Brain magnetic resonance image registration method fusing t1w-mri and fMRI information

Abstract

The application relates to a brain magnetic resonance image registration method based on fusion of T1w-MRI and fMRI information, and belongs to the technical field of brain image processing. The method simultaneously utilizes fine structural information of a brain T1 weighted magnetic resonance image and functional information of a functional magnetic resonance image, fuses different important information contained in the two modalities, and realizes correspondence of brain structure regions and functional regions. For registration of the two modalities, the registration accuracy is improved to a certain extent. For registration of the fMRI image, cumulative errors generated by step-by-step registration with the aid of the structural image are reduced, meanwhile, deep learning is referenced for global optimization, GPU accelerated calculation is utilized, and the calculation time of the registration algorithm is greatly shortened.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of brain imaging image processing, and particularly relates to a brain magnetic resonance image registration method based on fusion of T1w-MRI and fMRI information. BACKGROUND

[0002] The brain is one of the most important parts of the human body, which dominates human thinking and action through a large number of neurons and synaptic connections, and can be regarded as the most complex and delicate product in nature. Therefore, it is of great significance to study the human brain. By revealing the operation mode and nature of the brain, the understanding of human beings can be deepened, and the principles can also be used to benefit mankind, for example, artificial neural networks are products of bionic research based on in-depth understanding of the working mechanism of the brain, which greatly promotes the development of deep learning, thereby promoting the progress of human society. In addition, the research on the pathology of brain mental diseases is also an important means to prevent and treat mental diseases.

[0003] The research technology of human brain is also constantly progressing. In the early stage, people understood the structure and function of the brain by dissecting human corpses, which had great limitations and was difficult to conduct experiments on living people. With the development of medical imaging technology, various imaging technologies such as electroencephalogram, positron emission computed tomography, electron computed tomography, and nuclear magnetic resonance imaging have developed rapidly in brain research. Among them, the nuclear magnetic resonance imaging technology (Magnetic Resonance Imaging, MRI) has been widely favored by scholars due to its non-invasive advantage to the human body.

[0004] In the nuclear magnetic resonance imaging mode, common ones are T1 weighted magnetic resonance image (T1 weighted Magnetic Resonance Imaging, T1w-MRI) and functional magnetic resonance image (functional Magnetic Resonance Imaging, fMRI). Preprocessing of the collected images is an important prerequisite for subsequent research, and image registration is one of the links. For example, in the standardization step in preprocessing, the MRI image is registered to the standard space using a registration algorithm. In addition, for the analysis of differences between brain function networks of different populations, group analysis of functional magnetic resonance images is required, so registration between different individual images is the most basic requirement. Due to the complexity of the brain structure, different modal images contain different information, and it is difficult to establish structural correspondence and functional correspondence at the same time, so the registration algorithm for brain MRI images needs further research. SUMMARY

[0005] The application aims to provide a brain magnetic resonance image registration method fusing T1w-MRI and fMRI information, which utilizes the fine structural information contained in brain T1 weighted magnetic resonance images and the functional region information contained in functional magnetic resonance images, adopts a convolutional neural network as a framework, and perfects and develops a brain magnetic resonance image registration algorithm.

[0006] In order to achieve the above-mentioned purpose, the application is realized by the following technical scheme:

[0007] On one hand, the application discloses a brain magnetic resonance image registration method fusing T1w-MRI and fMRI information, which comprises the following steps:

[0008] Image acquisition: MRI image groups of multiple subjects are collected, and each MRI image group of a subject comprises a T1w-MRI image and a corresponding resting state fMRI image of the subject;

[0009] Image preprocessing: the collected MRI image groups of multiple subjects are preprocessed to obtain multiple preprocessed MRI image groups of subjects, wherein the size of the preprocessed resting state fMRI image in each preprocessed MRI image group of a subject is 1 / t of the size of the preprocessed T1w-MRI image, and t is in the range of 2-5;

[0010] Image division: the preprocessed MRI image groups of subjects are divided to obtain a training set and a validation set;

[0011] Model construction: a registration model comprising a convolutional neural network and a semi-supervised loss function is constructed;

[0012] Model training: the registration model is iteratively trained by using the training set to obtain a trained registration model;

