MRI (Magnetic Resonance Imaging) image segmentation method and imaging method based on frequency-space collaboration

Abstract

The invention discloses an MRI (Magnetic Resonance Imaging) image segmentation method based on frequency-space collaboration. The method comprises the following steps: acquiring MRI image data information of a target area and preprocessing the MRI image data information to construct a training data set; based on a random mask scheme, fast Fourier convolution and a Unet network, constructing an MRI image segmentation initial model based on frequency-space collaboration, and training to obtain an MRI image segmentation model based on frequency-space collaboration; and segmenting the MRI image of the target area by using the obtained MRI image segmentation model based on frequency-space coordination. The invention also discloses an imaging method comprising the MRI image segmentation method based on frequency-space cooperation. Through the random mask scheme, the fast Fourier convolution and the Unet network, the frequency-space collaboration is realized based on the fast Fourier convolution, so that the MRI image segmentation can be realized, the reliability is higher, and the accuracy is better.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of image processing, and particularly relates to a MRI image segmentation method based on frequency-space cooperation and an imaging method. BACKGROUND

[0002] Magnetic resonance images (MRI images) have an irreplaceable position in the process of clinical medical application and basic medical research. Therefore, the segmentation of MRI images is of great significance to clinical medicine and basic medicine.

[0003] At present, the segmentation scheme of MRI images still adopts the manual segmentation scheme, that is, medical imaging professionals manually segment the target region on the MRI image according to their own technical level. However, this manual segmentation scheme is not only time-consuming and laborious, but also has relatively poor reliability and accuracy.

[0004] At present, with the rapid development of artificial intelligence technology, researchers have proposed a self-supervised MRI image segmentation scheme and a semi-supervised MRI image segmentation scheme. The self-supervised learning scheme can effectively learn the general features in the MRI image to realize image segmentation by mining the information of the data itself. However, the self-supervised learning scheme often relies on large-scale high-quality training data sets, and the acquisition cost of MRI image data is high, the image complexity is high, and the researchers cannot obtain enough and high-quality MRI images due to the privacy protection problem. Moreover, the performance of the self-supervised learning scheme in highly targeted tasks (such as image segmentation) is relatively low. Similarly, the semi-supervised learning scheme improves the segmentation performance by combining a small amount of labeled data and a large amount of unlabeled data, but this scheme mainly relies on label or pseudo-label learning, ignores the rich anatomical information contained in medical images, and results in poor segmentation accuracy. SUMMARY

[0005] One of the purposes of the present application is to provide a MRI image segmentation method based on frequency-space cooperation with high reliability and good accuracy.

[0006] The second purpose of the present application is to provide an imaging method comprising the MRI image segmentation method based on frequency-space cooperation.

[0007] The MRI image segmentation method based on frequency-space cooperation provided by the present application comprises the following steps:

[0008] S1. Obtain MRI image data information of a target region;

[0009] S2. Preprocess the MRI image data information obtained in step S1 to construct a training data set;

[0010] S3. Constructing an initial model of MRI image segmentation based on frequency-space collaboration based on a random mask scheme, fast Fourier convolution and a Unet network;

[0011] S4. Training the initial model of MRI image segmentation based on frequency-space collaboration constructed in step S3 using the training data set obtained in step S2 to obtain an MRI image segmentation model based on frequency-space collaboration;

[0012] S5. Using the MRI image segmentation model based on frequency-space collaboration obtained in step S4 to perform segmentation of the target region of the MRI image.

[0013] The preprocessing of step S2 specifically includes the following steps:

[0014] The MRI image data information obtained in step S1 includes labeled MRI image data information and unlabeled MRI image data information;

[0015] The MRI image data information obtained in step S1 is preprocessed; the preprocessing includes image data cleaning and image data filtering.

[0016] The initial model of MRI image segmentation based on frequency-space collaboration constructed in step S3 based on a random mask scheme, fast Fourier convolution and a Unet network includes the following steps:

[0017] The initial model of MRI image segmentation based on frequency-space collaboration constructed includes an image mask module, an image inpainting module and an image segmentation module;

[0018] The image mask module is constructed based on a random mask scheme and is used to process the input image information to generate a damaged image with missing regions;

[0019] The image inpainting module is constructed based on fast Fourier convolution and a Unet network and is used to generate an image reconstruction result of the damaged image with missing regions output by the image mask module;

[0020] The image segmentation module is constructed based on fast Fourier convolution and a Unet network and is used to generate a segmentation result of the input image;

[0021] The processing process of the initial model of MRI image segmentation based on frequency-space collaboration constructed includes the following steps:

[0022] The unlabeled MRI image data information is generated into a damaged image with missing regions through the image mask module;

[0023] The damaged image with the missing area is processed by an image inpainting module to obtain an image reconstruction result;

[0024] The labeled MRI image and the unlabeled MRI image are collectively taken as input images, and are processed by an image segmentation module to obtain an image segmentation result;

[0025] The image reconstruction result is processed by the image segmentation module to obtain a segmentation result of the image reconstruction result;

[0026] Based on the image segmentation result and the segmentation result of the image reconstruction result, an initial MRI image segmentation model based on frequency-space collaboration is trained;

[0027] After the training is completed, a separate image segmentation module is used as a final MRI image segmentation model based on frequency-space collaboration.

[0028] The processing process of the image mask module specifically includes the following steps:

[0029] A binary mask m∈R H×W is generated by using a random mask scheme, where H is the height of the input image, and W is the width of the input image.

[0030] The mask m includes a plurality of patch blocks h is the height of the patch block, and w is the width of the patch block; the area of the missing area is controlled by setting a mask ratio parameter, and the position of the missing area is ensured to be a random position.

[0031] Finally, a damaged image x c with a missing area is obtained. c x

[0032] The processing process of the image inpainting module specifically includes the following steps:

[0033] The image inpainting module includes an inpainting base network submodule and a reconstruction submodule.

[0034] The processing process of the reconstruction submodule is represented as wherein is the image reconstruction result output by the reconstruction submodule, and x′ is an image prediction result output by the inpainting base network submodule.

[0035] A Unet network is used as the basic network of the inpainting base network submodule, and a plurality of convolution modules in the Unet network are replaced by fast Fourier transform modules to construct the inpainting base network submodule.

