Multi-site brain image index data coordination system based on deep learning

Abstract

The invention discloses a multi-site brain image index data coordination system based on deep learning, and the system comprises an fMRI three-dimensional feature index preprocessing module which carries out the standardized preprocessing and metadata protection of the fMRI three-dimensional feature index, so as to meet the requirements of a deep learning model for the matching of input and output dimensions. And the bidirectional coordination model construction module and the bidirectional coordination model training module are used for constructing an unsupervised cross-site three-dimensional bidirectional coordination model, realizing domain invariant feature decoupling through non-paired cross-site data confrontation training, completing X-to-Y or Y-to-X conversion of a data set, and eliminating a site effect between X and Y. And the NIfTI format file reconstruction module performs NIfTI format file reconstruction on the coordinated data to generate a standardized output file strictly compatible with the original data structure. According to the system, the site effect caused by scanning parameter heterogeneity can be eliminated, and meanwhile, the topological structure and metadata integrity of the original medical image file are completely reserved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical artificial intelligence brain imaging, and particularly relates to a multi-site brain imaging index data coordination system based on deep learning. Background Art

[0002] Feature indexes such as regional homogeneity (ReHo) derived from resting-state functional magnetic resonance imaging (resting-state fMRI) have become important indexes in multi-center imaging research of brain science due to their biological specificity in characterizing spontaneous neural activity characteristics. However, the heterogeneity of cross-site scanning parameters will form a significant site-dominated distribution shift in the feature space of feature indexes, which is called the "site effect". This site effect not only interferes with the extraction of physiological-related features and the analysis of inter-group differences as a strong confounding factor, but also leads to a decline in the cross-site generalization performance of classification models.

[0003] Currently, the domain adaptation method based on generative adversarial networks provides a new idea for alleviating the heterogeneity of medical image data. It has been maturely applied in the style conversion of two-dimensional natural images and shows great potential in mining the non-linear features of data. How to use deep neural networks to mine deep non-linear features and construct a three-dimensional bidirectional (i.e., site X to Y and site Y to X coordination) coordination model for unpaired fMRI feature indexes - establish a cross-site bidirectional mapping function through adaptive feature decoupling Eliminating the site effect while completely retaining the original data structure is an urgent problem to be solved. Summary of the Invention

[0004] Aiming at the problems existing in the prior art, the present invention provides a multi-site brain imaging index data coordination system based on deep learning, which can eliminate the site effect caused by the heterogeneity of scanning parameters while completely retaining the topological structure and metadata integrity of the original medical image file.

[0005] To solve the above technical problems, the present invention provides the following technical solutions: A multi-site brain imaging index data coordination system based on deep learning, comprising:

[0006] An fMRI three-dimensional feature index preprocessing module, which is used to define the source domain X and the target domain Y as the image acquisition site data sets with device parameter differences; based on the received source domain X and target domain Y, perform cross-site data loading and parsing, brain region mask generation, and coordination model input construction, and construct an efficient data stream adapted to the encoding and decoding architecture of the bidirectional coordination model;

[0007] The two-way coordination model construction module uses the source domain input X and the target domain input Y obtained by the fMRI three-dimensional feature index preprocessing module as data carriers to construct a two-way coordination transformation model with four-order collaborative effects of an encoding-decoding coordination module, a discriminant module, a distribution calculation module, and a feature matching module. Through the encoding-decoding coordination module and the discriminant module, unpaired cross-site data coordination asynchronous adversarial training is carried out, and the distribution calculation module, the feature matching module, and the cycle consistency constraint synchronously perform multi-level constraints on the encoding-decoding coordination module; train the encoding-decoding coordination module to achieve domain-invariant feature decoupling and two-way reversible mapping, eliminate the site effect and retain the original data topology structure;

[0008] The two-way coordination model training module trains the two-way coordination model according to the source domain input X and the target domain input Y obtained by the fMRI three-dimensional feature index preprocessing module. After the training is completed, the parameters of the encoding-decoding coordination module are fixed to obtain the coordination matrices of the source domain and the target domain;

[0009] The fMRI three-dimensional feature index NIfTI format file reconstruction module reconstructs the coordination matrix output by the two-way coordination model training module into a standardized three-dimensional feature index NIfTI format file and saves it through reverse space adaptation and topology restoration.

[0010] Furthermore, in the aforementioned fMRI three-dimensional feature index preprocessing module, the cross-site data loading and parsing are specifically as follows: load the fMRI three-dimensional feature index NIfTI format files of X and Y, and accurately extract and independently store the following elements through the medical image analysis tool NiBabel: the original three-dimensional feature index matrix V∈R D×H×W , where D represents depth, H represents height, and W represents width; the spatial positioning parameter: the affine matrix A; the NIfTI header file metadata;

[0011] In the fMRI three-dimensional feature index preprocessing module, the binary brain region mask is generated through the reference value 0 to generate the binary brain region mask

[0012] M∈{0,1} D×H×W The formula is as follows:

[0013]

[0014] In the formula, V ijk represents the voxel value at the position [i, j, k] in the matrix V;

[0015] In the fMRI three-dimensional feature index preprocessing module, the coordination model input is constructed as follows:

[0016] X = Padding / Cropping(V X )⊙M X

[0017] Y = Padding / Cropping(V Y ) ⊙ M Y

[0018] where Padding / Cropping refers to, according to the set target dimensions (D ′ , H ′ , W′), when D < D′, or H < H ′ , W < W′, padding the matrix V with zeros along the edges; when D > D′ or H > H ′ , W > W′, performing edge cropping on the matrix V; M is the brain region mask; ⊙ is element-wise multiplication.

