Magnetic resonance electrical characteristic imaging model construction method and device, equipment and medium

Abstract

The invention discloses a magnetic resonance electrical characteristic imaging model construction method and device, equipment and a medium, and the method comprises the steps: obtaining radio frequency emission field data of a human head, extracting auxiliary feature information, and constructing three-dimensional feature data of a three-dimensional voxel structure according to the auxiliary feature information and the radio frequency emission field data; constructing a dual-path neural network model based on structure perception according to the convolutional neural network and the artificial neural network; and inputting the three-dimensional characteristic data into the dual-path neural network model for training to obtain the magnetic resonance electrical characteristic imaging model. According to the scheme, the problem of noise amplification in an existing algorithm electrical characteristic imaging technology can be solved, and the accuracy of electrical characteristic mapping is improved.

Description

Technical Field

[0001] This invention relates to the field of magnetic resonance imaging technology, specifically to a method, apparatus, device, and medium for constructing a magnetic resonance electrical property imaging model. Background Art

[0002] MR EPT (Magnetic Resonance Electrical Properties Tomography) is a non-invasive measurement technique that obtains the electrical properties of tissues, including conductivity and relative permittivity, through a magnetic resonance scanning system. These properties are influenced by factors such as ion concentration, protein content, and the ratio of bound water to free water within the tissue. When the physiological or pathological state of a tissue changes, its electrical properties also change. Malignant tumor tissues differ significantly from their corresponding normal tissues in terms of electrical properties.

[0003] To obtain the electrical properties of tissues, several magnetic resonance electrical property tomography reconstruction algorithms have been developed based on Maxwell's equations, such as EPT (H-EPT) based on the Helmholtz equation or EPT (cr-EPT) based on the convection response equation. However, existing algorithms rely on numerical approximations of the derivative of the radio frequency transmission field, which easily amplifies noise, thereby introducing errors and artifacts during reconstruction, resulting in low accuracy of electrical property mapping. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a method, apparatus, device, and medium for constructing a magnetic resonance electrical characteristic imaging model, which can overcome the noise amplification problem in existing algorithmic electrical characteristic imaging techniques and improve the accuracy of electrical characteristic mapping.

[0005] This invention provides a method for constructing a magnetic resonance electrical property imaging model, the method comprising:

[0006] Acquire radio frequency emission field data of the human head, extract auxiliary feature information, and construct three-dimensional feature data of a three-dimensional voxel structure based on the auxiliary feature information and the radio frequency emission field data;

[0007] A structure-aware dual-path neural network model was built based on convolutional neural networks and artificial neural networks.

[0008] The stereo feature data is input into the dual-path neural network model for training to obtain the magnetic resonance electrical property imaging model.

[0009] Preferably, the step of acquiring radio frequency emission field data of the human head, extracting auxiliary feature information, and constructing three-dimensional voxel structure stereo feature data based on the auxiliary feature information and the radio frequency emission field data includes:

[0010] The amplitude and phase data of the radio frequency transmission field generated by finite difference time-domain electromagnetic simulation are obtained as the radio frequency transmission field data;

[0011] The amplitude data and the phase data are combined to obtain complex emission field data;

[0012] Random Gaussian noise is added to the real and imaginary parts of the complex emission field data;

[0013] Extract auxiliary feature information from the noise-added re-emission field data;

[0014] The real and imaginary parts of the noise-added complex emission field data, along with the auxiliary feature information, are divided into three-dimensional voxel blocks of a preset size to obtain the stereo feature data.

[0015] The auxiliary feature information includes gradient components, Laplacian operator, and T1-weighted image.

[0016] Preferably, the extraction process of the T1-weighted image includes:

[0017] Different T1 signal intensity values ​​are assigned to different tissues of the noise-added retransmission field data to obtain T1 image intensity maps;

[0018] The T1 image intensity map is element-wise multiplied with the amplitude data, and the result is normalized.

[0019] The normalized image is filtered using a preset three-dimensional smooth convolution kernel to obtain the T1-weighted image.

[0020] Preferably, the step of dividing the real and imaginary parts of the noise-added complex emission field data, along with the auxiliary feature information, into three-dimensional voxel blocks of a preset size to obtain the stereo feature data includes:

[0021] The real and imaginary parts of the noise-added complex emission field data, as well as the auxiliary feature information, are divided into three-dimensional voxel blocks of a preset size;

[0022] Based on the divided three-dimensional voxel blocks, a smooth perceptual kernel with a regular cubic structure and a differential coding kernel with a three-dimensional cross-shaped structure are constructed as the stereo feature data.

