Magnetic resonance image reconstruction and model training method and device

Abstract

The application discloses a magnetic resonance image reconstruction method and device and a model training method thereof. The model training method comprises the following steps: traversing sample data, the sample data comprising paired high-resolution three-dimensional structural images, thick layer target images under low-field and low-resolution, and corresponding high-resolution label images; generating a first feature map corresponding to the high-resolution three-dimensional structural image and a second feature map corresponding to the thick layer target image based on a preset feature coding network; obtaining a guided feature aligned with the spatial deformation of the thick layer target image through the first feature map and the second feature map; dynamically activating a target physical constraint based on the data type of the thick layer target image; performing image reconstruction according to the physical constraint and the guided feature to obtain a sample reconstructed image; and optimizing a multi-dimensional joint loss function by using the high-resolution label image and the sample reconstructed image. According to the embodiment of the application, the clinical performance of a low-field device in the detection of micro lesions and the tracking of nerve fibers is close to the level of a high-field device.

Description

Technical Field

[0001] This application relates to the field of smart medical technology, and in particular to a magnetic resonance image reconstruction and model training method and apparatus. Background Technology

[0002] In recent years, low-field magnetic resonance imaging (MRI) devices (such as 0.3T and 0.5T) have been rapidly developed and applied in primary healthcare, emergency medicine, and portable medical scenarios due to their advantages such as small footprint, low deployment cost, good compatibility with implants, and ease of mobility.

[0003] Low-field magnetic resonance imaging (MRI) devices are limited by their inherent physical field strength, resulting in extremely low signal baselines and severely inadequate signal-to-noise ratios (SNRs). Under this hardware constraint, conventional high-resolution three-dimensional structural imaging (such as T1-weighted sequences) can still achieve higher resolution images by extending acquisition time or multiple averaging. However, for sequences highly dependent on SNR and employing echo-planar imaging (EPI), such as diffusion-weighted imaging (DWI) and diffusion tensor imaging (DTI), extremely thick slices (e.g., 6mm to 10mm) are necessary to barely guarantee image usability. This leads to extremely low spatial resolution (Z-axis) in low-resolution, thick-slice images from low-field MRI devices, exhibiting severe partial volumetric effects and significantly limiting their clinical application value in detecting small infarcts, tracking nerve fibers, and analyzing brain networks.

[0004] In related technologies, high-resolution T1 images and low-resolution DWI images are directly stitched together in the channel dimension, and then input into a standard convolutional neural network to guide DWI image interpolation.

[0005] However, firstly, existing solutions fail to address the cross-modal misalignment problem caused by magnetic susceptibility distortion. Although the difference in tissue magnetic susceptibility is reduced at low field strengths, low-field devices typically have poor hardware gradient performance, and geometric distortion is still unavoidable in EPI sequences. Direct feature fusion can lead to the erroneous injection of T1 structural information into non-corresponding regions of the target image, resulting in false anatomical textures. Secondly, existing solutions severely compromise the physical positive definiteness and directionality of low-field diffusion tensor data. Low-field DTI data itself has extremely poor signal-to-noise ratio. Traditional single-channel scalar interpolation not only destroys the positive semidefiniteness of the tensor matrix but also easily causes the principal diffusion direction to deviate severely at low signal-to-noise ratios, rendering neural fiber tracking completely ineffective in low-field environments. Summary of the Invention

[0006] This application provides a method and apparatus for magnetic resonance image reconstruction and model training. To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments. Its sole purpose is to present some concepts in a simple form as a prelude to the detailed description that follows.

[0007] In a first aspect, embodiments of this application provide a method for training a magnetic resonance image reconstruction model, the method comprising: Traverse the sample data in the model training samples, which include paired high-resolution 3D structural images, thick-layer target images under low field and low resolution and their corresponding high-resolution label images. Based on a preset feature encoding network, a first feature map corresponding to a high-resolution 3D structural image and a second feature map corresponding to a thick-layer target image are generated. The first feature map and the second feature map are concatenated along the channel dimension and then input into a preset offset network to predict the three-dimensional spatial offset field. Based on the three-dimensional spatial offset field, the first feature map is resampled to obtain guiding features that are aligned with the spatial deformation of the thick target image. Based on the data type of thick target images, the physical constraints of the target are dynamically activated; image reconstruction is performed according to the physical constraints and guiding features to obtain the sample reconstructed image; A multi-dimensional joint loss function was optimized using high-resolution labeled images and sample reconstructed images to obtain a pre-trained magnetic resonance image reconstruction model.

[0008] Optionally, based on a preset feature encoding network, a first feature map corresponding to the high-resolution 3D structural image and a second feature map corresponding to the thick-layer target image are generated, including: The thick target image is upsampled to adjust its spatial dimension to the same dimension as the high-resolution three-dimensional structure image, thus obtaining the initial upsampled image. Load a preset feature encoding network, which includes a first feature encoding network and a second feature encoding network that are independent of each other. The network weights of the first feature encoding network and the second feature encoding network are not shared. The preset feature encoding network adopts a three-dimensional convolutional neural network architecture with residual connections and the number of channels of the output feature map is set to a preset value. Input the high-resolution 3D structural image into the first feature encoding network, and output the first feature map corresponding to the high-resolution 3D structural image; The initial upsampled image is input into the second feature encoding network, which outputs the second feature map corresponding to the thick target image.

[0009] Optionally, based on the data type of the thick target image, dynamically activate the target physical constraints, including: When the data type of the thick target image is a diffusion-weighted image, activate the edge-adaptive gated interpolation constraint. The edge-adaptive gated interpolation constraint is a feature fusion control mechanism based on local edge information of the image; or... When the data type of the thick target image is a diffuse tensor image, the physical positive definite constraint based on Joleski decomposition is activated. The physical positive definite constraint is a constraint mechanism that uses the mathematical properties of Joleski decomposition to ensure the physical validity of the diffuse tensor matrix.

[0010] Optionally, image reconstruction is performed based on physical constraints and guiding features to obtain sample reconstructed images, including: When the physical constraint is edge-adaptive gated interpolation constraint, high-frequency edge features are extracted from the guiding features to generate a spatial attention-gated map. By performing element-wise multiplication, the guiding features are multiplied with the spatial attention gating map to preserve the structural features of the high-resolution 3D structural image in the image edge region, thus obtaining the gating weighted guiding features. The gated weighted guiding features are added element-wise to the second feature map to obtain the fused feature map; The fused feature map is input into a preset decoder to restore spatial resolution and generate a final pixel-level high-resolution diffusion-weighted image, which is used as a sample reconstruction image.

[0011] Optionally, image reconstruction is performed based on physical constraints and guiding features to obtain sample reconstructed images, including: When the physical constraint is a positive definite physical constraint based on Joleski decomposition, the six elements of the lower triangular matrix are predicted based on the guiding feature and the second feature map, and the Softplus activation function is applied to the diagonal elements of the lower triangular matrix to ensure positive values, thus obtaining the target lower triangular matrix. The target lower triangular matrix is ​​reconstructed into a symmetric positive semidefinite tensor by differentiable matrix multiplication, resulting in a high-resolution diffuse tensor image, which is then used as a sample reconstruction image.