[0019] Furthermore, in the aforementioned bidirectional coordination model construction module, the encoding-decoding coordination module includes two three-dimensional bidirectional coordination mapping models H X→Y and H Y→X ; H X→Y from the source domain input X to the coordinated output Y ′ = H X→Y (X), which successively includes an input layer, an encoder, a decoder, and an output layer. Among them, the encoder and the decoder are symmetric structures;

[0020] The formula for the input layer is as follows:

[0021] X H_in = 3DConv(X)

[0022] where X is the source domain input obtained by the fMRI three-dimensional feature index preprocessing module, and 3DConv is the three-dimensional convolutional layer;

[0023] The encoder includes four downsampling layers. The first three layers are composed of an activation function, a three-dimensional convolutional layer, and an instance normalization layer, and the innermost layer is composed of an activation function and a three-dimensional convolutional layer. The formula is as follows:

[0024]

[0025] where ReLU is the activation function, IN is the instance normalization layer, X H_in is the output of the input layer, is the feature map output of the m-th layer of the encoder;

[0026] The decoder includes a total of four upsampling layers, which are composed of an activation function, a three-dimensional transposed convolutional layer, and an instance normalization layer. The formula is as follows:

[0027]

[0028] where 3DTransConv is the three-dimensional transposed convolutional layer, LeackyReLU is the activation function, is the input feature map of the n-th layer of the decoder, is the output feature map of the n-th layer of the decoder;

[0029] When n≠1, it is obtained by concatenating the output feature map of the (n - 1)-th layer of the decoder and the corresponding input feature map of the n-th layer of the encoder in the channel dimension. The formula is as follows:

[0030]

[0031] where Concat is the concatenation operation in the channel dimension;

[0032] The output layer performs transposed convolution on the output feature map of the decoder and performs a non-linear transformation. The formula is as follows:

[0033]

[0034] where Tahn is the hyperbolic tangent activation function, and y ′ is the coordinated output of X→Y obtained by inputting X to H X→Y ;

[0035] H Y→X From the input Y in the target domain to the coordinated output X ′ =H Y→X (Y), successively including an input layer, an encoder, a decoder, and an output layer. Among them, the encoder and the decoder are symmetric structures;

[0036] The formula of the input layer is as follows:

[0037] Y H_in =3DConv(Y)

[0038] where Y is the input in the target domain obtained by the fMRI three-dimensional feature index preprocessing module;

[0039] The encoder contains four downsampling layers. The first three layers are composed of an activation function, a three-dimensional convolutional layer, and an instance normalization layer. The innermost layer is composed of an activation function and a three-dimensional convolutional layer. The formula is as follows:

[0040]

[0041] where Y H_in is the output of the input layer, is the feature map output of the m-th layer of the encoder;

[0042] The decoder contains a total of four upsampling layers, which are composed of an activation function, a three-dimensional transposed convolutional layer, and an instance normalization layer. The formula is as follows:

[0043]

[0044] where, is the input feature map of the n-th layer of the decoder, is the output feature map of the n-th layer of the decoder;

[0045] When n≠1, is obtained by concatenating the output feature map of the (n-1)-th layer of the decoder and the corresponding input feature map of the n-th layer of the encoder in the channel dimension. The formula is as follows:

[0046]

[0047] The output layer performs transposed convolution on the output feature map of the decoder and performs a non-linear transformation. The formula is as follows:

[0048]

[0049] where X ′ is the coordinated output of Y→X obtained by inputting H into Y Y→X ;

[0050] Furthermore, in the aforementioned bidirectional coordination model construction module, the discrimination module includes two double-domain discrimination models D X , D Y ;

[0051] D X goes from the source domain input X to the output local discrimination result three-dimensional feature map D X (X), and successively includes an input layer, a downsampling layer, and an output layer.

[0052] The formula of the input layer is as follows:

[0053] d in = LeakyReLU(3DConv(X))

[0054] The formula of the downsampling layer is as follows:

[0055] d fe = LeakyReLU.BN(3DConv(d in )) /

[0056] where BN is the batch normalization layer;

[0057] The formula of the output layer is as follows:

[0058] d X = 3DConv(d fe )

[0059] where d x is the local discrimination result three-dimensional feature map.

[0060] D YFrom the input Y in the target domain to the output three-dimensional feature map D of the local discrimination result Y (Y), which successively includes an input layer, a downsampling layer, and an output layer.

[0061] The formula for the input layer is as follows:

[0062] d in = LeakyReLU(3DConv(Y))

[0063] The formula for the downsampling layer is as follows:

[0064] d fe = LeakyReLU.BN(3DConv(d in )) /

[0065] where BN is the batch normalization layer;

[0066] The formula for the output layer is as follows:

[0067] d Y = 3DConv(d fe )

[0068] where d Y is the three-dimensional feature map of the local discrimination result;

[0069] Furthermore, in the aforementioned bidirectional coordination model construction module, the distribution calculation module guides the distribution consistency within the implementation site in the encoding and decoding coordination module. This module is cascaded with the encoding and decoding coordination module, receiving the original input and coordinated output of the encoding and decoding coordination module; realizing distribution alignment by implicitly mapping both from the original space to the infinite-dimensional reproducing kernel Hilbert space through a multi-scale Gaussian kernel function; calculating the similarity within and between groups of both in the feature space; and obtaining the distribution difference between the original input and coordinated output of the encoding and decoding coordination module by calculating the mean embedding distance through weighted sum. The formula is as follows:

[0070] Dd X 2 = ‖E[k(X,·)] - E[k(H Y→X (Y),·)]‖ 2

[0071] Dd Y 2 = ‖E[k(Y,·)] - E[k(H X→Y (X),·)]‖ 2

[0072] where k() is the radial basis function kernel;

[0073] Furthermore, in the aforementioned two-way coordination model construction module, the feature matching module preserves the in-site fidelity of the coordinated fMRI three-dimensional indicators in terms of complex structure and spatial correlation between voxels through a high-dimensional feature alignment mechanism;

[0074] The feature matching module sequentially passes through a feature extraction layer and a feature matching layer from input to output, cascades with the discrimination module to extract intermediate layer feature maps, and calculates the feature matching degree using the mean absolute error. The formula is as follows:

[0075] FM X =‖D X (l) (H Y→X (Y)) - D X (l) (X)‖1

[0076] FM Y =‖D Y (l) (H X→Y (X)) - D Y (l) (Y)‖1

[0077] where l is the l-th layer of the discrimination model, and ‖‖1 represents the calculation of the mean absolute error.

[0078] Furthermore, in the aforementioned two-way coordination model construction module, the encoder-decoder coordination module performs unpaired cross-site data coordination asynchronous adversarial training with the discrimination module, and the adversarial loss is used as follows:

[0079] L adv (H X→Y , H Y→X , D X , D Y ) = L adv_H + L adv_D

[0080] The encoder-decoder coordination module makes the discrimination module misjudge the coordinated output domain by improving its own coordination ability, that is, it is considered that the coordinated source domain data belongs to the target domain; the formula is as follows:

[0081] L advH (H X→Y , H Y→X ) = MSE(D X (H Y→X (Y)), 1) + MSE(D Y (H X→Y ), 1)

[0082] Among them, 1 indicates belonging to the X / Y domain, 0 indicates not belonging to the X / Y domain, 1 and 0 are three-dimensional numerical matrices with the same dimension as the output of the discrimination module, and MSE is the mean square error;

[0083] By improving its own discrimination ability, the discrimination module correctly discriminates the original input and the coordinated output domain, that is, it is considered that both the original input and the coordinated output belong to the original domain. The formula is as follows:

[0084] L advD (D X ,D Y ) = MSE(D X (X), 1) + MSE(D X (H Y→X (Y)), 0) + MSE(D Y (Y), 1)

[0085] + MSE(D Y (H X→Y ), 0)

[0086] The encoding-decoding coordination module and the discrimination module play a game with each other, and finally maximize the performance under the balance condition.

