Image segmentation network training method, training device thereof and image segmentation method

Abstract

This invention discloses an image segmentation network training method and its training device, wherein the image segmentation network training method includes: acquiring an unlabeled target image I. t The corresponding deformation data pairs are respectively for the spectral image and the target image I. t Perform a Fourier transform to obtain the corresponding phase component and amplitude component {Pha(I t ),Amp(I t )}; combine the amplitude component and Amp(I t The weighted summation yields the amplitude component of the enhanced image, and the phase component is used as the phase component of the enhanced image. Based on the amplitude and phase components, the Fourier spectrum of the enhanced image is obtained. The Fourier spectrum is then subjected to inverse Fourier transform to obtain the enhanced image. The enhanced data pairs are used as training data for the segmentation network to train it. By performing spatial and style enhancement on the image through the above steps, high-quality training data can be obtained to optimize the training of the segmentation network and achieve a more accurate segmentation network.

Description

Technical Field

[0001] This invention belongs to the field of medical image segmentation technology, and more specifically, relates to an image segmentation network training method and training device, and an image segmentation method. Background Technology

[0002] In brain structure segmentation tasks, the complexity and numerous partitions of the brain make obtaining a large number of manually labeled samples extremely costly, thus hindering directly supervised training of segmentation networks. Atlas-based segmentation methods can reduce the need for a large number of labeled samples. The general process of atlas-based segmentation methods is as follows: given an atlas (composed of an image and segmentation labels) and a target image (the image to be segmented), the atlas image is registered to the target image, aligning their image contents spatially. Then, the segmentation labels from the atlas are mapped to the target image using the same spatial deformation, yielding the segmentation result. However, atlas-based methods are indirect segmentation methods; they learn the similarity between images rather than directly learning the mapping relationship between images and segmentation labels. Therefore, the training effect on segmentation networks is not ideal. Summary of the Invention

[0003] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides an image segmentation network training method and training device, and an image segmentation method, the purpose of which is to improve the segmentation accuracy of the segmentation network obtained by single-sample label supervised training.

[0004] To achieve the above objectives, according to one aspect of the present invention, an image segmentation network training method is provided, comprising:

[0005] Step S11: Obtain the unlabeled target image I t and corresponding deformation data pairs in, To generate a labeled spectral image I based on the deformation field Φ a The spectral image after deformation. To analyze the spectrum image I based on the deformation field Φ a Graph tag S a The transferred spectral label, Φ, represents the spectral image I. a To target image I t The deformation field of the registration;

[0006] Step S12: Process the spectral image Perform a Fourier transform to obtain the corresponding phase components. and amplitude components For target image I t Perform a Fourier transform to obtain the corresponding phase component Pha(I) t ) and amplitude component Amp(It );

[0007] Step S13: Convert the amplitude component and Amp(I t Weighted summation yields the enhanced image. amplitude components With phase components As an enhanced image phase components

[0008] Step S14: Based on the enhanced image amplitude components and phase components Obtain the Fourier spectrum of the enhanced image

[0009] Step S15: Calculate the Fourier spectrum The enhanced image is obtained by performing an inverse Fourier transform.

[0010] Step S16: Enhance data pairs As training data for the segmentation network, the image will be enhanced. Input an image segmentation network for training, so that the segmentation network predicts the spectral labels. Approaching map labels

[0011] In one embodiment, in step S16, a segmentation loss function combining the Dice coefficient and cross-entropy is used to calculate the segmentation network prediction result. With map labels The error between them, the segmentation loss function is:

[0012]

[0013] in, For the prediction results With map labels The Dice coefficient between them For the prediction results With map labels The cross-entropy between them, where,

[0014] In one embodiment, in step S11, deformation data pairs are acquired. The process includes:

[0015] Image I a and target image I tInputting the trained registration network obtains a deformation field Φ, and the atlas image is registered to the target image I through the deformation field Φ t , to obtain a deformed atlas image The atlas label S a is transferred through the deformation field Φ to obtain a deformed atlas label

[0016] In one of the embodiments, before step S11, further comprising:

[0017] Step S10: unsupervised training is performed on the registration network to obtain a deformation field Φ for registering the atlas image I a to the target image I t , and the atlas image I a is deformed according to the output deformation field Φ to obtain a deformed atlas image The deformed atlas image is made close to the target image I through the unsupervised training. t .

