WO2022041307A1

WO2022041307A1 - Method and system for constructing semi-supervised image segmentation framework

Info

Publication number: WO2022041307A1
Application number: PCT/CN2020/113496
Authority: WO
Inventors: 潘志方; 陈高翔; 茹劲涛
Original assignee: 温州医科大学
Priority date: 2020-08-31
Filing date: 2020-09-04
Publication date: 2022-03-03
Also published as: CN112150478B; CN112150478A

Abstract

A method for constructing a semi-supervised image segmentation framework, the method comprising: constructing a semi-supervised image segmentation framework which comprises a student model, a teacher model and a discriminator (S1); acquiring a marked MRI image and a corresponding gold standard thereof, so as to calculate a supervised segmentation loss; acquiring an original unmarked MRI image and an unmarked MRI noise image that is obtained after combining the original unmarked MRI image with pre-set Gaussian distribution noise, so as to obtain a corresponding student segmentation probability result graph and a corresponding teacher segmentation probability result graph, then respectively covering the original unmarked MRI image with same, and generating a student segmentation area and a teacher segmentation area and transmitting same together to the discriminator for similarity comparison, so as to calculate a consistency loss; and obtaining the total segmentation loss according to the supervised segmentation loss and the consistency loss, and optimizing the semi-supervised image segmentation framework according to the total segmentation loss (S4). By means of implementing the method, a universal semi-supervised segmentation framework capable of being used for 3D medical images is established by means of improving a mean teacher model, and no additional image-level mark is needed.

Description

A method and system for constructing a semi-supervised image segmentation framework

technical field

The invention relates to the technical field of image processing, in particular to a method and system for constructing a semi-supervised image segmentation framework.

Background technique

Medical image segmentation plays a vital role in clinical applications and scientific research. Accurate medical image segmentation can provide important quantitative measures for lesion grading, classification, and disease diagnosis, and further help clinicians evaluate treatment response to related diseases and provide a reliable basis for surgical planning and rehabilitation strategies.

In recent years, many computer-aided deep learning methods have emerged, such as convolutional neural networks that can automatically extract and learn image features, and their application in image segmentation has achieved great improvements in accuracy. However, these methods rely on large amounts of data with high-quality labels, especially in the medical imaging domain, where the process of labeling large-scale data is often more expensive and time-consuming due to the need for expert domain knowledge, making it difficult to obtain large amounts of hand-labeled labels. Furthermore, this segmentation may be subject to variations in labelers (e.g. clinicians) and is not reproducible.

To avoid the need for labeled data, unsupervised learning of medical images has been proposed. However, due to the very low segmentation accuracy, this fully unsupervised approach does not work well for complex anatomical structures or lesions with large variations in shape and size, both requiring the construction of appropriately sized and accurately labeled datasets for training Deep learning models, which are often difficult to implement in practical applications of medical imaging.

As another solution, weakly supervised learning does not require voxel-level labeled data, but uses image-level labeled data as weakly supervised signals in network training. Nonetheless, image-level labels or bounding boxes for medical images also require domain knowledge and are expensive to acquire, and the application of weakly supervised learning models in medical imaging is still limited, where image-level labels and bounding boxes are still required. Simple markup.

Therefore, it is necessary to design efficient semi-supervised learning methods that do not require additional auxiliary labels. This semi-supervised learning method utilizes both labeled and unlabeled data, and strikes a balance between tediously supervised and unsupervised to train a model to accurately segment medical images with only a small number of labeled samples, which is important for designing images in medicine A split frame might be a more meaningful option.

However, the existing semi-supervised segmentation methods not only utilize unlabeled data, but also require image-level labels (such as bounding box labels) to assist the training and learning of semi-supervised networks, which are not semi-supervised in the true sense, and in 3D The application effect on medical images has not been fully verified; at the same time, the mean teacher model used in the existing semi-supervised segmentation methods is almost only used for image classification, and has not been widely used in image segmentation.

SUMMARY OF THE INVENTION

The technical problem to be solved by the embodiments of the present invention is to provide a method and system for constructing a semi-supervised image segmentation framework, by improving the mean teacher model to establish a general semi-supervised segmentation framework that can be used for 3D medical images, and no additional images are required. level mark.

In order to solve the above technical problems, an embodiment of the present invention provides a method for constructing a semi-supervised image segmentation framework, including the following steps:

Step S1, constructing a semi-supervised image segmentation framework including a student model, a teacher model and a discriminator;

Step S2, obtaining a marked MRI image and its corresponding gold standard, and importing the marked MRI image as a first training set image into the student model for training, obtaining a segmentation probability map, and further combining the Gold standard to calculate supervised segmentation loss;

Step S3, obtain the original unlabeled MRI image and the noise unlabeled MRI image after it is combined with the noise of the preset Gaussian distribution, obtain a second training set image, and import the second training set image into the student The model and the teacher model are trained respectively to obtain the corresponding student segmentation probability result map and the teacher segmentation probability result map. Further, the student segmentation probability result map and the teacher segmentation probability result map are respectively overlaid on the original After the marked MRI image, the corresponding student segmentation area and teacher segmentation area are generated and passed to the discriminator for similarity comparison to calculate the consistency loss; wherein, the teacher model is based on the The weights of the student model use an exponential moving average strategy to update the model parameters;

Step S4: Obtain a total segmentation loss according to the supervised segmentation loss and the consistency loss, and optimize the semi-supervised image segmentation framework according to the total segmentation loss.