[0036] The input image is processed by an encoder first 3*3 convolution and an encoder second 3*3 convolution to obtain an encoder first feature; the encoder first feature is processed by an encoder first max pooling module, an encoder third 3*3 convolution, and an encoder first fast Fourier transform module to obtain an encoder second feature; the encoder second feature is processed by an encoder second max pooling module, an encoder fourth 3*3 convolution, and an encoder second fast Fourier transform module to obtain an encoder third feature; the encoder third feature is processed by an encoder third max pooling module, an encoder fifth 3*3 convolution, and an encoder third fast Fourier transform module to obtain an encoder fourth feature; and the encoder fourth feature is processed by an encoder fourth max pooling module, an encoder sixth 3*3 convolution, and an encoder fourth fast Fourier transform module to obtain an encoder fifth feature;

[0037] The decoder fifth feature is processed by a decoder first max pooling module and a decoder 1*1 convolution to obtain a decoder first feature; the decoder first feature is connected with the encoder fourth feature, and then processed by a decoder first upsampling module, a decoder first 3*3 convolution, and a decoder first fast Fourier transform module to obtain a decoder second feature; the decoder second feature is connected with the encoder third feature, and then processed by a decoder second upsampling module, a decoder second 3*3 convolution, and a decoder second fast Fourier transform module to obtain a decoder third feature; the decoder third feature is connected with the encoder second feature, and then processed by a decoder third upsampling module, a decoder third 3*3 convolution, and a decoder third fast Fourier transform module to obtain a decoder fourth feature; and the decoder fourth feature is connected with the encoder first feature, and then processed by a decoder fourth upsampling module, a decoder fourth 3*3 convolution, and a decoder softmax module to obtain an image prediction result output by the repair base network sub-module;

[0038] The processing procedures of the encoder first 3*3 convolution to the encoder sixth 3*3 convolution and the decoder first 3*3 convolution to the decoder fourth 3*3 convolution are the same, and each includes 3*3 convolution processing, batch normalization processing, and LeakyReLU activation function processing performed in sequence.

[0039] The processing procedure of the decoder 1*1 convolution includes 1*1 convolution processing, batch normalization processing, and LeakyReLU activation function processing performed in sequence.

[0040] The processing procedure of the image segmentation module specifically includes the following steps:

[0041] The image segmentation module includes a segmentation base sub-network module and a segmentation sub-module.

[0042] The processing procedure of the segmentation sub-module is as follows: the prediction result output by the segmentation basic sub-network module is obtained through an argmax operation to obtain the image segmentation result output by the image segmentation module;

[0043] The Unet network is used as the basic network of the segmentation basic sub-network module, and some convolution modules in the Unet network are replaced by fast Fourier transform modules to construct the segmentation basic sub-network module.

[0044] The encoder part: the input image is processed by the first 3*3 convolution of the encoder and the second 3*3 convolution of the encoder to obtain the first encoder feature; the first encoder feature is processed by the first maximum pooling module of the encoder, the third 3*3 convolution of the encoder, the first fast Fourier transform module of the encoder to obtain the second encoder feature; the second encoder feature is processed by the second maximum pooling module of the encoder, the fourth 3*3 convolution of the encoder, the second fast Fourier transform module of the encoder to obtain the third encoder feature; the third encoder feature is processed by the third maximum pooling module of the encoder, the fifth 3*3 convolution of the encoder, the third fast Fourier transform module of the encoder to obtain the fourth encoder feature; the fourth encoder feature is processed by the fourth maximum pooling module of the encoder, the sixth 3*3 convolution of the encoder, the fourth fast Fourier transform module of the encoder to obtain the fifth encoder feature.

[0045] The decoder part: the fifth encoder feature is processed by the first maximum pooling module of the decoder and the 1*1 convolution of the decoder to obtain the first decoder feature; the first decoder feature is connected with the fourth encoder feature, and then processed by the first upsampling module of the decoder, the first 3*3 convolution of the decoder and the first fast Fourier transform module of the decoder to obtain the second decoder feature; the second decoder feature is connected with the third encoder feature, and then processed by the second upsampling module of the decoder, the second 3*3 convolution of the decoder and the second fast Fourier transform module of the decoder to obtain the third decoder feature; the third decoder feature is connected with the second encoder feature, and then processed by the third upsampling module of the decoder, the third 3*3 convolution of the decoder and the third fast Fourier transform module of the decoder to obtain the fourth decoder feature; the fourth decoder feature is connected with the first encoder feature, and then processed by the fourth upsampling module of the decoder, the fourth 3*3 convolution of the decoder and the softmax module of the decoder to obtain the prediction result output by the segmentation basic sub-network module.

[0046] The processing procedures of the first 3*3 convolution of the encoder to the sixth 3*3 convolution of the encoder, and the first 3*3 convolution of the decoder to the fourth 3*3 convolution of the decoder are the same, and all include 3*3 convolution processing, batch normalization processing and LeakyReLU activation function processing in sequence.

[0047] The processing procedure of the decoder 1*1 convolution includes sequentially performing 1*1 convolution processing, batch normalization processing and LeakyReLU activation function processing;

[0048] The encoder part of the segmentation base subnetwork module of the image segmentation module is the same as the encoder part of the repair base network sub-module of the image repair module in structure and parameters.

[0049] The encoder first Fourier transform module to the encoder fourth Fourier transform module and the decoder first Fourier transform module to the decoder third Fourier transform module in the image segmentation module and the encoder first Fourier transform module to the encoder fourth Fourier transform module and the decoder first Fourier transform module to the decoder third Fourier transform module in the repair base network sub-module of the image repair module are the same in structure and processing procedure; the processing procedure includes the following steps:

[0050] The input feature map X is divided into local feature maps X local and global feature maps X global , where α in is a set percentage of channels allocated to the global feature part;

[0051] X local is processed through the first 3*3 convolution to obtain the first convolution feature Y local Y local =Conv 3×3 (X local ), where Conv 3×3 () represents 3*3 convolution processing;

[0052] X local is processed through the second 3*3 convolution to obtain the second convolution feature Y l→g Y l→g =Conv 3×3 (X local );

[0053] X global is processed through the third 3*3 convolution to obtain the third convolution feature Y g→l Y g→l =Conv 3×3 (X global );

[0054] X global is sequentially processed through the fourth 3*3 convolution, batch normalization and LeakyReLU activation function to obtain the Fourier pre-feature X global_pre X global_pre =ReLU(BN(Conv3×3 (X global )));

[0055] X global_pre The frequency domain feature X′ is obtained by converting it to the frequency domain through real two-dimensional fast Fourier transform global X′ global =Concat(RFFT2d(X global_pre )), where RFFT2d() represents a real two-dimensional fast Fourier transform operation, and Concat() concatenates the real and imaginary parts obtained by Fourier transform;

[0056] X′ global The inverse Fourier feature X′ is obtained by sequentially passing through the fourth 3*3 convolution, batch normalization and LeakyReLU activation function. global_pre X′ global_pre =ReLU(BN(Conv 3×3 (X′ global )));

[0057] X′ global_pre The spatial domain feature Y' is obtained by converting it to the spatial domain through the inverse real two-dimensional fast Fourier transform global Y' global =iRFFT2d(Split(X′ global_pre )), where Split() is the operation of splitting the features before inverse Fourier transform back into real and imaginary parts, and iRFFT2d() is the inverse real two-dimensional fast Fourier transform operation;

[0058] Will Y' global With X global_pre Perform element-by-element addition and then perform 1*1 convolution to obtain the fourth convolution feature Y global ;

[0059] Y local and Y g→l Perform element-by-element addition, and then process it through batch normalization and LeakyReLU activation function to obtain the local feature YY local , expressed as YY local =ReLU(BN(Y local +Y g→l ));

[0060] Y l→g and Y global Perform element-by-element addition, and then process it through batch normalization and LeakyReLU activation function to obtain the global feature YY global , expressed as YY global =ReLU(BN(Y global+Y l→g ));

[0061] YY local and YY global Splicing to get the final output Y.