[0018] In one of the embodiments, step S10 comprises:

[0019] Step S101: before training, the image information similarity between the deformed atlas image and the target image I t is taken as a pre-training loss function of the registration network;

[0020] Step S102: after a preset time, the features of the atlas image I and the features of the target image I t are extracted respectively, the sum of the image information similarity and the feature similarity between the deformed atlas image and the target image I t is taken as a post-training loss function, and the registration network is trained until the post-training loss function converges to an expectation.

[0021] In one of the embodiments, further comprising:

[0022] Step S17: inputting the target image I t into the currently trained segmentation network to obtain a predicted atlas label

[0023] Step S18: continuing to perform weak supervision training on the registration network to make the deformed atlas image close to the target image I t and make the deformed atlas label close to the atlas label Proceed to step S11; once both the segmentation network and the weakly supervised registration network converge to the expected values, training ends, and the trained segmentation network is obtained.

[0024] According to another aspect of the present invention, an image segmentation network training apparatus is provided, comprising a data augmentation processing module and a segmentation network, wherein the data augmentation processing module is used to train an unlabeled target image I. t and corresponding deformation data pairs Convert to augmented data pairs The data is then input into the segmentation network for training, where... To generate a labeled spectral image I based on the deformation field Φ a The spectral image after deformation. To analyze the spectrum image I based on the deformation field Φ a Graph tag S a The transferred spectral label, Φ, represents the spectral image I. a To target image I t Registration deformation field:

[0025] The data augmentation processing module includes:

[0026] Fourier transform unit, used for spectral image Perform a Fourier transform to obtain the corresponding phase components. and amplitude components For target image I t Perform a Fourier transform to obtain the corresponding phase component Pha(I) t ) and amplitude component Amp(I t );

[0027] Amplitude mixing unit, used to mix amplitude components and Amp(I t Weighted summation yields the enhanced image. amplitude components

[0028] Fourier spectrum building units for use in image enhancement amplitude components and phase components Obtain the Fourier spectrum of the enhanced image Among them, phase components With phase components same;

[0029] The inverse Fourier transform unit is used to convert the Fourier spectrum. The enhanced image is obtained by performing an inverse Fourier transform.

[0030] In one embodiment, an initial training data output module is also included for providing the target image I to the segmentation network. t Corresponding deformation data pairs include:

[0031] Registration network: used to output the spectral image I a Registered to target image I t The deformation field Φ;

[0032] Deformation conversion unit, used to convert the spectral image I through the deformation field Φ a Registered to target image I t The deformed spectral image was obtained. The map tag S is generated by the deformation field Φ. a The data is transferred to obtain the deformed map labels.

[0033] In one embodiment, it also includes

[0034] A dual-network iterative training control unit is used to transfer the target image I after each segmentation network training is completed. t Input the currently trained segmentation network to obtain the predicted atlas labels. The registration network is then trained under weak supervision, and the target image I is updated using the trained network. t Corresponding deformation data pairs And input the data augmentation processing module to convert it into augmented data pairs. The segmentation network is then fed back into the training network, and the training continues until both the segmentation network and the weakly supervised registration network converge to the expected values. The training then ends. The goal of the weakly supervised training of the registration network is to ensure that the deformed atlas image... Approaching the target image I t And make the deformed map label Approaching map labels

[0035] According to another aspect of the present invention, an image segmentation method is provided, wherein an image to be segmented is input into an image segmentation network to achieve image segmentation, wherein the image segmentation network is trained by the image segmentation network training method described above.