[0023] Preferably, the dual-path neural network model includes:

[0024] A convolutional neural network for extracting features from the smooth perceptual kernel in the stereo feature data;

[0025] An artificial neural network for feature extraction from the differential coding kernel in the stereo feature data;

[0026] A concatenated layer that fuses the features extracted by the convolutional neural network and the artificial neural network.

[0027] Preferably, the step of inputting the stereo feature data into the dual-path neural network model for training to obtain the magnetic resonance electrical property imaging model includes:

[0028] The smoothing kernel in the stereo feature data is input into the convolution kernel of the convolutional neural network for processing to obtain convolutional features. The convolutional features are processed through an activation function, and the processed features are input into the flattening layer of the convolutional neural network to flatten them into a one-dimensional vector. The one-dimensional vector is then input into the fully connected layer of the convolutional neural network to obtain the extracted feature output.

[0029] The differential encoding kernel in the stereo feature data is input into the fully connected hidden layer of the artificial neural network for processing to obtain hidden features. The hidden features are then processed by an activation function to obtain extracted feature output.

[0030] The features extracted by the convolutional neural network and the artificial neural network are input into the concatenated layer for feature fusion, and the fused features are processed by a fully connected layer to obtain the prediction output.

[0031] The loss value between the predicted output and the preset label in the stereo feature data is calculated by a preset loss function. The learnable parameters in the convolutional neural network and the artificial neural network are adjusted according to the loss value. The training continues based on the adjusted dual-path neural network model, and the loss value is recalculated until the preset early stopping mechanism condition is met or the preset training rounds are reached, so as to obtain the magnetic resonance electrical property imaging model.

[0032] The preset labels include conductivity images and relative permittivity images.

[0033] Preferably, the method further includes:

[0034] Acquire real-time radio frequency data of the body part to be tested;

[0035] Input data for constructing a three-dimensional voxel structure based on the real-time radio frequency data;

[0036] The input data is input into the magnetic resonance electrical property imaging model to obtain the output of the conductivity image and the relative permittivity image.

[0037] Another embodiment of the present invention provides a magnetic resonance electrical property imaging model construction device, the device comprising:

[0038] The extraction module is used to acquire radio frequency emission field data of the human head, extract auxiliary feature information, and construct three-dimensional feature data of a three-dimensional voxel structure based on the auxiliary feature information and the radio frequency emission field data.

[0039] The building block is used to construct a structure-aware dual-path neural network model based on convolutional neural networks and artificial neural networks;

[0040] The training module is used to input the stereo feature data into the dual-path neural network model for training, so as to obtain the magnetic resonance electrical property imaging model.

[0041] Another embodiment of the present invention provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the magnetic resonance electrical characteristic imaging model construction method described in any of the above embodiments.

[0042] Another embodiment of the present invention provides a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the magnetic resonance electrical characteristic imaging model construction method described in any one of the above embodiments.

[0043] This invention provides a method, apparatus, device, and medium for constructing a magnetic resonance electrical characteristic imaging model. The method involves acquiring radio frequency emission field data from the human head, extracting auxiliary feature information, and constructing three-dimensional voxel structure stereoscopic feature data based on the auxiliary feature information and the radio frequency emission field data. A structure-aware dual-path neural network model is built using convolutional neural networks and artificial neural networks. The stereoscopic feature data is input into the dual-path neural network model for training to obtain the magnetic resonance electrical characteristic imaging model. This solution overcomes the noise amplification problem in existing electrical characteristic imaging algorithms and improves the accuracy of electrical characteristic mapping. Attached Figure Description

[0044] Figure 1 This is a flowchart illustrating a method for constructing a magnetic resonance electrical property imaging model according to an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of the structure of the dual-path neural network model provided in an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of the training process of the convolutional neural network provided in an embodiment of the present invention;

[0047] Figure 4 This is a schematic diagram of the training process of an artificial neural network provided in an embodiment of the present invention;

[0048] Figure 5 This is a schematic diagram of the conductivity image calculated by the magnetic resonance electrical property imaging model provided in this embodiment of the invention;

[0049] Figure 6 This is a schematic diagram of the relative permittivity image calculated by the magnetic resonance electrical property imaging model provided in this embodiment of the invention.

