicates the first The current high-field clean image estimate obtained from the diffusion model prediction in the next iteration. This represents the high-field image estimation after manifold geometry gating optimization. This indicates that the data consistency correction direction is along the tangential component of the image manifold. This indicates the normal component of the data consistency correction direction along the image manifold. Indicates the first The adaptive gating coefficients of the next iteration are used to control the intensity of the introduction of the normal component; The gating coefficient is dynamically calculated based on measurement consistency and structural similarity. ; in, Indicates the first Normalized measurement mismatch at the next iteration The representation is defined as follows: This represents the magnetic resonance forward measurement operator, used to simulate the acquisition process from image to observation data. This represents a high-field encoder used to estimate the current high-field clean image. Mapped to the high-field latent space This refers to a target domain decoder or measurement consistency projection decoder, used to map latent space features to an image domain consistent with the observed data. This represents the smallest positive integer that prevents the denominator from being zero. ; in, express and cosine similarity, Indicates the first The overall correction direction obtained from data consistency in the next iteration This represents the inner product operation. Represents the L2 norm; ; in, This represents the temperature coefficient of the measurement mismatch, used to control the strength of the suppression of data consistency errors on the gating coefficient. This represents the geometric similarity between the correction direction and the tangent direction of the manifold. Represents the geometric similarity threshold. The temperature coefficient representing the geometric similarity term is used to control the gating function. The smoothness of the change This represents the Sigmoid function.

[0018] Optionally, the MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints, wherein learning the detail-enhanced image based on the residual diffusion model and compensating for remaining high-frequency residual information to output the final high-field MRI reconstruction result specifically includes: Constructing high-frequency residuals using high-field images: ; in, Indicates high-frequency residuals, Indicates high-pass filtering; A true high-field image; Enhance the image for structure; Learning high-frequency residual distributions using a residual diffusion model: ; in, Indicates the high-frequency residual distribution, It is Gaussian noise; Image detail enhancement based on a residual diffusion model learned from high-frequency residual distribution. The remaining high-frequency residual information is learned and compensated, and the final high-field MRI reconstruction result is output. .

[0019] Furthermore, to achieve the above objectives, the present invention also provides an MRI image reconstruction system based on a shared latent space diffusion model and physical consistency constraints, wherein the MRI image reconstruction system based on a shared latent space diffusion model and physical consistency constraints is used to implement the MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints, and the MRI image reconstruction system based on a shared latent space diffusion model and physical consistency constraints includes: The latent variable extraction module is used to acquire low-field MRI images, preprocess the low-field MRI images to obtain preprocessed low-field MRI images, input the preprocessed low-field MRI images into the low-field encoder, and extract cross-modal shared latent variables. The image enhancement module is used to input the cross-modal shared latent variables into the high-field decoder for decoding, generate an initial high-field baseline image, input the initial high-field baseline image into the structure-aware diffusion enhancement module for structure recovery, obtain a structure-enhanced image, and input it into the detail diffusion model; The image correction module is used to introduce the latent space data consistency proximal optimization module, correct the current reconstruction result according to the low field MRI image, obtain the corrected image, and input the corrected image into the detail diffusion model; The detail enhancement module is used to perform detail restoration processing based on the structure enhancement image and the correction image according to the detail diffusion model, to restore the local texture and high-frequency structure in high-field MRI, obtain the detail enhancement image, and input it into the residual diffusion model; The image reconstruction module is used to learn the detail-enhanced image based on the residual diffusion model and compensate for the remaining high-frequency residual information, and output the final high-field MRI reconstruction result.

[0020] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and an MRI image reconstruction program based on a shared latent space diffusion model and physical consistency constraints stored in the memory and executable on the processor. When the MRI image reconstruction program based on a shared latent space diffusion model and physical consistency constraints is executed by the processor, it implements the steps of the MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints as described above.

[0021] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an MRI image reconstruction program based on a shared latent space diffusion model and physical consistency constraints, wherein when the MRI image reconstruction program based on a shared latent space diffusion model and physical consistency constraints is executed by a processor, it implements the steps of the MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints as described above.