[0036] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0037] This invention uses a graph registration method to obtain a large amount of training data for segmentation networks, requiring only a single labeled graph {I}. a ,S a} is obtained by the registration network a and a large number of unlabeled target images, the atlas image I a is deformed by the deformation field Φ, and for each unlabeled target image, the deformed data pair of the atlas a is obtained by the registration network as training data of the segmentation network, thereby reducing the demand for a large number of labeled samples. However, and I are transformed by the same deformation field Φ, so and I are completely aligned in space, but due to the existence of registration errors, even if the atlas image I a is made close to the target image I t by the deformation field, it is difficult to make them completely the same, and there is a spatial misalignment between them, so that I t and I are not completely aligned, therefore, in the traditional technology, the segmentation network is directly trained using , training the segmentation network with this "target image-pseudo label" pair will introduce noise, causing the segmentation network to learn the wrong "image-segmentation label" correspondence, and the obtained segmentation precision is not ideal.

[0038] The present application proposes a Fourier transform-based style transformation (FST) method, which preserves the phase component of I as the phase component of I , and mixes the amplitude components of I t and I to obtain the amplitude component of I . On the one hand, the phase component of I is preserved, so the spatial structure information of I can be consistent with I , thereby ensuring that the correspondence between I and the segmentation label I is correct; on the other hand, the amplitude component of I introduces the amplitude component of the real image, so that the image style of I can be similar to that of the target image while achieving style enhancement. Finally, through the above two steps of spatial enhancement and style enhancement of the atlas, a pair of high-quality training data I is obtained to optimize the training of the segmentation network and obtain a segmentation network with higher precision. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 is a step flowchart of the image segmentation network training method of the first embodiment;

[0040] Figure 2 A flow chart of a Fourier transform image style enhancement method based on aligned images according to an embodiment;

[0041] Figure 3 A flow chart of steps of an image segmentation network training method according to a second embodiment;

[0042] Figure 4 A schematic diagram of a hybrid loss function combining image and feature similarity according to an embodiment;

[0043] Figure 5 A flow chart of steps of an image segmentation network training method according to a third embodiment;

[0044] Figure 6 A schematic diagram of iterative training of a segmentation network and a registration network according to an embodiment;

[0045] Figure 7 A schematic diagram of the structure of an image segmentation network training apparatus according to an embodiment. DETAILED DESCRIPTION

[0046] In order to make the objects, technical solutions and advantages of the present application clearer, further detailed description will be given to the present application in combination with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0047] As shown in Figure 1 is a flow chart of steps of an image segmentation network training method according to an embodiment, mainly comprising the following steps:

[0048] Step S110: obtaining a target image I t and corresponding deformation data pair

[0049] Among them, is a graph image I a deformed according to the deformation field Φ, is a graph label S a transferred according to the deformation field Φ of the graph image I a , and Φ is a deformation field for registering the graph image I a to the target image I t . Specifically, the graph image I a and the target image I t are input into the trained registration network to obtain the deformation field Φ, and the graph image is registered to the target image I t, to obtain the deformed atlas image by the deformation field Φ a , to obtain the deformed atlas label

[0050] Step S120: Fourier transform the atlas image to obtain the corresponding phase component and amplitude component Fourier transform the target image I t to obtain the corresponding phase component Pha(I t ) and amplitude component Amp(I t ).

[0051] Given an image where WxHxD represents the image size. Fourier transform of it can be expressed as:

[0052]

[0053] Let R(x) and I(x) represent the real and imaginary parts of F(x). The amplitude component and phase component of the Fourier spectrum can be expressed as:

[0054]

[0055]

[0056] Referring to FIG. 4, Fourier transform is performed on R Figure 2 and I t to obtain the corresponding phase and amplitude components, denoted as and {Pha(I t ), Amp(I t )}

[0057] Step S130: Weighted sum the amplitude components and Amp(I t ) to obtain the amplitude component of the enhanced image Use the phase component as the phase component of the enhanced image

[0058] The case reserves the phase component of I as the phase component of I , i.e. Then mix the amplitude components of R and I t in the following way to obtain the amplitude component of I :​​​

[0059]

[0060] Where r is a random number uniformly distributed between 0 and 1.

