A method, system, device, and medium for cross-domain translation of multicenter diffusion-weighted magnetic resonance imaging of the brain based on generative adversarial networks.

Abstract

A method, system, device, and medium for cross-domain conversion of multicenter diffusion-weighted magnetic resonance imaging (MRI) images of the brain based on generative adversarial networks (GANs) are disclosed. The method includes: acquiring a training sample set; data preprocessing; constructing a GAN model; training the GAN model; and processing the output data. The system, device, and medium are used to implement the cross-domain conversion method for multicenter diffusion-weighted MRI images of the brain based on GANs. This invention applies GANs to the homogenization of multicenter diffusion MRI and makes a series of optimizations to GANs, effectively solving technical problems such as blurred brain details, weak spatial continuity, and poor model versatility in existing multicenter diffusion MRI homogenization methods. It features increased data volume and improved data effectiveness in medical imaging.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method, system, device, and medium for cross-domain conversion of multicenter brain diffusion-weighted magnetic resonance images based on generative adversarial networks. Background Technology

[0002] Magnetic resonance imaging (MRI) is a common medical imaging technique that provides excellent soft tissue density and contrast resolution, offering higher diagnostic accuracy compared to CT images. However, diffusion MRI is highly sensitive to differences in acquisition settings and is easily affected by these settings. In multi-center studies, variations in scanning equipment and methods across different medical centers can lead to inherent differences in the diffusion MRI images acquired from different centers. Therefore, in multi-center studies, it is usually not possible to directly integrate diffusion MRI images from different centers. It is necessary to pre-process the diffusion MRI images from each center to increase data reliability and improve the ability of joint statistical analysis, thereby integrating more image data for algorithm training and improving the reliability of disease diagnosis and analysis. Diffusion MRI includes diffusion-weighted imaging; however, cross-domain conversion of diffusion-weighted imaging is complex and more difficult to handle. The current mainstream approach is to preprocess the acquired diffusion-weighted imaging data, compressing the amount of data to be processed during the conversion process to reduce the storage requirements.

[0003] Mirzaalian et al. proposed a cross-domain transformation method for diffusion-weighted imaging based on spherical harmonic function representation, which represents the original diffusion-weighted imaging with spherical harmonic coefficients, effectively reducing the differences between multi-center diffusion MRI images (Mirzaalian H, Ning L, Savadjiev P, et al. Inter-site and inter-scanner diffusionMRI data harmonization[J]. Neuroimage. 2016; 135(1):311–23). Koppers et al. proposed SHResNet, which, based on the design idea of ​​residual connections, constructs a neural network with a Res Block structure, representing the original diffusion-weighted MRI images with spherical harmonic rotation invariant features, thereby reducing the number of parameters in the training process and effectively reducing the differences between the acquired multi-center diffusion-weighted imaging data (Koppers S, Bloy L, Berman JI, et al. Spherical harmonic residual network for diffusion signal harmonization[C]. International Conference on Medical Image Computing and Computer-Assisted Intervention. 2019:173-182.).

[0004] However, existing technologies have the following problems:

[0005] (1) Diffusion magnetic resonance imaging data is four-dimensional. Existing diffusion magnetic resonance imaging is based on two-dimensional cranial slices for feature extraction and constraint, without effectively utilizing the spatial information in the data, and information is lost during the conversion process.

[0006] (2) Existing multicenter diffusion magnetic resonance imaging conversion methods are all based on traditional supervised conversion frameworks, and the conversion results are prone to problems such as blurred brain tissue edges; methods based on single supervised constraints usually have a certain degree of ambiguity in the conversion results, and the occurrence of ambiguity reflects the loss of information during the conversion process.

[0007] (3) When using generative adversarial networks to generate images for natural images, Markov discriminators are often used to make equal-probability judgments on each small patch in the image to guide the generator to generate images. However, for brain diffusion magnetic resonance imaging, the brain regions are concentrated in the central region of the image. At this time, the probability of the region of interest appearing in different small patches of the image is different. Making equal-probability judgments can easily lead to the loss of some brain detail information. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, the present invention aims to provide a method, system, device, and medium for cross-domain conversion of multicenter brain diffusion-weighted magnetic resonance imaging based on generative adversarial networks (GANs). By applying GANs to the homogenization of multicenter diffusion-weighted magnetic resonance imaging and making a series of optimizations to GANs, the invention can effectively solve the technical problems existing in the homogenization of existing multicenter diffusion-weighted magnetic resonance imaging, such as blurred brain details, weak spatial continuity, and poor model universality. It has the characteristics of increasing the amount of medical image data and improving data effectiveness.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A multi-center brain diffusion-weighted magnetic resonance imaging cross-domain translation method based on generative adversarial networks includes the following steps:

[0011] Step 1: Obtain the training sample set, which includes training sample groups of different cases. The training sample groups include source domain brain diffusion magnetic resonance imaging (MRI) data and target domain brain diffusion MRI data of the same case collected by scanners at two acquisition centers.

[0012] Step 2: Data preprocessing. Calculate the source domain RISH eigenvalues ​​and target domain RISH eigenvalues ​​based on the source domain brain diffusion MRI data and target domain brain diffusion MRI data obtained in Step 1.

