A tumor tissue segmentation method, system and storage medium based on MR images

Abstract

The present invention relates to the field of medical image processing technology, and specifically to a tumor tissue segmentation method, system, and storage medium based on MR images, comprising the following steps: using a neural network to respectively construct a tumor edge segmentation module and a tumor non-edge region segmentation module; setting a mutual learning module between the tumor edge segmentation module and the tumor non-edge region segmentation module; and training a model structure composed of the tumor edge segmentation module, the tumor non-edge region segmentation module, and the mutual learning module to obtain a tumor tissue segmentation model. By constructing the tumor edge segmentation module, the tumor non-edge region segmentation module, and the mutual learning module, the present invention enables interactive learning and mutual guidance between edge segmentation and non-edge region segmentation, thereby accurately segmenting the main area of the tumor tissue and accurately segmenting the complex boundaries of the tumor tissue, thereby improving the accurate segmentation effect of the tumor tissue.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a tumor tissue segmentation method, system and storage medium based on MR images. Background Art

[0002] Tumor segmentation aims to automatically segment tumor regions from multimodal magnetic resonance (MR) images captured by advanced medical imaging equipment. By segmenting the tumor, tumor volume, shape, and location can be provided, which plays a vital role in tumor diagnosis and monitoring.

[0003] In the existing technology, deep learning algorithms are usually used to segment tumor tissue in MR images. Such tumor tissue segmentation models based on deep learning tend to focus on the segmentation and recognition of the entire tumor tissue area. Most of them can accurately segment the main area of the tumor tissue, but it is difficult to accurately segment the complex boundaries of the tumor tissue, thus limiting the accurate segmentation effect of the tumor tissue. Summary of the Invention

[0004] The purpose of the present invention is to provide a tumor tissue segmentation method, system and storage medium based on MR images to solve the technical problem in the prior art that it is difficult to accurately segment the complex boundaries of tumor tissue.

[0005] In order to solve the above technical problems, the present invention specifically provides the following technical solutions:

[0006] A tumor tissue segmentation method based on MR images comprises the following steps:

[0007] Acquiring MR images containing tumor tissue;

[0008] A tumor edge segmentation module for segmenting the edge of tumor tissue in MR images and a tumor non-edge region segmentation module for segmenting the non-edge region of tumor tissue in MR images are constructed using neural networks.

[0009] Between the tumor edge segmentation module and the tumor non-edge area segmentation module, a mutual learning module is set up to control the mutual learning between the tumor tissue edge segmentation and the tumor tissue non-edge area segmentation;

[0010] The model structure composed of the tumor edge segmentation module, the tumor non-edge area segmentation module and the mutual learning module is trained to obtain a tumor tissue segmentation model for segmenting tumor tissue in MR images.

[0011] As a preferred embodiment of the present invention, the method for constructing the tumor edge segmentation module includes:

[0012] The Mask F-CNN network is used as the model structure of the tumor edge segmentation module;

[0013] The MR image is used as the input of the Mask F-CNN network, and the edge mask of the tumor tissue in the MR image is used as the output of the Mask F-CNN network. The tumor edge segmentation module is obtained as follows:

[0014] G rough =Mask F-CNN(MR);

[0015] Where G rough is the edge mask, MR is the MR image, and Mask F-CNN is the Mask F-CNN network.

[0016] As a preferred embodiment of the present invention, the method for constructing the tumor non-edge area segmentation module includes:

[0017] The Mask F-CNN network is used as the model structure of the tumor non-edge segmentation module;

[0018] The MR image is used as the input of the Mask F-CNN network, and the non-edge area mask of the tumor tissue in the MR image is used as the output of the Mask F-CNN network. The tumor non-edge area segmentation module is obtained as follows:

[0019] G body =Mask F-CNN(MR);

[0020] Where G body is the non-edge region mask, MR is the MR image, and Mask F-CNN is the Mask F-CNN network.