[0022] In this invention, a low-field MRI image is acquired, preprocessed to obtain a preprocessed low-field MRI image, and input into a low-field encoder to extract cross-modal shared latent variables. The cross-modal shared latent variables are then input into a high-field decoder for decoding to generate an initial high-field baseline image. This initial high-field baseline image is input into a structure-aware diffusion enhancement module for structure recovery, resulting in a structure-enhanced image, which is then input into a detail diffusion model. A latent space data consistency proximal optimization module is introduced to correct the current reconstruction result based on the low-field MRI image, obtaining a corrected image, which is then input into the detail diffusion model. Based on the detail diffusion model, detail recovery processing is performed on the structure-enhanced image and the corrected image to recover local textures and high-frequency structures in the high-field MRI, resulting in a detail-enhanced image, which is then input into a residual diffusion model. Based on the residual diffusion model, the detail-enhanced image is learned and the remaining high-frequency residual information is compensated, outputting the final high-field MRI reconstruction result. This invention first uses a shared latent variable encoder to map low-field MRI to a shared latent space, generates an initial high-field image through a high-field decoder, then uses a structure diffusion model to recover high-field structure information, and then uses a high-frequency residual diffusion model to recover edge and texture details. After that, k-space data consistency optimization is performed through an MRI physical forward model, and the generated results are iteratively corrected by a manifold geometry gating mechanism, thereby obtaining the final high-field MRI reconstruction result. Attached Figure Description

[0023] Figure 1 This is a flowchart of a preferred embodiment of the MRI image reconstruction method based on the shared latent space diffusion model and physical consistency constraints of the present invention; Figure 2 This is a schematic diagram of the overall process for reconstructing a high-field magnetic resonance image from a low-field magnetic resonance image in a preferred embodiment of the MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints of the present invention. Figure 3 This is a structural diagram of a preferred embodiment of the MRI image reconstruction system based on the shared latent space diffusion model and physical consistency constraints of the present invention; Figure 4 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0025] Magnetic Resonance Imaging (MRI) is an important medical imaging technique with advantages such as being radiation-free, providing high soft tissue contrast, and possessing strong multi-parameter imaging capabilities. Traditional clinical MRI systems often use 1.5T or 3T high-field magnets. Due to the higher magnetic field strength, high-field MRI can obtain higher signal-to-noise ratios, higher spatial resolution, and richer tissue contrast information. Therefore, it is widely used in imaging examinations of the brain, spine, joints, and other areas, as well as for detailed observation of soft tissue structures.

[0026] However, high-field MRI systems suffer from high equipment costs, complex installation, high operation and maintenance costs, and demanding infrastructure requirements, making them difficult to deploy widely in primary healthcare, remote areas, and mobile healthcare scenarios. In recent years, low-field MRI devices have gradually gained attention due to their small size, low cost, low power consumption, and high portability. For example, portable 0.064T MRI and 0.3T MRI systems have begun to be used in emergency rooms, ICUs, and bedside imaging.

[0027] However, low-field MRI also has significant drawbacks. Due to the reduced magnetic field strength, the magnetic resonance signal weakens, leading to a significant decrease in the image signal-to-noise ratio. Furthermore, low-field equipment is typically limited by scan time and hardware performance, further resulting in insufficient image resolution, loss of texture detail, and blurring of high-frequency edges. These problems severely impact image quality, reducing the discernibility of anatomical structures in the images and the reliability of subsequent image processing, structural measurements, and quantitative analysis.

[0028] Traditional convolutional neural network (CNN) methods typically employ supervised learning to map low-quality MRI images to high-quality ones. For example, the U-Net architecture uses an encoder and decoder to extract multi-scale features and utilizes skip connections to preserve spatial information. These methods have achieved some success in super-resolution and denoising tasks. However, they usually rely on strictly paired low-field and high-field data, which are difficult to obtain in real-world clinical settings, thus limiting their generalization ability.

[0029] This paper proposes a GAN-based unpaired image translation framework. Its core idea is to achieve bidirectional mapping between different domains through cycle consistency loss. The main structure includes: a low-field to high-field generator; a high-field to low-field generator; two discriminators; and cycle consistency constraints. Cycle GANs can complete image translation without paired data and are therefore widely used in medical imaging. However, this method has the following problems: it is essentially a single-step mapping; it has weak high-frequency detail recovery capability; it is prone to generating pseudo-structures; it lacks MRI physical constraints; and it is difficult to guarantee that the generated results are consistent with the actual samples.

[0030] UNIT (Unsupervised Image-to-Image Translation Networks) proposes the shared latent space hypothesis, meaning that images from different domains share the same content representation in the latent space. Its main structure includes: two variational autoencoders (VAEs); a shared latent variable space; a cross-domain decoder; and joint optimization of adversarial and reconstruction losses. This method has better latent space representation capabilities than CycleGAN, but it is still a one-step image translation method with limited ability to recover high-frequency textures.

[0031] In recent years, diffusion models have made groundbreaking progress in image generation. Diffusion models learn the data distribution through progressive noise addition and a reverse denoising process, resulting in high-quality generated images. In the MRI field, existing research has applied diffusion models to MRI reconstruction and super-resolution. For example, methods such as Score-based MRI Reconstruction and SynDiff utilize the diffusion process to recover high-quality images. SynDiff, in particular, combines diffusion models with medical image translation, generating high-quality images through conditional diffusion. However, existing diffusion models still have the following shortcomings: lack of explicit MRI physical consistency; the generated results are prone to deviating from the original measurements; unstable high-frequency detail recovery; slow diffusion sampling speed; and insufficient generalization ability to real low-field data.