[0013] Step 3: Construct a generative adversarial network model, which includes a generator, a discriminator, and a loss function;

[0014] Step 4: Train the generative adversarial network model. Input the source domain RISH feature values ​​and target domain RISH feature values ​​obtained in Step 2 into the generative adversarial network model constructed in Step 3 for training, and output the target domain synthesized RISH feature values.

[0015] Step 5: Process the output data. Based on the target domain RISH feature value obtained in Step 4 and the source domain RISH feature value obtained in Step 2, calculate the converted brain diffusion-weighted magnetic resonance imaging signal.

[0016] The specific process of step 2 is as follows:

[0017] Step 2.1: Using the source domain brain diffusion MRI data and the target domain brain diffusion MRI data obtained in Step 1, calculate the spherical harmonic coefficients of the source domain and the target domain, respectively. The spherical harmonic coefficients can be solved using the following formula:

[0018]

[0019] In the formula, This represents the direction vector in spherical harmonic coordinates. Representing coordinates The spherical harmonic function is given by the expression denoted by l, where l represents the angular frequency of the spherical harmonic function and m represents the phase. This represents the original brain diffusion weighting function collected;

[0020] Step 2.2: Calculate the RISH eigenvalues ​​of the source domain and the target domain using the spherical harmonic coefficients of the source domain obtained in Step 2.1. The specific formulas are shown below:

[0021]

[0022] In the formula, ||c l || 2 This represents the RISH eigenvalue at angular frequency l. This represents the spherical harmonic coefficients of phase m at frequency l.

[0023] The generator in step 3 includes multiple convolutional modules, and the convolutional layers in each convolutional module use three-dimensional convolutional kernels.

[0024] The discriminator in step 3 is a Markov discriminator capable of adaptive weight allocation. The specific method for adaptive weight allocation is as follows:

[0025] Adaptively calculate the number n of voxels representing brain regions contained in the current image patch based on brain region masks in the dataset. brain And the total number of voxels n in the current image patch. sum And calculate n bain With n sum The ratio of w i As the initial weights for the current image patch:

[0026]

[0027] In the formula, n byain n represents the number of voxel points representing the brain region contained in the current image patch. sum w represents the total number of voxels in the current image patch. i Indicates the initial weight of the current image patch;

[0028] The sum of the weights of each image patch is 1. After obtaining the output weights of each image patch, the initial weights are normalized using the softmax function to obtain the final adaptive weights for each image patch.

[0029]

[0030] In the formula, w i This represents the final adaptive weight for each image patch.

[0031] The loss function in step 3 includes a first loss function, a second loss function, and a third loss function. The first and second loss functions are the loss functions of the generator, and the third loss function is the loss function of the discriminator.

[0032] The first loss function is the difference between the RISH feature values ​​in the target domain and the generator's prediction results:

[0033]

[0034] In the formula, y represents the RISH feature value of the target domain, and y′ represents the prediction result of the generator;

[0035] The second loss function is the difference between the generator's prediction and the segmentation results between images in the target domain with RISH feature values ​​of order 0:

[0036]

[0037] In the formula, seg target This represents the segmentation result between images where the RISH eigenvalue order in the target domain is 0, seg harmonized This represents the segmentation results between images where the RISH feature value order is 0, as predicted by the generator.

[0038] The third loss function is a gradient-penalized Wasserstein loss function with the Lipschitz constant K set to 1. The Lipschitz continuity of the discriminator is constrained through gradient penalty. The specific calculation method is as follows:

[0039]

[0040] In the formula, f w (·) denotes a continuous function fitted by a parameterized discriminator. p g and p γ The random interpolation sampling distribution on the connection line, specifically, let x γ ~P γ x g ~p g , ε~Uniform[0,1], there are:

[0041]

[0042] The loss function used for training the generative adversarial network is:

[0043]

[0044] The specific process of step 4 is as follows:

[0045] Step 4.1: Fix the discriminator parameters and input the source domain RISH feature values ​​obtained in Step 2 into the generator to obtain the generated homogenized RISH feature values;

[0046] Step 4.2: Fix the generator parameters, input the homogenized RISH feature values ​​obtained in Step 4.1 and the target domain RISH feature values ​​obtained in Step 2 into the discriminator to distinguish local image patches, and output the authenticity judgment result of the local image patches;

[0047] Step 4.3: Update the network parameters of the generative adversarial network using the loss function set in Step 3;

[0048] Step 4.4: Repeat steps 4.1-4.3 until the preset number of training iterations is reached or the error between the generated result and the RISH feature value of the target domain reaches the preset value;

[0049] Step 4.5: Use the generator to fix the parameters of the generated results after training in Step 4.4, and output the RISH feature values ​​of the target domain after homogenization of the source domain.

[0050] The specific process of step 5 is as follows:

[0051] Step 5.1: Calculate the target mapping coefficient l using the target domain RISH eigenvalues ​​synthesized from the output of Step 4 and the source domain RISH eigenvalues ​​obtained in Step 2. s :

[0052] Target mapping coefficient l s :

[0053]

[0054] In the formula, l s Represents the target mapping coefficient, П(||c i || 2 ) represents the target domain synthesized RISH eigenvalues, ||c i || 2 Represents the RISH eigenvalues ​​of the source domain;

[0055] Step 5.2: Use the target mapping coefficients l obtained in Step 5.1 s Calculate the target spherical harmonic coefficients required to reconstruct the diffusion-weighted magnetic resonance imaging signal of the brain.