[0050] Figure 7 This is a schematic diagram of a magnetic resonance electrical property imaging model construction device provided in an embodiment of the present invention;

[0051] Figure 8 This is a schematic diagram of the structure of a terminal device provided in an embodiment of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] This application provides a method for constructing a magnetic resonance electrical property imaging model, see [link to relevant documentation]. Figure 1 This is a flowchart illustrating a method for constructing a magnetic resonance electrical characteristic imaging model according to an embodiment of the present invention. The method includes steps S1 to S3:

[0054] Step S1: Obtain radio frequency emission field data of the human head, extract auxiliary feature information, and construct three-dimensional feature data of a three-dimensional voxel structure based on the auxiliary feature information and the radio frequency emission field data.

[0055] Step S2: Construct a structure-aware dual-path neural network model based on convolutional neural networks and artificial neural networks;

[0056] Step S3: Input the stereo feature data into the dual-path neural network model for training to obtain the magnetic resonance electrical property imaging model.

[0057] In this specific implementation, specialized radio frequency measurement equipment is used to collect radio frequency emission field data from the human head, provided that safety standards are met.

[0058] Auxiliary feature information is extracted and fused with the preprocessed RF transmission field data. Following the rules of a three-dimensional voxel structure, the data is divided into corresponding voxel units to construct the three-dimensional feature data of the voxel structure. During the construction process, the size and number of voxels need to be determined to balance data accuracy and computational complexity.

[0059] Radio frequency (RF) emission field data of the human head contains information on the electromagnetic properties of head tissues, which are closely related to the electrical properties of the tissues, such as conductivity and dielectric constant. By combining RF emission field data with auxiliary feature information, three-dimensional voxel structure feature data can be constructed, enabling a more comprehensive and accurate description of the electromagnetic properties and structural features of the human head.

[0060] Convolutional Neural Networks (CNNs) possess powerful feature extraction capabilities, particularly adept at processing data with grid structures, such as 2D images or 3D voxel data. Through convolutional layers, pooling layers, and other structures, they can automatically extract local and hierarchical features from the data. Artificial Neural Networks (ANNs), on the other hand, have excellent nonlinear mapping capabilities, enabling them to learn complex data relationships. A structure-aware dual-path neural network model combines CNNs and ANNs. It utilizes the CNN path to extract spatial structural features from 3D feature data, while using the ANN path to learn nonlinear relationships and complex patterns in the data, achieving comprehensive perception and accurate modeling of the electromagnetic properties of the human head. The dual-path structure allows the model to process and analyze data from different perspectives, complementing each other and improving the model's performance and accuracy.

[0061] A CNN structure containing multiple convolutional layers, pooling layers, and activation function layers is constructed. Convolutional layers extract local features from the stereo feature data; different kernel sizes and numbers can extract features of different scales and types. Pooling layers downsample the data, reducing the amount of data while retaining key features. Activation function layers introduce non-linearity to enhance the model's expressive power. Based on the characteristics of the stereo feature data and model requirements, the parameters and connection methods of each layer are rationally designed. An ANN structure containing multiple hidden layers is then constructed. The number of neurons and layers in the hidden layers are adjusted according to the complexity of the data and the model's performance requirements. The ANN path receives features extracted by the CNN path or directly receives a portion of the stereo feature data. Through the connection weights and activation functions between neurons, it learns the non-linear relationships and complex patterns in the data.

[0062] By employing dual-path fusion, a suitable fusion method is designed to combine the outputs of the CNN and ANN paths. Methods such as direct connection and weighted summation can be used to integrate the features extracted from the two paths, forming the final model output. Simultaneously, the overall architecture and parameter initialization method of the dual-path neural network model are determined to prepare for model training.

[0063] The constructed 3D feature data is divided into a training set, a validation set, and a test set. The training set is used for training the model's parameters; the validation set is used to evaluate the model's performance during training, select the optimal model parameters, and prevent overfitting; the test set is used to finally evaluate the generalization ability of the trained model on unknown data. A reasonable division ratio can be adjusted according to the amount of data and actual needs, but a ratio of 7:1:2 is generally suitable.

[0064] The training set data is input into the constructed dual-path neural network model, and appropriate training parameters are set, such as learning rate, batch size, and number of training epochs. During training, the difference between the model output and the true label is calculated based on the loss function, and the model parameters are updated using optimization algorithms to continuously adjust the model's performance. Simultaneously, the model's performance metrics are periodically evaluated on the validation set, and the training parameters are adjusted based on the evaluation results to optimize the model training process.