[0062] The training described in step S4 specifically includes the following steps:

[0063] The first loss L is calculated using the following formula: ipt :

[0064] L ipt =α*L per +β*L ssim

[0065] Where α is the first weight set; β is the second weight set; L per is the perceptual loss; L ssim For structural losses;

[0066] The perceptual loss is to input the image reconstruction result into the VGG19 feature extractor, compare the feature maps of pool1 to pool5 in the VGG19 feature extractor with the feature maps of the original image corresponding to the image reconstruction result, and obtain L per The calculation formula is Where N is the number of activation layers required to calculate the VGG19 feature extractor, n i is the activation layer of the corresponding level, The product of the height, width and number of channels of the feature map calculated for the original image, is the product of the height, width and number of channels of the feature map calculated by the image reconstruction result, x is the original image corresponding to the image reconstruction result, is the image reconstruction result;

[0067] L ssim The calculation formula is where μ x is the mean of x, for The mean of , C1 is the first constant set, For x and The covariance of x is the standard deviation of x, for The standard deviation of , C2 is the second constant set;

[0068] The second loss L is calculated using the following formula: sup-seg :

[0069]

[0070] Wherein Dice() is a Dice loss function; y is a real segmentation result of a labeled image; is an image segmentation result output by the image segmentation module for the labeled image;

[0071] The third loss L is calculated by using the following formula self-con :

[0072]

[0073] Wherein is an image segmentation result output by the image segmentation module for the unlabeled image; is an image segmentation result output by the image segmentation module for the image reconstruction result;

[0074] The total loss L is calculated by using the following formula during training

[0075]

[0076] Wherein N is the number of labeled images; M is the number of unlabeled images.

[0077] The application further provides an imaging method comprising the MRI image segmentation method based on frequency-space cooperation.

[0078] A. Obtain a target MRI image;

[0079] B. Perform image segmentation on the target MRI image obtained in step A by using the MRI image segmentation method based on frequency-space cooperation to obtain an image segmentation result;

[0080] C. Label and perform secondary imaging on the image segmentation result obtained in step B on the target MRI image to obtain a target MRI image with the segmentation result.

[0081] The MRI image segmentation method and the imaging method based on frequency-space cooperation provided by the application can realize the segmentation of MRI images by using the random mask scheme, fast Fourier convolution and Unet network, and realize the cooperation of frequency-space based on fast Fourier convolution, so that the application not only can realize the segmentation of MRI images, but also has higher reliability and better accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0082] Figure 1 Fig. 1 is a method flow diagram of the segmentation method of the application.

[0083] Figure 2 Fig. 2 is a network structure diagram of the repair base network submodule and the segmentation base network submodule of the segmentation method of the application.

[0084] Figure 3 Schematic diagram of the structure of the fast Fourier transform module of the segmentation method of the present invention.

[0085] Figure 4 Schematic diagram of qualitative comparison between the segmentation method of the present invention and other methods on the 10% annotated ACDC dataset.

[0086] Figure 5 Schematic diagram of qualitative comparison between the segmentation method of the present invention and other methods on the 10% annotated PROMISE12 dataset.

[0087] Figure 6 This is a schematic diagram of the t-SNE visualization results of the last convolutional layer feature representation of the encoder trained on the ACDC dataset for the segmentation method of the present invention.

[0088] Figure 7 Schematic diagram of the process flow of the imaging method of the present invention. DETAILED DESCRIPTION

[0089] like Figure 1 The figure shows a flow chart of the segmentation method of the present invention: The MRI image segmentation method based on frequency-space synergy disclosed by the present invention comprises the following steps:

[0090] S1. Acquire MRI image data information of the target area.

[0091] S2. Preprocess the MRI image data information obtained in step S1 to construct a training data set; specifically, the steps include:

[0092] The MRI image data information obtained in step S1 includes MRI image data information with labels and MRI image data information without labels;

[0093] The MRI image data information obtained in step S1 is preprocessed; the preprocessing includes image data cleaning and image data filtering.

[0094] S3. Construct an initial model for MRI image segmentation based on frequency-space synergy using a random masking scheme, fast Fourier convolution, and a Unet network. This includes the following steps:

[0095] The constructed initial model of MRI image segmentation based on frequency-space synergy includes image mask module, image restoration module and image segmentation module;

[0096] An image masking module is constructed based on a random masking scheme to process the input image information to generate a damaged image with missing areas. In a specific implementation, the input of the image masking module is an unlabeled MRI image.

[0097] An image inpainting module is constructed based on a fast Fourier convolution and a Unet network, and is used to generate an image reconstruction result of a damaged image with a missing area output by the image mask module;

[0098] An image segmentation module is constructed based on a fast Fourier convolution and a Unet network, and is used to generate a segmentation result of an input image; in a specific implementation, the input of the image segmentation module is labeled MRI image data information and unlabeled MRI image data information;

[0099] The processing procedure of the constructed MRI image segmentation initial model based on frequency-space cooperation includes the following steps:

[0100] The unlabeled MRI image data information is input into the image mask module to generate a damaged image with a missing area;

[0101] The damaged image with the missing area is input into the image inpainting module to obtain an image reconstruction result;

[0102] The labeled MRI image and the unlabeled MRI image are jointly input into the image segmentation module to obtain an image segmentation result;

[0103] In order to fully utilize the unlabeled MRI image, the image reconstruction result is input into the image segmentation module to obtain a segmentation result of the image reconstruction result;

[0104] Based on the image segmentation result and the segmentation result of the image reconstruction result, the constructed MRI image segmentation initial model based on frequency-space cooperation is trained;

[0105] After the training is completed, a separate image segmentation module is used as a final MRI image segmentation model based on frequency-space cooperation.

[0106] In a specific implementation, the processing procedure of the image mask module specifically includes the following steps:

[0107] A binary mask m is generated by using a random mask scheme, and m e R H×W where H is the height of the input image, and W is the width of the input image;

[0108] m includes a plurality of patch blocks h is the height of the patch block, and w is the width of the patch block; a mask ratio parameter (with a value range of [0, 1]) is set to control the area of the missing area, and the position of the missing area is ensured to be a random position;

[0109] Finally, a damaged image x with a missing area is obtained c x is c= x ￠ (1 - m), wherein x is an input image, and ￠ is an element-wise multiplication operation on pixels in corresponding positions.