[0056]

[0057] In the formula, This represents the target spherical harmonic coefficients after conversion. l represents the spherical harmonic coefficients before the conversion.s Represents the target mapping coefficient;

[0058] Step 5.3: Using the target spherical harmonic coefficients obtained in Step 5.2 and the gradient direction information of the target domain brain diffusion magnetic resonance imaging data, calculate the converted brain diffusion-weighted magnetic resonance imaging signal:

[0059]

[0060] In the formula, This represents the direction vector in spherical harmonic coordinates. Target spherical harmonic coefficients, Representing coordinates The spherical harmonic function is given by l, where l represents the angular frequency of the spherical harmonic function, m represents the phase, and n represents the cutoff order set when calculating the spherical harmonic coefficients.

[0061] A multi-center brain diffusion-weighted magnetic resonance imaging cross-domain translation system based on generative adversarial networks includes:

[0062] Data acquisition module: used to acquire training sample set, which includes training sample groups of different cases. The training sample group includes source domain brain diffusion magnetic resonance data and target domain brain diffusion magnetic resonance data of the same case acquired by scanners at two acquisition centers.

[0063] Data preprocessing module: Converts source domain brain diffusion magnetic resonance imaging (DMRI) data and target domain brain diffusion MRI data in the training sample set into source domain RISH feature values ​​and target domain RISH feature values;

[0064] Generative Adversarial Network (GAN) module: Input source domain RISH feature values ​​and target domain RISH feature values ​​into the GAN model for training, and output target domain synthesized RISH feature values;

[0065] Data processing module: Calculates the synthesized RISH feature values ​​of the output target domain to obtain the converted brain diffusion-weighted magnetic resonance imaging signal.

[0066] A multi-center intracranial diffusion-weighted magnetic resonance imaging (MRI) transdomain conversion device based on generative adversarial networks includes:

[0067] Memory: for storing computer programs that implement the multicenter brain diffusion-weighted magnetic resonance imaging cross-domain conversion method based on generative adversarial networks as described in claims 1-7;

[0068] Processor: Used to implement the cross-domain conversion method for multicenter brain diffusion-weighted magnetic resonance imaging based on generative adversarial networks according to claims 1-7 when executing the computer program.

[0069] A computer-readable storage medium comprising:

[0070] The computer-readable storage medium stores a computer program that, when executed by a processor, enables the implementation of the multi-center brain diffusion-weighted magnetic resonance imaging cross-domain conversion method based on generative adversarial networks as described in claims 1-7.

[0071] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0072] 1. The generator in the generative adversarial network in step 3 of the present invention extracts features through a five-layer convolutional module based on three-dimensional convolutional kernels, which can effectively utilize the spatial information of the input data.

[0073] 2. In step 3 of this invention, the generative adversarial network model uses a second loss function for the generator to optimize and update the network parameters of the generative adversarial network. The auxiliary brain region segmentation task is used to further constrain the conversion result. The neural network has a strong feature extraction capability, and the extracted deep features can effectively represent the high-level features of the input image. The deep supervision mechanism can constrain the deep features of the input image, thereby placing higher requirements on the conversion result in terms of high-level semantics.

[0074] 3. In step 3 of this invention, a Markov discriminator with adaptive weight allocation based on the distribution characteristics of brain regions is used. In this weighting strategy, image patches distributed at the edges are assigned smaller weights because they involve fewer brain regions. In the final integration of discrimination probabilities, the discrimination probability of these image patches will play a smaller role. On the other hand, image patches distributed in the center are areas where brain regions are highly concentrated and are assigned larger weights. The discrimination probability of these image patches will play a larger role in the final discrimination probability of the entire image. During the image transformation process, the generator is guided to adaptively focus on the transformation of each small image patch, maintaining better local information, thereby effectively extracting detailed brain information features.

[0075] 4. This invention proposes a diffusion magnetic resonance imaging (DMRI) conversion method based on generative adversarial networks (GANs) and 3D convolution. This method effectively solves the problem that the conversion results in the traditional supervised conversion framework are prone to blurring of brain tissue edges. In the conversion process, a Markov discriminator is used to introduce local unsupervised constraints. The spatial information of the image data is utilized by using 3D convolution. The conversion quality is evaluated unsupervised by the Markov discriminator, which assists the generator in completing the conversion of RISH features. This improves the conversion quality of RISH features and thus enhances the image generation quality.

[0076] In summary, compared with the prior art, this invention applies generative adversarial networks to the homogenization of multicenter diffusion magnetic resonance imaging and makes a series of optimizations to the generator, discriminator and loss function in the generative adversarial network. This can solve the technical problems existing in the homogenization of multicenter diffusion magnetic resonance imaging, such as blurred brain details, weak spatial continuity and poor model universality. It has the characteristics of increasing the amount of medical image data and improving data effectiveness. Attached Figure Description

[0077] Figure 1 This is a flowchart of the method of the present invention.

[0078] Figure 2 This is a flowchart of the process for generating an adversarial network in this invention.

[0079] Figure 3 This is a schematic diagram of a brain region segmentation loss network.