[0021] As a preferred solution of the present invention, the method for constructing the mutual learning module includes:

[0022] Constructing a first path for the tumor edge segmentation module to learn from the tumor non-edge area segmentation module, and quantifying the learning rate of the first path;

[0023] The learning rate of the first path is:

[0024] ;

[0025] Where, is the learning rate of the first path, is the edge mask output by the tumor edge segmentation module on the i-th sample in the dataset used to train the model structure, is the non-edge area mask output by the tumor non-edge area segmentation module on the i-th sample in the data set used to train the model structure, m is the total number of samples in the data set used to train the model structure, for and The KL divergence between

[0026] Constructing a second path for the tumor non-edge region segmentation module to learn from the tumor edge segmentation module, and quantifying the learning rate of the second path;

[0027] The learning rate of the second path is:

[0028] ;

[0029] Where, is the learning rate of the second path, is the edge mask output by the tumor edge segmentation module on the i-th sample in the dataset used to train the model structure, is the non-edge area mask output by the tumor non-edge area segmentation module on the i-th sample in the data set used to train the model structure, m is the total number of samples in the data set used to train the model structure, for and The KL divergence between

[0030] in, ;

[0031] .

[0032] As a preferred solution of the present invention, the training loss function of the model structure is:

[0033] ;

[0034] in, ;

[0035] ;

[0036] ;

[0037] Where, is the total loss, To predict losses, is the mutual learning loss, To rebuild the losses, 、 and Balanced 、 and The hyperparameters of is the edge mask output by the tumor edge segmentation module on the i-th sample in the dataset used to train the model structure, is the non-edge area mask output by the tumor non-edge area segmentation module on the i-th sample in the data set used to train the model structure, m is the total number of samples in the data set used to train the model structure, The true value of the non-edge area mask on the i-th sample in the dataset used to train the model structure, The true value of the edge mask on the i-th sample in the dataset used to train the model structure, is the true value of the overall mask of the tumor tissue on the i-th sample in the data set used to train the model structure, is the learning rate of the first path, is the learning rate of the second path, 、 and Both are L2 norm forms.

[0038] As a preferred embodiment of the present invention, the method for training the model structure to obtain the tumor tissue segmentation model includes:

[0039] Dividing the data set into a training set and a test set;

[0040] Based on the loss function, the model structure is trained on the training set to obtain a tumor tissue segmentation model;

[0041] Based on the model evaluation indicators, the performance of the tumor tissue segmentation model was evaluated on the test set.

[0042] As a preferred embodiment of the present invention, the method for obtaining tumor tissue segmentation results of MR images using the tumor tissue segmentation model includes:

[0043] Input the MR image into the tumor tissue segmentation model to obtain the edge mask and non-edge area mask;

[0044] The edge mask and the non-edge area mask are superimposed and fused to obtain the overall tumor tissue mask as the tumor tissue segmentation result of the MR image.

[0045] As a preferred solution of the present invention, the hyperparameter ,in,

[0046] ;

[0047] ;

[0048] Where, for The balanced hyperparameters of for The balanced hyperparameters of is the edge mask output by the tumor edge segmentation module on the i-th sample in the dataset used to train the model structure, is the non-edge area mask output by the tumor non-edge area segmentation module on the i-th sample in the data set used to train the model structure, m is the total number of samples in the data set used to train the model structure, The true value of the non-edge area mask on the i-th sample in the dataset used to train the model structure, The true value of the edge mask on the i-th sample in the dataset used to train the model structure, is the learning rate of the first path, is the learning rate of the second path, 、 Both are L2 norm forms.

[0049] As a preferred embodiment of the present invention, the present invention provides a tumor tissue segmentation system based on MR images, which is applied to a tumor tissue segmentation method based on MR images. The system includes:

[0050] a data acquisition unit, configured to acquire an MR image containing tumor tissue;

[0051] a model building unit configured to use a neural network to construct a tumor edge segmentation module for segmenting tumor tissue edges in MR images, and a tumor non-edge region segmentation module for segmenting tumor tissue non-edge regions in MR images; a mutual learning module configured to control mutual learning between the tumor edge segmentation module and the tumor non-edge region segmentation module; and a model structure composed of the tumor edge segmentation module, the tumor non-edge region segmentation module, and the mutual learning module to obtain a tumor tissue segmentation model for segmenting tumor tissue in MR images.

[0052] The segmentation output unit is used to obtain tumor tissue segmentation results of MR images using the tumor tissue segmentation model.