[0065] The model is optimized by adjusting its architecture, parameters, and training methods to improve its performance and accuracy. After training, the optimized model is evaluated using test set data to comprehensively measure its performance in magnetic resonance electrical characteristic imaging and ensure that it meets the needs of practical applications.

[0066] By combining radio frequency emission field data and auxiliary feature information to construct three-dimensional feature data, and by adopting a structure-aware dual-path neural network model, the problem of noise amplification in traditional numerical differentiation-based electrical characteristic imaging technology is overcome. It also solves the problem of ignoring the inherent three-dimensional properties of the underlying electromagnetic equations in existing end-to-end deep learning methods. This improves the accuracy of mapping the electrical characteristics of human tissues, helps to more accurately understand the physiological and pathological state of the human head, and provides a more reliable basis for disease diagnosis and treatment.

[0067] In another embodiment provided by the present invention, step S1 specifically includes the following steps:

[0068] The amplitude and phase data of the radio frequency transmission field generated by finite difference time-domain electromagnetic simulation are obtained as the radio frequency transmission field data;

[0069] The amplitude data and the phase data are combined to obtain complex emission field data;

[0070] Random Gaussian noise is added to the real and imaginary parts of the complex emission field data;

[0071] Extract auxiliary feature information from the noise-added re-emission field data;

[0072] The real and imaginary parts of the noise-added complex emission field data, along with the auxiliary feature information, are divided into three-dimensional voxel blocks of a preset size to obtain the stereo feature data.

[0073] The auxiliary feature information includes gradient components, Laplacian operator, and T1-weighted image.

[0074] In this specific implementation, the finite-difference time-domain (FDTD) electromagnetic simulation program Sim4Life was used to generate all 84 simulation data.

[0075] A birdcage coil operating in orthogonal and anti-orthogonal modes at 128MHz was used for simulation. A rectangular region was defined around the coil's center point as the output area for all data. This region encompasses the "brain," measuring 256mm in the X and Y directions and 160mm in the Z direction. The generated... Field amplitude data and phase data Combined to form complex launch site data Random Gaussian noise with a mean of 0 and adjustable variance is added to the real and imaginary parts respectively to simulate noise interference during magnetic resonance acquisition.

[0076] Auxiliary feature information is extracted from the noise-added complex emission field data, mainly including the magnetic field gradient, Laplacian operator, and T1 weighted image;

[0077] Among them, the calculation of complex The gradient components of the field in the three spatial dimensions (X, Y, Z):

[0078]

[0079] For The Laplaces of the fields are calculated separately for their real and imaginary parts, i.e.:

[0080] The real and imaginary parts of the noise-added resonant emission field data, along with the auxiliary feature information, are divided into three-dimensional voxel blocks of a preset size, and the data is adjusted to the preset size to adapt to the subsequent network input format.

[0081] In another embodiment provided by the present invention, the extraction process of the T1-weighted image includes:

[0082] Different T1 signal intensity values ​​are assigned to different tissues of the noise-added retransmission field data to obtain T1 image intensity maps;

[0083] The T1 image intensity map is element-wise multiplied with the amplitude data, and the result is normalized.

[0084] The normalized image is filtered using a preset three-dimensional smooth convolution kernel to obtain the T1-weighted image.

[0085] In this specific implementation, in order to generate T1-weighted images, representative T1 signal intensity values ​​(e.g., 1.0 for white matter, 0.8 for gray matter, and 0.5 for cerebrospinal fluid) are specified for different tissue categories. These values ​​simulate the relative brightness performance of different tissues in T1-weighted imaging.

[0086] Different T1 signal intensity values ​​are assigned to different tissues of the noise-added retransmission field data to obtain T1 image intensity maps;

[0087] The generated T1 image intensity map was then compared with... The amplitude map of the field is multiplied element-wise to account for the modulation effect of radio frequency field non-uniformity on signal strength.

[0088] The results are then normalized by dividing by the maximum value of the image to unify the dynamic range of the image.

[0089] A 3×3×3 three-dimensional smooth convolution kernel is used to filter the T1 image to obtain a T1 weighted image with anatomical structure.

[0090] In another embodiment of the present invention, the step of dividing the real and imaginary parts of the noise-added complex emission field data, as well as the auxiliary feature information, into three-dimensional voxel blocks of a preset size to obtain the stereo feature data includes:

[0091] The real and imaginary parts of the noise-added complex emission field data, as well as the auxiliary feature information, are divided into three-dimensional voxel blocks of a preset size;

[0092] Based on the divided three-dimensional voxel blocks, a smooth perceptual kernel with a regular cubic structure and a differential coding kernel with a three-dimensional cross-shaped structure are constructed as the stereo feature data.