[0110] The processing procedure of the image inpainting module includes the following steps:

[0111] The image inpainting module includes an inpainting base network submodule and a reconstruction submodule.

[0112] The processing procedure of the reconstruction submodule is represented as wherein is an image reconstruction result output by the reconstruction submodule, and x' is an image prediction result output by the inpainting base network submodule.

[0113] The Unet network is used as the base network of the inpainting base network submodule, and several convolution modules in the Unet network are replaced by fast Fourier transform modules to construct the inpainting base network submodule (structure as shown in Figure 2

[0114] The input image is processed by the first 3*3 convolution of the encoder and the second 3*3 convolution of the encoder to obtain the first encoder feature; the first encoder feature is processed by the first maximum pooling module of the encoder, the third 3*3 convolution of the encoder, the first fast Fourier transform module of the encoder to obtain the second encoder feature; the second encoder feature is processed by the second maximum pooling module of the encoder, the fourth 3*3 convolution of the encoder, the second fast Fourier transform module of the encoder to obtain the third encoder feature; the third encoder feature is processed by the third maximum pooling module of the encoder, the fifth 3*3 convolution of the encoder, the third fast Fourier transform module of the encoder to obtain the fourth encoder feature; the fourth encoder feature is processed by the fourth maximum pooling module of the encoder, the sixth 3*3 convolution of the encoder, the fourth fast Fourier transform module of the encoder to obtain the fifth encoder feature;

[0115] ​The decoder part: the fifth feature of the encoder is obtained by the first maximum pooling module of the decoder and the 1*1 convolution processing of the decoder; the first feature of the decoder is connected with the fourth feature of the encoder, and then processed by the first up-sampling module of the decoder, the first 3*3 convolution of the decoder and the first fast Fourier transform module of the decoder to obtain the second feature of the decoder; the second feature of the decoder is connected with the third feature of the encoder, and then processed by the second up-sampling module of the decoder, the second 3*3 convolution of the decoder and the second fast Fourier transform module of the decoder to obtain the third feature of the decoder; the third feature of the decoder is connected with the second feature of the encoder, and then processed by the third up-sampling module of the decoder, the third 3*3 convolution of the decoder and the third fast Fourier transform module of the decoder to obtain the fourth feature of the decoder; the fourth feature of the decoder is connected with the first feature of the encoder, and then processed by the fourth up-sampling module of the decoder, the fourth 3*3 convolution of the decoder and the softmax module of the decoder to obtain the image prediction result output by the repair basic network sub-module;

[0116] The processing procedures of the first 3*3 convolution of the encoder to the sixth 3*3 convolution of the encoder and the first 3*3 convolution of the decoder to the fourth 3*3 convolution of the decoder are the same, and all include 3*3 convolution processing, batch normalization processing and LeakyReLU activation function processing in sequence.

[0117] The processing procedure of the 1*1 convolution of the decoder includes 1*1 convolution processing, batch normalization processing and LeakyReLU activation function processing in sequence.

[0118] The processing procedure of the image segmentation module specifically includes the following steps:

[0119] The image segmentation module includes a segmentation basic sub-network module and a segmentation sub-module.

[0120] The processing procedure of the segmentation sub-module is that the prediction result output by the segmentation basic sub-network module is obtained by argmax operation to obtain the image segmentation result output by the image segmentation module.

[0121] The Unet network is used as the basic network of the segmentation basic sub-network module, and some convolution modules therein are replaced by fast Fourier transform modules to construct the segmentation basic sub-network module (structure as shown in Figure 2

[0122] ​The input image is processed by an encoder first 3*3 convolution and an encoder second 3*3 convolution to obtain an encoder first feature; the encoder first feature is processed by an encoder first max pooling module, an encoder third 3*3 convolution, and an encoder first fast Fourier transform module to obtain an encoder second feature; the encoder second feature is processed by an encoder second max pooling module, an encoder fourth 3*3 convolution, and an encoder second fast Fourier transform module to obtain an encoder third feature; the encoder third feature is processed by an encoder third max pooling module, an encoder fifth 3*3 convolution, and an encoder third fast Fourier transform module to obtain an encoder fourth feature; and the encoder fourth feature is processed by an encoder fourth max pooling module, an encoder sixth 3*3 convolution, and an encoder fourth fast Fourier transform module to obtain an encoder fifth feature;

[0123] The decoder fifth feature is processed by a decoder first max pooling module and a decoder 1*1 convolution to obtain a decoder first feature; the decoder first feature is connected with the encoder fourth feature, and then processed by a decoder first upsampling module, a decoder first 3*3 convolution, and a decoder first fast Fourier transform module to obtain a decoder second feature; the decoder second feature is connected with the encoder third feature, and then processed by a decoder second upsampling module, a decoder second 3*3 convolution, and a decoder second fast Fourier transform module to obtain a decoder third feature; the decoder third feature is connected with the encoder second feature, and then processed by a decoder third upsampling module, a decoder third 3*3 convolution, and a decoder third fast Fourier transform module to obtain a decoder fourth feature; and the decoder fourth feature is connected with the encoder first feature, and then processed by a decoder fourth upsampling module, a decoder fourth 3*3 convolution, and a decoder softmax module to obtain a prediction result output by the segmentation basic subnetwork module;

[0124] The processing procedures of the encoder first 3*3 convolution to the encoder sixth 3*3 convolution and the decoder first 3*3 convolution to the decoder fourth 3*3 convolution are the same, and each includes 3*3 convolution processing, batch normalization processing, and LeakyReLU activation function processing performed in sequence.

[0125] The processing procedure of the decoder 1*1 convolution includes 1*1 convolution processing, batch normalization processing, and LeakyReLU activation function processing performed in sequence.

[0126] The encoder part of the segmentation basic subnetwork module of the image segmentation module is the same as the encoder part of the repair basic network subnetwork module of the image repair module in structure and parameters.