[0053] As a preferred embodiment of the present invention, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions. When a processor executes the computer-executable instructions, a tumor tissue segmentation method based on MR images is implemented.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] The present invention constructs a tumor edge segmentation module, a tumor non-edge area segmentation module and a mutual learning module, so that the edge segmentation and non-edge area segmentation can interactively learn and guide each other, which can achieve accurate segmentation of the main area of ​​tumor tissue and accurate segmentation of the complex boundaries of tumor tissue, thereby improving the accurate segmentation effect of tumor tissue. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other implementation drawings based on the provided drawings without inventive effort.

[0057] Figure 1 A flow chart of a tumor tissue segmentation method based on MR images provided in an embodiment of the present invention;

[0058] Figure 2 A block diagram of a tumor tissue segmentation system based on MR images provided in an embodiment of the present invention;

[0059] Figure 3 A schematic diagram of the tumor tissue segmentation model structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0061] like Figure 1 and Figure 3 As shown, the present invention provides a tumor tissue segmentation method based on MR images, comprising the following steps:

[0062] Acquiring MR images containing tumor tissue;

[0063] A tumor edge segmentation module for segmenting the edge of tumor tissue in MR images and a tumor non-edge region segmentation module for segmenting the non-edge region of tumor tissue in MR images are constructed using neural networks.

[0064] Between the tumor edge segmentation module and the tumor non-edge area segmentation module, a mutual learning module is set up to control the mutual learning between the tumor tissue edge segmentation and the tumor tissue non-edge area segmentation;

[0065] The model structure composed of the tumor edge segmentation module, the tumor non-edge area segmentation module and the mutual learning module is trained to obtain a tumor tissue segmentation model for segmenting tumor tissue in MR images.

[0066] The present invention divides the overall segmentation of tumor tissue in MR images into two local segmentations, namely tumor tissue edge segmentation and tumor tissue non-edge area segmentation, and finally superimposes and fuses the two local segmentation results to obtain the overall tumor tissue segmentation result.

[0067] After the overall segmentation is divided into two local segmentations, the overall segmentation process needs to balance the two focus points of edge segmentation and main body segmentation, that is, the edge segmentation and main body segmentation have mutual compromise in the overall segmentation process, so that the overall segmentation is difficult to achieve the most accurate performance in edge segmentation and main body segmentation. The present invention transforms the overall segmentation into two segmentation processes with independent focus points, establishes a tumor edge segmentation module and a tumor non-edge area segmentation module, which respectively correspond to achieving the most accurate performance in edge segmentation and the most accurate performance in main body area segmentation, so that the superimposed overall segmentation effect also has the most accurate performance of edge segmentation and main body segmentation, realizing high-precision segmentation of tumor tissue in MR images.

[0068] Since the tumor edge and the tumor non-edge area are adjacent to each other and not completely separated, there will be some correlation and constraints in the segmentation of the tumor edge and the tumor non-edge area. In other words, the edge segmentation can assist or constrain the segmentation of the peripheral boundary of the tumor non-edge area, and the corresponding peripheral boundary of the tumor non-edge area can also assist or constrain the edge segmentation. The results of the edge segmentation and the segmentation of the peripheral boundary of the tumor non-edge area are similar and can learn from each other. Therefore, the present invention sets a mutual learning unit between the tumor edge segmentation module and the tumor non-edge area segmentation module to learn the similarity process between the tumor edge segmentation module and the tumor non-edge area segmentation module, corresponding to the first path and the second path, and quantifies the learning rate, that is, the degree of difference between the results of the tumor edge segmentation module and the tumor non-edge area segmentation module. The lower the degree of difference, the higher the learning rate of the first path or the second path, and the learning effect of the first path or the second path meets the expectations. The higher the degree of difference, the lower the learning rate of the first path or the second path, and the learning effect of the first path or the second path does not meet the expectations.

[0069] When training a tumor tissue segmentation model including a tumor edge segmentation module, a tumor non-edge area segmentation module, and a mutual learning module, the present invention adopts a loss function consisting of three parts. The first part is the independent prediction loss of the tumor edge segmentation module and the tumor non-edge area segmentation module, that is, the loss between the output result and the true value of the edge mask and the main area mask. Using this as the loss training model, the most accurate performance in edge segmentation and the most accurate performance in main area segmentation can be achieved.

[0070] The second part is the reconstruction loss of the fusion of the tumor edge segmentation module and the tumor non-edge area segmentation module, that is, the loss between the fusion mask of the output results of the tumor edge segmentation module and the tumor non-edge area segmentation module and the true value of the overall mask of the tumor tissue. Using this as the loss training model can ensure that the overall segmentation effect after superposition also achieves the most accurate performance of the overall segmentation.