[0080] Figure 4 The value of b is 2000s / mm 2 Visualization results of RISH eigenvalues ​​at that time. Detailed Implementation

[0081] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0082] See Figure 1 and Figure 2 A multi-center brain diffusion-weighted magnetic resonance imaging cross-domain translation method based on generative adversarial networks includes the following steps:

[0083] Step 1: Obtain the training sample set, which includes training sample groups of different cases. The training sample groups include source domain brain diffusion magnetic resonance imaging (MRI) data and target domain brain diffusion MRI data of the same case collected by scanners at two acquisition centers.

[0084] Step 2: Data preprocessing. Based on the source domain brain diffusion MRI data and target domain brain diffusion MRI data obtained in Step 1, calculate the source domain RISH eigenvalues ​​and target domain RISH eigenvalues, specifically:

[0085] Step 2.1: Using the source domain brain diffusion MRI data and the target domain brain diffusion MRI data obtained in Step 1, calculate the spherical harmonic coefficients of the source domain and the target domain, respectively. The spherical harmonic coefficients can be solved using the following formula:

[0086]

[0087] In the formula, This represents the direction vector in spherical harmonic coordinates. Representing coordinates The spherical harmonic function is given by the expression denoted by l, where l represents the angular frequency of the spherical harmonic function and m represents the phase. This represents the original brain diffusion weighting function collected;

[0088] Spherical harmonics (SH) are a set of complex functions derived from mappings of points on the unit sphere. The set of spherical harmonics can be used as basis functions to approximate any function distributed on the unit sphere. The specific function expressions are shown below:

[0089]

[0090] In the formula, This represents the direction vector in spherical harmonic coordinates. Representing coordinates The spherical harmonic function is given by the expression denoted by l, where l represents the angular frequency of the spherical harmonic function and m represents the phase. Represents the Legendre polynomial with angular frequency l and phase m;

[0091] Step 2.2: Calculate the RISH eigenvalues ​​of the source domain and the target domain using the spherical harmonic coefficients of the source domain obtained in Step 2.1. The specific formulas are shown below:

[0092]

[0093] In the formula, ||c l || 2 This represents the RISH eigenvalue at angular frequency l. This represents the spherical harmonic coefficients of phase m at frequency l.

[0094] For brain diffusion-weighted magnetic resonance images acquired in the source and target domains, each image yields 10 RISH eigenvalues ​​after approximating the original data using the spherical harmonic function representation. For each image acquired under a diffusion factor b value, RISH eigenvalues ​​are obtained at five even orders: {0, 2, 4, 6, 8}.

[0095] Step 3: Construct a generative adversarial network model, which includes a generator, a discriminator, and a loss function;

[0096] The generator in step 3 includes multiple convolutional modules, and the convolutional layers in each module use three-dimensional convolutional kernels. This generator network extracts features through five layers of convolutional modules based on three-dimensional convolutional kernels, effectively utilizing the spatial information of the input data. Each convolutional module consists of a convolutional layer, an instance normalization layer, and a ReLU activation function. In the output layer, the hyperbolic tangent function is used as the activation function.

[0097] The discriminator in step 3 is a Markov discriminator capable of adaptive weight allocation. The specific method for adaptive weight allocation is as follows:

[0098] Adaptively calculate the number n of voxels representing brain regions contained in the current image patch based on brain region masks in the dataset. brain And the total number of voxels n in the current image patch. sum And calculate n bain With n sum The ratio of w i As the initial weights for the current image patch:

[0099]

[0100] In the formula, n brain n represents the number of voxel points representing the brain region contained in the current image patch. sum w represents the total number of voxels in the current image patch. i Indicates the initial weight of the current image patch;

[0101] The sum of the weights of each image patch is 1. After obtaining the output weights of each image patch, the initial weights are normalized using the softmax function to obtain the final adaptive weights for each image patch.

[0102]

[0103] In the formula, w i This represents the final adaptive weight for each image patch.

[0104] In this weighting strategy, image patches distributed at the edges are assigned smaller weights because they involve fewer brain regions. In the final integration of discrimination probabilities, the discrimination probability of these image patches will play a smaller role. On the other hand, image patches distributed in the center, which are areas where brain regions are highly concentrated, are assigned larger weights. The discrimination probability of these image patches will play a larger role in the final discrimination probability of the entire image. During the image transformation process, the generator is guided to adaptively focus on the transformation of each small image patch to maintain better local information.

[0105] The loss function in step 3 includes a first loss function, a second loss function, and a third loss function. The first and second loss functions are the loss functions of the generator, and the third loss function is the loss function of the discriminator.

[0106] The first loss function is the difference between the RISH feature values ​​in the target domain and the generator's prediction results:

[0107]

[0108] In the formula, y represents the RISH feature value of the target domain, and y′ represents the prediction result of the generator;

[0109] The second loss function is the difference between the generator's prediction and the segmentation results between images in the target domain with RISH feature values ​​of order 0:

[0110] The segmentation network was trained using RISH features of order 0 from the target domain and the brain region masks provided in the dataset. The segmentation network selected was 3DU-Net, and the loss function was the Dice loss function.