[0127] The encoder first fast Fourier transform module to the encoder fourth fast Fourier transform module and the decoder first fast Fourier transform module to the decoder third fast Fourier transform module in the image segmentation module, as well as the encoder first fast Fourier transform module to the encoder fourth fast Fourier transform module and the decoder first fast Fourier transform module to the decoder third fast Fourier transform module in the restoration basic network submodule of the image restoration module, have the same structure and processing process (such as Figure 3 The processing process includes the following steps:

[0128] The input feature map X is divided into local feature maps X along the channel dimension local and the global feature map X global ,in α in The percentage of channels allocated to the global feature part is set, and the preferred value is 0.5;

[0129] X local Through the first 3*3 convolution process, the first convolution feature Y is obtained local Y local =Conv 3×3 (X local ), where Conv 3×3 () indicates 3*3 convolution processing;

[0130] X local Through the second 3*3 convolution process, the second convolution feature Y is obtained l→g Y l→g =Conv 3×3 (X local );

[0131] X global Through the third 3*3 convolution process, the third convolution feature Y is obtained g→l Y g→l =Conv 3×3 (X global );

[0132] X global The Fourier feature X is obtained by sequentially passing through the fourth 3*3 convolution, batch normalization and LeakyReLU activation function. global_pre For X global_pre =ReLU(BN(Conv 3×3 (X global )));

[0133] X global_pre The frequency domain feature X′ is obtained by converting it to the frequency domain through real two-dimensional fast Fourier transform global X′ global=Concat(RFFT2d(X global_pre )), where RFFT2d() represents a real two-dimensional fast Fourier transform operation, and Concat() concatenates the real and imaginary parts obtained by Fourier transform;

[0134] X′ global The inverse Fourier feature X′ is obtained by sequentially passing through the fourth 3*3 convolution, batch normalization and LeakyReLU activation function. global_pre X′ global_pre =ReLU(BN(Conv 3×3 (X′ global )));

[0135] X′ global_pre The spatial domain feature Y' is obtained by converting it to the spatial domain through the inverse real two-dimensional fast Fourier transform global Y' global =iRFFT2d(Split(X′ global_pre )), where Split() is the operation of splitting the features before inverse Fourier transform back into real and imaginary parts, and iRFFT2d() is the inverse real two-dimensional fast Fourier transform operation;

[0136] Will Y' global With X global_pre Perform element-by-element addition and then perform 1*1 convolution to obtain the fourth convolution feature Y global ;

[0137] Y local and Y g→l Perform element-by-element addition, and then process it through batch normalization and LeakyReLU activation function to obtain the local feature YY local , expressed as YY local =ReLU(BN(Y local +Y g→l ));

[0138] Y l→g and Y global Perform element-by-element addition, and then process it through batch normalization and LeakyReLU activation function to obtain the global feature YY global , expressed as YY global =ReLU(BN(Y global +Y l→g ));

[0139] YY local and YY global Splicing to get the final output Y.

[0140] S4. Training the initial model of MRI image segmentation based on frequency-space cooperation built in step S3 by using the training data set obtained in step S2 to obtain the model of MRI image segmentation based on frequency-space cooperation.

[0141] The training process specifically comprises the following steps:

[0142] The first loss L is calculated by using the following formula ipt :

[0143] L ipt =α*L per +β*L ssim

[0144] In the formula, α is the first weight set; β is the second weight set; L per is the perceptual loss; L ssim is the structural loss; the perceptual loss and the structural loss are used to better generate a more realistic and structurally reasonable repair image;

[0145] The perceptual loss is that the image reconstruction result is input into the VGG19 feature extractor, the feature maps of pool1~pool5 in the VGG19 feature extractor are compared with the feature maps of the original image corresponding to the image reconstruction result, and the L per is obtained. The calculation formula of L per is Where N is the number of activation layers required for calculation of the VGG19 feature extractor, n i is the activation layer corresponding to the level, is the product of the height, width and channel number of the feature map calculated from the original image, is the product of the height, width and channel number of the feature map calculated from the image reconstruction result, x is the original image corresponding to the image reconstruction result, is the image reconstruction result;

[0146] The calculation formula of L ssim is Where μ x is the mean of x, is the mean of , C1 is the first constant set, is the covariance of x and , σ x is the standard deviation of x, is the standard deviation of , C2 is the second constant set; L ssim compares the repair image with the original image from the aspects of brightness, contrast and structure;

[0147] The second loss L sup-seg is calculated by using the following formula:

[0148]

[0149] wherein Dice() is a Dice loss function; y is a real segmentation result of a labeled image; is an image segmentation result output by the image segmentation module for the labeled image;

[0150] The third loss L is calculated by using the following formula self-con :

[0151]

[0152] wherein is an image segmentation result output by the image segmentation module for the unlabeled image; is an image segmentation result output by the image segmentation module for the image reconstruction result;

[0153] The total loss L is calculated by using the following formula during the training

[0154]

[0155] wherein N is the number of labeled images; and M is the number of unlabeled images.

[0156] S5. The MRI image segmentation model based on the frequency-space collaboration obtained in step S4 is used to perform segmentation of the MRI image of the target region.

[0157] The effect of the segmentation method of the present application is described below in combination with an embodiment:

[0158] The segmentation method of the present application is compared with the prior art in terms of effect on the two public data sets of ACDC and PROMISE12.

[0159] The ACDC data set is a data set for the segmentation task of the endocardium of the left ventricle of the heart. The PROMISE12 data set is a data set for the segmentation task of the prostate.

[0160] The segmentation method of the present application is compared with other methods in terms of performance on the 10% labeled and 20% labeled ACDC data sets. The performance indicators are shown in Tables 1 and 2:

[0161] Table 1: Performance indicator table on the 10% labeled ACDC data set

[0162]

[0163] Table 2: Performance indicator table on the 20% labeled ACDC data set

[0164]

[0165] ↑ in the table indicates that the larger the index value, the better the performance, and ↓ indicates that the smaller the index value, the better the performance.

[0166] From the experimental results in Table 1, it can be seen that the segmentation method of the present application has a significant technical advantage on the 10% labeled ACDC data set. In terms of segmentation accuracy, the DSC value of the segmentation method of the present application reaches 87.91%, which is 0.56 percentage points higher than the 87.35% of the existing optimal method SS-Net; the Jaccard index reaches 79.15%, which is 0.78 percentage points higher than the 78.37% of SS-Net. This shows that the segmentation method of the present application can achieve higher segmentation precision under the condition of less labeled data. In terms of segmentation boundary quality, although XNet performs best in 95HD and ASD indicators, the 95HD of the segmentation method of the present application is 3.49, and the ASD is 1.24, which still remains at a relatively optimal level, and is significantly better than other comparison methods. Table 2 shows that as the proportion of labeled data increases to 20%, the technical advantage of the segmentation method of the present application is further highlighted, with the DSC value increasing to 88.83%, which is 0.59 percentage points higher than the 88.24% of the sub-optimal method XNet; the Jaccard index reaches 80.61%, which is 1.00 percentage points higher than the 79.61% of PACAS-WA. It is worth noting that under the condition of 20% labeled data, the performance of each method generally improves, but the segmentation method of the present application can still maintain the optimal segmentation accuracy, which reflects its stability and adaptability under different data conditions.