[0071] The third part is the mutual learning loss, which is composed of the learning rates of the first path and the second path. Using this loss can ensure that the learning rates of the first path and the second path reach the highest, and the learning effects of the first path and the second path meet expectations. It fully utilizes the correlation and constraints that also exist in the segmentation of tumor edges and tumor non-edge areas, and completes the segmentation of similar parts of the tumor edge segmentation module and the tumor non-edge area segmentation module more quickly, thereby improving segmentation efficiency.

[0072] Furthermore, in order to balance the learning progress of the two paths, the present invention sets the hyperparameter , is the hyperparameter of the first path, is the hyperparameter of the second path, It is quantified by the training progress of the tumor edge segmentation module (represented by the prediction loss). The smaller the prediction loss, the faster the training progress is completed. Similarly, It is quantified by the training progress of the tumor non-edge area segmentation module (characterized by prediction loss). The smaller the prediction loss, the faster the training progress is completed. The present invention expects that the learning progress of the two paths can be kept consistent as much as possible, so as to utilize the correlation and constraints existing in the segmentation of the tumor edge and the tumor non-edge area. Therefore, the training progress of the tumor non-edge area segmentation module and the tumor edge segmentation module needs to be relatively consistent. If it is inconsistent, it will cause deviations in the correlation and constraints, resulting in ineffective application.

[0073] In this regard, the present invention and It can be achieved that when the training progress of the tumor edge segmentation module or the tumor non-edge area segmentation module is fast, the first path or the second path learning rate is given a low weight, or when the training progress of the tumor edge segmentation module or the tumor non-edge area segmentation module is slow, the first path or the second path learning rate is given a high weight. For example, when the training progress of the tumor edge segmentation module is fast, the first path learning rate is given a low weight, and when the training progress of the tumor non-edge area segmentation module is slow, the second path learning rate is given a high weight, thereby achieving the balance of the learning rate of the first path and the learning rate of the second path in the mutual learning loss, achieving the effect of balancing the learning progress of the two paths, and ensuring the effectiveness of mutual learning.

[0074] The present invention divides the overall segmentation into two segmentation processes with independent focus points, establishing a tumor edge segmentation module and a tumor non-edge area segmentation module, which can respectively achieve the most accurate performance in edge segmentation and the most accurate performance in bulk area segmentation. Therefore, the superimposed overall segmentation effect also has the most accurate performance of both edge segmentation and bulk segmentation, realizing high-precision segmentation of tumor tissue in MR images, as follows:

[0075] The construction method of the tumor edge segmentation module includes:

[0076] The Mask F-CNN network is used as the model structure of the tumor edge segmentation module;

[0077] The MR image is used as the input of the Mask F-CNN network, and the edge mask of the tumor tissue in the MR image is used as the output of the Mask F-CNN network. The tumor edge segmentation module is obtained as follows:

[0078] G rough =Mask F-CNN(MR);

[0079] Where G rough is the edge mask, MR is the MR image, and Mask F-CNN is the Mask F-CNN network.

[0080] The construction method of the tumor non-marginal area segmentation module includes:

[0081] The Mask F-CNN network is used as the model structure of the tumor non-edge segmentation module;

[0082] The MR image is used as the input of the Mask F-CNN network, and the non-edge area mask of the tumor tissue in the MR image is used as the output of the Mask F-CNN network. The tumor non-edge area segmentation module is obtained as follows:

[0083] G body =Mask F-CNN(MR);

[0084] Where G body is the non-edge region mask, MR is the MR image, and Mask F-CNN is the Mask F-CNN network.