[0032] Physical model-based MRI reconstruction methods embed a physical forward model of the MRI MRI into a deep network and implement k-space constraints through a data consistency layer. These methods typically include: Fourier transform; sampling mask; coil sensitivity mapping; and a data consistency iterative module. While physical model methods can guarantee consistency between the reconstruction results and the measurement data, they are primarily geared towards undersampled MRI reconstruction, rather than low-field to high-field image enhancement tasks.

[0033] The Neural Schrödinger Bridge (UNSB) method models image translation as a multi-step probabilistic bridging process, improving generation stability through progressive refinement. This method achieves more stable generation results compared to GANs, but the original method was not designed for MRI physical consistency and high-frequency texture restoration.

[0034] Existing low-field MRI enhancement methods mainly have the following drawbacks: 1. Heavy reliance on paired data: Most supervised learning methods require strictly registered low-field and high-field MRI as training labels, but it is difficult to obtain strong paired data of the same patient and the same location in real clinical practice, thus limiting the application of the model.

[0035] 2. Insufficient ability to recover high-frequency details: Methods such as CycleGAN focus more on overall style transfer and have difficulty recovering real high-frequency textures, resulting in blurry generated results.

[0036] 3. Lack of MRI physical consistency constraints: Most existing image translation methods only optimize in the image domain and do not incorporate MRI sampling physical models, so the output results may be inconsistent with the real k-space sampling.

[0037] 4. Prone to generating pseudo-structures: GAN-type methods may generate false textures and incorrect structures during unsupervised training, thus affecting the reliability of detection results.

[0038] 5. Poor generalization ability: Many methods are only effective on specific devices or specific datasets, and are prone to failure when faced with real 0.3T low-field data.

[0039] 6. Limited single-stage generation capability: One-step generation methods cannot simultaneously achieve both structural and detail restoration, which can easily lead to models with correct structure but missing textures.

[0040] To address the aforementioned issues, this invention proposes a method for reconstructing high-field MRI images from low-field MRI based on a shared latent space diffusion model and physical consistency constraints. This method achieves high-fidelity reconstruction of the style and structure of low-field MRI to high-field MRI by means of shared latent variable encoding, structural diffusion enhancement, high-frequency residual diffusion recovery, and joint optimization of k-space data consistency.

[0041] This invention aims to solve the following technical problems: (1) Achieve stable mapping from low-field MRI to high-field MRI in the absence of strictly paired high-field labels.

[0042] (2) Restore missing high-frequency textures and edge details in low-field MRI.

[0043] (3) Introduce MRI physical measurement consistency constraints during the generation process. (4) Reduce the structural drift problem in the generation of diffusion models.

[0044] (5) Improve the model’s ability to generalize to real low-field data (such as 0.3T data).

[0045] (6) Construct a low-field MRI enhancement framework that balances image quality and physical reliability.

[0046] Through the above technical solution, the present invention can generate image results that are closer to high-field MRI in real low-field MRI scenarios, while maintaining the consistency of anatomical structure and measurement physics, thereby improving the clinical application value of low-field MRI.

[0047] The preferred embodiment of the MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints described in this invention, such as... Figure 1 and Figure 2 As shown, the MRI image reconstruction method based on the shared latent space diffusion model and physical consistency constraints includes the following steps: Step S10: Acquire a low-field MRI image, preprocess the low-field MRI image to obtain a preprocessed low-field MRI image, input the preprocessed low-field MRI image into a low-field encoder, and extract cross-modal shared latent variables.

[0048] Specifically, a low-field MRI image is acquired, and the low-field MRI image is preprocessed to obtain a preprocessed low-field MRI image. The preprocessing includes intensity normalization, center cropping, size unification, Fourier transform standardization, multi-coil reconstruction, and noise standardization. The preprocessed low-field MRI image is then input into a low-field encoder for encoding to obtain shared latent variables. ; in, To share latent variables across modalities (i.e.) Figure 2 Hidden variables in ), The image is a preprocessed low-field MRI image. For low field encoders (i.e.) Figure 2 Low field encoder in E A ).

[0049] The low-field encoder employs a convolutional coding structure; latent variables are shared across modalities. It indicates cross-field shared content; the low-field encoder mainly includes: a 7×7 convolutional layer; two downsampling modules; a residual block module; and a mean and variance output layer.