Wherein, the step S3 specifically includes:

obtaining the original unlabeled MRI image and the noise unlabeled MRI image obtained by combining the original unlabeled MRI image with the preset Gaussian distribution noise, to obtain a second training set image;

Import the original unlabeled MRI image of the second training set image into the student model for training to obtain a corresponding student segmentation probability result map, and import the noise unlabeled MRI image of the second training set image into the student model. Training is performed in the teacher model, and the teacher model uses the exponential moving average strategy to update the model parameters based on the weight of the student model in the training process to obtain a teacher segmentation probability result map;

Multiply the student segmentation probability result map and the teacher segmentation probability result map with the original unlabeled MRI image pixel by pixel, respectively, to obtain the corresponding student segmentation area and teacher segmentation area;

The student segmentation area and the teacher segmentation area are passed to the discriminator for similarity comparison, and the student multi-scale features and teacher multi-scale features are extracted respectively, and according to the student multi-scale features and the teacher multi-scale features Scale features, and calculate the consistency loss.

Wherein, the model parameter updated by the teacher model is the weight, which is realized by the formula θ' _t =αθ' _t-1 +(1-α)θ _t ; wherein, θ' is the weight of the teacher model, and θ is The weight of the student model, α is a hyperparameter controlling the decay of the exponential moving average strategy, and t is the number of training steps.

Wherein, the calculation formula of the consistency loss is:

in,

is the consistency loss;

is the voxel-by-voxel multiplication of the two images;

a student segmentation region obtained by multiplying the original unlabeled MRI image by the student segmentation probability result map;

is the teacher segmentation area obtained by multiplying the original unlabeled MRI image and the teacher segmentation probability result map; X _u is the original unlabeled MRI image; S(X _u ) is the student segmentation probability result Figure; R(X _u ) is the teacher segmentation probability result map; f( ) is the hierarchical feature map extracted from the corresponding segmentation area; h, w, d are the height, width and length of each image; δ _mae for

K is the number of network layers in the discriminator; f(x _i ) is the feature vector output by the ith layer.

Wherein, the calculation formula of the supervised segmentation loss is:

in,

is the supervised segmentation loss; Y _l is the gold standard for labeled images; h, w, d are the height, width, and length of each image; C is the number of label categories; c is a certain number of label categories in C a class; X _l is the labeled MRI image; S(X _l ) is the segmentation probability map.

Wherein, the method further includes:

According to the student segmentation probability result map and its corresponding gold standard, the self-training loss of the discriminator is calculated, the adversarial loss of the discriminator is obtained, and the self-training loss of the discriminator and its corresponding loss are obtained. The adversarial loss is combined with the supervised segmentation loss and the consistency loss, the total segmentation loss is updated, and the semi-supervised image segmentation framework is optimized according to the updated total segmentation loss.

Wherein, the calculation formula of the self-training loss of the discriminator is:

in,

is the self-training loss of the discriminator;

is the concatenation of the student segmentation probability result map and the corresponding segmentation region, where || represents the cascade operation of the two images; A( ) is the

The corresponding confidence map generated; μ _self is the confidence threshold;

is the one-hot encoding of the ground truth generated from argmax _c S(X _u );

The gold standard set for the student segmentation probability result map correspondingly.

The formula for calculating the adversarial loss of the discriminator is:

in,

is the adversarial loss of the discriminator; X _n is the image set formed by the labeled MRI image X _l and the original unlabeled MRI image _Xu , X _n ={X _l , _Xu }.

The embodiment of the present invention also provides a system for constructing a semi-supervised image segmentation framework, including an image segmentation framework construction unit, a supervised segmentation loss calculation unit, a consistency loss calculation unit, and an image segmentation framework optimization unit; wherein,

The image segmentation framework building unit is used to construct a semi-supervised image segmentation framework including a student model, a teacher model and a discriminator;

A supervised segmentation loss calculation unit is used to obtain the labeled MRI image and its corresponding gold standard, and import the labeled MRI image as the first training set image into the student model for training to obtain a segmentation probability map , and further combined with the gold standard to calculate the supervised segmentation loss;