[0085] The present invention sets up a mutual learning unit between the tumor edge segmentation module and the tumor non-edge area segmentation module to mutually learn the similarity process between the tumor edge segmentation module and the tumor non-edge area segmentation module, corresponding to the first path and the second path, and quantifies the learning rate as follows:

[0086] The construction method of the mutual learning module includes:

[0087] Constructing a first path for the tumor edge segmentation module to learn from the tumor non-edge area segmentation module, and quantifying the learning rate of the first path;

[0088] The learning rate for the first path is:

[0089] ;

[0090] Where, is the learning rate of the first path, is the edge mask output by the tumor edge segmentation module on the i-th sample in the dataset used to train the model structure, is the non-edge area mask output by the tumor non-edge area segmentation module on the i-th sample in the data set used for training the model structure, m is the total number of samples in the data set used for training the model structure, for and The KL divergence between

[0091] Constructing a second path for the tumor non-edge region segmentation module to learn from the tumor edge segmentation module, and quantifying the learning rate of the second path;

[0092] The learning rate for the second path is:

[0093] ;

[0094] Where, is the learning rate of the second path, is the edge mask output by the tumor edge segmentation module on the i-th sample in the dataset used to train the model structure, is the non-edge area mask output by the tumor non-edge area segmentation module on the i-th sample in the data set used for training the model structure, m is the total number of samples in the data set used for training the model structure, for and The KL divergence between

[0095] in, ;

[0096] .

[0097] The training loss function of the model structure is:

[0098] ;

[0099] in, ;

[0100] ;

[0101] ;

[0102] Where, is the total loss, To predict losses, is the mutual learning loss, To rebuild the losses, 、 and Balance 、 and The hyperparameters of is the edge mask output by the tumor edge segmentation module on the i-th sample in the dataset used to train the model structure, is the non-edge area mask output by the tumor non-edge area segmentation module on the i-th sample in the data set used for training the model structure, m is the total number of samples in the data set used for training the model structure, The true value of the non-edge area mask on the i-th sample in the dataset used to train the model structure, The true value of the edge mask on the i-th sample in the dataset used to train the model structure, is the true value of the overall mask of the tumor tissue on the i-th sample in the data set used to train the model structure, is the learning rate of the first path, is the learning rate of the second path, 、 and All are L2 norms, set here , , and can be reset according to actual needs during actual use.

[0103] When training a tumor tissue segmentation model including a tumor edge segmentation module, a tumor non-edge area segmentation module, and a mutual learning module, the present invention adopts a loss function consisting of three parts. The first part is the independent prediction loss of the tumor edge segmentation module and the tumor non-edge area segmentation module, that is, the loss between the output result and the true value of the edge mask and the main area mask. Using this as the loss training model, the most accurate performance in edge segmentation and the most accurate performance in main area segmentation can be achieved.

[0104] The second part is the reconstruction loss of the fusion of the tumor edge segmentation module and the tumor non-edge area segmentation module, that is, the loss between the fusion mask of the output results of the tumor edge segmentation module and the tumor non-edge area segmentation module and the true value of the overall mask of the tumor tissue. Using this as the loss training model can ensure that the overall segmentation effect after superposition also achieves the most accurate performance of the overall segmentation.

[0105] The third part is the mutual learning loss, which is composed of the learning rates of the first path and the second path. Using this loss can ensure that the learning rates of the first path and the second path reach the highest, and the learning effects of the first path and the second path meet expectations. It fully utilizes the correlation and constraints that also exist in the segmentation of tumor edges and tumor non-edge areas, and completes the segmentation of similar parts of the tumor edge segmentation module and the tumor non-edge area segmentation module more quickly, thereby improving segmentation efficiency.

[0106] Methods for training the model structure to obtain a tumor tissue segmentation model include:

[0107] Divide the dataset into training and testing sets;

[0108] Based on the loss function, the model structure is trained on the training set to obtain the tumor tissue segmentation model;

[0109] Based on the model evaluation indicators, the performance of the tumor tissue segmentation model was evaluated on the test set.

[0110] The method for obtaining tumor tissue segmentation results of MR images by the tumor tissue segmentation model includes:

[0111] Input the MR image into the tumor tissue segmentation model to obtain the edge mask and non-edge area mask;

[0112] The edge mask and the non-edge area mask are superimposed and fused to obtain the overall tumor tissue mask, which is used as the tumor tissue segmentation result of the MR image.

[0113] Hyperparameters ,in,

[0114] ;

[0115] ;

[0116] Where, for The balanced hyperparameters of for The balanced hyperparameters of is the edge mask output by the tumor edge segmentation module on the i-th sample in the dataset used to train the model structure, is the non-edge area mask output by the tumor non-edge area segmentation module on the i-th sample in the data set used for training the model structure, m is the total number of samples in the data set used for training the model structure, The true value of the non-edge area mask on the i-th sample in the dataset used to train the model structure, The true value of the edge mask on the i-th sample in the dataset used to train the model structure, is the learning rate of the first path, is the learning rate of the second path, 、 Both are L2 norm forms.