[0087] Furthermore, in the aforementioned bidirectional coordination model construction module, the distribution calculation module constrains the encoding-decoding coordination module through the distribution alignment loss. The formula is as follows:

[0088] L Dd (H X→Y ,H Y→X ) = Dd X 2 + Dd Y 2

[0089] Among them, Dd X 2 、Dd Y 2 are the outputs obtained by the distribution calculation module for the original input and the output of the coordination module;

[0090] The feature matching module constrains the encoding-decoding coordination module through the feature matching loss. The formula is as follows:

[0091] L FM (H X→Y ,H Y→X ) = FM X + FM Y

[0092] Among them, FM X 、FM Y are the feature matching degrees of the intermediate layer of the discrimination module output by the feature matching module;

[0093] The cyclic consistency loss formula is as follows:

[0094] L Cyc (H X→Y , H Y→X ) = ‖H Y→X (H X→Y (X)) - X‖1 + ‖H X→Y (H Y→X (Y)) - Y‖1

[0095] Among them, the encoding - decoding coordination module outputs to the cyclic input reverse encoding - decoding coordination module for cyclic cross - domain inverse mapping reconstruction, and performs cyclic consistency constraints on the reconstructed data and the original input through voxel - level alignment;

[0096] Furthermore, for the aforementioned bidirectional coordination model training module, the parameters of the encoding - decoding coordination module and the discriminator module are trained asynchronously. During the optimization stage of the encoding - decoding coordination module, the parameters of the discriminator module are fixed, and the n:1 asynchronous training method with the encoding - decoding module taking precedence is executed. The parameters of the encoding - decoding coordination module are iterated n times, and the parameters of the discriminator module are iterated 1 time. The model objective function is as follows:

[0097]

[0098] In the formula, λ1, λ2, λ3, and λ4 are the coefficients of the adversarial loss, distribution alignment loss, feature matching loss, and cyclic consistency loss respectively.

[0099] Furthermore, the aforementioned fMRI three - dimensional feature index NIfTI - format file reconstruction module is configured to perform the following steps:

[0100] S1. Perform symmetric spatial dimension adjustment on the coordination matrix output by the bidirectional coordination model training module: When D < D′, or H < H ′ , W < W′, perform edge cropping; when D > D′, or H > H ′ , W > W′, perform zero - value padding along the edges;

[0101] S2. Perform voxel - level masking based on the brain region mask M and restore the original coordinate system based on the affine matrix A. The formula is as follows:

[0102] V ′ =(M X ⊙X ′ )·A

[0103] Among them, · represents affine transformation;

[0104] S3. Normalize the matrix V ′Fuse with the original header file metadata extracted by the fMRI three-dimensional feature index preprocessing module to generate a NIfTI file that conforms to medical imaging standards and completes data coordination to eliminate site effects;

[0105] S4. Cascade the fMRI three-dimensional feature index preprocessing and metadata protection in the fMRI three-dimensional feature index preprocessing module, the fixed codec coordination module parameters obtained from the two-way coordination model training module, and the fMRI three-dimensional feature index NIfTI format file reconstruction to obtain a multi-site brain imaging index data coordination system based on deep learning.

[0106] Compared with the prior art, the beneficial technical effects of the present invention adopting the above technical solutions are as follows:

[0107] (1) The present invention proposes a multi-site brain imaging index data coordination system based on deep learning, which can realize the automatic directional coordination of relevant indicators for site effects. This system supports the independent processing of single-sample data without relying on the unified coordination of multi-site batch data, significantly improving the data processing efficiency;

[0108] (2) For the fMRI three-dimensional feature index, the present invention constructs a standardized processing flow for the deep learning coordination network, including a two-stage processing mechanism of input data preprocessing and output result postprocessing. On the one hand, it effectively improves the adaptability of data features to the codec network architecture. On the other hand, it generates a coordinated correction file that removes site effects while completely retaining the original metadata and data structure, which is beneficial for directly supporting downstream tasks such as subsequent multi-center joint analysis;

[0109] (3) For the non-paired cross-center data features of medical images, the present invention constructs an unsupervised cross-site three-dimensional two-way coordination model. Based on the classical adversarial training mechanism and the idea of cycle consistency constraint, it integrates the distribution calculation module and the feature matching module to realize the asynchronous training of the codec module and the discriminant module, effectively capture the non-linear dependence relationship in the three-dimensional space, establish a cross-site two-way mapping, and achieve more accurate data coordination to eliminate site effects. Brief Description of the Drawings

[0110] Figure 1 It is a schematic diagram of a multi-site brain imaging index data coordination system based on deep learning.

[0111] Figure 2 It is a schematic diagram of the structure and training of the two-way coordination conversion model.

[0112] Figure 3 It is a schematic diagram of the internal module structure of the two-way coordination conversion model. In the figure, (a) is a schematic diagram of the codec coordination module structure, (b) is a schematic diagram of the discriminant module structure, (c) is a schematic diagram of the distribution calculation module structure, and (d) is a schematic diagram of the feature matching module structure.

[0113] Figure 4 It is a schematic diagram of the test results of a multi-site brain imaging index data coordination system based on deep learning. Detailed implementation manners

[0114] In order to better understand the technical content of the present invention, specific embodiments are given below in conjunction with the accompanying drawings for illustration.

[0115] In the present invention, various aspects of the present invention are described with reference to the accompanying drawings, and many illustrative embodiments are shown in the drawings. The embodiments of the present invention are not limited to those described in the drawings. It should be understood that the present invention can be implemented by any one of the various concepts and embodiments introduced above, and the concepts and embodiments described in detail below, because the concepts and embodiments disclosed in the present invention are not limited to any embodiment. In addition, some aspects disclosed in the present invention can be used alone, or in any suitable combination with other aspects disclosed in the present invention.

[0116] The present invention provides a multi-site brain imaging index data coordination system based on deep learning, which uses fMRI three-dimensional feature indexes from the source domain and the target domain to perform data coordination from the source domain to the target domain and from the target domain to the source domain. First, the input fMRI-derived three-dimensional feature indexes are preprocessed as the model input and the metadata is saved. Then, a bidirectional coordination conversion model is trained using the dual-domain input data. Finally, the output of the bidirectional coordination conversion model is reconstructed into a NIfTI format file to obtain a coordinated file without site effects. The deep learning data coordination model is trained with a training set and the parameters of the encoding and decoding coordination modules in the model are saved and applied to the system. The coordination effect of the model and the system is tested with a test set.