[0061] Step S140: Based on the enhanced image amplitude components and phase components Obtain the Fourier spectrum of the enhanced image

[0062] Based on its amplitude components and phase components It can be obtained Fourier spectrum:

[0063]

[0064] Step S150: Calculate the Fourier spectrum The enhanced image is obtained by performing an inverse Fourier transform.

[0065] Will The enhanced image is obtained by performing an inverse Fourier transform.

[0066] Through FST The advantages of style enhancement are: on the one hand, it preserves... The phase component, therefore Spatial structural information can be combined with Maintain consistency, thereby ensuring With split labels The correspondence is correct; on the other hand, The amplitude components of the real image are introduced into the amplitude components, thus enabling style enhancement while simultaneously achieving... The image style is similar to that of the target image.

[0067] Finally, by performing spatial augmentation and style enhancement on the graph through the above two steps, a pair of high-quality training data can be obtained.

[0068] Step S160: Enhance data pair As training data for the segmentation network, the image will be enhanced. Input an image segmentation network for training, so that the segmentation network predicts the spectral labels. Approaching map labels

[0069] Specifically, after obtaining the training data for the segmentation network... After that, The corresponding prediction results are obtained by inputting the network respectively The loss function combining the Dice coefficient and the cross entropy (CE) is used to calculate the prediction results of the segmentation network in this embodiment and the segmentation label :

[0070]

[0071]

[0072] The smaller the error is, the closer the prediction result is to the segmentation label . The segmentation network gradually converges.

[0073] In an embodiment, as Figure 3 shown, before step S110, the registration network is also trained, that is, comprising:

[0074] Step S100: unsupervised training is performed on the registration network to make the deformed atlas image close to the target image I t .

[0075] The deformation field Φ that registers the atlas image I a to the target image I t is obtained through unsupervised training, and the atlas image I a is deformed according to the output deformation field Φ to obtain the deformed atlas image Through the unsupervised training, the deformed atlas image is made close to the target image I t .

[0076] During the unsupervised training of the traditional registration network, only the image itself information is used to calculate the similarity for training the supervised registration network. However, since there are differences in the image styles of the target image and the atlas image, using only the image itself gray scale information to calculate the similarity may introduce errors. Therefore, this embodiment proposes a hybrid loss function combining image and feature similarity to train the registration network. The important features irrelevant to the image style are extracted by the registration network convolution layer, and the feature similarity is added in the loss function to reduce the influence of the image style difference on the registration network training and improve the registration accuracy.

[0077] Specifically, as Figure 4 shown, the loss function for training the registration network is divided into two aspects.

[0078] On the one hand, the image-based similarity between I t and is calculated:

[0079]

[0080] LNCC stands for Locally Normalized Mutual Information.

[0081] On the other hand, I is extracted separately using the registration network convolutional layers. t and The characteristics of F are obtained t and F a Then calculate F t and F a Similarities between them:

[0082] SIM(F t ,F a ) = LNCC(F t ,F a )#(9)

[0083] Furthermore, a regularization term, Loss, will be added to ensure the smoothness of the deformation field. smooth :

[0084]

[0085] Where, p i Ω represents voxel, and Ω represents voxel space. Indicated in voxel p i Gradient of the deformation field

[0086] The overall unsupervised training hybrid loss function for the registration network is:

[0087]

[0088] λ reg The weight of the regularization term is set to 1.0 in this embodiment.