[0012] Optionally, a multi-dimensional joint loss function is optimized using high-resolution labeled images and sample reconstructed images to obtain a pre-trained magnetic resonance image reconstruction model, including: Calculate the basic reconstruction loss between the sample reconstructed image and the high-resolution labeled image; Calculate the perceptual loss of the reconstructed sample image and the high-resolution labeled image in the feature space of the pre-trained convolutional neural network; The mean square error between the anisotropy scores of the reconstructed sample image and the high-resolution label image is calculated and used as the tensor eigenvalue loss. The basic reconstruction loss, perception loss, and tensor eigenvalue loss are summed to obtain the total loss value for the current iteration. This total loss value is the result of the calculation of the preset joint loss function in the current iteration. The network parameters are updated by backpropagation based on the overall loss value until convergence is obtained to obtain the pre-trained magnetic resonance image reconstruction model.

[0013] Optionally, the network parameters are updated via backpropagation based on the overall loss value until convergence is obtained to obtain a pre-trained magnetic resonance image reconstruction model, including: If the overall loss value has not reached its minimum, continue to iterate through the sample data in the model training samples until the overall loss value reaches its minimum, and obtain the pre-trained magnetic resonance image reconstruction model; or, if the overall loss value reaches its minimum, stop iterating and generate the pre-trained magnetic resonance image reconstruction model.

[0014] Secondly, embodiments of this application provide a magnetic resonance image reconstruction method, the method comprising: Receive a magnetic resonance image reconstruction request, which carries the current high-resolution three-dimensional structural image of the target object and the current low-resolution thick-layer target image. Bicubic interpolation upsampling is performed on the current low-resolution thick-layer target image to make the spatial size of the current low-resolution thick-layer target image consistent with the current high-resolution three-dimensional structure image, thus obtaining the current initial upsampled image; The current high-resolution 3D structural image and the current initial upsampled image are input into the pre-trained magnetic resonance image reconstruction model; the pre-trained magnetic resonance image reconstruction model is trained using the above-mentioned magnetic resonance image reconstruction model training method. Output the current high-resolution reconstructed image of the target object. The current high-resolution reconstructed image is either the current pixel-level high-resolution diffusion-weighted image or the current high-resolution diffusion tensor image.

[0015] Thirdly, embodiments of this application provide a magnetic resonance image reconstruction model training device, the device comprising: The sample data traversal module is used to traverse the sample data in the model training samples. The sample data includes paired high-resolution 3D structural images, thick-layer target images under low field and low resolution and their corresponding high-resolution label images. The feature map generation module is used to generate a first feature map corresponding to a high-resolution three-dimensional structural image and a second feature map corresponding to a thick-layer target image based on a preset feature encoding network. The guided feature generation module is used to concatenate the first feature map and the second feature map along the channel dimension and input them into a preset offset network to predict the three-dimensional spatial offset field; based on the three-dimensional spatial offset field, the first feature map is resampled to obtain guided features aligned with the spatial deformation of the thick target image; The physical constraint dynamic routing module is used to dynamically activate the physical constraints of the target based on the data type of thick target images; and to reconstruct the image based on the physical constraints and guiding features to obtain the sample reconstructed image. The joint optimization module is used to perform multi-dimensional joint loss function optimization using high-resolution labeled images and sample reconstructed images to obtain a pre-trained magnetic resonance image reconstruction model.

[0016] Fourthly, embodiments of this application provide a magnetic resonance image reconstruction apparatus, the apparatus comprising: The request receiving module is used to receive magnetic resonance image reconstruction requests, which carry the current high-resolution three-dimensional structural image of the target object and the current low-resolution thick-layer target image. The preprocessing module is used to perform bicubic interpolation upsampling on the current low-resolution thick-layer target image so that the spatial size of the current low-resolution thick-layer target image is consistent with that of the current high-resolution three-dimensional structure image, thus obtaining the current initial upsampled image; The image reconstruction module is used to input the current high-resolution 3D structural image and the current initial upsampled image into the pre-trained magnetic resonance image reconstruction model; the pre-trained magnetic resonance image reconstruction model is trained using the above-mentioned magnetic resonance image reconstruction model training method. The resolution reconstruction image output module is used to output the current high-resolution reconstructed image of the target object. The current high-resolution reconstructed image is either the current pixel-level high-resolution diffusion-weighted image or the current high-resolution diffusion tensor image.

[0017] The technical solutions provided in this application embodiment may include the following beneficial effects: In this embodiment, on the one hand, by stitching the first feature map and the second feature map along the channel dimension and inputting them into a preset offset network, a three-dimensional spatial offset field is predicted. Based on this offset field, the first feature map is resampled, thereby achieving the acquisition of guiding features aligned with the spatial deformation of the thick target image. This mechanism can accurately compensate for the local geometric distortion of the EPI sequence at the feature level, fundamentally avoiding the misplacement of T1 structural information, thus eliminating the generation of false anatomical textures and improving the structural realism and edge sharpness of the reconstructed image. On the other hand, by dynamically activating the target physical constraints based on the data type of the thick target image, when the data type is a diffusion tensor image, the physical positive definite constraint based on Joleski decomposition is activated, mathematically ensuring that the output tensor always satisfies the positive semi-definite property. This mechanism strictly maintains the physical validity of the diffusion tensor during the reconstruction process, ensuring the accuracy of the main diffusion direction, thereby enabling effective neural fiber tracking in low field strength environments.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] Figure 1 This is a schematic flowchart of a magnetic resonance image reconstruction model training method provided in an embodiment of this application; Figure 2 This is a schematic flowchart of a magnetic resonance image reconstruction method provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a magnetic resonance image reconstruction model training device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a magnetic resonance image reconstruction device provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0021] The following description and accompanying drawings fully illustrate specific embodiments of this application to enable those skilled in the art to practice them.

[0022] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0023] In the following description, when referring to the accompanying drawings, the same numbers in different drawings denote the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0024] In the description of this application, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances. Furthermore, in the description of this application, unless otherwise stated, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship.

[0025] This application provides a magnetic resonance image reconstruction method and apparatus for model training to address the problems existing in the aforementioned related technologies. In the embodiments of this application, on the one hand, by stitching a first feature map and a second feature map along the channel dimension and inputting them into a preset offset network, a three-dimensional spatial offset field is predicted. Based on this offset field, the first feature map is resampled, achieving the acquisition of guiding features aligned with the spatial deformation of the thick target image. This mechanism can accurately compensate for local geometric distortions of the EPI sequence at the feature level, fundamentally avoiding the misplacement of T1 structural information, thereby eliminating the generation of false anatomical textures and improving the structural realism and edge sharpness of the reconstructed image. On the other hand, by dynamically activating the target physical constraints based on the data type of the thick target image, when the data type is a diffusion tensor image, a positive definite physical constraint based on Joleskiel decomposition is activated, mathematically ensuring that the output tensor always satisfies positive semi-definiteness. This mechanism strictly maintains the physical validity of the diffusion tensor during reconstruction, ensuring the accuracy of the main diffusion direction, thereby enabling effective neural fiber tracking in low-field environments. The following describes this in detail using exemplary embodiments.