[0013] Model verification: the trained registration model is verified by using the validation set, and the final registration model is obtained by adjusting hyperparameters;

[0014] Registration after image acquisition: the registration after image acquisition of the MRI image group of a subject to be registered is obtained according to the final registration model;

[0015] Wherein, the registration after image acquisition of the MRI image group of a subject to be registered is obtained according to the final registration model, which comprises:

[0016] Selecting a target subject and a subject to be registered to form a subject pair, and performing the image preprocessing on the MRI image groups of the subject pair;

[0017] The T1w-MRI image of the target subject and the pre-processed T1w-MRI image of the subject to be registered are simultaneously input into the final registration model to output a deformation field of the T1w-MRI image of the subject to be registered from the final registration model, deform the T1w-MRI image of the subject to be registered through the spatial transformation network to obtain a T1w-MRI image of the subject to be registered after registration, down-sample the deformation field of the T1w-MRI image of the subject to be registered to obtain a deformation field of the resting-state fMRI image of the subject to be registered, and deform the resting-state fMRI image of the subject to be registered through the spatial transformation network to obtain a resting-state fMRI image of the subject to be registered after registration.

[0018] In some embodiments, the convolutional neural network is a U-net convolutional neural network.

[0019] In some embodiments, the semi-supervised loss function is defined as follows:

[0020] L = L T1-sim + λL f-sim + γL smooth .

[0021] L is the semi-supervised loss function, L T1-sim is a similarity loss function of the T1w-MRI image of the subject to be registered after registration and the T1w-MRI image of the target subject, L f-sim is a similarity function of the fMRI image of the subject to be registered after registration and the fMRI image of the target subject; L smooth is a deformation field smoothing function; and λ and γ are weight coefficients.

[0022] An initial value is given to the weight coefficients λ and γ in the semi-supervised loss function according to experience to obtain an initial semi-supervised loss function.

[0023] In some embodiments, the initial values of the weight coefficients λ and γ are both 0.2.

[0024] In some embodiments, the step of iteratively training the registration model using the training set to obtain the trained registration model comprises:

[0025] S1, randomly select two MRI image groups of different subjects from the training set as a subject pair of MRI image groups, input the T1w-MRI images corresponding to the two different subjects in the subject pair of MRI image groups into the convolutional neural network as the T1w-MRI image to be registered and the target T1w-MRI image respectively, and after the calculation of the convolutional neural network and the obtaining of the deformation field φ, down-sample the deformation field φ to obtain a 1 / t deformation field φ 1 / t; wherein the deformation field φ is a deformation field of the T1w-MRI image, and the deformation field φ 1 / t is a deformation field of the resting-state T1w-MRI image;

[0026] Step S2, inputting the deformation field φ and the deformation field φ 1 / t to the corresponding to-be-registered images of the subject pair respectively through the spatial transformation network to obtain the registered images;

[0027] Step S3, according to the similarity of the registered images and the target image and the gradient of the deformation field, calculating the value of the semi-supervised loss function, and then updating the parameters of the convolutional neural network by using the optimizer according to the value of the semi-supervised loss function to obtain the updated convolutional neural network.

[0028] Step S4, continuing to input the data of the training set into the updated convolutional neural network, repeating steps S1-S3 until the maximum iteration number is reached, to obtain the trained registration model.

[0029] In some embodiments, the Adam optimizer is used to update the parameters of the convolutional neural network, and the learning rate is set to 10 -4 to obtain the updated convolutional neural network.

[0030] In some embodiments, in the image preprocessing step, the preprocessing step of the MRI image group collected from the plurality of subjects respectively includes:

[0031] The T1w-MRI image of the subject is subjected to skull removal, intensity standardization, affine alignment, cropping and normalization processing, and the resting-state fMRI image of the subject is subjected to removal of the first 10 time points, time layer correction, head motion correction, covariate regression, removal of linear drift, filtering, smoothing, affine alignment and cropping processing.

[0032] In some embodiments, the affine alignment of the T1w-MRI image of the subject is to affine align the T1w-MRI image after intensity standardization processing to a standard space template with a voxel size of 1mm, and the affine alignment of the resting-state T1w-MRI image of the subject is to affine align the resting-state fMRI image after smoothing processing to a standard space template with a voxel size of 1mm.