[0167] The performance of the segmentation method of the present application and other schemes on the 10% labeled and 20% labeled PROMISE2 data set is compared, and the performance indicators are shown in Tables 3 and 4:

[0168] Table 3 Performance indicator table on 10% labeled PROMISE2 data set

[0169]

[0170] Table 4 Performance indicator table on 20% labeled PROMISE2 data set

[0171]

[0172] From the experimental results in Table 3, it can be clearly seen that the segmentation method of the present application has more significant technical advantages on the PROMISE12 dataset. In terms of segmentation accuracy, the DSC value of the segmentation method of the present application reaches 78.37%, which is 10.47 percentage points higher than 67.90% of the existing optimal method PACAS-WA, which is a very significant performance improvement; the Jaccard index reaches 67.02%, which is 8.77 percentage points higher than 58.25% of SS-Net. In terms of boundary accuracy, the segmentation method of the present application also achieves the optimal performance, with a 95HD value of 10.27, which is slightly improved compared to 10.62 of SS-Net; the ASD value is 1.87, which is 57% higher than 4.34 of SS-Net, and is significantly better than all the comparison methods. It is worth noting that on the PROMISE12 dataset, the overall performance of the existing methods is relatively low (the highest DSC is only 67.90%), while the segmentation method of the present application can reach 78.37%, indicating that it has significant technical advantages in processing more challenging medical image segmentation tasks. Table 4 shows that under the condition of 20% labeled data, the segmentation method of the present application continues to maintain overall leadership, with a DSC value of 80.44%, which is 5.87 percentage points higher than 74.57% of the suboptimal method PACAS-WA; the Jaccard index reaches 69.33%, which is 4.55 percentage points higher than 64.78% of PACAS-WA. In terms of boundary quality, 95HD is 11.72 and ASD is 2.63, which although slightly increases compared to 10% labeled data, still remains at the optimal level among all methods.

[0173] Figure 4 and Figure 5 Qualitative comparison results of the segmentation method of the present application and other methods on the 10% labeled ACDC and PROMISE12 datasets are shown.

[0174] wherein, Figure 4 (a) is MT (NeurIPS’17) Qualitative schematic diagram of the scheme; Figure 4 (b) is UA-MT (MICCAI’19) Qualitative schematic diagram of the scheme; Figure 4 (c) is EM (CVPR’19) Qualitative schematic diagram of the scheme; Figure 4 (d) is CCT (CVPR’20) Qualitative schematic diagram of the scheme; Figure 4 (e) is URPC (MICCAI’21) Qualitative schematic diagram of the scheme;

[0175] Figure 4 (f) is CPS (CVPR’21) Qualitative schematic diagram of the scheme; Figure 4 (g) is SS-Net(MICCAI’22) qualitative illustration of the scheme; Figure 4 (h) is a qualitative illustration of the scheme; Figure 4 (i) is XNet (ICCV’23) qualitative illustration of the scheme; Figure 4 (j) is PACAS-WA (ICASSP’24) qualitative illustration of the scheme; Figure 4 (k) is a qualitative illustration of the segmentation method of the application; Figure 4 (l) is a real label.

[0176] wherein, Figure 5 (a) is MT (NeurIPS’17) qualitative illustration of the scheme; Figure 5 (b) is UA-MT (MICCAI’19) qualitative illustration of the scheme; Figure 5 (c) is EM (CVPR’19) qualitative illustration of the scheme; Figure 5 (d) is CCT (CVPR’20) qualitative illustration of the scheme; Figure 5 (e) is URPC (MICCAI’21) qualitative illustration of the scheme;

[0177] Figure 5 (f) is CPS (CVPR’21) qualitative illustration of the scheme; Figure 5 (g) is SS-Net (MICCAI’22) qualitative illustration of the scheme; Figure 5 (h) is a qualitative illustration of the scheme; Figure 5 (i) is XNet (ICCV’23) qualitative illustration of the scheme; Figure 5 (j) is PACAS-WA (ICASSP’24) qualitative illustration of the scheme; Figure 5 (k) is a qualitative illustration of the segmentation method of the application; Figure 5 (l) is a real label.

[0178] Figure 4 and Figure 5 The visualization results of the application segmentation method clearly show that, compared with other methods, the application segmentation method can more accurately locate and segment the relevant region.

[0179] At the same time, in order to more intuitively show the advantages of the segmentation method of the application, t-SNE visualization method is used for in-depth analysis. Figure 6The visualization results on the ACDC dataset test set are shown, and the following phenomena can be observed through comparative analysis: Notably, the feature representation generated by the baseline method (a) is chaotic, and there is significant overlap between different categories (right ventricle, myocardium, left ventricle), indicating that the model is difficult to distinguish between these categories, resulting in poor feature discriminability. When trained using only the single-task method (b), the feature representation is improved. The dual-task method (c) that integrates segmentation and repair tasks achieves the best feature separation effect, with clear boundaries between feature distributions of different categories and minimized overlap areas. This shows that the segmentation method of the present application significantly enhances the model's ability to capture richer and more meaningful feature representations, thereby achieving better class distinction.

[0180] The segmentation method of the present application proposes an innovative dual-task feature mining framework that integrates semi-supervised learning and self-supervised learning. By simultaneously training the image segmentation and image inpainting tasks, the model can analyze the MRI image from multiple angles, thereby learning more rich and meaningful feature representations. To maximize the value of unlabeled data, a consistency loss is introduced, which uses the image generated by the image inpainting task as a natural disturbance source to ensure the consistency between the original image and the inpainted image, thereby eliminating the need for an additional image enhancement step. In addition, the present application designs a hybrid receptive field model based on frequency-space synergy as the backbone network of the dual-task feature mining framework. This model uses Fourier convolution instead of traditional spatial convolution to achieve a hybrid receptive field, effectively improving the ability to capture global information while maintaining high computational efficiency and low storage requirements. Experiments have verified the effectiveness of the segmentation method of the present application on multiple MRI image datasets, proving that the combination of semi-supervised and self-supervised learning methods based on a hybrid receptive field can improve the performance of MRI image segmentation while maintaining advantages in computational efficiency and model size.

[0181] As Figure 7 The method flowchart of the imaging method of the present application is shown: The imaging method disclosed in the present application, which includes the MRI image segmentation method based on frequency-space synergy, comprises the following steps:

[0182] A. Obtain the target MRI image;

[0183] B. Perform image segmentation on the target MRI image obtained in step A using the MRI image segmentation method based on frequency-space synergy to obtain an image segmentation result;

[0184] C. Label and re-image the image segmentation result obtained in step B on the target MRI image to obtain a target MRI image with segmentation results.

[0185] The imaging method provided by the application can be directly applied to an existing MRI medical image device or a terminal (such as a computer); in a specific application, an actual medical MRI image is acquired by using an existing scheme, and then the acquired data is input into a corresponding machine device or terminal; at this time, the machine device or terminal can obtain a segmentation result of the actual MRI image according to the imaging method disclosed by the application, mark and display the segmentation result on the original image through different types of representations (such as colors), and then perform secondary imaging and output; at this time, the output image is a medical MRI image with the segmentation result, which can reflect the actual medical image and the corresponding segmentation result content, thereby greatly facilitating subsequent work of clinical medical personnel and laboratory experimental personnel.