[0026] This application provides a method and apparatus for magnetic resonance image reconstruction and model training to solve the problems existing in the aforementioned related technologies. The following will be discussed in conjunction with the appendix... Figure 1 - Appendix Figure 2 This application provides a detailed description of the magnetic resonance image reconstruction model training method provided in its embodiments. This method can be implemented using a computer program and can run on a von Neumann-based magnetic resonance image reconstruction model training device. This computer program can be integrated into applications or run as a standalone utility application.

[0027] Please see Figure 1 This is a flowchart illustrating a magnetic resonance image reconstruction model training method provided in this application embodiment. Figure 1 As shown, the method in this application embodiment may include the following steps: S101, Traverse the sample data in the model training samples. The sample data includes paired high-resolution 3D structural images, thick-layer target images under low field and low resolution and their corresponding high-resolution label images. High-resolution 3D structural images are isotropic, high spatial resolution anatomical images acquired using T1-weighted 3D magnetic resonance imaging (MRI) sequences. Low-field, low-resolution thick-slice target images are low spatial resolution images with slice thicknesses of 6mm to 10mm acquired using diffusion-weighted imaging or diffusion tensor imaging sequences on low-field MRI equipment (e.g., 0.3T or 0.5T), serving as the raw target data for reconstruction. High-resolution labeled images are high spatial resolution diffusion-weighted or diffusion tensor images of the same patient and scanned area as the thick-slice target images.

[0028] In one possible implementation, a clinical dataset collected by a 0.3T low-field magnetic resonance imaging (MRI) device in a hospital is selected as the source of training samples.

[0029] For example, for each patient, historical image data from the following three types of historical scans were acquired as a paired sample: High-resolution 3D T1 structural images were acquired using a 3D magnetized preparation fast gradient echo sequence (MPRAGE), with scanning parameters including voxel size 1mm×1mm×1mm, repetition time TR=2300ms, echo time TE=3.5ms, and flip angle 9°. Low-resolution thick-slice diffusion-weighted images were acquired using a single-shot excitation spin echo planar echo imaging sequence, with scanning parameters including slice thickness 8mm, slice spacing 0mm, voxel size 2mm×2mm×8mm, b-values ​​of 0s / mm² and 1000s / mm², and diffusion gradient direction number 30. Diffusion-weighted images with higher resolution parameters could be acquired using high-field magnetic resonance imaging (MRI) or other methods as high-resolution labeled images.

[0030] S102, Based on the preset feature coding network, generate the first feature map corresponding to the high-resolution three-dimensional structural image and the second feature map corresponding to the thick-layer target image; In some embodiments of this application, the specific process of generating a first feature map corresponding to a high-resolution three-dimensional structural image and a second feature map corresponding to a thick-layer target image based on a preset feature encoding network includes: upsampling the thick-layer target image to adjust its spatial dimension to the same dimension as the high-resolution three-dimensional structural image, thereby obtaining an initial upsampled image; loading a preset feature encoding network, which includes a first feature encoding network and a second feature encoding network that are independent of each other, wherein the network weights of the first feature encoding network and the second feature encoding network are not shared, and the preset feature encoding network adopts a three-dimensional convolutional neural network architecture with residual connections and the number of output feature map channels is set to a preset value; inputting the high-resolution three-dimensional structural image into the first feature encoding network and outputting the first feature map corresponding to the high-resolution three-dimensional structural image; and inputting the initial upsampled image into the second feature encoding network and outputting the second feature map corresponding to the thick-layer target image.

[0031] Upsampling refers to interpolating and magnifying a low-resolution, thick-layer target image along the slice direction (Z-axis) (usually using bicubic interpolation) to make its voxel count, layer count, and spatial dimensions completely consistent with the high-resolution 3D structural image, thus obtaining an initial upsampled image with matching spatial dimensions but blurred details. The preset feature encoding network is a dual-branch encoder structure composed of two sets of 3D convolutional neural networks with non-shared weights, used to extract deep features from different modalities of the image. The 3D convolutional neural network architecture is constructed by stacking 3D residual blocks in the encoder. Each residual block contains at least two 3D convolutional layers and one identity shortcut connection to alleviate the gradient vanishing problem in deep networks and enhance the extraction capability of multi-scale anatomical and diffuse features.

[0032] It's important to note that not sharing network weights allows the first and second feature encoding networks to each learn the features best suited to their respective modalities. Residual blocks in residual connections make the network easier to train, ensuring that the extracted features contain rich original information. Setting the number of feature map channels to a preset value maps feature maps from two different sources to the same number of channels, ensuring they are computed within the same feature space.

[0033] In one possible implementation, a high-resolution three-dimensional structural image can be denoted as: A thick-layer target image can be denoted as The bicubic interpolation method can be used to interpret thick target images as follows: Upsampling is performed to obtain an initial upsampled image, denoted as . Load the preset feature encoding network (Encoder) to extract high-resolution 3D structural images. Initial upsampled image The feature map, the expression of the encoder network is:

[0034]

[0035] in, It is the first feature map corresponding to the high-resolution three-dimensional structural image. It is the second feature map corresponding to the thick target image. It is the first feature coding network. It is the second feature encoding network.

[0036] S103, the first feature map and the second feature map are spliced ​​along the channel dimension and then input into a preset offset network to predict the three-dimensional spatial offset field; based on the three-dimensional spatial offset field, the first feature map is resampled to obtain the guiding feature aligned with the spatial deformation of the thick target image; The pre-set offset network refers to a neural network sub-module that is pre-built before the model training in this application. It consists of three-dimensional convolutional layers with learnable parameters and is used to optimize its parameters through backpropagation during training in order to learn from the spliced ​​features and output the three-dimensional spatial offset field. The architecture parameters of the network are pre-designed and selected, but its convolutional kernel weights are not fixed and are updated with training iterations.

[0037] In one specific implementation, the preset offset network can adopt the following structure: Input layer: Receives the feature map after channel concatenation (number of channels is...) First 3D convolutional layer: kernel size is The number of output channels is ( The number of sampling points for deformable convolution, such as correspond The activation function is LeakyReLU; the normalization layer is either Instance Norm or Batch Norm; the output layer outputs a 3D offset field with the same spatial resolution as the input through linear convolution. Its channel dimension corresponds to The offsets in three directions. The network parameters are initialized orthogonally or zero, and are included in the joint loss function. End-to-end training is conducted under supervision.

[0038] It should be noted that EPI sequences from low-field MRI are susceptible to magnetic field inhomogeneities, leading to image stretching or distortion (especially at the skull base). Using rigid transformations (such as affine transformations) for alignment results in inadequate fit. This application utilizes deformable convolution to learn an offset field, dynamically applying structural features from high-resolution 3D structural images to the correct locations on distorted low-field DWI images. Deformable convolution allows each pixel to adaptively find its correct position, making it suitable for handling localized deformations.