[0033] In some embodiments, in the image division step: the ratio of the training set and the validation set is 7:3-9:1.

[0034] In another aspect, the application also discloses a computer readable storage medium, which stores a computer program, and is characterized by that when the computer program product is executed, the brain magnetic resonance image registration method for fusing T1w-MRI and fMRI information according to any one of claims 1-9 is performed.

[0035] Compared with the prior art, the above-mentioned application has at least one of the following advantages or beneficial effects:

[0036] The application simultaneously utilizes the fine structural information of the brain T1 weighted magnetic resonance image and the functional information of the functional magnetic resonance image, fuses different important information contained in the two modalities, and realizes the correspondence of the brain structural region and the functional region. For the registration of the two modalities, the registration accuracy is improved to a certain extent. For the registration of the fMRI image, the cumulative error generated by the step-by-step registration with the aid of the structural image is reduced, and deep learning is used for global optimization, GPU acceleration calculation is used, and the calculation time of the registration algorithm is greatly shortened. BRIEF DESCRIPTION OF DRAWINGS

[0037] The present application and its features, shapes and advantages will become more apparent from the following detailed description of non-limiting embodiments, with reference to the accompanying drawings. The same reference signs indicate the same parts throughout the drawings. The drawings are not necessarily drawn to scale, the emphasis being on illustrating the principle of the application.

[0038] Figure 1 The flow chart of the brain magnetic resonance image registration method for fusing T1w-MRI and fMRI information in the embodiments of the application. DETAILED DESCRIPTION

[0039] The application will be further described below in conjunction with the drawings and specific embodiments, but not as a limitation of the application.

[0040] As shown in the drawings, Figure 1 The application discloses a brain magnetic resonance image registration method for fusing T1w-MRI and fMRI information, and specifically, the method comprises the following steps:

[0041] Step 1, image acquisition: acquiring a plurality of (for example, 250) subject MRI image groups, and each subject MRI image group comprises the subject's T1w-MRI image and corresponding resting state fMRI image, wherein the voxel size of the acquired T1w-MRI image is 1mm. And the corresponding resting state fMRI image, the voxel size of the corresponding resting state fMRI image is 3mm.

[0042] Step 2, image preprocessing: the collected MRI image groups of multiple subjects are respectively preprocessed to obtain multiple preprocessed MRI image groups of subjects, wherein the T1w-MRI image of the subject is subjected to skull removal, intensity standardization, affine alignment, cropping and normalization processing, and the resting state fMRI image of the subject is subjected to removal of the first 10 time points, time layer correction, head motion correction, covariate regression, removal of linear drift, filtering, smoothing, affine alignment and cropping processing.

[0043] Specifically, the T1w-MRI image of the subject is subjected to skull removal and intensity standardization using Freesurfer software; the resting state fMRI image of the subject is subjected to removal of the first 10 time points, time layer correction, head motion correction, covariate regression, removal of linear drift, filtering and smoothing using DPARSF software; the T1w-MRI image of the subject is subjected to affine alignment using SPM software to affine align the T1w-MRI image subjected to intensity standardization to a standard space template with a voxel size of 1mm; and the resting state T1w-MRI image of the subject is subjected to affine alignment using SPM software to affine align the resting state fMRI image subjected to covariate regression to a standard space template with a voxel size of 3mm.

[0044] In each preprocessed MRI image group of the subject, the size of the preprocessed resting state fMRI image is 1 / 3 of the size of the preprocessed T1w-MRI image, specifically, the T1w-MRI is cropped to a size of 144x192x192, and the preprocessed resting state fMRI image is cropped to a size of 48x64x64, and then the T1w-MRI image is subjected to normalization processing.

[0045] Step 3, image division: the preprocessed MRI image groups of the subjects are divided in proportion to obtain a training set and a validation set; specifically, the preprocessed MRI image groups of the subjects are divided in proportion of 9:1 to obtain a training set and a validation set, wherein 225 are the training set and 25 are the test set; of course, in other embodiments of the present application, other proportions can also be used for division, which does not affect the purpose of the present application.