[0050] Step S20: Input the cross-modal shared latent variable into the high-field decoder for decoding to generate an initial high-field baseline image. Input the initial high-field baseline image into the structure-aware diffusion enhancement module for structure recovery to obtain a structure-enhanced image, and input it into the detail diffusion model.

[0051] Specifically, sharing latent variables across modalities The input is fed into the high-field decoder for decoding, generating the initial high-field baseline image: ; in, For the initial high-field baseline image (i.e. Figure 2 (initial baseline image in the image). For high field decoders (i.e.) Figure 2 High field decoder in G B The high-field decoder is responsible for recovering the high-field structure.

[0052] Then the initial high-field baseline image The image is input into the structure-aware diffusion enhancement module (which is mainly responsible for recovering low-frequency structural information in high-field MRI) for structure restoration processing, resulting in an enhanced structure image. In the process of enhancing the structure of the image Input into the detail diffusion model.

[0053] The structure-aware diffusion enhancement module adopts a conditional diffusion model, and the forward diffusion process is as follows: ; in, Represents a true high-field MRI image. Indicates the first The noisy image corresponding to each diffusion time step This represents the conditional probability distribution during the forward diffusion process, derived from real high-field MRI images. Gradually adding noise to obtain the first Noisy images at different times The probability distribution, Indicates a Gaussian distribution. Indicates the first Signal retention coefficients for each diffusion time step Indicates the first The noise intensity at each diffusion time step The identity matrix is ​​used to represent the addition of independent and identically distributed Gaussian noise in each pixel / voxel dimension.

[0054] Meanwhile, the reverse denoising process is as follows: ; in, Indicates the first The noisy image corresponding to each diffusion time step Indicates the current noisy image and initial high-field baseline image As a condition, through learnable parameters The denoising network predicts the noisy image from the previous time step. The conditional probability distribution.

[0055] In other words, the forward process is responsible for gradually adding noise, while the reverse process gradually removes noise under the guidance of the initial high-field baseline image, ultimately restoring a reconstruction result with clearer structure and contrast closer to that of high-field MRI.

[0056] Step S30: Introduce the latent space data consistency proximal optimization module, correct the current reconstruction result based on the low-field MRI image to obtain a corrected image, and input the corrected image into the detail diffusion model.

[0057] Specifically, to ensure that the generated results conform to the physical process of MRI sampling, a k-space data consistency constraint is introduced. The latent space data consistency proximal optimization module adopts the MRI forward model, which is represented as follows: ; in, For sampling mask, For Fourier transform, For coil sensitivity, MRI images in the image domain; For observing k-space data (i.e., real observation data).

[0058] This invention achieves data consistency by: transforming the generated image to k-space; retaining the data at the actual sampling locations; replacing the corresponding frequency domain sampling points; performing a new inverse Fourier transform; and iterative optimization.

[0059] Data consistency loss for: ; in: To generate the result; The current reconstruction result is corrected based on the low-field MRI image to obtain a corrected image, which is then input into the detail diffusion model.

[0060] Furthermore, a manifold geometry-gated optimization mechanism is introduced. Its core idea is to decompose the data consistency correction direction into the manifold tangential direction and the manifold normal direction. The expression for this manifold geometry-gated optimization mechanism is: ; in, Indicates the first The current high-field clean image estimate obtained from the diffusion model prediction in the next iteration. This represents the high-field image estimation after manifold geometry gating optimization. This indicates that the data consistency correction direction is along the tangential component of the image manifold. This indicates the normal component of the data consistency correction direction along the image manifold. Indicates the first The adaptive gating coefficients of the next iteration are used to control the intensity of the introduction of the normal component.

[0061] Subsequently, the gating coefficients are dynamically calculated based on measurement consistency and structural similarity: ; in, Indicates the first Normalized measurement mismatch at the next iteration The representation is defined as follows: This represents the magnetic resonance forward measurement operator, used to simulate the acquisition process from image to observation data. This represents a high-field encoder used to estimate the current high-field clean image. Mapped to the high-field latent space This refers to a target domain decoder or measurement consistency projection decoder, used to map latent space features to an image domain consistent with the observed data. This represents the smallest positive integer that prevents the denominator from being zero.

[0062] Geometric alignment (the closer the alignment, the more innovation is allowed), cosine similarity: ; in, express and cosine similarity, Indicates the first The overall correction direction obtained from data consistency in the next iteration This represents the inner product operation. This represents the L2 norm.

[0063] when A higher value indicates that the current data consistency (DC) correction is mainly located in the manifold tangential direction, which is a relatively safe structural refinement; when... A lower value indicates that the correction contains more normal components, which may lead to structural shifts, artifacts, or unreliable details. Therefore, it is necessary to use gating coefficients. Suppress normal updates.