[0039] In one possible implementation, the first feature map Second feature map The concatenation is performed along the channel dimension and then fed into an offset network. The default offset network contains at least one convolutional kernel of size [size missing]. Convolutional layers (i.e.) The network predicts the three-dimensional spatial migration field. The mathematical expression is:

[0040] in, It is a channel-level splicing operation. It is an offset network.

[0041] Secondly, using deformable convolution mechanisms, based on the predicted three-dimensional spatial offset field... The first feature map is resampled to obtain guiding features aligned with the spatial deformation of the thick target image. For any spatial location on the first feature map The output is calculated as follows:

[0042] in, The number of sampling points for the convolution kernel. For the fixed sampling offset position of standard convolution, The prediction of the offset network, corresponding to the first Variable offset of each sampling point The attention weights for prediction (within a range of values) ), used to weight the sampled features, These are the weights of the deformable convolution.

[0043] S104, based on the data type of thick target image, dynamically activates the target physical constraints; performs image reconstruction according to physical constraints and guiding features to obtain the sample reconstructed image; In some embodiments of this application, the specific process of dynamically activating target physical constraints based on the data type of the thick target image includes: when the data type of the thick target image is a diffusion-weighted image, activating edge-adaptive gated interpolation constraints, which are feature fusion control mechanisms based on local edge information of the image; or, when the data type of the thick target image is a diffusion tensor image, activating physical positive definite constraints based on Joleski decomposition, which are constraint mechanisms that utilize the mathematical properties of Joleski decomposition to ensure the physical validity of the diffusion tensor matrix.

[0044] In some embodiments of this application, the specific process of image reconstruction based on physical constraints and guiding features to obtain sample reconstructed images includes: when the physical constraint is an edge adaptive gating interpolation constraint, extracting high-frequency edge features from the guiding features to generate a spatial attention gating map; multiplying the guiding features and the spatial attention gating map element-wise to retain the structural features of the high-resolution three-dimensional structure image in the image edge region, obtaining gated weighted guiding features; adding the gated weighted guiding features and the second feature map element-wise to obtain a fused feature map; and inputting the fused feature map into a preset decoder to restore the spatial resolution and generate the final pixel-level high-resolution diffusion-weighted image as the sample reconstructed image.

[0045] The pre-defined decoder refers to a neural network submodule pre-built before the model training in this application, consisting of a 3D transposed convolutional layer (or an upsampling layer superimposed with a 3D convolutional layer) with learnable parameters and a 3D residual connection block. It is used to optimize its parameters through backpropagation during training so as to restore the fused feature map or the physically constrained feature map to the target high-resolution spatial size and output a pixel-level or tensor-level image. The macroscopic architecture of the decoder (such as the Encoder-Decoder symmetric structure and residual skip connections) is pre-designed, but its convolutional kernel weights are not fixed, but are updated iteratively with the optimization of the joint loss function.

[0046] In one possible implementation, the edge-adaptive gated interpolation branch for DWI scalar data (diffusion-weighted image type data) is calculated as follows: Gating weight calculation: from guiding features Extract high-frequency edge features to generate a spatial attention gating map. The calculation formula is:

[0047] in, It's a convolution operation. It is an activation function.

[0048] Feature fusion and reconstruction: T1 structural features are selectively injected into the edge regions of DWI through element-wise multiplication, avoiding contamination of the lesion's internal signal. The reconstruction formula is defined as:

[0049] in, It is a pixel-level high-resolution diffusion-weighted image. This indicates element-wise multiplication. A preset decoder for DWI scalar data. It is the second feature map, that is, the feature map of DWI scalar data.

[0050] In some embodiments of this application, the specific process of image reconstruction based on physical constraints and guiding features to obtain sample reconstructed images includes: when the physical constraint is a positive definite physical constraint based on Joleski decomposition, predicting the six elements of the lower triangular matrix based on guiding features and the second feature map, and applying the Softplus activation function to the diagonal elements of the lower triangular matrix to ensure positive values, thereby obtaining the target lower triangular matrix; reconstructing the target lower triangular matrix into a symmetric positive semi-definite tensor through differentiable matrix multiplication to obtain a high-resolution diffusion tensor image, which serves as the sample reconstructed image.

[0051] In one possible implementation, a physically positive definite branch based on Cholesky decomposition is used for DTI tensor data (data of the diffuse tensor image type): Lower triangular matrix regression: targeting the second feature map , guiding features The lower triangular matrix is ​​predicted through the decoder network. The six elements. Specifically, considering that DTI voxels are composed of... Symmetric diffuse tensor matrix The representation (containing 6 independent elements) would violate the positive semidefiniteness if directly regressed on all 6 elements. Therefore, the network branches do not directly predict the tensor. Instead, it predicts a feature map containing 6 channels, which corresponds to a lower triangular matrix. The 6 elements: ; Differentiable matrix reconstruction and physical constraints: To ensure that the diagonal elements are greater than zero, the network output... Apply the Softplus activation function, i.e. Add a physical constraint layer at the end of the network and use matrix multiplication to recover the high-resolution tensor graph: The calculation process is fully differentiable, allows gradient backpropagation, and mathematically guarantees 100% positive semidefiniteness of the output tensor.

[0052] in, It is the raw logit value (inactive value) directly output by the network branch, corresponding to the lower triangular matrix. The Middle The linear prediction result of the positions of the diagonal elements, i.e. =1,2,3, respectively corresponding to . Based on the natural constant Exponential operations with base 0.1 map the original values ​​of the network output to the positive real number field. It is the natural logarithm function, which performs a smooth inverse transformation on positive expressions.

[0053] S105 uses high-resolution labeled images and sample reconstructed images to perform multi-dimensional joint loss function optimization to obtain a pre-trained magnetic resonance image reconstruction model.

[0054] In some embodiments of this application, the specific process of optimizing a pre-trained magnetic resonance image reconstruction model using a multi-dimensional joint loss function with high-resolution label images and sample reconstructed images includes: calculating the basic reconstruction loss between the sample reconstructed image and the high-resolution label image; calculating the perceptual loss between the sample reconstructed image and the high-resolution label image in the feature space of the pre-trained convolutional neural network; calculating the mean square error between the anisotropy scores of the sample reconstructed image and the high-resolution label image as the tensor eigenvalue loss; summing the basic reconstruction loss, perceptual loss, and tensor eigenvalue loss to obtain the overall loss value of the current iteration, which is the calculation result of the preset joint loss function in the current iteration; and updating the network parameters based on the overall loss value through backpropagation until convergence is obtained to obtain the pre-trained magnetic resonance image reconstruction model.

[0055] Specifically, the process of updating network parameters based on backpropagation of the overall loss value until convergence to obtain the pre-trained magnetic resonance image reconstruction model includes: if the overall loss value has not reached the minimum, continue to execute the step of traversing the sample data in the model training samples until the overall loss value reaches the minimum, and obtain the pre-trained magnetic resonance image reconstruction model; or, if the overall loss value reaches the minimum, stop iterating and generate the pre-trained magnetic resonance image reconstruction model.