[0046] Step 4, model construction: a registration model including a U-net convolutional neural network and a semi-supervised loss function is constructed, specifically, the U-net convolutional neural network is constructed using Tensorflow.

[0047] Specifically, the semi-supervised loss function is obtained by the following formula:

[0048] L = L T1-sim + λL f-sim + γL smooth

[0049] wherein, L is a semi-supervised loss function, L T1-sim is a similarity loss function of the T1w-MRI post-registration image of the subject to be registered and the T1w-MRI image of the target subject, L f-sim is a similarity function of the fMRI post-registration image of the subject to be registered and the fMRI image of the target subject; L smooth is a deformation field smoothing function; λ and γ are weight coefficients; an initial value is given to the weight coefficients λ and γ in the semi-supervised loss function according to experience to obtain an initial semi-supervised loss function; specifically, the weight coefficients λ and γ are given an initial value of 0.2 and 0.2 according to experience to obtain an initial semi-supervised loss function.

[0050] Specifically, the construction steps for each term in the semi-supervised loss function are as follows:

[0051] Step A, L T1-sim is a similarity loss function of the T1w-MRI post-registration image of the subject to be registered and the T1w-MRI image of the target subject, which is calculated using the mean square error of the gray value of the two images (the T1w-MRI post-registration image of the subject to be registered and the T1w-MRI image of the target subject), as shown in the following formula:

[0052]

[0053] wherein, p is the voxel position; Ω is the image domain of the entire image; F T1 (p) is the gray value of the T1w-MRI post-registration image of the subject to be registered at the voxel position p; R T1 (p) is the gray value of the T1w-MRI image of the target subject at the voxel position p.

[0054] Step B, in the semi-supervised loss function L, the second term L f-sim is a similarity function of the fMRI post-registration image of the subject to be registered and the fMRI image of the target subject, which is calculated using the Bhattacharyya coefficient of the local functional connectivity histogram of the two images (the fMRI post-registration image of the subject to be registered and the fMRI image of the target subject), and the local functional connectivity histogram is calculated by a local functional connectivity vector. The calculation method of the local functional connectivity vector is as follows: for a voxel point n, take a cubic neighborhood with the point as the center and the side length of 7 voxels, calculate the Pearson correlation coefficient of all voxels m i,i＝1～343 in the neighborhood with the center point n in any order, and then obtain a local functional connectivity vector C(n) with the voxel point n as the center. For a pair of time series (the time series I(n) of the voxel point n and the time series I(m i), their Pearson correlation coefficients C(n,m i ) are calculated as follows:

[0055]

[0056] At a voxel position n, the local functional connectivity vector of a 7-voxel cube neighborhood of n is denoted as C(n) = [C(n,m1), C(n,m2), …, C(n,m 343 ), where m i ,i = 1 ~ 343, are the voxel points in the neighborhood of n.

[0057] Next, in order to eliminate the influence of relative position and calculation order, the obtained local functional connectivity vector is converted into a local functional connectivity distribution histogram. The Pearson correlation coefficient value ranges from -1 to 1, and this range is divided into 20 intervals with an interval of 0.1. The number of correlation coefficients in the calculation vector in each interval is counted, and finally the correlation coefficient distribution histogram, that is, the local functional connectivity distribution histogram, is obtained. The vector is denoted as H(n) = [H(n,1), H(n,2), …, H(n,20)], where H(n,i), i = 1 ~ 20, is the number of values in the local functional connectivity vector centered at the voxel point n falling into the i-th interval.

[0058] Finally, the Bhattacharyya Coefficient (BC) is used to calculate the difference between the local functional connectivity distribution histograms H F (n) and H R (n) of the same region of the fMRI image of the target subject and the fMRI image of the subject to be registered after registration. The difference in the local functional connectivity distribution histogram of all voxel neighborhoods Ω is calculated, and the average of all obtained values is taken to obtain the final loss value, which is as follows.

[0059] Step C, in actual situations, physically unrealistic non-smooth deformation fields may be generated, so it is necessary to smooth the deformation field to prevent spatial folding. Here, some applications in the diffusion model are used to constrain the spatial gradient of the displacement field u using regularization, where L smooth (φ) is a smoothing function.

[0060]

[0061] where, is the spatial gradient of the displacement field u at voxel position p.