[0064] This method can: reduce over-correction; reduce pseudo-structures; improve generation stability; and maintain structural authenticity.

[0065] ; in, This represents the temperature coefficient of the measurement mismatch, used to control the strength of the suppression of data consistency errors on the gating coefficient. This represents the geometric similarity between the correction direction and the tangent direction of the manifold. Represents the geometric similarity threshold, when When the correction direction is closer to the manifold tangential direction, more normal information can be allowed to enter; The temperature coefficient representing the geometric similarity term is used to control the gating function. The smoothness of the change This represents the Sigmoid function.

[0066] The total loss function of this invention includes: pixel reconstruction loss, diffusion noise prediction loss, high frequency loss, edge loss, data consistency loss, adversarial loss, and latent variable consistency loss.

[0067] Step S40: Based on the detail diffusion model, perform detail restoration processing on the structure enhancement image and the correction image to restore the local texture and high-frequency structure in the high-field MRI, obtain the detail enhancement image, and input it into the residual diffusion model.

[0068] Specifically, such as Figure 2 As shown, the image is enhanced based on the structure according to the detail diffusion model. The corrected image is then subjected to detail restoration processing to recover the local texture and high-frequency structure in the high-field MRI, resulting in a detail-enhanced image. And input it into the residual diffusion model.

[0069] Step S50: Based on the residual diffusion model, learn the detail enhancement image and compensate for the remaining high-frequency residual information, and output the final high-field MRI reconstruction result.

[0070] Specifically, high-frequency residuals are constructed using high-field images: ; in, Indicates high-frequency residuals, Indicates high-pass filtering; A true high-field image; Enhance the image for structure; Learning high-frequency residual distributions using a residual diffusion model: ; in, Indicates the high-frequency residual distribution, It is Gaussian noise; A residual diffusion model based on the learned high-frequency residual distribution (used to restore missing texture and edge details in low-field MRI, focusing on restoring: brain sulcus texture, cartilage edges, target region details, tissue texture, and hyperspectral information) is used to enhance the detail of images. The remaining high-frequency residual information is learned and compensated, and the final high-field MRI reconstruction result is output. .

[0071] This invention first uses a shared latent variable encoder to map low-field MRI to a shared latent space, and then generates an initial high-field image using a high-field decoder. Subsequently, a structure diffusion model is used to recover high-field structural information, and a high-frequency residual diffusion model is used to recover edge and texture details. Afterwards, this invention optimizes k-space data consistency using an MRI physical forward model and iteratively corrects the generated results using a manifold geometry gating mechanism, thereby obtaining the final high-field MRI reconstruction result.

[0072] This invention simultaneously supports: supervised paired training; unsupervised unpaired training; zero-shot generalization of low-field real data; multi-coil k-space input; single-channel image input; and multimodal MRI extension.

[0073] The key innovations of this invention include: (1) Low-field to high-field MRI mapping mechanism based on shared latent space: A unified structural representation between low-field and high-field domains is achieved through shared latent variables.

[0074] (2) Two-stage recovery framework based on structural diffusion and high-frequency residual diffusion: adopting a separate design for structural recovery and detail recovery to improve the high-frequency texture recovery capability.

[0075] (3) High-frequency residual diffusion learning mechanism: Diffusion learning is performed using high-frequency residuals instead of direct images to improve the stability of detail generation.

[0076] (4) MRI physical consistency constraint mechanism: k-space data consistency optimization is introduced during the generation process.

[0077] (5) Manifold geometry gated optimization mechanism: physical constraints and structural constraints are jointly optimized through tangential and normal decomposition.

[0078] (6) Unpaired low-field MRI enhancement training strategy: High-quality low-field to high-field conversion can be completed without strict pairing data.

[0079] (7) Multi-stage joint training and inference framework: to achieve unified and collaborative optimization of shared latent variables, diffusion model and physical consistency.

[0080] (8) Multi-coil k-space compatible structure: supports input of real MRI raw k-space data.

[0081] Compared with the prior art, the present invention has the following advantages: (1) No strict pairing data is required. Compared with traditional supervised methods, this invention can be trained without strict low field / high field pairing data.

[0082] (2) Stronger high-frequency detail recovery capability: The present invention has specially designed a high-frequency residual diffusion module, which can recover more edge and texture information.

[0083] (3) It has MRI physical consistency. Through k-space data consistency constraints, it ensures that the generated results conform to the real MRI sampling physical process, reduces the generation of pseudo structures, and the manifold geometry gating mechanism can reduce the structural drift problem in GAN and diffusion models. Multi-stage generation is more stable. The separation design of structural diffusion and detail diffusion can improve training stability, has stronger generalization ability, and can adapt to real 0.3T low-field MRI data.