[0056] In one possible implementation, high-resolution label images are used. As a monitoring benchmark, a composite total loss function is constructed and optimized. This guides the network to generate reconstructed images that combine high fidelity and physical plausibility. The total loss function is composed of a weighted sum of the following three sub-loss terms:

[0057] in, This is the total loss value. It is a loss from basic reconstruction. It is perceived loss. It is tensor eigenvalue loss. It is the first preset weighting coefficient, used to adjust the perceptual loss. The proportion of contribution to the joint total loss function. It is the second preset weighting coefficient, used to adjust the tensor eigenvalue loss. The proportion of contribution to the joint total loss function.

[0058] To replace traditional L1 or L2 losses and better preserve high-frequency edge details while suppressing artifacts, robust Charbonnier Loss is introduced. Its calculation formula is defined as follows:

[0059] in, This is a very small constant used to ensure numerical stability and to give the loss function better differential properties. It is a loss from basic reconstruction. It is a sample reconstructed image. It is the L2 norm (Euclidean norm), which is the magnitude of the error vector over all pixels of the image.

[0060] Among them, perceived loss The expression is:

[0061] in, This indicates the first [unit / item] in the pre-trained 3D VGG-19 network. The output of each feature extraction layer These represent the number of channels, height, width, and depth of the feature map for that layer, respectively.

[0062] This application preferably uses the average perceptual loss of three feature maps: conv3_1, conv4_1, and conv5_1.

[0063] Among them, tensor eigenvalue loss The expression is:

[0064] in, The total number of voxels, and The first The diffusion tensor matrix of the individual element is reconstructed from the output and the label. For anisotropic fractions, the formula is:

[0065] in These are the three eigenvalues ​​of the diffuse tensor matrix.

[0066] To improve the high-level semantic consistency of the reconstructed image, a pre-trained VGG19 deep convolutional neural network is used to extract the output image. With label image The high-dimensional feature map is obtained, and the difference between the two in the feature space is calculated to obtain the perceptual loss, which is used as a supervision constraint at the perceptual level.

[0067] Specifically, for the physical properties of tensor data (such as DTI), the mean square error (MSE) between the anisotropy scores (FA) of the output tensor and the true label tensor is calculated as the tensor eigenvalue loss, thereby forcibly constraining the accuracy of the network output at the physical feature level.

[0068] In this embodiment, on the one hand, by stitching the first feature map and the second feature map along the channel dimension and inputting them into a preset offset network, a three-dimensional spatial offset field is predicted. Based on this offset field, the first feature map is resampled, thereby achieving the acquisition of guiding features aligned with the spatial deformation of the thick target image. This mechanism can accurately compensate for the local geometric distortion of the EPI sequence at the feature level, fundamentally avoiding the misplacement of T1 structural information, thus eliminating the generation of false anatomical textures and improving the structural realism and edge sharpness of the reconstructed image. On the other hand, by dynamically activating the target physical constraints based on the data type of the thick target image, when the data type is a diffusion tensor image, the physical positive definite constraint based on Joleski decomposition is activated, mathematically ensuring that the output tensor always satisfies the positive semi-definite property. This mechanism strictly maintains the physical validity of the diffusion tensor during the reconstruction process, ensuring the accuracy of the main diffusion direction, thereby enabling effective neural fiber tracking in low field strength environments.

[0069] Please see Figure 2 This is a schematic flowchart illustrating a magnetic resonance image reconstruction method provided in this application. Figure 2 As shown, the method in this application embodiment may include the following steps: S201, Receive magnetic resonance image reconstruction request, the magnetic resonance image reconstruction request carries the current high-resolution three-dimensional structural image of the target object and the current low-resolution thick-layer target image; S202, Perform bicubic interpolation upsampling on the current low-resolution thick-layer target image to make the spatial size of the current low-resolution thick-layer target image consistent with the current high-resolution three-dimensional structure image, and obtain the current initial upsampled image; S203, input the current high-resolution 3D structural image and the current initial upsampled image into the pre-trained magnetic resonance image reconstruction model; the pre-trained magnetic resonance image reconstruction model is trained by the above-mentioned magnetic resonance image reconstruction model training method; The pre-trained magnetic resonance image reconstruction model includes a pre-trained feature encoding network, an offset network, and a dynamic routing decoder.

[0070] In some embodiments of this application, the current high-resolution three-dimensional structural image is... and the current initial upsampled image When the trained model is input, it performs the following operations in sequence: Dual-branch feature encoding: Two non-weight-sharing 3D residual convolutional encoders extract features separately. and Multiscale feature maps and ; Cross-modal elastic alignment: and The spliced ​​data is input into an offset network to predict the 3D spatial offset field. and through deformable convolution pairs Resampling is performed to obtain a guided feature map aligned with the spatial deformation of the target image. ; Dynamic routing decoding: Based on the data type (DWI or DTI) of the current low-resolution thick-layer target image, the corresponding physical constraint decoding branch is automatically activated to output a high-resolution reconstructed image.

[0071] S204, outputs the current high-resolution reconstructed image of the target object. The current high-resolution reconstructed image is either the current pixel-level high-resolution diffusion-weighted image or the current high-resolution diffusion tensor image.

[0072] In some embodiments of this application, high-resolution reconstructed images are output, and the high-resolution reconstructed images are saved in standard DICOM format and pushed to the PACS system for radiologists to read and diagnose.

[0073] In this embodiment, on the one hand, by stitching the first feature map and the second feature map along the channel dimension and inputting them into a preset offset network, a three-dimensional spatial offset field is predicted. Based on this offset field, the first feature map is resampled, thereby achieving the acquisition of guiding features aligned with the spatial deformation of the thick target image. This mechanism can accurately compensate for the local geometric distortion of the EPI sequence at the feature level, fundamentally avoiding the misplacement of T1 structural information, thus eliminating the generation of false anatomical textures and improving the structural realism and edge sharpness of the reconstructed image. On the other hand, by dynamically activating the target physical constraints based on the data type of the thick target image, when the data type is a diffusion tensor image, the physical positive definite constraint based on Joleski decomposition is activated, mathematically ensuring that the output tensor always satisfies the positive semi-definite property. This mechanism strictly maintains the physical validity of the diffusion tensor during the reconstruction process, ensuring the accuracy of the main diffusion direction, thereby enabling effective neural fiber tracking in low field strength environments.

[0074] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0075] Please see Figure 3 This illustration shows a schematic diagram of a magnetic resonance image reconstruction model training device provided in an exemplary embodiment of this application. The magnetic resonance image reconstruction model training device can be implemented as all or part of an electronic device through software, hardware, or a combination of both. The device 1 includes a sample data traversal module 10, a feature map generation module 20, a guided feature generation module 30, a physical constraint dynamic routing module 40, and a joint optimization module 50.