[0117] Figure 1 It is a schematic diagram of the system of the present invention. The system of the present invention includes:

[0118] An fMRI three-dimensional feature index preprocessing module, which is used to define the source domain X and the target domain Y as image acquisition site data sets with different device parameters; based on the received source domain X and target domain Y, perform cross-site data loading and parsing, brain region mask generation, and construction of the coordination model input to construct an efficient data stream suitable for the encoding and decoding architecture of the bidirectional coordination model;

[0119] The two-way coordination model construction module uses the source domain input X and the target domain input Y obtained by the fMRI three-dimensional feature index preprocessing module as data carriers to construct a two-way coordination transformation model with four-order collaborative effects of an encoding-decoding coordination module, a discriminant module, a distribution calculation module, and a feature matching module. Through the encoding-decoding coordination module and the discriminant module, unpaired cross-site data coordination asynchronous adversarial training is carried out, and the distribution calculation module, the feature matching module, and the cycle consistency constraint synchronously perform multi-level constraints on the encoding-decoding coordination module; train the encoding-decoding coordination module to achieve domain-invariant feature decoupling and two-way reversible mapping, eliminate the site effect and retain the original data topology structure;

[0120] The two-way coordination model training module trains the two-way coordination model according to the source domain input X and the target domain input Y obtained by the fMRI three-dimensional feature index preprocessing module. After the training is completed, the parameters of the encoding-decoding coordination module are fixed to obtain the coordination matrices of the source domain and the target domain;

[0121] The fMRI three-dimensional feature index NIfTI format file reconstruction module reconstructs the coordination matrix output by the two-way coordination model training module into a standardized three-dimensional feature index NIfTI format file and saves it through reverse space adaptation and topology restoration.

[0122] Furthermore, as a preferred embodiment of the present invention, in the fMRI three-dimensional feature index preprocessing module, the cross-site data loading and parsing are specifically as follows: load the fMRI three-dimensional feature index NIfTI format files of X and Y, and accurately extract and independently store the following elements through the medical image analysis tool NiBabel: the original three-dimensional feature index matrix V∈R D×H×W , where D represents the depth, H represents the height, and W represents the width; the spatial positioning parameter: the affine matrix A; the NIfTI header file metadata;

[0123] The binary brain region mask is generated through the reference value 0 in the brain region mask generation process

[0124] M∈{0,1} D×H×W The formula is as follows:

[0125]

[0126] In the formula, V ijk represents the voxel value at the position [i, j, k] in the matrix V;

[0127] The coordination model input is constructed as follows:

[0128] X = Padding / Cropping(V X )⊙M X

[0129] Y = Padding / Cropping(VY ) ⊙M Y

[0130] Among them, Padding / Cropping means that according to the set target dimensions (D ′ , H ′ , W′), when D < D′, or H < H ′ , W < W′, zero-padding is performed on the edges of the matrix V; when D > D′ or H > H ′ , W > W′, edge cropping is performed on the matrix V; M is a brain region mask; ⊙ is element-wise multiplication.

[0131] In this embodiment, the fMRI three-dimensional feature index preprocessing module performs data coordination on the fMRI-derived three-dimensional feature index, which is the regional homogeneity (ReHo) index. The source domain X is a dataset collected using a GE device, with a total of 66 healthy subjects; the target domain Y is a dataset collected using a SIEMENS device, with a total of 126 healthy subjects. There is a significant site effect between X and Y. The ReHo index is divided into time periods with a window length of 30 and a step size of 10 by sliding windows, resulting in augmented datasets X and Y, with sample sizes of 1452 and 1512 respectively.

[0132] The system receives X and Y, obtains the 61×73×61 matrix of the original ReHo through data loading and parsing, generates a brain region mask based on this; saves the affine matrix and NIfTI header file metadata; performs padding and cropping on the original matrix to obtain the 72×72×72 matrices X and Y of ReHo as the input to the coordination model.

[0133] As Figure 2 shown, the bidirectional coordination model construction module includes an encoding-decoding coordination module, a discriminant module, a distribution calculation module, and a feature matching module. Using the source domain input X and target domain input Y obtained by the fMRI three-dimensional feature index preprocessing module as data carriers, non-paired cross-site data coordination asynchronous adversarial training is performed through the encoding-decoding coordination module and the discriminant module. The distribution calculation module, the feature matching module, and the cycle consistency constraint perform multi-level constraints on the encoding-decoding coordination module synchronously. The encoding-decoding coordination module is trained to achieve domain-invariant feature decoupling and bidirectional reversible mapping, ultimately eliminating the site effect and retaining the original data topological structure. After the model training is completed, the parameters of the encoding-decoding coordination module are fixed for system construction;

[0134] As Figure 3 shown in (a) of X→Y and H Y→XFrom the input to the output, it sequentially includes an input layer, an encoder, a decoder, and an output layer. Among them, the encoder and the decoder are symmetric structures. The number of convolutional kernels in the input layer is 64, and the size of the convolutional kernels is 4x4x4. The number of convolutional kernels in the downsampling layers of the encoder is sequentially 128, 256, 512, 512, and the sizes of the convolutional kernels are 4x4x4, 4x4x4, 3x3x3, 3x3x3 respectively. The decoder has a symmetric structure with the encoder. The number of convolutional kernels in the upsampling layers is sequentially 512, 256, 128, 64, and the sizes of the convolutional kernels are 3x3x3, 3x3x3, 4x4x4, 4x4x4 respectively. The number of convolutional kernels in the output layer is 1, and the size of the convolutional kernels is 4x4x4. Except for the innermost layer, the output feature map of the (n - 1)-th layer of the decoder and the corresponding input feature map of the n-th layer of the encoder are concatenated in the channel dimension and input into the n-th layer of the decoder. All downsampling layers use ReLU as the activation function, and all upsampling layers use LeackyReLU as the activation function. Instance normalization is used.