[0111]

[0112] Where X represents the segmentation result and Y represents the target output;

[0113] Next, the parameters of the 3D U-Net are fixed, and the segmentation maps of the transformation result and the target domain R0 feature values ​​are obtained using the 3D U-Net with fixed weights. By using the segmentation result of the target domain R0 feature values ​​as the target value, the Dice loss is used to implicitly constrain the depth features of the transformation result. That is, during the model training process, the segmentation map of the transformation result is required to be as consistent as possible with the segmentation map of the target domain R0 feature values. During the training process, the loss... dice for:

[0114]

[0115] In the formula, seg target This represents the segmentation result between images where the RISH eigenvalue order in the target domain is 0, seg harmonized This represents the segmentation results between images where the RISH feature value order is 0, as predicted by the generator.

[0116] The third loss function is a gradient-penalized Wasserstein loss function with the Lipschitz constant K set to 1. This gradient penalty constrains the Lipschitz continuity of the discriminator, making the distribution of network parameters more uniform and effectively improving the parameter distribution in the discriminator. The specific calculation method is as follows:

[0117]

[0118] In the formula, f w (·) denotes a continuous function fitted by a parameterized discriminator. p g and p r The random interpolation sampling distribution on the connection line, specifically, let x r ~p r x g ~pg , ε~Uniform[0,1], there are:

[0119]

[0120] The loss function used for training the generative adversarial network is:

[0121]

[0122] Step 4: Train the generative adversarial network model. Input the source domain RISH feature values ​​and target domain RISH feature values ​​obtained in Step 2 into the generative adversarial network model constructed in Step 3 for training, and output the target domain synthesized RISH feature values, specifically:

[0123] Step 4.1: Fix the discriminator parameters and input the source domain RISH feature values ​​obtained in Step 2 into the generator to obtain the generated homogenized RISH feature values;

[0124] Step 4.2: Fix the generator parameters, input the homogenized RISH feature values ​​obtained in Step 4.1 and the target domain RISH feature values ​​obtained in Step 2 into the discriminator to distinguish local image patches, and output the authenticity judgment result of the local image patches;

[0125] Step 4.3: Update the network parameters of the generative adversarial network using the loss function set in Step 3;

[0126] Step 4.4: Repeat steps 4.1-4.3 until the preset number of training iterations is reached or the error between the generated result and the RISH feature value of the target domain reaches the preset value;

[0127] Step 4.5: Use the generator to fix the parameters of the generated results after training in Step 4.4, and output the RISH feature values ​​of the target domain after homogenization of the source domain.

[0128] Step 5: Process the output data. Based on the target domain RISH feature values ​​obtained in Step 4 and the source domain RISH feature values ​​obtained in Step 2, calculate the converted brain diffusion-weighted magnetic resonance imaging signal, specifically:

[0129] Step 5.1: Calculate the target mapping coefficient l using the target domain RISH eigenvalues ​​synthesized from the output of Step 4 and the source domain RISH eigenvalues ​​obtained in Step 2. s :

[0130] Target mapping coefficient l s :

[0131]

[0132] In the formula, l s Denotes the target mapping coefficient, ∏(||ci || 2 ) represents the target domain synthesized RISH eigenvalues, ||c i || 2 Represents the RISH eigenvalues ​​of the source domain;

[0133] Step 5.2: Use the target mapping coefficients l obtained in Step 5.1 s Calculate the target spherical harmonic coefficients required to reconstruct the diffusion-weighted magnetic resonance imaging signal of the brain.

[0134]

[0135] In the formula, This represents the target spherical harmonic coefficients after conversion. l represents the spherical harmonic coefficients before the conversion. s Represents the target mapping coefficient;

[0136] Step 5.3: Using the target spherical harmonic coefficients obtained in Step 5.2 and the gradient direction information of the target domain brain diffusion magnetic resonance imaging data, calculate the converted brain diffusion-weighted magnetic resonance imaging signal:

[0137]

[0138] In the formula, This represents the direction vector in spherical harmonic coordinates. Represents the target spherical harmonic coefficients.

[0139] Representing coordinates The spherical harmonic function is given by l, where l represents the angular frequency of the spherical harmonic function, m represents the phase, and n represents the cutoff order set when calculating the spherical harmonic coefficients.

[0140] See Figure 3 The third loss function is an implicit loss function based on brain region segmentation task supervision. Since each case has a diffusion factor b of 1000 s / mm, 2 2000s / mm 2 Data is collected, so each case contains an R0 feature under different diffusion factors. The R0 features are concatenated along the channel dimension. The generator's prediction and the R0 feature value of the target domain are respectively input into the pre-trained segmentation network to obtain their corresponding segmentation maps. Finally, the Dice loss function is used to calculate the segmentation error between the two, and backpropagation is performed to update the parameters, guiding the segmentation map of the prediction result obtained by the generator to be as consistent as possible with the segmentation map of the target domain.

[0141] See Figure 4In this model, b-value is the diffusion factor set during acquisition, and 3T and 7T represent the magnetic field strength used by the scanner during acquisition. In this invention, the source domain dataset uses DWI data acquired with a magnetic field strength of 3T, and the target domain dataset uses DWI data acquired with a magnetic field strength of 7T. RISH eigenvalues ​​represent different aspects of diffuse magnetic resonance data. RISH eigenvalues ​​of different orders differ significantly and decrease numerically with increasing order. R0 is the eigenvalue when the RISH eigenvalue order is 0, containing rich tissue structure information. The middle of each column in the image indicates the display range of the image. This invention uses DWI data with b-values ​​of 1000 and 2000, and magnetic field strengths of 3T and 7T, respectively, to achieve homogenization of multi-center DWI data.