[0093] In this specific implementation, in order to effectively model the local and directional characteristics of tissue electrical properties in the spatial dimension, the present invention designs two voxel structure sensing units, namely the smooth sensing kernel SAK and the differential coding kernel DEK.

[0094] The real and imaginary parts of the noise-added complex emission field data, as well as the auxiliary feature information, are divided into three-dimensional voxel blocks of a preset size;

[0095] Based on the divided three-dimensional voxel blocks, a smooth perceptual kernel with a regular cubic structure and a differential coding kernel with a three-dimensional cross-shaped structure are constructed as the stereo feature data.

[0096] The Smoothing Perception Kernel (SAK) is a regular cubic structure designed to combine with a convolutional neural network to simulate a low-pass filter effect, thus smoothing noise-enhanced regions. The Differential Encoding Kernel (DEK) employs a three-dimensional cross-shaped structure extending along the X, Y, and Z axes, inspired by the central difference scheme widely used in the traditional MR-EPT algorithm. This structure emphasizes the ability to model the derivatives along each principal axis, combining this with an artificial neural network to extract direction-sensitive information. End-to-end training automatically learns the weights and biases of the differential operators, thereby improving the model's robustness and generalization ability in complex structural regions.

[0097] In another embodiment provided by the present invention, the dual-path neural network model includes:

[0098] A convolutional neural network for extracting features from the smooth perceptual kernel in the stereo feature data;

[0099] An artificial neural network for feature extraction from the differential coding kernel in the stereo feature data;

[0100] A concatenated layer that fuses the features extracted by the convolutional neural network and the artificial neural network.

[0101] In the specific implementation of this embodiment, when building the magnetic resonance electrical property imaging model based on structure-aware dual-branch modeling, the dual-path neural network model is composed of a convolutional neural network and an artificial neural network.

[0102] See Figure 2 This is a schematic diagram of the dual-path neural network model provided in an embodiment of the present invention. The input 3D feature data are all radio frequency transmission fields. The real and imaginary parts, the corresponding gradients and Laplacian operators, and the T1-weighted image;

[0103] The neural network model employs a dual-path structure, processing two types of structural sensing input units: smoothing kernels and differential coding kernels, to achieve joint modeling of local and directional features. Following the two network structures is a cascaded layer that fuses the features extracted by the convolutional neural network and the artificial neural network.

[0104] The model is trained by using convolutional neural networks, artificial neural networks, and input, hidden, and output layers composed of cascaded layers, and the training is achieved by backpropagating information layer by layer.

[0105] In another embodiment of the present invention, the step of inputting the stereo feature data into the dual-path neural network model for training to obtain the magnetic resonance electrical property imaging model includes:

[0106] The smoothing kernel in the stereo feature data is input into the convolution kernel of the convolutional neural network for processing to obtain convolutional features. The convolutional features are processed through an activation function, and the processed features are input into the flattening layer of the convolutional neural network to flatten them into a one-dimensional vector. The one-dimensional vector is then input into the fully connected layer of the convolutional neural network to obtain the extracted feature output.

[0107] The differential encoding kernel in the stereo feature data is input into the fully connected hidden layer of the artificial neural network for processing to obtain hidden features. The hidden features are then processed by an activation function to obtain extracted feature output.

[0108] The features extracted by the convolutional neural network and the artificial neural network are input into the concatenated layer for feature fusion, and the fused features are processed by a fully connected layer to obtain the prediction output.

[0109] The loss value between the predicted output and the preset label in the stereo feature data is calculated by a preset loss function. The learnable parameters in the convolutional neural network and the artificial neural network are adjusted according to the loss value. The training continues based on the adjusted dual-path neural network model, and the loss value is recalculated until the preset early stopping mechanism condition is met or the preset training rounds are reached, so as to obtain the magnetic resonance electrical property imaging model.

[0110] The preset labels include conductivity images and relative permittivity images.

[0111] In the specific implementation of this embodiment, please refer to Figure 3 This is a schematic diagram of the training process of the convolutional neural network provided in this embodiment of the invention. The convolutional neural network module has a smooth perceptual kernel input of size 5×5×5. This structure corresponds to a regular cubic region and aims to model local spatial continuity and suppress noise. The network structure consists of two three-dimensional convolutional layers with 32 and 64 channels respectively, and the convolutional kernel size is 3×3×3 for each. Each convolutional layer is followed by a ReLU activation function, and then the features are flattened into a one-dimensional vector through a flattening layer and input into a fully connected layer with 40 neurons.