[0084] (4) It is suitable for real clinical deployment and can be deployed on low-field MRI workstations and edge GPU devices. It is compatible with various MRI data formats and supports image domain, k-space domain and multi-coil input.

[0085] This invention has been experimentally validated based on publicly available MRI datasets and real low-field MRI data. The experimental data include: M4Raw low-field MRI dataset and 0.5T brain low-field MRI dataset.

[0086] Experimental results show that the present invention outperforms methods such as CycleGAN, UNIT, SynDiff, and UNSB in terms of image quality, contrast, resolution, and high-frequency recovery capabilities. In terms of high-frequency detail recovery, the present invention can recover more edge and texture structures. Experimental results on real low-field 0.3T data demonstrate that the present invention has good generalization ability.

[0087] Ablation experiments further demonstrate that: the high-frequency residual diffusion module can significantly improve texture recovery capability; the data consistency module can improve physical reliability; the manifold gating mechanism can reduce pseudo-structure generation; and the multi-stage diffusion structure is superior to the single-stage generation structure. Experimental results indicate that this invention has good feasibility and application value.

[0088] Furthermore, such as Figure 3As shown, based on the above-mentioned MRI image reconstruction method based on the shared latent space diffusion model and physical consistency constraints, the present invention also provides an MRI image reconstruction system based on the shared latent space diffusion model and physical consistency constraints. The MRI image reconstruction system based on the shared latent space diffusion model and physical consistency constraints is used to implement the MRI image reconstruction method based on the shared latent space diffusion model and physical consistency constraints. The MRI image reconstruction system based on the shared latent space diffusion model and physical consistency constraints includes: The latent variable extraction module 51 is used to acquire low-field MRI images, preprocess the low-field MRI images to obtain preprocessed low-field MRI images, input the preprocessed low-field MRI images into the low-field encoder, and extract cross-modal shared latent variables. Image enhancement module 52 is used to input the cross-modal shared latent variable into the high-field decoder for decoding, generate an initial high-field baseline image, input the initial high-field baseline image into the structure-aware diffusion enhancement module for structure recovery, obtain a structure-enhanced image, and input it into the detail diffusion model; Image correction module 53 is used to introduce latent space data consistency proximal optimization module, correct the current reconstruction result according to the low field MRI image to obtain a corrected image, and input the corrected image into detail diffusion model; The detail enhancement module 54 is used to perform detail restoration processing based on the structure enhancement image and the correction image according to the detail diffusion model, to restore the local texture and high-frequency structure in the high-field MRI, to obtain the detail enhancement image, and input it into the residual diffusion model; The image reconstruction module 55 is used to learn the detail-enhanced image based on the residual diffusion model and compensate for the remaining high-frequency residual information, and output the final high-field MRI reconstruction result.

[0089] Furthermore, such as Figure 4 As shown, based on the above-mentioned MRI image reconstruction method and system based on the shared latent space diffusion model and physical consistency constraints, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 4 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0090] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores an MRI image reconstruction program 40 based on a shared latent space diffusion model and physical consistency constraints. This MRI image reconstruction program 40 can be executed by the processor 10 to implement the MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints described in this application.

[0091] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the MRI image reconstruction method based on the shared latent space diffusion model and physical consistency constraints.

[0092] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The terminal's processor 10, memory 20, and display 30 communicate with each other via a system bus.

[0093] In one embodiment, when the processor 10 executes the MRI image reconstruction program 40 based on the shared latent space diffusion model and physical consistency constraints in the memory 20, it implements the steps of the MRI image reconstruction method based on the shared latent space diffusion model and physical consistency constraints as described above.

[0094] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an MRI image reconstruction program based on a shared latent space diffusion model and physical consistency constraints, wherein when the MRI image reconstruction program based on a shared latent space diffusion model and physical consistency constraints is executed by a processor, it implements the steps of the MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints as described above.

[0095] In summary, this invention provides an MRI image reconstruction method, system, terminal, and computer-readable storage medium based on a shared latent space diffusion model and physical consistency constraints. The method includes: acquiring a low-field MRI image; preprocessing the low-field MRI image to obtain a preprocessed low-field MRI image; inputting the preprocessed low-field MRI image into a low-field encoder to extract cross-modal shared latent variables; inputting the cross-modal shared latent variables into a high-field decoder for decoding to generate an initial high-field baseline image; inputting the initial high-field baseline image into a structure-aware diffusion enhancement module for structure recovery to obtain a structure-enhanced image, and inputting it into a detail diffusion model; introducing a latent space data consistency proximal optimization module to correct the current reconstruction result based on the low-field MRI image to obtain a corrected image, and inputting the corrected image into the detail diffusion model; performing detail recovery processing based on the detail diffusion model according to the structure-enhanced image and the corrected image to recover local textures and high-frequency structures in the high-field MRI to obtain a detail-enhanced image, and inputting it into a residual diffusion model; learning the detail-enhanced image based on the residual diffusion model and compensating for remaining high-frequency residual information, and outputting the final high-field MRI reconstruction result. This invention first uses a shared latent variable encoder to map low-field MRI to a shared latent space, generates an initial high-field image through a high-field decoder, then uses a structure diffusion model to recover high-field structure information, and then uses a high-frequency residual diffusion model to recover edge and texture details. After that, k-space data consistency optimization is performed through an MRI physical forward model, and the generated results are iteratively corrected by a manifold geometry gating mechanism, thereby obtaining the final high-field MRI reconstruction result.