[0142] Depend on Figure 3 , Figure 4 As can be seen, compared with the prior art, this invention uses the spherical harmonic function representation method to represent the acquired diffusion-weighted magnetic resonance data as rotational invariant spherical harmonic features in image preprocessing, such as... Figure 4 As shown, this reduces the number of parameters during training; secondly, the obtained RISH feature values ​​are input into a 3D convolution-based generator to generate images, effectively utilizing the spatial continuity information in the images. Furthermore, a weight redistribution Markov discriminator based on brain region distribution characteristics guides the generator's training, enabling it to better focus on brain region features. Simultaneously, for problems with single constraints, such as... Figure 3 As shown, the accuracy of segmentation using the spherical harmonic rotation invariant feature R0 of the output image and the target image, i.e., Dice loss, is introduced as an auxiliary supervisory constraint. This guides the generator to focus on the consistency between the transformation result and the depth features extracted from the target output in the loss network during the transformation process, thereby further improving the transformation effect of diffusion magnetic resonance imaging. Finally, the transformed diffusion-weighted magnetic resonance imaging is reconstructed using the RISH features obtained from the output and the gradient information of the target image, thus achieving homogenization of information acquired from multiple centers. This solves the technical problems existing in the homogenization of multi-center diffusion magnetic resonance imaging, such as blurred brain details, weak spatial continuity, and poor model versatility, which is beneficial to the analysis and processing of downstream tasks.

[0143] A multi-center brain diffusion-weighted magnetic resonance imaging cross-domain translation system based on generative adversarial networks includes:

[0144] Data acquisition module: This module is used to acquire the training sample set and corresponds to step 1.

[0145] Data preprocessing module: Converts source domain brain diffusion magnetic resonance data and target domain brain diffusion magnetic resonance data in the training sample set into source domain RISH feature values ​​and target domain RISH feature values. This module corresponds to step 2.

[0146] Generative Adversarial Network Module: Input the source domain RISH feature values ​​and the target domain RISH feature values ​​into the generative adversarial network model for training, and output the target domain synthetic RISH feature values. This module corresponds to steps 3 and 4.

[0147] Data processing module: Calculates the RISH feature value of the output target domain to obtain the converted brain diffusion-weighted magnetic resonance imaging signal. This module corresponds to step 5.

[0148] A multi-center intracranial diffusion-weighted magnetic resonance imaging (MRI) transdomain conversion device based on generative adversarial networks includes:

[0149] Memory: for storing computer programs that implement the multicenter brain diffusion-weighted magnetic resonance imaging cross-domain conversion method based on generative adversarial networks as described in claims 1-7;

[0150] Processor: Used to implement the cross-domain conversion method for multicenter brain diffusion-weighted magnetic resonance imaging based on generative adversarial networks according to claims 1-7 when executing the computer program.

[0151] The processor referred to 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. This processor is the control center of the multi-center brain diffusion-weighted magnetic resonance imaging cross-domain conversion device based on generative adversarial networks, connecting all parts of the device via various interfaces and lines.

[0152] When the processor executes the computer program, it implements the steps of the above-described multi-center brain diffusion-weighted magnetic resonance imaging cross-domain conversion method based on generative adversarial networks, for example: acquiring a training sample set: the training sample set includes training sample groups of different cases, the training sample groups include source domain brain diffusion magnetic resonance data and target domain brain diffusion magnetic resonance data acquired by scanners at two acquisition centers for the same case; data preprocessing: calculating source domain RISH feature values ​​and target domain RISH feature values ​​based on source domain brain diffusion magnetic resonance data and target domain brain diffusion magnetic resonance data; constructing a generative adversarial network model: the generative adversarial network includes a generator, a discriminator, and a loss function; training the generative adversarial network model: inputting the source domain RISH feature values ​​and target domain RISH feature values ​​into the constructed generative adversarial network model for training, and outputting target domain synthetic RISH feature values; processing the output data: calculating the converted brain diffusion-weighted magnetic resonance imaging signal based on the target domain synthetic RISH feature values ​​and the source domain RISH feature values.

[0153] Alternatively, when the processor executes the computer program, it implements the functions of each module in the above system, for example: a data preprocessing module: converts the source domain brain diffusion magnetic resonance imaging data and the target domain brain diffusion magnetic resonance imaging data in the training sample set into source domain RISH feature values ​​and target domain RISH feature values; a generative adversarial network module: inputs the source domain RISH feature values ​​and the target domain RISH feature values ​​into a generative adversarial network model for training, and outputs target domain synthetic RISH feature values; a data processing module: calculates the output target domain synthetic RISH feature values ​​to obtain the converted brain diffusion-weighted magnetic resonance imaging signal.

[0154] For example, the computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing preset functions. These instruction segments describe the execution process of the computer program in the multi-center brain diffusion-weighted magnetic resonance imaging cross-domain conversion device based on generative adversarial networks. For example, the computer program can be divided into a data acquisition module, a data preprocessing module, a generative adversarial network module, and a data processing module. The specific functions of each module are as follows: Data acquisition module: used to acquire a training sample set; Data preprocessing module: converts source domain brain diffusion magnetic resonance data and target domain brain diffusion magnetic resonance data in the training sample set into source domain RISH feature values ​​and target domain RISH feature values; Generative adversarial network module: inputs the source domain RISH feature values ​​and target domain RISH feature values ​​into a generative adversarial network model for training, and outputs target domain synthetic RISH feature values; Data processing module: calculates the output target domain synthetic RISH feature values ​​to obtain the converted brain diffusion-weighted magnetic resonance imaging signal.