Claims

1. A method for MRI image segmentation based on frequency-space synergy, comprising the following steps: S1. Acquire MRI image data information of the target area; S2. Preprocessing the MRI image data information obtained in step S1 to construct a training data set; S3. Construct an initial model for MRI image segmentation based on frequency-space synergy using a random masking scheme, fast Fourier convolution, and Unet network. S4. Using the training data set obtained in step S2, the initial MRI image segmentation model based on frequency-space collaboration constructed in step S3 is trained to obtain an MRI image segmentation model based on frequency-space collaboration; S5. Use the frequency-space synergy-based MRI image segmentation model obtained in step S4 to segment the MRI image of the target area.

2. The MRI image segmentation method based on frequency-space synergy according to claim 1, characterized in that The pre-processing described in step S2 specifically includes the following steps: The MRI image data information obtained in step S1 includes MRI image data information with labels and MRI image data information without labels; The MRI image data information obtained in step S1 is preprocessed; the preprocessing includes image data cleaning and image data filtering.

3. The MRI image segmentation method based on frequency-space synergy according to claim 1 or 2, characterized in that The initial model for MRI image segmentation based on frequency-space collaboration is constructed based on the random mask scheme, fast Fourier convolution and Unet network described in step S3, including the following steps: The constructed initial model of MRI image segmentation based on frequency-space synergy includes image mask module, image restoration module and image segmentation module; An image mask module is constructed based on a random mask scheme to process the input image information to generate a damaged image with missing areas; An image restoration module is constructed based on fast Fourier convolution and Unet network to generate image reconstruction results of damaged images with missing areas output by the image mask module; An image segmentation module is constructed based on fast Fourier convolution and Unet network to generate the segmentation results of the input image; The processing process of constructing the initial model of MRI image segmentation based on frequency-space synergy includes the following steps: The unlabeled MRI image data information is used to generate a damaged image with missing areas through the image mask module; The damaged image with missing areas is processed by the image restoration module to obtain the image reconstruction result; The labeled MRI image and the unlabeled MRI image are taken as input images and processed by the image segmentation module to obtain the image segmentation result; The image reconstruction result is processed by an image segmentation module to obtain a segmentation result of the image reconstruction result; The constructed MRI image segmentation initial model based on frequency-space synergy is trained based on the segmentation results of the image segmentation results and the image reconstruction results; After the training is completed, a separate image segmentation module is used as the final MRI image segmentation model based on frequency-space synergy.

4. The MRI image segmentation method based on frequency-space synergy according to claim 3, characterized in that The processing of the image mask module includes the following steps: Generate binary mask m∈R using random masking scheme H×W , where H is the height of the input image and W is the width of the input image; m includes several patch blocks h is the height of the patch, and w is the width of the patch. The area of ​​the missing region is controlled by setting the mask ratio parameter, and the position of the missing region is guaranteed to be random. Finally, the damaged image x with missing areas is obtained c is x c =x⊙(1-m), where x is the input image and ⊙ is the element-by-element multiplication of the pixels at the corresponding position.

5. The MRI image segmentation method based on frequency-space synergy according to claim 4, characterized in that The processing of the image restoration module includes the following steps: The image restoration module includes a restoration basic network submodule and a reconstruction submodule; The process of rebuilding the submodule is expressed as in is the image reconstruction result output by the reconstruction submodule, and x′ is the image prediction result output by the restoration basic network submodule; The Unet network is used as the basic network of the repair basic network submodule, and several convolution modules are replaced with fast Fourier transform modules to construct the repair basic network submodule: Encoder part: The input image is processed by the first 3*3 convolution of the encoder and the second 3*3 convolution of the encoder to obtain the first feature of the encoder; the first feature of the encoder is processed by the first maximum pooling module of the encoder, the third 3*3 convolution of the encoder, and the first fast Fourier transform module of the encoder to obtain the second feature of the encoder; the second feature of the encoder is processed by the second maximum pooling module of the encoder, the fourth 3*3 convolution of the encoder, and the second fast Fourier transform module of the encoder to obtain the third feature of the encoder; the third feature of the encoder is processed by the third maximum pooling module of the encoder, the fifth 3*3 convolution of the encoder, and the third fast Fourier transform module of the encoder to obtain the fourth feature of the encoder; the fourth feature of the encoder is processed by the fourth maximum pooling module of the encoder, the sixth 3*3 convolution of the encoder, and the fourth fast Fourier transform module of the encoder to obtain the fifth feature of the encoder; Decoder part: The fifth feature of the encoder is processed by the first maximum pooling module of the decoder and the 1*1 convolution of the decoder to obtain the first feature of the decoder; After the decoder's first feature is concatenated with the encoder's fourth feature, it is processed through the decoder's first upsampling module, the decoder's first 3*3 convolution, and the decoder's first fast Fourier transform module to obtain the decoder's second feature. After the decoder's second feature is concatenated with the encoder's third feature, it is processed through the decoder's second upsampling module, the decoder's second 3*3 convolution, and the decoder's second fast Fourier transform module to obtain the decoder's third feature. The decoder's third feature is concatenated with the encoder's second feature and then processed through the decoder's third upsampling module, the decoder's third 3*3 convolution, and the decoder's third fast Fourier transform module to obtain the decoder's fourth feature. After the fourth feature of the decoder is concatenated with the first feature of the encoder, it is processed by the fourth upsampling module of the decoder, the fourth 3*3 convolution of the decoder, and the softmax module of the decoder to obtain the image prediction result output by the restoration basic network submodule; Among them, the processing process of the encoder's first 3*3 convolution to the encoder's sixth 3*3 convolution and the decoder's first 3*3 convolution to the decoder's fourth 3*3 convolution is the same, both including the sequential 3*3 convolution processing, batch normalization processing and LeakyReLU activation function processing; The decoder 1*1 convolution processing process includes 1*1 convolution processing, batch normalization processing and LeakyReLU activation function processing in sequence.