[0096] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0097] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0098] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. An MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints, characterized in that, The MRI image reconstruction method based on the shared latent space diffusion model and physical consistency constraints includes: Acquire low-field MRI images, preprocess the low-field MRI images to obtain preprocessed low-field MRI images, input the preprocessed low-field MRI images into a low-field encoder, and extract cross-modal shared latent variables. The cross-modal shared latent variables are input into the high-field decoder for decoding to generate an initial high-field baseline image. The initial high-field baseline image is then input into the structure-aware diffusion enhancement module for structure recovery to obtain a structure-enhanced image, which is then input into the detail diffusion model. A latent space data consistency proximal optimization module is introduced to correct the current reconstruction result based on the low-field MRI image, obtain a corrected image, and input the corrected image into the detail diffusion model; Based on the detail diffusion model, detail restoration processing is performed on the structure-enhanced image and the corrected image to restore the local texture and high-frequency structure in high-field MRI, thereby obtaining a detail-enhanced image, which is then input into the residual diffusion model. The detailed enhancement image is learned based on the residual diffusion model, and the remaining high-frequency residual information is compensated to output the final high-field MRI reconstruction result.

2. The MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints according to claim 1, characterized in that, The process of acquiring low-field MRI images, preprocessing the low-field MRI images to obtain preprocessed low-field MRI images, and inputting the preprocessed low-field MRI images into a low-field encoder to extract cross-modal shared latent variables specifically includes: A low-field MRI image is acquired, and the low-field MRI image is preprocessed to obtain a preprocessed low-field MRI image. The preprocessing includes intensity normalization, center cropping, size unification, Fourier transform standardization, multi-coil reconstruction, and noise standardization. The preprocessed low-field MRI image is input into a low-field encoder for encoding to obtain shared latent variables: ； in, To share latent variables across modalities, The image is a preprocessed low-field MRI image. It is a low-field encoder.

3. The MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints according to claim 2, characterized in that, The process of inputting the cross-modal shared latent variables into a high-field decoder for decoding to generate an initial high-field baseline image, inputting the initial high-field baseline image into a structure-aware diffusion enhancement module for structure recovery to obtain a structure-enhanced image, and then inputting it into a detail diffusion model, specifically includes: Sharing latent variables across modalities The input is fed into the high-field decoder for decoding, generating the initial high-field baseline image: ； in, For the initial high-field baseline image, For high-field decoders; Initial high-field baseline image The image is input into the structure-aware diffusion enhancement module for structure reconstruction processing, resulting in an enhanced structure image. ; Enhanced image Input into the detail diffusion model.

4. The MRI image reconstruction method based on the shared latent space diffusion model and physical consistency constraints according to claim 3, characterized in that, The structure-aware diffusion enhancement module adopts a conditional diffusion model, and the forward diffusion process is as follows: ； in, Represents a true high-field MRI image. Indicates the first The noisy image corresponding to each diffusion time step This represents the conditional probability distribution during the forward diffusion process. Indicates a Gaussian distribution. Indicates the first Signal retention coefficients for each diffusion time step Indicates the first The noise intensity at each diffusion time step Represents the identity matrix; The reverse denoising process is as follows: ； in, Indicates the first The noisy image corresponding to each diffusion time step Indicates the current noisy image and initial high-field baseline image As a condition, through learnable parameters The denoising network predicts the noisy image from the previous time step. The conditional probability distribution.

5. The MRI image reconstruction method based on the shared latent space diffusion model and physical consistency constraints according to claim 3, characterized in that, The introduced latent space data consistency proximal optimization module corrects the current reconstruction result based on the low-field MRI image to obtain a corrected image, and then inputs the corrected image into the detail diffusion model, specifically including: Introducing k-space data consistency constraints, the latent space data consistency proximal optimization module adopts the MRI forward model, which is represented as follows: ； in, For sampling mask, For Fourier transform, For coil sensitivity, MRI images in the image domain; To observe k-space data; Data consistency loss for: ； in: To generate the result; The current reconstruction result is corrected based on the low-field MRI image to obtain a corrected image, which is then input into the detail diffusion model.