[0076] The sample data traversal module 10 is used to traverse the sample data in the model training samples. The sample data includes paired high-resolution three-dimensional structural images, thick-layer target images under low field and low resolution and their corresponding high-resolution label images. The feature map generation module 20 is used to generate a first feature map corresponding to a high-resolution three-dimensional structural image and a second feature map corresponding to a thick-layer target image based on a preset feature coding network. The guided feature generation module 30 is used to stitch the first feature map and the second feature map along the channel dimension and input them into a preset offset network to predict the three-dimensional spatial offset field; based on the three-dimensional spatial offset field, the first feature map is resampled to obtain guided features aligned with the spatial deformation of the thick target image; The physical constraint dynamic routing module 40 is used to dynamically activate the target physical constraints based on the data type of the thick target image; and to reconstruct the image according to the physical constraints and guiding features to obtain the sample reconstructed image. The joint optimization module 50 is used to perform multi-dimensional joint loss function optimization using high-resolution labeled images and sample reconstructed images to obtain a pre-trained magnetic resonance image reconstruction model.

[0077] It should be noted that the magnetic resonance image reconstruction model training device provided in the above embodiments is only illustrated by the division of the above functional modules when executing the magnetic resonance image reconstruction model training method. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the magnetic resonance image reconstruction model training device and the magnetic resonance image reconstruction model training method embodiments provided in the above embodiments belong to the same concept, and the implementation process is detailed in the method embodiments, which will not be repeated here.

[0078] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0079] In this embodiment, on the one hand, by stitching the first feature map and the second feature map along the channel dimension and inputting them into a preset offset network, a three-dimensional spatial offset field is predicted. Based on this offset field, the first feature map is resampled, thereby achieving the acquisition of guiding features aligned with the spatial deformation of the thick target image. This mechanism can accurately compensate for the local geometric distortion of the EPI sequence at the feature level, fundamentally avoiding the misplacement of T1 structural information, thus eliminating the generation of false anatomical textures and improving the structural realism and edge sharpness of the reconstructed image. On the other hand, by dynamically activating the target physical constraints based on the data type of the thick target image, when the data type is a diffusion tensor image, the physical positive definite constraint based on Joleski decomposition is activated, mathematically ensuring that the output tensor always satisfies the positive semi-definite property. This mechanism strictly maintains the physical validity of the diffusion tensor during the reconstruction process, ensuring the accuracy of the main diffusion direction, thereby enabling effective neural fiber tracking in low field strength environments.

[0080] Please see Figure 4 This illustration shows a schematic diagram of a magnetic resonance image reconstruction apparatus provided in an exemplary embodiment of this application. The magnetic resonance image reconstruction model training apparatus can be implemented as all or part of an electronic device through software, hardware, or a combination of both. The apparatus 2 includes a request receiving module 60, a preprocessing module 70, an image reconstruction module 80, and a resolution reconstructed image output module 90.

[0081] The request receiving module 60 is used to receive a magnetic resonance image reconstruction request, which carries the current high-resolution three-dimensional structural image of the target object and the current low-resolution thick-layer target image. Preprocessing module 70 is used to perform bicubic interpolation upsampling on the current low-resolution thick-layer target image so that the spatial size of the current low-resolution thick-layer target image is consistent with the current high-resolution three-dimensional structure image, and to obtain the current initial upsampled image; The image reconstruction module 80 is used to input the current high-resolution three-dimensional structural image and the current initial upsampled image into the pre-trained magnetic resonance image reconstruction model; the pre-trained magnetic resonance image reconstruction model is trained by the above-mentioned magnetic resonance image reconstruction model training method. The resolution reconstruction image output module 90 is used to output the current high-resolution reconstructed image of the target object. The current high-resolution reconstructed image is either the current pixel-level high-resolution diffusion-weighted image or the current high-resolution diffusion tensor image.

[0082] It should be noted that the magnetic resonance image reconstruction model training device provided in the above embodiments is only illustrated by the division of the above functional modules when executing the magnetic resonance image reconstruction model training method. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the magnetic resonance image reconstruction model training device and the magnetic resonance image reconstruction model training method embodiments provided in the above embodiments belong to the same concept, and the implementation process is detailed in the method embodiments, which will not be repeated here.

[0083] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0084] In this embodiment, on the one hand, by stitching the first feature map and the second feature map along the channel dimension and inputting them into a preset offset network, a three-dimensional spatial offset field is predicted. Based on this offset field, the first feature map is resampled, thereby achieving the acquisition of guiding features aligned with the spatial deformation of the thick target image. This mechanism can accurately compensate for the local geometric distortion of the EPI sequence at the feature level, fundamentally avoiding the misplacement of T1 structural information, thus eliminating the generation of false anatomical textures and improving the structural realism and edge sharpness of the reconstructed image. On the other hand, by dynamically activating the target physical constraints based on the data type of the thick target image, when the data type is a diffusion tensor image, the physical positive definite constraint based on Joleski decomposition is activated, mathematically ensuring that the output tensor always satisfies the positive semi-definite property. This mechanism strictly maintains the physical validity of the diffusion tensor during the reconstruction process, ensuring the accuracy of the main diffusion direction, thereby enabling effective neural fiber tracking in low field strength environments.

[0085] This application also provides a computer-readable medium having program instructions stored thereon, which, when executed by a processor, implement the magnetic resonance image reconstruction model training method provided in the above-described method embodiments.

[0086] This application also provides a computer program product containing instructions that, when run on a computer, cause the computer to execute the magnetic resonance image reconstruction model training method of the above-described method embodiments.

[0087] Please see Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 5 As shown, the electronic device 1000 may include: at least one processor 1001, at least one network interface 1004, a user interface 1003, a memory 1005, and at least one communication bus 1002.

[0088] The communication bus 1002 is used to realize the connection and communication between these components.

[0089] The user interface 1003 may include a display screen and a camera. Optionally, the user interface 1003 may also include a standard wired interface and a wireless interface.

[0090] The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0091] The processor 1001 may include one or more processing cores. The processor 1001 connects to various parts within the electronic device 1000 using various interfaces and lines. It executes various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 1005, and by calling data stored in the memory 1005. Optionally, the processor 1001 may be implemented using at least one hardware form selected from Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). The processor 1001 may integrate one or more of the following: a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), and a modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed on the screen; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip, without being integrated into the processor 1001.

[0092] The memory 1005 may include random access memory (RAM) or read-only memory. Optionally, the memory 1005 may include a non-transitory computer-readable storage medium. The memory 1005 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 1005 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 1005 may also be at least one storage system located remotely from the aforementioned processor 1001. Figure 5 As shown, the memory 1005, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a magnetic resonance image reconstruction model training application.

[0093] exist Figure 5 In the illustrated electronic device 1000, the user interface 1003 is mainly used to provide an input interface for the user and to acquire user input data; while the processor 1001 can be used to call the magnetic resonance image reconstruction model training application stored in the memory 1005, and specifically perform the following operations: Traverse the sample data in the model training samples, which include paired high-resolution 3D structural images, thick-layer target images under low field and low resolution and their corresponding high-resolution label images. Based on a preset feature encoding network, a first feature map corresponding to a high-resolution 3D structural image and a second feature map corresponding to a thick-layer target image are generated. The first feature map and the second feature map are concatenated along the channel dimension and then input into a preset offset network to predict the three-dimensional spatial offset field. Based on the three-dimensional spatial offset field, the first feature map is resampled to obtain guiding features that are aligned with the spatial deformation of the thick target image. Based on the data type of thick target images, the physical constraints of the target are dynamically activated; image reconstruction is performed according to the physical constraints and guiding features to obtain the sample reconstructed image; A multi-dimensional joint loss function was optimized using high-resolution labeled images and sample reconstructed images to obtain a pre-trained magnetic resonance image reconstruction model.