[0135] H X→Y From the source domain input X to the coordinated output Y ′ = H X→Y (X), which sequentially includes an input layer, an encoder, a decoder, and an output layer. The formula for the input layer is as follows:

[0136] X H_in = 3DConv(X)

[0137] Among them, X is the source domain input obtained by the fMRI three-dimensional feature index preprocessing module, and 3DConv is the three-dimensional convolutional layer;

[0138] The encoder contains four downsampling layers. The first three layers are composed of an activation function, a three-dimensional convolutional layer, and an instance normalization layer, and the innermost layer is composed of an activation function and a three-dimensional convolutional layer. The formula is as follows:

[0139]

[0140] Among them, X H_in is the output of the input layer, is the feature map output of the m-th layer of the encoder;

[0141] The decoder contains a total of four upsampling layers, which are composed of an activation function, a three-dimensional transposed convolutional layer, and an instance normalization layer. The formula is as follows:

[0142]

[0143] Among them, is the input feature map of the n-th layer of the decoder, is the output feature map of the n-th layer of the decoder;

[0144] When n ≠ 1, The output feature map of the (n-1)-th layer of the decoder and the corresponding input feature map of the n-th layer of the encoder are concatenated in the channel dimension, and the formula is as follows:

[0145]

[0146] The output layer performs transposed convolution on the output feature map of the decoder and performs a non-linear transformation, and the formula is as follows:

[0147]

[0148] Among them, Y ′ is the coordinated output of X→Y obtained by inputting X into H X→Y ;

[0149] H Y→X from the input Y in the target domain to the coordinated output X ′ =H Y→X (Y), and successively includes an input layer, an encoder, a decoder, and an output layer. The formula of the input layer is as follows:

[0150] Y H_in =3DConv(Y)

[0151] Among them, Y is the input in the target domain obtained by the fMRI three-dimensional feature index preprocessing module;

[0152] The encoder includes four downsampling layers. The first three layers are composed of an activation function, a three-dimensional convolutional layer, and an instance normalization layer, and the innermost layer is composed of an activation function and a three-dimensional convolutional layer. The formula is as follows:

[0153]

[0154] Among them, Y H_in is the output of the input layer, is the feature map output of the m-th layer of the encoder;

[0155] The decoder includes a total of four upsampling layers, which are composed of an activation function, a three-dimensional transposed convolutional layer, and an instance normalization layer. The formula is as follows:

[0156]

[0157] Among them, is the input feature map of the n-th layer of the decoder, is the output feature map of the n-th layer of the decoder;

[0158] When n≠1, The output feature map of the (n-1)-th layer of the decoder and the corresponding input feature map of the n-th layer of the encoder are concatenated in the channel dimension, and the formula is as follows:

[0159]

[0160] The output layer performs transposed convolution on the decoder output feature map and performs a non - linear transformation. The formula is as follows:

[0161]

[0162] Where X ′ is the coordinated output of Y→X obtained by inputting H Y→X ;

[0163] As shown in (b) of Figure 3 , the discriminant module contains two double - domain discriminant models D X , D Y with the same structure. From the input to the output of the three - dimensional feature map of the local discriminant result, it successively includes an input layer, a downsampling layer, and an output layer. The number of convolution kernels of the three three - dimensional convolution layers is 64, 64, and 1 respectively, the convolution kernel size is 4x4x4, and the size of the output three - dimensional feature map of the local discriminant result is 9×9×9.

[0164] D X From the source domain input X to the output of the three - dimensional feature map of the local discriminant result D X (X), it successively includes an input layer, a downsampling layer, and an output layer.

[0165] The formula of the input layer is as follows:

[0166] d in = LeakyReLU(3DConv(X))

[0167] The formula of the downsampling layer is as follows:

[0168] d fe = LeakyReLU.BN(3DConv(d in )) /

[0169] Where BN is the batch normalization layer;

[0170] The formula of the output layer is as follows:

[0171] d X = 3DConv(d fe )

[0172] Where d X is the three - dimensional feature map of the local discriminant result.

[0173] D Y From the target domain input Y to the output of the three - dimensional feature map of the local discriminant result D Y (Y), it successively includes an input layer, a downsampling layer, and an output layer.

[0174] The formula of the input layer is as follows:

[0175] d in = LeakyReLU(3DConv(Y))

[0176] The formula for the downsampling layer is as follows:

[0177] d fe = LeakyReLU.BN(3DConv(d in )) /

[0178] where BN is the batch normalization layer;

[0179] The formula for the output layer is as follows:

[0180] d Y = 3DConv(d fe )

[0181] where d Y is the three-dimensional feature map of the local discrimination result.

[0182] The framework structure of the distribution calculation module is as shown in (c) of Figure 3 . The distribution calculation module guides the distribution consistency within the implementation site in the encoding and decoding coordination module. This module is cascaded with the encoding and decoding coordination module, receiving the original input and coordinated output of the encoding and decoding coordination module; realizing distribution alignment by implicitly mapping both from the original space to the infinite-dimensional reproducing kernel Hilbert space through a multi-scale Gaussian kernel function; calculating the similarity within and between groups of both in the feature space; and obtaining the distribution difference between the original input and coordinated output of the encoding and decoding coordination module by calculating the mean embedding distance through weighted sum based on the mean difference of the similarities. The formula is as follows:

[0183] Dd X 2 = ‖E[k(X,·)] - E[k(H Y→X (Y),·)]‖ 2

[0184] Dd Y 2 = ‖E[k(Y,·)] - E[k(H X→Y (X),·)]‖ 2

[0185] where k() is the radial basis function kernel, the basic bandwidth of the radial basis function kernel is dynamically calculated according to the characteristics of the input data, the number of kernels is set to 5, and the multiplier factor of the kernel bandwidth is set to 2.

[0186] The framework structure of the feature matching module is as shown in (d) of Figure 3 . The feature matching module sequentially passes through the feature extraction layer and the feature matching layer from input to output, is cascaded with the discrimination module, extracts the intermediate layer feature map, and calculates the feature matching degree. The formula is as follows:

[0187] FM X =‖D X (l) (H Y→X (Y)) - D X (l) (X)‖1

[0188] FM Y =‖D Y (l) (H X→Y (X)) - D Y (l) (Y)‖1

[0189] Where l is the l-th layer of the discrimination model.

[0190] As a preferred embodiment of the present invention, in the two-way coordination model construction module, the encoding-decoding coordination module and the discrimination module perform unpaired cross-site data coordination asynchronous adversarial training, and the formula using the adversarial loss is as follows:

[0191] L adv (H X→Y , H Y→X , D X , D Y ) = L adv_H +L adv_D

[0192] The encoding-decoding coordination module makes the discrimination module misjudge the coordinated output domain by improving its own coordination ability, that is, it is considered that the coordinated source domain data belongs to the target domain; the formula is as follows:

[0193] L advH (H X→Y , H Y→X ) = MSE(D X (H Y→X (Y)), 1) + MSE(D Y (H X→Y ), 1)

[0194] Where 1 indicates belonging to the X / Y domain, 0 indicates not belonging to the X / Y domain, 1 and 0 are three-dimensional numerical matrices with the same dimension size as the output of the discrimination module, and MSE is the mean square error;

[0195] The discrimination module makes correct judgments on the original input and the coordinated output domain by improving its own discrimination ability, that is, it is considered that both the original input and the coordinated output belong to the original domain, and the formula is as follows:

[0196] L advD (D X , D Y ) = MSE(D X(X),1)+MSE(D X (H Y→X (Y)),0)+MSE(D Y (Y),1)

[0197] +MSE(D Y (H X→Y ),0)

[0198] The encoding and decoding coordination module and the discrimination module play games with each other, and finally maximize the performance under the balanced condition.