[0117] In the present invention and It can be achieved that when the training progress of the tumor edge segmentation module or the tumor non-edge area segmentation module is fast, the first path or the second path learning rate is given a low weight, or when the training progress of the tumor edge segmentation module or the tumor non-edge area segmentation module is slow, the first path or the second path learning rate is given a high weight. For example, when the training progress of the tumor edge segmentation module is fast, the first path learning rate is given a low weight, and when the training progress of the tumor non-edge area segmentation module is slow, the second path learning rate is given a high weight, thereby achieving the balance of the learning rate of the first path and the learning rate of the second path in the mutual learning loss, achieving the effect of balancing the learning progress of the two paths, and ensuring the effectiveness of mutual learning.

[0118] like Figure 2 As shown, the present invention provides a tumor tissue segmentation system based on MR images, which is applied to a tumor tissue segmentation method based on MR images. The system includes:

[0119] a data acquisition unit, configured to acquire an MR image containing tumor tissue;

[0120] a model building unit configured to use a neural network to construct a tumor edge segmentation module for segmenting tumor tissue edges in MR images, and a tumor non-edge region segmentation module for segmenting tumor tissue non-edge regions in MR images; a mutual learning module configured to control mutual learning between the tumor edge segmentation module and the tumor non-edge region segmentation module; and a model structure composed of the tumor edge segmentation module, the tumor non-edge region segmentation module, and the mutual learning module to obtain a tumor tissue segmentation model for segmenting tumor tissue in MR images.

[0121] The segmentation output unit is used to obtain tumor tissue segmentation results of MR images using the tumor tissue segmentation model.

[0122] The present invention provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, a tumor tissue segmentation method based on MR images is implemented.

[0123] The present invention constructs a tumor edge segmentation module, a tumor non-edge area segmentation module and a mutual learning module, so that the edge segmentation and non-edge area segmentation can interactively learn and guide each other, which can achieve accurate segmentation of the main area of ​​tumor tissue and accurate segmentation of the complex boundaries of tumor tissue, thereby improving the accurate segmentation effect of tumor tissue.

[0124] The above embodiments are merely exemplary embodiments of the present application and are not intended to limit the scope of the present application. The scope of protection of the present application is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present application within the essence and scope of protection of the present application, and such modifications or equivalent substitutions shall also be deemed to fall within the scope of protection of the present application.

Claims

1. A tumor tissue segmentation method based on MR images, characterized in that: The following steps are involved: Acquiring MR images containing tumor tissue; A tumor edge segmentation module for segmenting the edge of tumor tissue in MR images and a tumor non-edge region segmentation module for segmenting the non-edge region of tumor tissue in MR images are constructed using neural networks. Between the tumor edge segmentation module and the tumor non-edge area segmentation module, a mutual learning module is set up to control the mutual learning between the tumor tissue edge segmentation and the tumor tissue non-edge area segmentation; A model structure composed of a tumor edge segmentation module, a tumor non-edge area segmentation module, and a mutual learning module is trained to obtain a tumor tissue segmentation model for segmenting tumor tissue in MR images; The method for constructing the mutual learning module includes: Constructing a first path for the tumor edge segmentation module to learn from the tumor non-edge area segmentation module, and quantifying the learning rate of the first path; The learning rate of the first path is: ； Where, is the learning rate of the first path, is the edge mask output by the tumor edge segmentation module on the i-th sample in the dataset used to train the model structure, is the non-edge area mask output by the tumor non-edge area segmentation module on the i-th sample in the data set used to train the model structure, m is the total number of samples in the data set used to train the model structure, for and The KL divergence between Constructing a second path for the tumor non-edge region segmentation module to learn from the tumor edge segmentation module, and quantifying the learning rate of the second path; The learning rate of the second path is: ； Where, is the learning rate of the second path, is the edge mask output by the tumor edge segmentation module on the i-th sample in the dataset used to train the model structure, is the non-edge area mask output by the tumor non-edge area segmentation module on the i-th sample in the data set used to train the model structure, m is the total number of samples in the data set used to train the model structure, for and The KL divergence between in, ; ； The training loss function of the model structure is: ； in, ; ； ； Where, is the total loss, To predict losses, is the mutual learning loss, To rebuild the losses, 、 and Balance 、 and The hyperparameters of is the edge mask output by the tumor edge segmentation module on the i-th sample in the dataset used to train the model structure, is the non-edge area mask output by the tumor non-edge area segmentation module on the i-th sample in the data set used to train the model structure, m is the total number of samples in the data set used to train the model structure, The true value of the non-edge area mask on the i-th sample in the dataset used to train the model structure, The true value of the edge mask on the i-th sample in the dataset used to train the model structure, is the true value of the overall mask of the tumor tissue on the i-th sample in the data set used to train the model structure, is the learning rate of the first path, is the learning rate of the second path, 、 and All are L2 norm forms; The hyperparameters ,in, ； ； Where, for The balanced hyperparameters of for The balanced hyperparameters of is the edge mask output by the tumor edge segmentation module on the i-th sample in the dataset used to train the model structure, is the non-edge area mask output by the tumor non-edge area segmentation module on the i-th sample in the data set used to train the model structure, m is the total number of samples in the data set used to train the model structure, The true value of the non-edge area mask on the i-th sample in the dataset used to train the model structure, The true value of the edge mask on the i-th sample in the dataset used to train the model structure, is the learning rate of the first path, is the learning rate of the second path, 、 Both are L2 norm forms.