[0094] In one embodiment, when the processor 1001 generates a first feature map corresponding to a high-resolution three-dimensional structural image and a second feature map corresponding to a thick-layer target image based on a preset feature encoding network, it specifically performs the following operations: The thick target image is upsampled to adjust its spatial dimension to the same dimension as the high-resolution three-dimensional structure image, thus obtaining the initial upsampled image. Load a preset feature encoding network, which includes a first feature encoding network and a second feature encoding network that are independent of each other. The network weights of the first feature encoding network and the second feature encoding network are not shared. The preset feature encoding network adopts a three-dimensional convolutional neural network architecture with residual connections and the number of channels of the output feature map is set to a preset value. Input the high-resolution 3D structural image into the first feature encoding network, and output the first feature map corresponding to the high-resolution 3D structural image; The initial upsampled image is input into the second feature encoding network, which outputs the second feature map corresponding to the thick target image.

[0095] In one embodiment, when the processor 1001 executes data types based on thick target images and dynamically activates target physical constraints, it specifically performs the following operations: When the data type of the thick target image is a diffusion-weighted image, activate the edge-adaptive gated interpolation constraint. The edge-adaptive gated interpolation constraint is a feature fusion control mechanism based on local edge information of the image; or... When the data type of the thick target image is a diffuse tensor image, the physical positive definite constraint based on Joleski decomposition is activated. The physical positive definite constraint is a constraint mechanism that uses the mathematical properties of Joleski decomposition to ensure the physical validity of the diffuse tensor matrix.

[0096] In one embodiment, when processor 1001 performs image reconstruction based on physical constraints and guiding features to obtain a sample reconstructed image, it specifically performs the following operations: When the physical constraint is edge-adaptive gated interpolation constraint, high-frequency edge features are extracted from the guiding features to generate a spatial attention-gated map. By performing element-wise multiplication, the guiding features are multiplied with the spatial attention gating map to preserve the structural features of the high-resolution 3D structural image in the image edge region, thus obtaining the gating weighted guiding features. The gated weighted guiding features are added element-wise to the second feature map to obtain the fused feature map; The fused feature map is input into a preset decoder to restore spatial resolution and generate a final pixel-level high-resolution diffusion-weighted image, which is used as a sample reconstruction image.

[0097] In one embodiment, when processor 1001 performs image reconstruction based on physical constraints and guiding features to obtain a sample reconstructed image, it specifically performs the following operations: When the physical constraint is a positive definite physical constraint based on Joleski decomposition, the six elements of the lower triangular matrix are predicted based on the guiding feature and the second feature map, and the Softplus activation function is applied to the diagonal elements of the lower triangular matrix to ensure positive values, thus obtaining the target lower triangular matrix. The target lower triangular matrix is ​​reconstructed into a symmetric positive semidefinite tensor by differentiable matrix multiplication, resulting in a high-resolution diffuse tensor image, which is then used as a sample reconstruction image.

[0098] In one embodiment, when the processor 1001 performs multi-dimensional joint loss function optimization using high-resolution labeled images and sample reconstructed images to obtain a pre-trained magnetic resonance image reconstruction model, it specifically performs the following operations: Calculate the basic reconstruction loss between the sample reconstructed image and the high-resolution labeled image; Calculate the perceptual loss of the reconstructed sample image and the high-resolution labeled image in the feature space of the pre-trained convolutional neural network; The mean square error between the anisotropy scores of the reconstructed sample image and the high-resolution label image is calculated and used as the tensor eigenvalue loss. The basic reconstruction loss, perception loss, and tensor eigenvalue loss are summed to obtain the total loss value for the current iteration. This total loss value is the result of the calculation of the preset joint loss function in the current iteration. The network parameters are updated by backpropagation based on the overall loss value until convergence is obtained to obtain the pre-trained magnetic resonance image reconstruction model.

[0099] In one embodiment, when processor 1001 performs backpropagation based on the overall loss value to update network parameters until a pre-trained magnetic resonance image reconstruction model converges, it specifically performs the following operations: If the overall loss value has not reached its minimum, continue to iterate through the sample data in the model training samples until the overall loss value reaches its minimum, and obtain the pre-trained magnetic resonance image reconstruction model; or, if the overall loss value reaches its minimum, stop iterating and generate the pre-trained magnetic resonance image reconstruction model.

[0100] In this embodiment, on the one hand, by stitching the first feature map and the second feature map along the channel dimension and inputting them into a preset offset network, a three-dimensional spatial offset field is predicted. Based on this offset field, the first feature map is resampled, thereby achieving the acquisition of guiding features aligned with the spatial deformation of the thick target image. This mechanism can accurately compensate for the local geometric distortion of the EPI sequence at the feature level, fundamentally avoiding the misplacement of T1 structural information, thus eliminating the generation of false anatomical textures and improving the structural realism and edge sharpness of the reconstructed image. On the other hand, by dynamically activating the target physical constraints based on the data type of the thick target image, when the data type is a diffusion tensor image, the physical positive definite constraint based on Joleski decomposition is activated, mathematically ensuring that the output tensor always satisfies the positive semi-definite property. This mechanism strictly maintains the physical validity of the diffusion tensor during the reconstruction process, ensuring the accuracy of the main diffusion direction, thereby enabling effective neural fiber tracking in low field strength environments.

[0101] 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. The program for training the magnetic resonance image reconstruction model can be stored in a computer-readable storage medium. When executed, the program can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory, or random access memory, etc.

[0102] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A method for training a magnetic resonance image reconstruction model, characterized in that, The method includes: Traverse the sample data in the model training samples, the sample data including paired high-resolution three-dimensional structural images, thick-layer target images under low field and low resolution and their corresponding high-resolution label images; Based on a preset feature coding network, a first feature map corresponding to the high-resolution three-dimensional structural image and a second feature map corresponding to the thick-layer target image are generated. The first feature map and the second feature map are concatenated along the channel dimension and then input into a preset offset network to predict the three-dimensional spatial offset field; based on the three-dimensional spatial offset field, the first feature map is resampled to obtain guiding features aligned with the spatial deformation of the thick target image; Based on the data type of the thick target image, the target physical constraints are dynamically activated; image reconstruction is performed according to the physical constraints and the guiding features to obtain the sample reconstructed image; The high-resolution labeled image and the sample reconstructed image are used to perform multi-dimensional joint loss function optimization to obtain a pre-trained magnetic resonance image reconstruction model.