[0112] In the figure, c represents the number of channels, d represents the depth of a single-channel stereo feature map, w represents the width of a single-channel feature map, h represents the height of a single-channel feature map, and n represents the number of small perceptual input units into which each 3D data is segmented.

[0113] During training, each weight ω of the convolution kernel i,j,k ∈R are learnable parameters that are automatically optimized through backpropagation. Essentially, they are:

[0114] Where x represents the input 3D cube voxel, y represents the output value of the convolution operation, (p,q,r) are the center voxel coordinates of the input x, and ω i,j,k Let i,j,k∈{-1,0,1} be the parameters of the 3D convolution kernel, where i,j,k∈{-1,0,1} represent the offsets in the three directions within the kernel. These parameters are continuously updated during training to achieve optimal suppression and smoothing of noise at different locations.

[0115] See Figure 4 This is a schematic diagram of the training process of the artificial neural network provided in this embodiment of the invention. The artificial neural network module receives input from a differential encoding kernel of the same size (5×5×5) but with a different structure. Its structure consists of cross-shaped voxels symmetrically extended along the X / Y / Z axes, simulating the numerical structure of the traditional central difference scheme. This branch consists of three fully connected hidden layers, with ReLU activation functions inserted between each layer. The number of neurons is 40, 80, and 40 respectively. To simulate the first derivative operation, each input channel in the artificial neural network represents a difference structure in a certain direction, which can be regarded as a learnable differential kernel, in the form of:

[0116]

[0117] in, This represents a unit vector with direction, where i is the offset step index, indicating the point sampled along the x, y, or z direction, and θ... i ∈R represents the learnable parameters of the neural network, indicating the weighting coefficients in the difference direction, and f(·) represents the feature tensor input into the network, such as the complex... Features such as the real and imaginary parts of the field are used, and p represents the current position coordinates of the voxel. Through end-to-end training, this structure can adaptively learn the forms of all directions of the derivatives, thereby enhancing its sensitivity and robustness to changes in the electromagnetic properties of the tissue structure.

[0118] Following the two network structures is a cascaded layer used to fuse the extracted features, which is then connected to a fully connected layer containing 80 neurons. Finally, a fully connected layer with 2 neurons and a regression layer output two prediction results: conductivity and relative permittivity. The regression layer calculates the mean squared error (MSE) of the regression task as a loss function to guide model optimization.

[0119] Mean square error

[0120] N is the number of input samples, P i For the network prediction output corresponding to the i-th sample, GTi This corresponds to the true value of the i-th sample, i.e., the target output. The preset labels include the conductivity image and the relative permittivity image.

[0121] The constructed network model was trained, and an early stopping mechanism was used as the criterion for training termination. Training was terminated automatically when the loss function (MSE) on the validation set stopped decreasing within 10 consecutive epochs; otherwise, training continued for a maximum of 1000 epochs. Finally, the magnetic resonance electrical property imaging model was obtained.

[0122] Accurate electrical property imaging is achieved using a magnetic resonance electrical property imaging model.

[0123] In another embodiment provided by the present invention, the method further includes:

[0124] Acquire real-time radio frequency data of the body part to be tested;

[0125] Input data for constructing a three-dimensional voxel structure based on the real-time radio frequency data;

[0126] The input data is input into the magnetic resonance electrical property imaging model to obtain the output of the conductivity image and the relative permittivity image.

[0127] In this specific implementation, an radio frequency detection device is used to attach to or maintain a suitable distance from the part of the human body to be detected according to a predetermined position and method to obtain real-time radio frequency data of the part of the human body to be detected.

[0128] When acquiring real-time radio frequency data, radio frequency signals of a specific frequency and intensity are emitted to the part of the human body to be tested, and reflected and scattered radio frequency signals that change due to differences in the electromagnetic properties of tissues are received in real time.

[0129] Based on the anatomical structure and imaging accuracy requirements of the area to be detected, the input data for the three-dimensional voxel structure is determined;

[0130] The constructed 3D voxel structure input data is converted into a format that the magnetic resonance electrical property imaging model can recognize and process. Based on the model's input requirements, operations such as dimensional adjustment and data type conversion are performed on the data.