2. The tumor tissue segmentation method based on MR images according to claim 1, characterized in that: The method for constructing the tumor edge segmentation module includes: The Mask F-CNN network is used as the model structure of the tumor edge segmentation module; The MR image is used as the input of the Mask F-CNN network, and the edge mask of the tumor tissue in the MR image is used as the output of the MaskF-CNN network. The tumor edge segmentation module is obtained as follows: G rough =Mask F-CNN(MR); Where G rough is the edge mask, MR is the MR image, and Mask F-CNN is the Mask F-CNN network.

3. The method for tumor tissue segmentation based on MR images according to claim 1, characterized in that: The method for constructing the tumor non-edge area segmentation module includes: The Mask F-CNN network is used as the model structure of the tumor non-edge segmentation module; The MR image is used as the input of the Mask F-CNN network, and the non-edge area mask of the tumor tissue in the MR image is used as the output of the Mask F-CNN network. The tumor non-edge area segmentation module is obtained as follows: G body =Mask F-CNN(MR); Where G body is the non-edge region mask, MR is the MR image, and Mask F-CNN is the Mask F-CNN network.

4. The method for tumor tissue segmentation based on MR images according to claim 1, characterized in that: The method of training the model structure to obtain the tumor tissue segmentation model includes: Dividing the data set into a training set and a test set; Based on the loss function, the model structure is trained on the training set to obtain a tumor tissue segmentation model; Based on the model evaluation indicators, the performance of the tumor tissue segmentation model was evaluated on the test set.

5. The method for tumor tissue segmentation based on MR images according to claim 1, characterized in that: The method for obtaining the tumor tissue segmentation result of the MR image by the tumor tissue segmentation model includes: Input the MR image into the tumor tissue segmentation model to obtain the edge mask and non-edge area mask; The edge mask and the non-edge area mask are superimposed and fused to obtain the overall tumor tissue mask as the tumor tissue segmentation result of the MR image.

6. A tumor tissue segmentation system based on MR images, characterized in that: A method for tumor tissue segmentation based on MR images as described in any one of claims 1 to 5, wherein the system comprises: a data acquisition unit, configured to acquire an MR image containing tumor tissue; a model building unit configured to use a neural network to construct a tumor edge segmentation module for segmenting tumor tissue edges in MR images, and a tumor non-edge region segmentation module for segmenting tumor tissue non-edge regions in MR images; a mutual learning module configured to control mutual learning between the tumor edge segmentation module and the tumor non-edge region segmentation module; and a model structure composed of the tumor edge segmentation module, the tumor non-edge region segmentation module, and the mutual learning module to obtain a tumor tissue segmentation model for segmenting tumor tissue in MR images. The segmentation output unit is used to obtain tumor tissue segmentation results of MR images using the tumor tissue segmentation model.

7. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions, and when a processor executes the computer-executable instructions, the method according to any one of claims 1 to 5 is implemented.

Images