6. The MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints according to claim 5, characterized in that, The MRI image reconstruction method based on the shared latent space diffusion model and physical consistency constraints also includes: A manifold geometry-gated optimization mechanism is introduced, which decomposes the data consistency correction direction into the manifold tangential direction and the manifold normal direction. The expression of the manifold geometry-gated optimization mechanism is as follows: ； in, Indicates the first The current high-field clean image estimate obtained from the diffusion model prediction in the next iteration. This represents the high-field image estimation after manifold geometry gating optimization. This indicates that the data consistency correction direction is along the tangential component of the image manifold. This indicates the normal component of the data consistency correction direction along the image manifold. Indicates the first The adaptive gating coefficients of the next iteration are used to control the intensity of the introduction of the normal component; The gating coefficient is dynamically calculated based on measurement consistency and structural similarity. ； in, Indicates the first Normalized measurement mismatch at the next iteration The representation is defined as follows: This represents the magnetic resonance forward measurement operator, used to simulate the acquisition process from image to observation data. This represents a high-field encoder used to estimate the current high-field clean image. Mapped to the high-field latent space This refers to a target domain decoder or measurement consistency projection decoder, used to map latent space features to an image domain consistent with the observed data. This represents the smallest positive integer that prevents the denominator from being zero. ； in, express and cosine similarity, Indicates the first The overall correction direction obtained from data consistency in the next iteration This represents the inner product operation. Represents the L2 norm; ； in, This represents the temperature coefficient of the measurement mismatch, used to control the strength of the suppression of data consistency errors on the gating coefficient. This represents the geometric similarity between the correction direction and the tangent direction of the manifold. Represents the geometric similarity threshold. The temperature coefficient representing the geometric similarity term is used to control the gating function. The smoothness of the change This represents the Sigmoid function.

7. The MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints according to claim 5, characterized in that, The process of learning and compensating for residual high-frequency residual information in the detail-enhanced image based on the residual diffusion model, and outputting the final high-field MRI reconstruction result, specifically includes: Constructing high-frequency residuals using high-field images: ； in, Indicates high-frequency residuals, Indicates high-pass filtering; A true high-field image; Enhance the image for structure; Learning high-frequency residual distributions using a residual diffusion model: ； in, Indicates the high-frequency residual distribution, It is Gaussian noise; Image detail enhancement based on a residual diffusion model learned from high-frequency residual distribution. The remaining high-frequency residual information is learned and compensated, and the final high-field MRI reconstruction result is output. .

8. An MRI image reconstruction system based on a shared latent space diffusion model and physical consistency constraints, characterized in that, The MRI image reconstruction system based on the shared latent space diffusion model and physical consistency constraints is used to implement the MRI image reconstruction method based on the shared latent space diffusion model and physical consistency constraints as described in any one of claims 1-7. The MRI image reconstruction system based on the shared latent space diffusion model and physical consistency constraints comprises: The latent variable extraction module is used to acquire low-field MRI images, preprocess the low-field MRI images to obtain preprocessed low-field MRI images, input the preprocessed low-field MRI images into the low-field encoder, and extract cross-modal shared latent variables. The image enhancement module is used to input the cross-modal shared latent variables into the high-field decoder for decoding, generate an initial high-field baseline image, input the initial high-field baseline image into the structure-aware diffusion enhancement module for structure recovery, obtain a structure-enhanced image, and input it into the detail diffusion model; The image correction module is used to introduce the latent space data consistency proximal optimization module, correct the current reconstruction result according to the low field MRI image, obtain the corrected image, and input the corrected image into the detail diffusion model; The detail enhancement module is used to perform detail restoration processing based on the structure enhancement image and the correction image according to the detail diffusion model, to restore the local texture and high-frequency structure in high-field MRI, obtain the detail enhancement image, and input it into the residual diffusion model; The image reconstruction module is used to learn the detail-enhanced image based on the residual diffusion model and compensate for the remaining high-frequency residual information, and output the final high-field MRI reconstruction result.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and an MRI image reconstruction program based on a shared latent space diffusion model and physical consistency constraints, stored in the memory and executable on the processor. When the MRI image reconstruction program based on a shared latent space diffusion model and physical consistency constraints is executed by the processor, it implements the steps of the MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an MRI image reconstruction program based on a shared latent space diffusion model and physical consistency constraints, which, when executed by a processor, implements the steps of the MRI image reconstruction method based on a shared latent space diffusion model and physical consistency constraints as described in any one of claims 1-7.

Images