[0062] The spatial gradient can be approximated using the difference between adjacent voxels, specifically where the gradient in the x direction can be approximated by the following equation.

[0063]

[0064] For the y direction and the z direction, Similarly, it is not described here.

[0065] Step 5, training the model: iteratively training the registration model using the training set to obtain the trained registration model.

[0066] In a preferred embodiment of the present application, the above step of iteratively training the registration model using the training set to obtain the trained registration model comprises:

[0067] Step S1, randomly selecting two different MRI image groups of subjects from the training set as the MRI image groups of the subject pair, inputting the T1w-MRI images corresponding to the two different subjects in the MRI image groups of the subject pair into the convolutional neural network as the to-be-registered T1w-MRI image and the target T1w-MRI image respectively, and after the calculation of the convolutional neural network (encoder and decoder), obtaining the deformation field φ (represented as the displacement field u) after the deformation field φ is down-sampled to obtain the deformation field φ 1 / 3 (represented as the displacement field u 1 / 3 ); wherein the deformation field φ is the deformation field of the T1w-MRI image, and the deformation field φ 1 / 3 is the deformation field of the resting state T1w-MRI image.

[0068] Step S2, applying the deformation field φ and the deformation field φ 1 / 3 to the corresponding to-be-registered images of the subject pair through the spatial transformation network to obtain the registered images.

[0069] Specifically, the step of applying the deformation field φ and the deformation field φ 1 / 3 to the corresponding to-be-registered images of the subject pair through the spatial transformation network to obtain the registered images comprises the following steps:

[0070] Step S21, for the to-be-registered subject's T1w-MRI image (M T1 ), the deformation field φ is applied to the to-be-registered image, denoted as so that the to-be-registered image is distorted to obtain the to-be-registered subject's T1w-MRI registered image (R T1 ). R T1 The intensity calculation formula of the image obtained by interpolation calculation is:

[0071]

[0072] wherein, the intensity of the voxel at the transformed position the voxel position of the neighboring region, q is one of the voxel positions of the neighboring region, d represents three directions of x, y, and z in the three-dimensional space.

[0073] Step S22, similarly, the obtained deformation field φ 1 / 3 is applied to the fMRI (M f ) image of the selected to-be-registered subject, and a corresponding registered fMRI image (R f ) is obtained, and the calculation formula is as follows.

[0074]

[0075] wherein, the intensity of the voxel at the transformed position the voxel position of the neighboring region, one of the voxel positions of the neighboring region, d represents three directions of x, y, and z in the three-dimensional space.

[0076] Step S3, after obtaining the value of the semi-supervised loss function according to the similarity of the registered image and the target image and the gradient of the deformation field, the parameters of the convolutional neural network are updated by using an optimizer according to the value of the semi-supervised loss function to obtain an updated convolutional neural network.

[0077] Specifically, the target of the model is to minimize the loss function, and then an optimal convolutional neural network is obtained.

[0078]

[0079] wherein θ is a parameter in the convolutional neural network. Then, the parameters θ of the convolutional neural network are iterated by using an Adam optimizer, and the learning rate is set to 10 -4 , and an updated convolutional neural network is obtained.

[0080] Step S4, continue to input the data of the training set into the updated convolutional neural network, repeat steps S1-S3, until the maximum iteration number is reached, to obtain a trained registration model, and specifically, the maximum iteration number is set to 500 times.

[0081] ​​​​​Step 5, verifying the model: verifying the trained registration model by using the verification set and adjusting the hyperparameters including the weight coefficients λ and γ, learning rate, and iteration number, etc. to obtain the final registration model.

[0082] Specifically, the trained registration model is verified by using the verification set, where one subject in the verification set is selected as a target, and the other subjects in the verification set are respectively registered to the target subject by using the trained model, and then the registration results are observed. Then, the hyperparameters including the weight coefficients λ and γ, iteration number, learning rate, etc. are adjusted, the model is retrained, and the optimal result in the verification set is selected as the final registration model.