[0199] As a preferred embodiment of the present invention, in the bidirectional coordination model construction module, the distributed calculation module constrains the encoding and decoding coordination module through the distributed alignment loss, and the formula is as follows:

[0200] L Dd (H X→Y ,H Y→X )=Dd X 2 +Dd Y 2

[0201] Where Dd X 2 、Dd Y 2 are the outputs obtained by calculating the input distribution of the original input and the output of the coordination module by the distribution calculation module;

[0202] The feature matching module constrains the encoding and decoding coordination module through the feature matching loss, and the formula is as follows:

[0203] L FM (H X→Y ,H Y→X )=FM X +FM Y

[0204] Where FM X 、FM Y are the feature matching degrees of the intermediate layer of the discrimination module output by the feature matching module;

[0205] The cyclic consistency constraint formula is as follows:

[0206] L Cyc (H X→Y ,H Y→X )=‖H Y→X (H X→Y (X))-X‖1+‖H X→Y (H Y→X (Y))-Y‖1

[0207] Among them, the encoding and decoding coordination module outputs to the cyclic input reverse encoding and decoding coordination module for cyclic cross-domain inverse mapping reconstruction, and performs cyclic consistency constraint between the reconstructed data and the original input through voxel-level alignment;

[0208] As a preferred embodiment of the present invention, for the two-way coordination model training module, the parameters of the encoding and decoding coordination module and the discriminant module are trained asynchronously. During the optimization stage of the encoding and decoding coordination module, the parameters of the discriminant module are fixed, and a 5:1 asynchronous training method with the encoding and decoding module taking precedence is executed. The parameters of the encoding and decoding coordination module are iterated 5 times, and the parameters of the discriminant module are iterated 1 time. The model objective function is as follows:

[0209]

[0210] In the formula, λ1, λ2, λ3, and λ4 are 1, 1, 1, and 100 respectively.

[0211] After the model training is completed, the parameters of the encoding and decoding coordination module are fixed. 41 source domain samples X and 41 target domain samples Y for testing are input into the encoding and decoding coordination module, and the target domain coordination matrix Y ′ = H Y→X (X) and the source domain coordination matrix X ′ = H X→Y (Y);

[0212] As a preferred embodiment of the present invention, the fMRI three-dimensional feature index NIfTI format file reconstruction module is configured to perform the following steps:

[0213] S1. Perform symmetric spatial dimension adjustment on the coordination matrix output by the two-way coordination model training module: when D < D′, or H < H ′ , W < W′, perform edge clipping; when D > D′, or H > H ′ , W > W′, perform zero-value padding along the edge. In the embodiment, the size of 72×72×72 is restored to 61×73×71.

[0214] S2. Perform voxel-level masking based on the brain region mask M, and restore the original coordinate system based on the affine matrix A. The formula is as follows:

[0215] V ′ = (M X ⊙ X ′ ) · A

[0216] Among them, ⊙ represents element-wise multiplication, and · represents affine transformation;

[0217] S3. Fuse the standardized matrix V ′ with the original header metadata extracted by the fMRI three-dimensional feature index preprocessing module to generate a NIfTI file that conforms to the medical image standard and completes data coordination to eliminate the site effect;

[0218] S4. Reconstruct and cascade the fMRI three-dimensional feature index preprocessing and metadata protection in the fMRI three-dimensional feature index preprocessing module, the fixed codec coordination module parameters obtained from the two-way coordination model training module, and the NIfTI file in step S3 to obtain a multi-site brain imaging index data coordination system based on deep learning.

[0219] The present invention also includes evaluating the data coordination results of the test set. The evaluation indicators include data t-SNE dimensionality reduction visualization, maximum mean distribution difference (MMD) between datasets, and the voxel volume of significantly different brain regions between sites.

[0220] t-SNE maps the ReHo data to a two-dimensional space through non-linear dimensionality reduction;

[0221] The formula for the maximum mean distribution difference is as follows:

[0222] MMD = ‖E[k(X,·)] - E[k(Y,·)]‖ 2

[0223] The significantly different brain regions between sites are obtained through cross-domain two-sample t-tests, and the Gaussian random field correction method is used to perform multiple comparison corrections on the F-statistics across the whole brain. The correction thresholds are set as follows: the cluster-level significance threshold for two-sided tests Pcluster < 0.05, and the initial voxel-level threshold is Pvoxel < 0.001.

[0224] The results of this embodiment are as Figure 4 shown. Before coordination, t-SNE dimensionality reduction visualization is performed on X and Y, and the data distribution shows a site-dominated clustering effect, with a maximum mean distribution difference of 0.8210 and a voxel volume of 33,633 for significantly different brain regions between sites; after using the H of the codec coordination module X→Y to achieve coordination of data from GE to SIEMENS, the data distributions of X and Y show a consistent fusion trend, with a maximum mean distribution difference of 0.0442 and a voxel volume of 0 for significantly different brain regions between sites; after using the H of the codec coordination module Y→X to achieve coordination of data from SIEMENS to GE, the t-SNE dimensionality reduction visualization of the data distributions of X and Y shows a consistent fusion trend, with a maximum mean distribution difference of 0.0746 and a voxel volume of 0 for significantly different brain regions between sites.

[0225] Although the present invention has been described above with preferred embodiments, it is not intended to limit the present invention. Those with ordinary knowledge in the technical field to which the present invention pertains can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be determined by the scope defined in the claims.