6. The MRI image segmentation method based on frequency-space synergy according to claim 5, characterized in that The processing process of the image segmentation module includes the following steps: The image segmentation module includes a segmentation basic sub-network module and a segmentation sub-module; The processing process of the segmentation submodule is as follows: the prediction result output by the segmentation basic subnetwork module is subjected to the argmax operation to obtain the image segmentation result output by the image segmentation module; The Unet network is used as the basic network of the segmentation basic sub-network module, and several convolution modules are replaced with fast Fourier transform modules to construct the segmentation basic sub-network module: Encoder part: The input image is processed by the first 3*3 convolution of the encoder and the second 3*3 convolution of the encoder to obtain the first feature of the encoder; the first feature of the encoder is processed by the first maximum pooling module of the encoder, the third 3*3 convolution of the encoder, and the first fast Fourier transform module of the encoder to obtain the second feature of the encoder; the second feature of the encoder is processed by the second maximum pooling module of the encoder, the fourth 3*3 convolution of the encoder, and the second fast Fourier transform module of the encoder to obtain the third feature of the encoder; the third feature of the encoder is processed by the third maximum pooling module of the encoder, the fifth 3*3 convolution of the encoder, and the third fast Fourier transform module of the encoder to obtain the fourth feature of the encoder; the fourth feature of the encoder is processed by the fourth maximum pooling module of the encoder, the sixth 3*3 convolution of the encoder, and the fourth fast Fourier transform module of the encoder to obtain the fifth feature of the encoder; Decoder part: The fifth feature of the encoder is processed by the first maximum pooling module of the decoder and the 1*1 convolution of the decoder to obtain the first feature of the decoder; After the decoder's first feature is concatenated with the encoder's fourth feature, it is processed through the decoder's first upsampling module, the decoder's first 3*3 convolution, and the decoder's first fast Fourier transform module to obtain the decoder's second feature. After the decoder's second feature is concatenated with the encoder's third feature, it is processed through the decoder's second upsampling module, the decoder's second 3*3 convolution, and the decoder's second fast Fourier transform module to obtain the decoder's third feature. The decoder's third feature is concatenated with the encoder's second feature and then processed through the decoder's third upsampling module, the decoder's third 3*3 convolution, and the decoder's third fast Fourier transform module to obtain the decoder's fourth feature. After the fourth feature of the decoder is concatenated with the first feature of the encoder, it is processed by the fourth upsampling module of the decoder, the fourth 3*3 convolution of the decoder, and the softmax module of the decoder to obtain the prediction result output by the segmentation basic sub-network module; Among them, the processing process of the encoder's first 3*3 convolution to the encoder's sixth 3*3 convolution and the decoder's first 3*3 convolution to the decoder's fourth 3*3 convolution is the same, both including the sequential 3*3 convolution processing, batch normalization processing and LeakyReLU activation function processing; The decoder 1*1 convolution process includes 1*1 convolution processing, batch normalization processing and LeakyReLU activation function processing in sequence; The encoder part of the segmentation basic sub-network module of the image segmentation module and the encoder part of the restoration basic sub-module of the image restoration module have the same structure and parameters.

7. The MRI image segmentation method based on frequency-space synergy according to claim 6, characterized in that The encoder first fast Fourier transform module to the encoder fourth fast Fourier transform module and the decoder first fast Fourier transform module to the decoder third fast Fourier transform module in the image segmentation module, and the encoder first fast Fourier transform module to the encoder fourth fast Fourier transform module and the decoder first fast Fourier transform module to the decoder third fast Fourier transform module in the restoration basic network submodule of the image restoration module have the same structure and processing process; the processing process includes the following steps: The input feature map X is divided into local feature maps X along the channel dimension local and the global feature map X global ,in α in is the percentage of channels allocated to the global feature part; X local Through the first 3*3 convolution process, the first convolution feature Y is obtained local Y local =Conv 3×3 (X local ), where Conv 3×3 () indicates 3*3 convolution processing; X local Through the second 3*3 convolution process, the second convolution feature Y is obtained l→g Y l→g =Conv 3×3 (X local ); X global Through the third 3*3 convolution process, the third convolution feature Y is obtained g→l Y g→l =Conv 3×3 (X global ); X global The Fourier feature X is obtained by sequentially passing through the fourth 3*3 convolution, batch normalization and LeakyReLU activation function. global_pre For X global_pre =ReLU(BN(Conv 3×3 (X global ))); X global_pre The frequency domain feature X′ is obtained by converting it to the frequency domain through real two-dimensional fast Fourier transform global X′ global =Concat(RFFT2d(X global_pre )), where RFFT2d() represents a real two-dimensional fast Fourier transform operation, and Concat() concatenates the real and imaginary parts obtained by Fourier transform; X′ global The inverse Fourier feature X′ is obtained by sequentially passing through the fourth 3*3 convolution, batch normalization and LeakyReLU activation function. global_pre X′ global_pre =ReLU(BN(Conv 3×3 (X′ global ))); X′ global_pre The spatial domain feature Y' is obtained by converting it to the spatial domain through the inverse real two-dimensional fast Fourier transform global Y' global =iRFFT2d(Split(X′ global_pre )), where Split() is the operation of splitting the features before inverse Fourier transform back into real and imaginary parts, and iRFFT2d() is the inverse real two-dimensional fast Fourier transform operation; Will Y' global With X global_pre Perform element-by-element addition and then perform 1*1 convolution to obtain the fourth convolution feature Y global ; Y local and Y g→l Perform element-by-element addition, and then process it through batch normalization and LeakyReLU activation function to obtain the local feature YY local , expressed as YY local =ReLU(BN(Y local +Y g→l )); Y l→g and Y global Perform element-by-element addition, and then process it through batch normalization and LeakyReLU activation function to obtain the global feature YY global , expressed as YY global =ReLU(BN(Y global +Y l→g )); YY local and YY global Splicing to get the final output Y.

8. The MRI image segmentation method based on frequency-space synergy according to claim 7, characterized in that The training described in step S4 specifically includes the following steps: The first loss L is calculated using the following formula: ipt : L ipt =α*L per +β*L ssim Where α is the first weight set; β is the second weight set; L per is the perceptual loss; L ssim For structural losses; The perceptual loss is to input the image reconstruction result into the VGG19 feature extractor, compare the feature maps of pool1 to pool5 in the VGG19 feature extractor with the feature maps of the original image corresponding to the image reconstruction result, and obtain L per The calculation formula is Where N is the number of activation layers required to calculate the VGG19 feature extractor, n i is the activation layer of the corresponding level, The product of the height, width and number of channels of the feature map calculated for the original image, is the product of the height, width and number of channels of the feature map calculated by the image reconstruction result, x is the original image corresponding to the image reconstruction result, is the image reconstruction result; L ssim The calculation formula is where μ x is the mean of x, for The mean of , C1 is the first constant set, For x and The covariance of x is the standard deviation of x, for The standard deviation of , C2 is the second constant set; The second loss L is calculated using the following formula: sup-seg : Where Dice() is the Dice loss function; y is the true segmentation result of the labeled image; The image segmentation result output by the image segmentation module for the labeled image; The third loss L is calculated using the following formula: self-con : In the formula The image segmentation result output by the image segmentation module for the unlabeled image; The image reconstruction result is the image segmentation result output by the image segmentation module; Training, the total loss L is calculated using the following formula: Where N is the number of labeled images; M is the number of unlabeled images.

9. An imaging method comprising the MRI image segmentation method based on frequency-space synergy according to any one of claims 1 to 8, characterized in that The steps include: A. Acquire target MRI images; B. performing image segmentation on the target MRI image obtained in step A using the MRI image segmentation method based on frequency-space synergy according to any one of claims 1 to 8 to obtain an image segmentation result; C. The image segmentation result obtained in step B is marked and re-imaged on the target MRI image to obtain a target MRI image with the segmentation result.

Images