[0155] The multi-center brain diffusion-weighted magnetic resonance imaging (MRI) transdomain conversion device based on generative adversarial networks (GANs) 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, processors and memory. Those skilled in the art will understand that the above examples of a multi-center brain diffusion-weighted MRI transdomain conversion device based on GANs do not constitute a limitation on such devices. The device may include more components than described above, or combine certain components, or use different components. For example, the device may also include input / output devices, network access devices, buses, etc.

[0156] The memory can be used to store the computer program and / or modules. The processor realizes various functions of the multi-center brain diffusion-weighted magnetic resonance imaging cross-domain conversion device based on generative adversarial networks by running or executing the computer program and / or modules stored in the memory and calling the data stored in the memory.

[0157] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function (such as sound playback or image playback). The data storage area may store data created based on the use of the mobile phone (such as audio data or phonebook entries). Furthermore, the memory may include high-speed random access memory (RAM) and non-volatile memory, such as hard disks, RAM, plug-in hard disks, SmartMediaCards (SMC), Secure Digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0158] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the multi-center brain diffusion-weighted magnetic resonance imaging cross-domain conversion method based on generative adversarial networks.

[0159] If the modules / units integrated by the multicenter brain diffusion-weighted magnetic resonance imaging cross-domain conversion system based on generative adversarial networks are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0160] This invention implements all or part of the processes in the above-described method for cross-domain conversion of multicenter brain diffusion-weighted magnetic resonance imaging based on generative adversarial networks. It can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of the above-described method for cross-domain conversion of multicenter brain diffusion-weighted magnetic resonance imaging based on generative adversarial networks. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or a preset intermediate form, etc.

[0161] The computer-readable storage medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0162] It should be noted that the content contained in the computer-readable storage medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable storage medium does not include electrical carrier signals and telecommunication signals.

Claims

1. A multi-center brain diffusion-weighted magnetic resonance image cross-domain conversion method based on a generative adversarial network, characterized in that, The method comprises the following steps: Step 1: obtaining a training sample set, wherein the training sample set comprises training sample groups of different cases, and each training sample group comprises source domain brain diffusion magnetic resonance data and target domain brain diffusion magnetic resonance data of a same case collected by using scanners in two collection centers; Step 2: data preprocessing, wherein source domain RISH characteristic values and target domain RISH characteristic values are calculated according to the source domain brain diffusion magnetic resonance data and the target domain brain diffusion magnetic resonance data obtained in step 1; Step 3: constructing a generative adversarial network model, wherein the generative adversarial network comprises a generator, a discriminator and a loss function; The discriminator in step 3 is a Markov discriminator capable of realizing adaptive weight allocation, and the specific method of adaptive weight allocation is as follows: Adaptively calculate the number n of voxels representing brain regions contained in the current image patch based on brain region masks in the dataset. brain And the total number of voxels n in the current image patch. sum And calculate n brain With n sum The ratio of w i As the initial weights for the current image patch: In the formula, n brain represents the number of voxel points of the brain region contained in the current image block, n sum represents the total number of voxel points in the current image block, w i represents the initial weight of the current image block; The sum of the weights of each image block is 1, the initial weights are obtained, the initial weights are normalized by using a softmax function, and the final adaptive weights of each image block are obtained: In the formula, w i denotes the final adaptive weight of each image block; The loss function in step 3 comprises a first loss function, a second loss function and a third loss function, the first loss function and the second loss function are loss functions of the generator, and the third loss function is a loss function of the discriminator; The first loss function is the difference between the target domain RISH characteristic values and the prediction results of the generator: In the formula, y represents the target domain RISH characteristic values, and y' represents the prediction results of the generator; The second loss function is the difference between the prediction results of the generator and the segmentation results of the images with the order of 0 of the target domain RISH characteristic values: In the formula, seg target The inter-image segmentation result of the target domain RISH characteristic value order 0 is represented as seg harmonized The inter-image segmentation result of the generator prediction result RISH characteristic value order 0 is represented as seg The third loss function is a Wasserstein loss function based on gradient penalty, and the Lipschitz constant K is taken as 1, the Lipschitz continuity of the discriminator is constrained by the gradient penalty, and the specific calculation method is as follows: where f w (·) denotes the continuous function fitted by the parametrized discriminator, denotes p g and p r the random interpolation sampling distribution on the line segment, specifically, let x r ~ p r , x g ~ p g , and ε ~ Uniform[0, 1], then: The loss function for training the generative adversarial network is as follows: Step 4: training the generative adversarial network model, wherein the source domain RISH characteristic values and the target domain RISH characteristic values obtained in step 2 are input into the generative adversarial network model constructed in step 3 for training, and target domain synthetic RISH characteristic values are output; Step 5: processing the output data, wherein the converted brain diffusion weighted magnetic resonance imaging signals are calculated according to the target domain synthetic RISH characteristic values obtained in step 4 and the source domain RISH characteristic values obtained in step 2.