[0089] Considering that the convolutional layers of the registration network cannot effectively extract important features unrelated to image style in the early stages of training, the registration network is initially trained solely based on image similarity. After a period of training, a hybrid function combining image and feature similarity is used to train the registration network. In other words, the unsupervised training of the registration network can be considered to consist of the following two processes:

[0090] Step S101: In the early stage of training, use the deformed spectral image and target image I t The image information similarity between them is used as the initial loss function of the registration network, that is, Equation (8) + Equation (10) is used as the loss function;

[0091] Step S102: After a preset time period, extract the spectral images respectively. Features and target images I tcharacteristics of the atlas image after deformation and the target image I t as the late loss function, i.e., training the registration network with equation (11) as the late loss function until the late loss function converges to the expectation.

[0092] By virtue of the ability of the segmentation network to learn the "image-segmentation label" correspondence, after the training is completed, the segmentation result obtained by the segmentation network will be closer to the true result than the segmentation result obtained by the registration network (i.e. ). Therefore, in the subsequent training of the registration network, the segmentation result of the segmentation network can be used to guide the weakly supervised training of the registration network.

[0093] As shown in Figure 5 and Figure 6 , after obtaining the trained segmentation network through step S160, the following steps are further included:

[0094] Step S170: inputting the target image I t into the currently trained segmentation network to obtain the predicted atlas label

[0095] Step S180: continuing to weakly supervise the training of the registration network to make the atlas image after deformation tend to the target image I t and make the atlas label after deformation tend to the atlas label and jump to step S110; when the segmentation network and the weakly supervised registration network both converge to the expectation, the training is completed, and the trained segmentation network is obtained.

[0096] Specifically, the Dice coefficient between and can be used as the weakly supervised item for training the registration network:

[0097]

[0098] At this time, the total registration network loss function is:

[0099] Loss reg2 = Loss reg1 + λ weak_sup Loss weak_sup #(13)

[0100] wherein λ weak_sup is set to 1.0.

[0101] Then repeat steps S110 to S180 until both networks converge. In this way, through iterative training, the registration network and the segmentation network gradually improve the registration and segmentation accuracy by utilizing the additional information provided by the other network.

[0102] Accordingly, the present invention also relates to an image segmentation network training device, such as... Figure 7 As shown, it includes a data augmentation processing module and a segmentation network. The data augmentation processing module is used to augment the unlabeled target image I. t and corresponding deformation data pairs Convert to augmented data pairs The data is then input into the segmentation network for training. The data augmentation module includes:

[0103] Fourier transform unit, used for spectral image Perform a Fourier transform to obtain the corresponding phase components. and amplitude components For target image I t Perform a Fourier transform to obtain the corresponding phase component Pha(I) t ) and amplitude component Amp(I t );

[0104] Amplitude mixing unit, used to mix amplitude components and Amp(I t Weighted summation yields the enhanced image. amplitude components

[0105] Fourier spectrum building units for use in image enhancement amplitude components and phase components Obtain the Fourier spectrum of the enhanced image Among them, phase components With phase components same;

[0106] The inverse Fourier transform unit is used to convert the Fourier spectrum. The enhanced image is obtained by performing an inverse Fourier transform.

[0107] In one embodiment, an initial training data output module is further included, for providing the target image I to the segmentation network. t Corresponding deformation data pairs include:

[0108] Registration network: used to output the spectral image I a Registered to target image I t The deformation field Φ;

[0109] a deformation field Φ to the atlas image I a to the target image I t to obtain a deformed atlas image through the deformation field Φ to the atlas label S a to obtain a deformed atlas label

[0110] In an embodiment, further comprising:

[0111] a dual-network iterative training control unit configured to, whenever the segmentation network training is completed, input the target image I t to the currently trained segmentation network to obtain a predicted atlas label and continue to weakly supervise the training of the registration network, and update the target image I t corresponding deformation data pair and input the data augmentation processing module to convert into an augmented data pair and input the segmentation network to continue training the segmentation network until the segmentation network and the weakly supervised registration network are both converged to the expectation, and the training is ended, wherein the weakly supervised training of the registration network aims to make the deformed atlas image tend to the target image I t and make the deformed atlas label tend to the atlas label

[0112] Correspondingly, the present application also relates to an image segmentation method, which inputs an image to be segmented into an image segmentation network to realize image segmentation, wherein the image segmentation network is trained by the image segmentation network training method described above.