2. The method according to claim 1, characterized in that, The step of generating a first feature map corresponding to the high-resolution three-dimensional structural image and a second feature map corresponding to the thick-layer target image based on a preset feature coding network includes: The thick target image is upsampled to adjust its spatial dimension to the same dimension as the high-resolution three-dimensional structural image, thus obtaining an initial upsampled image. Load a preset feature encoding network, which includes a first feature encoding network and a second feature encoding network that are independent of each other. The network weights of the first feature encoding network and the second feature encoding network are not shared. The preset feature encoding network adopts a three-dimensional convolutional neural network architecture with residual connections and the number of channels of the output feature map is set to a preset value. The high-resolution three-dimensional structural image is input into the first feature encoding network, and the first feature map corresponding to the high-resolution three-dimensional structural image is output. The initial upsampled image is input into the second feature encoding network, which outputs the second feature map corresponding to the thick target image.

3. The method according to claim 1, characterized in that, The data type based on the thick target image, dynamically activating target physical constraints, includes: When the data type of the thick target image is a diffusion-weighted image, an edge-adaptive gated interpolation constraint is activated. This edge-adaptive gated interpolation constraint is a feature fusion control mechanism based on local edge information of the image; or... When the data type of the thick target image is a diffusion tensor image, a physical positive definite constraint based on Jolesky decomposition is activated. The physical positive definite constraint is a constraint mechanism that uses the mathematical properties of Jolesky decomposition to ensure the physical validity of the diffusion tensor matrix.

4. The method according to claim 1, characterized in that, The step of reconstructing the image based on the physical constraints and the guiding features to obtain the sample reconstructed image includes: When the physical constraint is an edge-adaptive gated interpolation constraint, high-frequency edge features are extracted from the guiding features to generate a spatial attention-gated map. By performing element-wise multiplication, the guiding features are multiplied with the spatial attention gating map to preserve the structural features of the high-resolution three-dimensional structural image in the image edge region, thus obtaining the gating weighted guiding features. The gated weighted guiding features are added element-by-element to the second feature map to obtain the fused feature map; The fused feature map is input into a preset decoder to restore spatial resolution and generate a final pixel-level high-resolution diffusion-weighted image, which is used as a sample reconstruction image.

5. The method according to claim 1, characterized in that, The step of reconstructing the image based on the physical constraints and the guiding features to obtain the sample reconstructed image includes: When the physical constraint is a positive definite physical constraint based on Joleski decomposition, based on the guiding feature and the second feature map, the six elements of the lower triangular matrix are predicted, and the Softplus activation function is applied to the diagonal elements of the lower triangular matrix to ensure positive values, thereby obtaining the target lower triangular matrix; The target lower triangular matrix is ​​reconstructed into a symmetric positive semidefinite tensor by differentiable matrix multiplication, resulting in a high-resolution diffuse tensor image, which is then used as a sample reconstruction image.

6. The method according to claim 1, characterized in that, The process of optimizing a pre-trained magnetic resonance image reconstruction model by using the high-resolution labeled image and the sample reconstructed image through a multi-dimensional joint loss function includes: Calculate the basic reconstruction loss between the reconstructed sample image and the high-resolution labeled image; Calculate the perceptual loss of the reconstructed sample image and the high-resolution label image in the feature space of the pre-trained convolutional neural network; The mean square error between the anisotropy scores of the reconstructed sample image and the high-resolution label image is calculated and used as the tensor eigenvalue loss. The basic reconstruction loss, the perception loss, and the tensor eigenvalue loss are summed to obtain the total loss value for the current iteration. This total loss value is the result of the calculation of the preset joint loss function in the current iteration. The network parameters are updated by backpropagation based on the overall loss value until convergence is obtained to obtain the pre-trained magnetic resonance image reconstruction model.

7. The method according to claim 6, characterized in that, The process of updating network parameters through backpropagation based on the overall loss value until convergence yields a pre-trained magnetic resonance image reconstruction model includes: If the overall loss value has not reached its minimum, continue to execute the step of traversing the sample data in the model training samples until the overall loss value reaches its minimum, and obtain the pre-trained magnetic resonance image reconstruction model; or, if the overall loss value reaches its minimum, stop the iteration and generate the pre-trained magnetic resonance image reconstruction model.

8. A magnetic resonance image reconstruction method, characterized in that, The method includes: Receive a magnetic resonance image reconstruction request, the magnetic resonance image reconstruction request carrying the current high-resolution three-dimensional structural image of the target object and the current low-resolution thick-layer target image; Bicubic interpolation upsampling is performed on the current low-resolution thick-layer target image to make the spatial size of the current low-resolution thick-layer target image consistent with that of the current high-resolution three-dimensional structure image, thus obtaining the current initial upsampled image; The current high-resolution three-dimensional structural image and the current initial upsampled image are input into a pre-trained magnetic resonance image reconstruction model; the pre-trained magnetic resonance image reconstruction model is obtained by training the magnetic resonance image reconstruction model training method according to any one of claims 1-7; Output the current high-resolution reconstructed image of the target object, wherein the current high-resolution reconstructed image is either the current pixel-level high-resolution diffusion-weighted image or the current high-resolution diffusion tensor image.

9. A magnetic resonance image reconstruction model training device, characterized in that, The device includes: The sample data traversal module is used to traverse the sample data in the model training samples. The sample data includes paired high-resolution three-dimensional structural images, thick-layer target images under low field and low resolution and their corresponding high-resolution label images. The feature map generation module is used to generate a first feature map corresponding to the high-resolution three-dimensional structural image and a second feature map corresponding to the thick-layer target image based on a preset feature coding network. The guided feature generation module is used to concatenate the first feature map and the second feature map along the channel dimension and input them into a preset offset network to predict the three-dimensional spatial offset field; and to resample the first feature map based on the three-dimensional spatial offset field to obtain guided features aligned with the spatial deformation of the thick target image. The physical constraint dynamic routing module is used to dynamically activate the target physical constraints based on the data type of the thick target image; and to perform image reconstruction according to the physical constraints and the guiding features to obtain the sample reconstructed image. The joint optimization module is used to perform multi-dimensional joint loss function optimization using the high-resolution labeled image and the sample reconstructed image to obtain a pre-trained magnetic resonance image reconstruction model.

10. A magnetic resonance image reconstruction device, characterized in that, The device includes: The request receiving module is used to receive a magnetic resonance image reconstruction request, which carries the current high-resolution three-dimensional structural image of the target object and the current low-resolution thick-layer target image. The preprocessing module is used to perform bicubic interpolation upsampling on the current low-resolution thick-layer target image so that the spatial size of the current low-resolution thick-layer target image is consistent with that of the current high-resolution three-dimensional structure image, thereby obtaining the current initial upsampled image; An image reconstruction module is used to input the current high-resolution three-dimensional structural image and the current initial upsampled image into a pre-trained magnetic resonance image reconstruction model; the pre-trained magnetic resonance image reconstruction model is obtained by training the magnetic resonance image reconstruction model training method according to any one of claims 1-7; The resolution reconstruction image output module is used to output the current high-resolution reconstructed image of the target object, wherein the current high-resolution reconstructed image is either the current pixel-level high-resolution diffusion-weighted image or the current high-resolution diffusion tensor image.

Images