[0083] Step 6, obtaining the registered image: obtaining the registered image of the MRI image group of the subject to be registered according to the final registration model;

[0084] Specifically, the above-mentioned obtaining the registered image of the MRI image group of the subject to be registered according to the final registration model includes:

[0085] Step a, selecting a target subject and a subject to be registered to form a subject pair, and performing the image preprocessing of step 2 on the MRI image group of the subject pair;

[0086] Step b, inputting the pre-processed T1w-MRI images of the subject to be registered and the target subject into the final registration model to output the deformation field of the T1w-MRI image of the subject to be registered from the final registration model, deforming the T1w-MRI image of the subject to be registered by using the spatial transformation network to obtain the T1w-MRI registered image of the subject to be registered, downsampling the deformation field of the T1w-MRI image of the subject to be registered to obtain the deformation field of the resting-state fMRI image of the subject to be registered, and deforming the resting-state fMRI image of the subject to be registered by using the spatial transformation network to obtain the resting-state fMRI registered image of the subject to be registered.

[0087] It should be noted that the order of the above-mentioned steps 1-6 can be adjusted as needed, which does not affect the purpose of the present application.

[0088] In addition, the present application also discloses a computer readable storage medium having a computer program stored thereon, when the computer program product is executed, the above-mentioned brain magnetic resonance image registration method of fusing T1w-MRI and fMRI information is performed.

[0089] In summary, the application inputs two T1w-MRI images of a subject pair into a registration model, obtains a deformation field through a series of transformations, then applies the deformation field to the image to be registered using a spatial transformation network, then uses mean square error as a similarity loss function to measure T1w-MRI, uses the Bhattacharyya coefficient of the local functional connection distribution histogram as a similarity loss function to measure fMRI, and simultaneously applies constraints to the deformation field, constructs a total loss function according to a certain proportion coefficient. Then the optimizer is optimized iteratively to continuously train the neural network, and finally a trained registration model is obtained, and image registration is performed according to the trained registration model. For the registration of two modalities, the registration accuracy is improved to a certain extent. For the registration of fMRI images, the cumulative error generated by the distribution registration with the aid of structural images is reduced, and deep learning is used for global optimization, GPU acceleration is used for calculation, and the calculation time of the registration algorithm is greatly shortened.

[0090] Those skilled in the art should understand that those skilled in the art can realize the variants in combination with the prior art and the above-mentioned embodiments, which are not described here. Such variants do not affect the essential content of the application, and are not described here.

[0091] The preferred embodiments of the application are described above. It should be understood that the application is not limited to the above specific embodiments, and the devices and structures not described in detail should be understood as being implemented in the ordinary way in the art; any person skilled in the art can make many possible changes and modifications to the technical solutions of the application, or modify them into equivalent embodiments with equivalent changes, without departing from the scope of the technical solutions of the application, which does not affect the essential content of the application. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the application, without departing from the scope of the technical solutions of the application, still belongs to the scope of protection of the technical solutions of the application.