[0113] The method of the present application is compared with other excellent few-labeled sample segmentation methods. The comparison models include Brainstorm using data augmentation, DeepAtlas using iterative training of registration network and segmentation network, LT-Net using a front-back consistency registration network, and SiB using a feature pyramid registration network. The segmentation accuracy (Dice value) of the present application (ours) and the above four models on the OASIS and CANDIShare datasets is compared, and the comparison results are shown in Table 1. Among them, the Brainstorm and DeepAtlas methods are realized by open source code. LT-Net and SiB have no open source code, so the results in the table are directly taken from their papers, which only provide results on the CANDIShare dataset.

[0114] Table 1: Segmentation accuracy of different models on OASIS and CANDIShare datasets

[0115]

[0116] According to the results in Table 1, it can be seen that the method in this paper achieves the highest accuracy segmentation results on both data sets. The average Dice is at most increased by 4.67% and 1.45% on the two data sets, respectively. Therefore, it can be concluded that in the segmentation task under the scenario of lacking labeled data, the method in this paper can better overcome the problem of lacking labeled data, improve the segmentation accuracy, and achieve better segmentation results than the current best method.

[0117] In addition, the results obtained by the method in this paper are also compared with those of other methods by sample T test analysis, and the difference between different methods is calculated. The smaller the p value, the more obvious the difference. Since the experimental results of LT-Net and SiB method for each test case cannot be obtained, the test analysis only compares the method in this paper with Brainstorm and DeepAtlas methods. The analysis results on the two data sets are shown in Table 2.

[0118] From the results in Table 2, it can be seen that the p value is less than 0.05, so there is a statistical difference between the Dice obtained by the method in this paper and those of other methods. Combined with the results in Table 1, it can be concluded that compared with Brainstorm and DeepAtlas methods, the method in this paper can significantly improve the segmentation accuracy.

[0119] Table 2: p value between the method in this paper and other algorithms on different data sets

[0120]

[0121] Those skilled in the art can easily understand that the above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for training an image segmentation network, characterized in that, include: Step S11: Obtain the unlabeled target image I t and corresponding deformation data pairs in, To generate a labeled spectral image I based on the deformation field Φ a The spectral image after deformation. To analyze the spectrum image I based on the deformation field Φ a Graph tag S a The transferred spectral label, Φ, represents the spectral image I. a To target image I t The deformation field of the registration; Step S12: Process the spectral image Perform a Fourier transform to obtain the corresponding phase components. and amplitude components For target image I t Perform a Fourier transform to obtain the corresponding phase component PCa(I) t ) and amplitude component Amp(I t ); Step S13: Convert the amplitude component and Amp(I t Weighted summation yields the enhanced image. amplitude components With phase components As an enhanced image phase components Step S14: Based on the enhanced image amplitude components and phase components Obtain the Fourier spectrum of the enhanced image Step S15: Calculate the Fourier spectrum The enhanced image is obtained by performing an inverse Fourier transform. Step S16: Enhance data pairs As training data for the segmentation network, the image will be enhanced. Input an image segmentation network for training, so that the segmentation network predicts the spectral labels. Approaching map labels 2. The image segmentation network training method as described in claim 1, characterized in that, In step S16, the segmentation loss function, which combines the Dice coefficient and cross-entropy, is used to calculate the prediction result of the segmentation network. With map labels The error between them, the segmentation loss function is: in, For the prediction results With map labels The Dice coefficient between them For the prediction results With map labels The cross-entropy between them, where, 3. The image segmentation network training method as described in claim 1, characterized in that, In step S11, deformation data pairs are acquired. The process includes: Image I a and target image I t The deformation field Φ is obtained by inputting the trained registration network, and the atlas image is registered to the target image I through the deformation field Φ. t The deformed spectral image was obtained. The map tag S is generated by the deformation field Φ. a The data is transferred to obtain the deformed map labels.