Claims

1. A multi-site brain imaging index data coordination system based on deep learning, characterized in that Including: The fMRI three-dimensional feature index preprocessing module is used to define the source domain X and the target domain Y as datasets of image acquisition sites with different device parameters; based on the received source domain X and target domain Y, it performs cross-site data loading and parsing, brain region mask generation, and coordinated model input construction to construct an efficient data stream adapted to the bidirectional coordinated model encoding and decoding architecture; The bidirectional coordinated model construction module uses the source domain input X and target domain input Y obtained by the fMRI three-dimensional feature index preprocessing module as data carriers to construct a bidirectional coordinated transformation model with the fourth-order collaborative action of the encoding and decoding coordination module, discriminant module, distribution calculation module, and feature matching module. Through the encoding and decoding coordination module and the discriminant module, it performs unpaired cross-site data coordination asynchronous adversarial training, and the distribution calculation module, feature matching module, and cycle consistency constraint synchronously perform multi-level constraints on the encoding and decoding coordination module; trains the encoding and decoding coordination module to achieve domain-invariant feature decoupling and bidirectional reversible mapping, eliminating the site effect and retaining the original data topology structure; The bidirectional coordinated model training module performs bidirectional coordinated model training based on the source domain input X and target domain input Y obtained by the fMRI three-dimensional feature index preprocessing module. After the training is completed, the parameters of the encoding and decoding coordination module are fixed to obtain the coordination matrices of the source domain and the target domain; The fMRI three-dimensional feature index NIfTI format file reconstruction module reconstructs the coordination matrix output by the bidirectional coordinated model training module into a standardized three-dimensional feature index NIfTI format file and saves it through reverse space adaptation and topology restoration.

2. The multi-site brain imaging index data coordination system based on deep learning according to claim 1, characterized in that In the fMRI three-dimensional feature index preprocessing module, the cross-site data loading and parsing are specifically as follows: Load the NIfTI format files of the fMRI three-dimensional feature indexes of X and Y, and accurately extract and independently store the following elements through the medical image parsing tool NiBabel: The original three-dimensional feature index matrix V ∈ R D×H×W , where D represents the depth, H represents the height, and W represents the width; Spatial positioning parameters: Affine matrix A; NIfTI header file metadata; In the fMRI three-dimensional feature index preprocessing module, the binary brain region mask M∈{0,1} is generated in the brain region mask generation process by referring to the reference value 0 D×H×W , and the formula is as follows: Wherein, V ijk represents the voxel value at the position [i, j, k] in the matrix V; In the fMRI three-dimensional feature index preprocessing module, the coordinated model input is constructed as follows: X = Padding / Cropping(V X ) ⊙ M X Y = Padding / Cropping(V Y ) ⊙ M Y Among them, Padding / Cropping means that according to the set target dimensions (D ′ , H ′ , W′), when D < D′, or H < H ′ , W < W′, zero-padding is performed on the matrix V along the edges; when D > D′ or H > H ′ , W > W′, edge cropping is performed on the matrix V; M is the brain region mask; ⊙ is element-wise multiplication.

3. The multi-site brain imaging index data coordination system based on deep learning according to claim 1, characterized in that In the two-way coordination model construction module, the encoding and decoding coordination module contains two three-dimensional two-way coordination mapping models H with the same structure X→Y and H Y→X ; H X→Y from the source domain input X to the coordinated output Y ′ = H X→Y (X), which successively includes an input layer, an encoder, a decoder, and an output layer. Among them, the encoder and the decoder are symmetric structures; The input layer formula is as follows: X H_in = 3DConv(X) Where X is the source domain input obtained by the fMRI three-dimensional feature index preprocessing module, and 3DConv is the three-dimensional convolutional layer; The encoder includes four downsampling layers. The first three layers are composed of an activation function, a three-dimensional convolutional layer, and an instance normalization layer, and the innermost layer is composed of an activation function and a three-dimensional convolutional layer. The formula is as follows: where ReLU is the activation function, IN is the instance normalization layer, and X H_in is the output of the input layer, is the feature map output of the m-th layer of the encoder; The decoder includes a total of four upsampling layers, which are composed of an activation function, a three-dimensional transposed convolutional layer, and an instance normalization layer. The formula is as follows: Among them, 3DTransConv is a three-dimensional transposed convolution layer, and LeackyReLU is an activation function. is the input feature map of the nth layer of the decoder. is the output feature map of the nth layer of the decoder. When n≠1, The output feature map of the (n - 1)-th layer of the decoder and the corresponding input feature map of the n-th layer of the encoder are concatenated in the channel dimension, and the formula is as follows: Where Concat is the channel dimension concatenation operation; The output layer performs transposed convolution on the decoder output feature map and performs a non-linear transformation. The formula is as follows: Among them, Tahn is the hyperbolic tangent activation function, and y ′ is the coordinated output of X→Y obtained by inputting X into H X→Y ; H Y→X From the target domain input Y to the coordinated output X ′ = H Y→X (Y), successively including an input layer, an encoder, a decoder, and an output layer, wherein the encoder and the decoder are symmetric structures; The input layer formula is as follows: Y H_in = 3DConv(Y) Where Y is the target domain input obtained by the fMRI three-dimensional feature index preprocessing module; The encoder includes four downsampling layers. The first three layers are composed of an activation function, a three-dimensional convolutional layer, and an instance normalization layer, and the innermost layer is composed of an activation function and a three-dimensional convolutional layer. The formula is as follows: Among them, Y H_in is the output of the input layer, and is the feature map output of the m-th layer of the encoder; The decoder includes a total of four upsampling layers, which are composed of an activation function, a three-dimensional transposed convolutional layer, and an instance normalization layer. The formula is as follows: Among them, is the input feature map of the n-th layer of the decoder, is the output feature map of the n-th layer of the decoder; When n≠1, The output feature map of the (n-1)-th layer of the decoder and the corresponding input feature map of the n-th layer of the encoder are concatenated in the channel dimension, and the formula is as follows: The output layer performs transposed convolution on the decoder output feature map and performs a non-linear transformation. The formula is as follows: Among them, X ′ is the Y input H Y→X to obtain the Y→X coordinated output.

4. The multi-site brain imaging index data coordination system based on deep learning according to claim 1, wherein In the two-way coordination model construction module, the discrimination module contains two dual-domain discrimination models D with the same structure X , D Y ; D X From the source domain input X to the output local discrimination result three-dimensional feature map D X (X), successively including an input layer, a downsampling layer, and an output layer. The input layer formula is as follows: d in = LeakyReLU(3DConv(X)) The downsampling layer formula is as follows: d fe = LeakyReLU.BN(3DConv(d in )) / Among them, BN is the batch normalization layer; The formula of the output layer is as follows: d X = 3DConv(d fe ) Among them, d x is the three-dimensional feature map of the local discrimination result. D Y From the input Y in the target domain to the output of the three-dimensional feature map D of the local discrimination result Y (Y), which successively includes an input layer, a downsampling layer, and an output layer. The formula of the input layer is as follows: d in = LeakyReLU(3DConv(Y)) The formula of the downsampling layer is as follows: d fe = LeakyReLU.BN(3DConv(d in )) / Among them, BN is the batch normalization layer; The formula of the output layer is as follows: d Y = 3DConv(d fe ) Among them, d Y is the three-dimensional feature map of the local discrimination result.