Claims

1. A method of brain magnetic resonance image registration that fuses T1w-MRI and fMRI information, characterized in that, The method comprises the following steps: The method comprises the following steps: Image acquisition: acquiring a plurality of MRI image groups of subjects, and each MRI image group of a subject comprises a T1w-MRI image and a corresponding resting-state fMRI image of the subject; Image preprocessing: preprocessing the acquired MRI image groups of the plurality of subjects to obtain a plurality of preprocessed MRI image groups of the subjects, wherein in each preprocessed MRI image group of a subject, the size of the preprocessed resting-state fMRI image is 1 / t of the size of the preprocessed T1w-MRI image, wherein t is in the range of 2-5; Image division: dividing the preprocessed MRI image groups of the subjects to obtain a training set and a validation set; Model construction: constructing a registration model comprising a convolutional neural network and a semi-supervised loss function; Model training: iteratively training the registration model using the training set to obtain a trained registration model; Step S1: Randomly select two different subjects' MRI image sets from the training set as the subject pair's MRI image set. The T1w-MRI images corresponding to the two different subjects in the subject pair's MRI image set are respectively input into the convolutional neural network as the T1w-MRI image to be registered and the target T1w-MRI image. After calculation by the convolutional neural network, the deformation field is obtained. Then, the deformation field Downsampling is performed to obtain the deformation field. 1 / t proportional deformation field Among them, the deformation field The deformation field of the T1w-MRI image. The deformation field of the resting-state T1w-MRI image; Step S2, applying the deformation field and the deformation field to the respective images to be registered of the pair of subjects respectively via the spatial transformation network to obtain the registered images; The step of iteratively training the registration model using the training set to obtain a trained registration model comprises: Step S3: After obtaining the value of the semi-supervised loss function according to the similarity of the registered image and the target image and the gradient of the deformation field, updating the parameters of the convolutional neural network using an optimizer according to the value of the semi-supervised loss function to obtain an updated convolutional neural network; Step S4: Continue inputting the data of the training set into the updated convolutional neural network, and repeat steps S1-S3 until the maximum number of iterations is reached, to obtain a trained registration model; Model validation: validating the trained registration model using the validation set and adjusting the hyperparameters to obtain a final registration model; Obtaining a registered image: obtaining a registered image of the MRI image group of a subject to be registered according to the final registration model; Wherein, obtaining a registered image of the MRI image group of a subject to be registered according to the final registration model comprises: Selecting a target subject and a subject to be registered to form a subject pair, and performing the image preprocessing on the MRI image groups of the subject pair; 2. The method of registering brain magnetic resonance images fusing Tl w-MRI and fMRI information according to claim 1, characterized in that, Inputting the preprocessed T1w-MRI images of the subject to be registered and the target subject into the final registration model at the same time, outputting the deformation field of the T1w-MRI image of the subject to be registered from the final registration model, deforming the T1w-MRI image of the subject to be registered through a spatial transformation network to obtain a T1w-MRI registered image of the subject to be registered, downsampling the deformation field of the T1w-MRI image of the subject to be registered to obtain the deformation field of the resting-state fMRI image of the subject to be registered, and deforming the resting-state fMRI image of the subject to be registered through a spatial transformation network to obtain a resting-state fMRI registered image of the subject to be registered.

3. The method of registering brain magnetic resonance images fusing Tl w-MRI and fMRI information of claim 2, wherein, The convolutional neural network is a U-net convolutional neural network. ； wherein, is the semi-supervised loss function, is a similarity loss function between the T1w-MRI post-registration image of the subject to be registered and the T1w-MRI image of the target subject, is a similarity function between the fMRI post-registration image of the subject to be registered and the fMRI image of the target subject; and is a weight coefficient; According to experience to give the weight coefficient in the semi-supervised loss function and An initial value to get a first semi-supervised loss function.

4. The method of registering brain magnetic resonance images fusing Tl w-MRI and fMRI information of claim 3, wherein, weight coefficients and The initial values of the weight coefficients are 0.

2.

5. The method of registering brain magnetic resonance images fusing Tl w-MRI and fMRI information as claimed in claim 1, wherein, The parameters of the convolutional neural network are updated using an Adam optimizer with a learning rate of 10 -4 to obtain an updated convolutional neural network.

6. The method of registering brain magnetic resonance images fusing Tl w-MRI and fMRI information as claimed in claim 1, wherein, The semi-supervised loss function is defined as the following formula: In the step of image preprocessing, the step of preprocessing the acquired MRI image groups of the plurality of subjects comprises: The T1w-MRI images of the subjects are skull-stripped, intensity normalized, affine-aligned, cropped and normalized, and the resting-state fMRI images of the subjects are processed by removing the first 10 time points, time layer correction, head motion correction, covariate regression, linear drift removal, filtering, smoothing, affine alignment and cropping.

7. The method of registering brain magnetic resonance images fusing Tl w-MRI and fMRI information as claimed in claim 6, wherein, The affine alignment of the T1w-MRI images of the subjects is to affine-align the intensity-normalized T1w-MRI images to a standard space template with a voxel size of 1mm, and the affine alignment of the resting-state T1w-MRI images of the subjects is to affine-align the smoothed resting-state fMRI images to a standard space template with a voxel size of t mm.

8. The method of registering brain magnetic resonance images fusing Tl w-MRI and fMRI information as claimed in claim 1, wherein, In the image division step, the ratio of the training set to the validation set is 7:3-9:

1.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed, the method for registering brain magnetic resonance images by fusing T1w-MRI and fMRI information according to any one of claims 1-8 is performed.

Images