4. The image segmentation network training method as described in claim 1, characterized in that, Before step S11, the method further includes: Step S10: Perform unsupervised training on the registration network to obtain the atlas image I a Registered to target image I t The deformation field Φ is used to modify the spectral image I based on the output deformation field Φ. a Deformation is performed to obtain the deformed spectral image. The deformed spectral image is obtained through the unsupervised training. Approaching the target image I t .

5. The image segmentation network training method as described in claim 4, characterized in that, Step S10 includes: Step S101: In the early stage of training, use the deformed spectral image and target image I t The similarity of image information between the two is used as the initial loss function of the registration network; Step S102: After a preset time period, extract the spectral images respectively. Features and target images I t The characteristics, in the deformed spectral image and target image I t The sum of image information similarity and feature similarity between the two is used as the late loss function to train the registration network until the late loss function converges to the expected value.

6. The image segmentation network training method as described in claim 4, characterized in that, Also includes: Step S17: Transfer the target image I t Input the currently trained segmentation network to obtain the predicted atlas labels. Step S18: Continue weakly supervised training of the registration network to improve the deformed spectral image. Approaching the target image I t And make the deformed map label Approaching map labels Proceed to step S11; once both the segmentation network and the weakly supervised registration network converge to the expected values, training ends, and the trained segmentation network is obtained.

7. An image segmentation network training device, characterized in that, It includes a data augmentation processing module and a segmentation network, wherein the data augmentation processing module is used to augment the unlabeled target image I t and corresponding deformation data pairs Convert to augmented data pairs The data is then input into the segmentation network for training, where... To generate a labeled spectral image I based on the deformation field Φ a The spectral image after deformation. To analyze the spectrum image I based on the deformation field Φ a Graph tag S a The transferred spectral label, Φ, represents the spectral image I. a To target image I t Registration deformation field: The data augmentation processing module includes: Fourier transform unit, used for spectral image Perform a Fourier transform to obtain the corresponding phase components. and amplitude components For target image I t Perform a Fourier transform to obtain the corresponding phase component PCa(I) t ) and amplitude component Amp(I t ); Amplitude mixing unit, used to mix amplitude components and Amp(I t Weighted summation yields the enhanced image. amplitude components Fourier spectrum building units for use in image enhancement amplitude components and phase components Obtain the Fourier spectrum of the enhanced image Among them, phase components With phase components same; The inverse Fourier transform unit is used to convert the Fourier spectrum. The enhanced image is obtained by performing an inverse Fourier transform.

8. The image segmentation network training apparatus as described in claim 7, characterized in that, It also includes an initial training data output module for providing the target image I to the segmentation network. t Corresponding deformation data pairs include: Registration network: used to output the spectral image I a Registered to target image I t The deformation field Φ; Deformation conversion unit, used to convert the spectral image I through the deformation field Φ a Registered to target image I t The deformed spectral image was obtained. The map tag S is generated by the deformation field Φ. a The data is transferred to obtain the deformed map labels.

9. The image segmentation network training apparatus as described in claim 8, characterized in that, It also includes a dual-network iterative training control unit, used to transfer the target image I after each segmentation network training is completed. t Input the currently trained segmentation network to obtain the predicted atlas labels. The registration network is then trained under weak supervision, and the target image I is updated using the trained network. t Corresponding deformation data pairs And input the data augmentation processing module to convert it into augmented data pairs. The segmentation network is then fed back into the training network, and the training continues until both the segmentation network and the weakly supervised registration network converge to the expected values. The training then ends. The goal of the weakly supervised training of the registration network is to ensure that the deformed atlas image... Approaching the target image I t And make the deformed map label Approaching map labels 10. An image segmentation method, characterized in that, The image to be segmented is input into an image segmentation network to achieve image segmentation, wherein the image segmentation network is trained by the image segmentation network training method according to any one of claims 1 to 6.

Images