2. The multi-center brain diffusion weighted magnetic resonance image cross-domain conversion method based on a generative adversarial network according to claim 1, characterized in that, The specific process of step 2 is as follows: Step 2.1: calculating source domain spherical harmonic coefficients and target domain spherical harmonic coefficients by using the source domain brain diffusion magnetic resonance data and the target domain brain diffusion magnetic resonance data obtained in step 1, wherein the spherical harmonic coefficients can be solved by the following formula: wherein is the direction vector in spherical coordinates, is the spherical harmonic function in coordinates , l is the angular frequency of the spherical harmonic function, and m is the phase, is the original acquired brain diffusion weighted function; Step 2.2: calculating source domain RISH characteristic values and target domain RISH characteristic values by using the source domain spherical harmonic coefficients and the target domain spherical harmonic coefficients obtained in step 2.1, and the specific expression formula is as follows: where ||c l || 2 denotes the RISH eigenvalue at angular frequency l, denotes the spherical harmonic coefficients of the m phase at frequency l.

3. The multi-center brain diffusion weighted magnetic resonance image cross-domain conversion method based on a generative adversarial network according to claim 1, characterized in that, The generator in step 3 comprises a plurality of convolution modules, and the convolution layers in the convolution modules all use three-dimensional convolution kernels.

4. The multi-center brain diffusion weighted magnetic resonance image cross-domain conversion method based on a generative adversarial network according to claim 1, characterized in that, The specific process of step 4 is as follows: Step 4.1: fixing the discriminator parameters, inputting the source domain RISH eigenvalues obtained in step 2 into the generator to obtain generated homogenized RISH eigenvalues; Step 4.2: fixing the generator parameters, inputting the homogenized RISH eigenvalues obtained in step 4.1 and the target domain RISH eigenvalues obtained in step 2 into the discriminator to perform local image block discrimination, and outputting the authenticity discrimination results of the local image blocks; Step 4.3: updating the network parameters of the generative adversarial network using the loss function set in step 3; Step 4.4: repeating steps 4.1-4.3 until the preset training times are reached or the error between the generated results and the target domain RISH eigenvalues reaches the preset value; Step 4.5: using the generator fixed parameters to output the source domain homogenized target domain synthetic RISH eigenvalues.

5. The multi-center brain diffusion weighted magnetic resonance image cross-domain conversion method based on a generative adversarial network according to claim 1, characterized in that, The specific process of step 5 is: Step 5.1: Calculate the target mapping coefficients l by using the target domain synthesized RISH feature values from step 4 output and the source domain RISH feature values from step 2 s : Target mapping coefficient l s : In the formula, l s denotes the target mapping coefficient, Π(||c i || 2 ) denotes the target domain synthesis RISH feature value, ||c i || 2 denotes the source domain RISH feature value; Step 5.2: Obtain the target mapping coefficient l using the result of step 5.1 s , compute the target spherical harmonic coefficients needed to reconstruct the brain diffusion weighted magnetic resonance imaging signal : wherein denotes the converted target spherical harmonic coefficients, denotes the spherical harmonic coefficients before conversion, l s denotes the target mapping coefficients; Step 5.3: calculating the converted brain diffusion weighted magnetic resonance imaging signal using the target spherical harmonic coefficients obtained in step 5.2 and the gradient direction information of the target domain brain diffusion magnetic resonance data: wherein represents a direction vector in spherical coordinates, represents a converted target spherical harmonic coefficient, represents a spherical harmonic function in coordinates l represents an angular frequency of the spherical harmonic function, m represents a phase, and n represents a truncation order set when the spherical harmonic coefficient is calculated.

6. The multi-center brain diffusion-weighted magnetic resonance image cross-domain conversion system based on the generative adversarial network based on the method of claim 1, characterized in that, It includes: Data acquisition module: used for acquiring a training sample set, wherein the training sample set includes training sample groups of different cases, and each training sample group includes source domain brain diffusion magnetic resonance data and target domain brain diffusion magnetic resonance data of the same case collected by a scanner at two acquisition centers; Data preprocessing module: converting the source domain brain diffusion magnetic resonance data and the target domain brain diffusion magnetic resonance data in the training sample set into source domain RISH eigenvalues and target domain RISH eigenvalues; Generative adversarial network module: inputting the source domain RISH eigenvalues and the target domain RISH eigenvalues into a generative adversarial network model for training, and outputting target domain synthetic RISH eigenvalues; Data processing module: calculating the output target domain synthetic RISH eigenvalues to obtain converted brain diffusion weighted magnetic resonance imaging signals.

7. A multi-center brain diffusion-weighted magnetic resonance image cross-domain conversion device based on a generative adversarial network, characterized in that, It includes: Memory: used for storing a computer program for implementing the multi-center brain diffusion weighted magnetic resonance image cross-domain conversion method based on the generative adversarial network according to any one of claims 1-5; Processor: used for executing the computer program to implement the multi-center brain diffusion weighted magnetic resonance image cross-domain conversion method based on the generative adversarial network according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, It includes: The computer readable storage medium stores a computer program, and the computer program is executed by the processor to implement the multi-center brain diffusion weighted magnetic resonance image cross-domain conversion method based on the generative adversarial network according to any one of claims 1-5.

Images