5. A multi-site brain imaging index data coordination system based on deep learning according to claim 4, characterized in that, In the two-way coordination model construction module, the distribution calculation module guides the distribution consistency within the implementation site in the encoding-decoding coordination module. This module is cascaded with the encoding-decoding coordination module and receives the original input and coordinated output of the encoding-decoding coordination module; the two are implicitly mapped from the original space to the infinite-dimensional reproducing kernel Hilbert space through a multi-scale Gaussian kernel function to achieve distribution alignment; the intra-group and inter-group similarities of the two in the feature space are calculated; based on the mean difference of the similarities, the mean embedding distance is calculated through weighted summation to obtain the distribution difference between the original input and the coordinated output of the encoding-decoding coordination module. The formula is as follows: Dd X 2 = ‖E[k(X,·)] - E[k(H Y→X (Y),·)]‖ 2 Dd Y 2 = ‖E[k(Y,·)] - E[k(H X→Y (X),·)]‖ 2 Among them, k() is the radial basis function kernel.

6. The multi-site brain imaging index data coordination system based on deep learning according to claim 4, characterized in that, In the two-way coordination model construction module, the feature matching module retains the in-site fidelity of the coordinated fMRI three-dimensional indicators in terms of the spatial correlation between complex structures and voxels through a high-dimensional feature alignment mechanism; The feature matching module sequentially passes through the feature extraction layer and the feature matching layer from input to output, is cascaded with the discriminant module, extracts the intermediate layer feature map, and calculates the feature matching degree using the mean absolute error. The formula is as follows: FM X = || D X (l) (H Y→X (Y)) - D X (l) (X) || 1 FM Y =‖D Y (l) (H X→Y (X)) - D Y (l) (Y)‖1 Among them, l is the l-th layer of the discriminant model, and ‖‖1 represents the calculation of the mean absolute error.

7. The multi-site brain imaging index data coordination system based on deep learning according to claim 4, characterized in that In the two-way coordination model construction module, the encoding-decoding coordination module and the discriminant module perform unpaired cross-site data coordination asynchronous adversarial training, and use the adversarial loss as follows: L adv (H X→Y ,H Y→X ,D X ,D Y ) = L adv_H +L adv_D The encoding-decoding coordination module makes the discriminant module misjudge the coordinated output domain by improving its own coordination ability, that is, it is considered that the coordinated source domain data belongs to the target domain; the formula is as follows: L advH (H X→Y ,H Y→X ) = MSE(D X (H Y→X (Y)), 1) + MSE(D Y (H X→Y ), 1) Among them, 1 means belonging to the X / Y domain, 0 means not belonging to the X / Y domain, 1 and 0 are three-dimensional numerical matrices with the same dimension as the output dimension of the discriminant module, and MSE is the mean square error; The discriminant module makes a correct judgment on the original input and the coordinated output domain by improving its own discriminant ability, that is, it is considered that both the original input and the coordinated output belong to the original domain. The formula is as follows: L advD (D X ,D Y ) = MSE(D X (X), 1) + MSE(D X (H Y→X (Y)), 0) + MSE(D Y (Y), 1) +MSE(D Y (H X→Y ),0) The encoding-decoding coordination module and the discriminant module play against each other, and finally maximize the performance under the balanced condition.

8. A multi-site brain imaging index data coordination system based on deep learning according to claim 4, characterized in that, In the two-way coordination model construction module, the distribution calculation module constrains the encoding-decoding coordination module through the distribution alignment loss. The formula is as follows: L Dd (H X→Y ,H Y→X ) = Dd X 2 + Dd Y 2 Among them, Dd X 2 and Dd Y 2 are the outputs obtained from the original input and the input distribution calculation module of the coordination module output; The feature matching module constrains the encoding-decoding coordination module through the feature matching loss. The formula is as follows: L FM (H X→Y ,H Y→X ) = FM X +FM Y Among them, FM X and FM Y are the discriminant module intermediate layer feature matching degrees output by the feature matching module; The formula of the cycle consistency loss is as follows: L Cyc (H X→Y ,H Y→X ) = ‖H Y→X (H X→Y (X)) - X‖1 + ‖H X→Y (H Y→X (Y)) - Y‖1 Among them, the encoding-decoding coordination module outputs the loop input to the reverse encoding-decoding coordination module, performs loop cross-domain inverse mapping reconstruction, and performs cycle consistency constraint on the reconstructed data and the original input through voxel-level alignment.

9. A multi-site brain imaging index data coordination system based on deep learning according to claim 8, characterized in that, In the two-way coordination model training module, the parameters of the encoding-decoding coordination module and the discriminant module are trained asynchronously. In the optimization stage of the encoding-decoding coordination module, the parameters of the discriminant module are fixed, and the n:1 asynchronous training method with the encoding-decoding module prioritized is executed. The parameters of the encoding-decoding coordination module are iterated n times, and the parameters of the discriminant module are iterated 1 time. The model objective function is as follows: In the formula, λ1, λ2, λ3, and λ4 are the coefficients of the adversarial loss, the distribution alignment loss, the feature matching loss, and the cycle consistency loss, respectively.

10. A multi-site brain imaging index data coordination system based on deep learning according to claim 2, characterized in that, The fMRI three-dimensional feature index NIfTI format file reconstruction module is configured to perform the following steps: S1. Perform symmetric spatial dimension adjustment on the coordination matrix output by the two-way coordination model training module: When D < D′, or H < H ′ , W < W′, perform edge clipping; when D > D′, or H > H ′ , W > W′, perform zero-padding along the edges; S2. Perform voxel-level masking based on the brain region mask M, and restore the original coordinate system based on the affine matrix A. The formula is as follows: V ′ = (M X ⊙ X ′ ) · A where · represents affine transformation; S3. Fuse the standardized matrix V ′ with the original header file metadata extracted by the fMRI three-dimensional feature index preprocessing module to generate a NIfTI file that conforms to the medical image standard and completes data coordination to eliminate the site effect; S4. Cascade the fMRI three-dimensional feature index preprocessing and metadata protection in the fMRI three-dimensional feature index preprocessing module, the fixed codec coordination module parameters obtained from the bidirectional coordination model training module, and the fMRI three-dimensional feature index NIfTI format file reconstruction to obtain a multi-site brain imaging index data coordination system based on deep learning.

Images