[0131] The adapted input data is fed into a pre-trained magnetic resonance electrical property imaging model. The model activates both a convolutional neural network (CNN) path and an artificial neural network (ANN) path to process the input data in parallel. The CNN path extracts spatial structural features from the data through multi-layer convolution and pooling operations; the ANN path, based on the connection weights and activation functions between neurons, mines nonlinear relationships and complex patterns in the data. The two paths interact and fuse information within the model, and through forward propagation calculations, progressively infer the conductivity and relative permittivity at each voxel location of the target area, resulting in the output of conductivity and relative permittivity images.

[0132] See Figure 5 This is a schematic diagram of the conductivity image calculated by the magnetic resonance electrical property imaging model provided in this embodiment of the invention. See also Figure 6 This is a schematic diagram of the relative permittivity image calculated by the magnetic resonance electrical property imaging model provided in this embodiment of the invention.

[0133] By acquiring radio frequency data of the human body to be detected in real time and quickly processing it into three-dimensional voxel structure data suitable for model input, the data can be promptly input into the imaging model to obtain output results.

[0134] Another embodiment of the present invention provides a magnetic resonance electrical property imaging model construction device, see [link to relevant documentation]. Figure 7 This is a schematic diagram of a magnetic resonance electrical characteristic imaging model construction device provided in an embodiment of the present invention. The device includes:

[0135] The extraction module is used to acquire radio frequency emission field data of the human head, extract auxiliary feature information, and construct three-dimensional feature data of a three-dimensional voxel structure based on the auxiliary feature information and the radio frequency emission field data.

[0136] The building block is used to construct a structure-aware dual-path neural network model based on convolutional neural networks and artificial neural networks;

[0137] The training module is used to input the stereo feature data into the dual-path neural network model for training, so as to obtain the magnetic resonance electrical property imaging model.

[0138] The magnetic resonance electrical characteristic imaging model construction device provided in this embodiment can perform all the steps and functions of the magnetic resonance electrical characteristic imaging model construction method provided in any of the above embodiments. The specific functions of the device will not be described in detail here.

[0139] See Figure 8This is a schematic diagram of a terminal device provided in an embodiment of the present invention. The terminal device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a magnetic resonance electrical characteristic imaging model construction program. When the processor executes the computer program, it implements the steps in the various embodiments of the magnetic resonance electrical characteristic imaging model construction method described above, for example... Figure 1 The steps S1 to S3 are shown. Alternatively, when the processor executes the computer program, it implements the functions of each module in the above-described device embodiments.

[0140] For example, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules can be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the magnetic resonance electrical characteristic imaging model construction device. For example, the computer program can be divided into a detection module, an output power control module, and a window control module. The specific functions of each module have been described in detail in the magnetic resonance electrical characteristic imaging model construction method provided in any of the above embodiments, and the specific functions of the device will not be repeated here.

[0141] The magnetic resonance electrical characteristic imaging model construction device can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. The device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the schematic diagram is merely an example of a magnetic resonance electrical characteristic imaging model construction device and does not constitute a limitation on such a device. It may include more or fewer components than illustrated, or combine certain components, or use different components. For example, the device may also include input / output devices, network access devices, buses, etc.

[0142] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the magnetic resonance electrical characteristic imaging model construction device, connecting all parts of the device via various interfaces and lines.

[0143] The memory can be used to store the computer program and / or modules. The processor implements various functions of the magnetic resonance electrical characteristic imaging model construction device by running or executing the computer program and / or modules stored in the memory and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0144] If the integrated module of the magnetic resonance electrical characteristic imaging model construction device is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0145] It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered to be within the scope of protection of this invention.

Claims

1. A method for constructing a magnetic resonance electrical property imaging model, characterized in that, The method includes: Acquire radio frequency emission field data of the human head, extract auxiliary feature information, and construct three-dimensional feature data of a three-dimensional voxel structure based on the auxiliary feature information and the radio frequency emission field data; A structure-aware dual-path neural network model was built based on convolutional neural networks and artificial neural networks. The stereo feature data is input into the dual-path neural network model for training to obtain the magnetic resonance electrical property imaging model.

2. The method for constructing a magnetic resonance electrical property imaging model according to claim 1, characterized in that, The process of acquiring radio frequency emission field data of the human head, extracting auxiliary feature information, and constructing three-dimensional voxel structure stereo feature data based on the auxiliary feature information and the radio frequency emission field data includes: The amplitude and phase data of the radio frequency transmission field generated by finite difference time-domain electromagnetic simulation are obtained as the radio frequency transmission field data; The amplitude data and the phase data are combined to obtain complex emission field data; Random Gaussian noise is added to the real and imaginary parts of the complex emission field data; Extract auxiliary feature information from the noise-added re-emission field data; The real and imaginary parts of the noise-added complex emission field data, along with the auxiliary feature information, are divided into three-dimensional voxel blocks of a preset size to obtain the stereo feature data. The auxiliary feature information includes gradient components, Laplacian operator, and T1-weighted image.

3. The method for constructing a magnetic resonance electrical property imaging model according to claim 2, characterized in that, The extraction process of the T1-weighted image includes: Different T1 signal intensity values ​​are assigned to different tissues of the noise-added retransmission field data to obtain T1 image intensity maps; The T1 image intensity map is element-wise multiplied with the amplitude data, and the result is normalized. The normalized image is filtered using a preset three-dimensional smooth convolution kernel to obtain the T1-weighted image.

4. The method for constructing a magnetic resonance electrical property imaging model according to claim 2, characterized in that, The step of dividing the real and imaginary parts of the noise-added complex emission field data, along with the auxiliary feature information, into three-dimensional voxel blocks of a preset size to obtain the stereo feature data includes: The real and imaginary parts of the noise-added complex emission field data, as well as the auxiliary feature information, are divided into three-dimensional voxel blocks of a preset size; Based on the divided three-dimensional voxel blocks, a smooth perceptual kernel with a regular cubic structure and a differential coding kernel with a three-dimensional cross-shaped structure are constructed as the stereo feature data.

5. The method for constructing a magnetic resonance electrical property imaging model according to claim 1, characterized in that, The dual-path neural network model includes: A convolutional neural network for extracting features from the smooth perceptual kernel in the stereo feature data; An artificial neural network for feature extraction from the differential coding kernel in the stereo feature data; A concatenated layer that fuses the features extracted by the convolutional neural network and the artificial neural network.

6. The method for constructing a magnetic resonance electrical property imaging model according to claim 5, characterized in that, The step of inputting the stereo feature data into the dual-path neural network model for training to obtain the magnetic resonance electrical property imaging model includes: The smoothing kernel in the stereo feature data is input into the convolution kernel of the convolutional neural network for processing to obtain convolutional features. The convolutional features are processed through an activation function, and the processed features are input into the flattening layer of the convolutional neural network to flatten them into a one-dimensional vector. The one-dimensional vector is then input into the fully connected layer of the convolutional neural network to obtain the extracted feature output. The differential encoding kernel in the stereo feature data is input into the fully connected hidden layer of the artificial neural network for processing to obtain hidden features. The hidden features are then processed by an activation function to obtain extracted feature output. The features extracted by the convolutional neural network and the artificial neural network are input into the concatenated layer for feature fusion, and the fused features are processed by a fully connected layer to obtain the prediction output. The loss value between the predicted output and the preset label in the stereo feature data is calculated by a preset loss function. The learnable parameters in the convolutional neural network and the artificial neural network are adjusted according to the loss value. The training continues based on the adjusted dual-path neural network model, and the loss value is recalculated until the preset early stopping mechanism condition is met or the preset training rounds are reached, so as to obtain the magnetic resonance electrical property imaging model. The preset labels include conductivity images and relative permittivity images.

7. The method for constructing a magnetic resonance electrical property imaging model according to claim 1, characterized in that, The method further includes: Acquire real-time radio frequency data of the body part to be tested; Input data for constructing a three-dimensional voxel structure based on the real-time radio frequency data; The input data is input into the magnetic resonance electrical property imaging model to obtain the output of the conductivity image and the relative permittivity image.

8. A magnetic resonance electrical property imaging model construction device, characterized in that, The device includes: An extraction module is used to acquire radio frequency emission field data of the human head, extract auxiliary feature information, and construct three-dimensional feature data of a three-dimensional voxel structure based on the auxiliary feature information and the radio frequency emission field data. The building block is used to construct a structure-aware dual-path neural network model based on convolutional neural networks and artificial neural networks; The training module is used to input the stereo feature data into the dual-path neural network model for training, so as to obtain the magnetic resonance electrical property imaging model.

9. A terminal device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the magnetic resonance electrical property imaging model construction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the magnetic resonance electrical characteristic imaging model construction method as described in any one of claims 1 to 7.

Images