The consistency loss calculation unit is used to obtain the original unlabeled MRI image and the noise unlabeled MRI image after it is combined with the noise of the preset Gaussian distribution to obtain a second training set image, and the second training set The image is imported into the student model and the teacher model for training respectively, and the corresponding student segmentation probability result map and teacher segmentation probability result map are obtained. Further, the student segmentation probability result map and the teacher segmentation probability result map are respectively covered On the original unlabeled MRI image, the corresponding student segmentation area and teacher segmentation area are generated and passed to the discriminator for similarity comparison, so as to calculate the consistency loss; wherein, the teacher model is trained during training. In the process, the model parameters are updated using an exponential moving average strategy based on the weight of the student model;

An image segmentation framework optimization unit, configured to obtain a total segmentation loss according to the supervised segmentation loss and the consistency loss, and optimize the semi-supervised image segmentation framework according to the total segmentation loss.

Among them, it also includes:

The image segmentation framework re-optimization unit is used to calculate the self-training loss of the discriminator according to the student segmentation probability result map and the gold standard set correspondingly, and obtain the confrontation loss of the discriminator, and further The self-training loss of the discriminator and its adversarial loss are combined with the supervised segmentation loss and the consistency loss to update the total segmentation loss, and according to the updated total segmentation loss, the semi-supervised image segmentation framework optimize.

Implementing the embodiment of the present invention has the following beneficial effects:

1. The present invention uses the consistency mechanism based on multi-scale features to improve the mean teacher model, and incorporates voxel-level regularization information into the semi-supervised model, thereby further improving the mean teacher model and making it more suitable for image segmentation;

2. The present invention is deeply integrated with adversarial networks (such as discriminators for adversarial learning), which can achieve semi-supervised segmentation without additional image-level labels, and the role of adversarial networks is not only to extract multi-scale images containing spatial context information features, and can be used to measure the confidence of the segmentation probability map that implements the self-training scheme;

3. The present invention establishes a general semi-supervised segmentation framework that can be used for various MRI images (medical images).

Description of drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the following briefly introduces the accompanying drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention, and for those of ordinary skill in the art, obtaining other drawings according to these drawings still belongs to the scope of the present invention without any creative effort.

1 is a flowchart of a method for constructing a semi-supervised image segmentation framework proposed by an embodiment of the present invention;

FIG. 2 is an application scene diagram before preprocessing of MRI images of four modalities in a method for constructing a semi-supervised image segmentation framework proposed by an embodiment of the present invention;

FIG. 3 is a schematic structural diagram of a system for constructing a semi-supervised image segmentation framework proposed by an embodiment of the present invention.

detailed description

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

As shown in FIG. 1, a method for constructing a semi-supervised image segmentation framework proposed in an embodiment of the present invention includes the following steps:

The specific process is that the constructed semi-supervised image segmentation framework is mainly composed of two modules: the mean teacher model and the adversarial network. In general, the framework deeply integrates the adversarial network into the improved mean teacher model, which mainly includes a mean teacher model formed by a student model S and a teacher model R and an adversarial network formed by a discriminator. All these models (including the discriminator) are based on CNN, especially the student and teacher models are based on the same segmentation network (such as U-Net).

Step S2, obtaining a marked MRI image and its corresponding gold standard, and importing the marked MRI image as a first training set image into the student model for training to obtain a segmentation probability map, and further combining the Gold standard to calculate supervised segmentation loss;

The specific process is as follows: for the labeled MRI image X _l , input it together with the corresponding gold standard Y _l into the student model S for training, and after obtaining the segmentation probability map S(X _l ), calculate it by formula (1) Supervised segmentation loss

h, w, d are the height, width and length of each image; C is the number of label categories; c is a certain category in the number of label categories C.

The specific process is that, since the traditional mean teacher model has two losses, one is segmentation loss and the other is consistency loss, which is usually calculated directly from the segmentation maps of the student model S and the teacher model R. . Therefore, in order to overcome the inaccuracy of the accuracy caused by the direct conversion of consistency loss in the traditional mean teacher model, a consistency mechanism based on multi-scale features is used to improve the traditional mean teacher model and make it more suitable for image segmentation. The process is as follows:

acquiring the original unlabeled MRI image _Xu and the noise unlabeled MRI image obtained by combining the original unlabeled MRI image _Xu with the noise of the preset Gaussian distribution, to obtain a second training set image;

Import the original unlabeled MRI image X _u of the second training set image into the student model S for training to obtain the corresponding student segmentation probability result map S(X _u ), and import the noisy unlabeled MRI image of the second training set image into The teacher model R is trained in the teacher model R, and the teacher model R uses the exponential moving average (EMA) strategy to update the model parameters (such as the weight θ') based on the weight θ' of the student model S during the training process, and obtains the teacher segmentation probability result graph R ( X _u ); wherein, the model weight θ' updated by the teacher model R is realized by the formula θ' _t =αθ' _t-1 +(1-α)θ _t ; α is a hyperparameter that controls the decay of the exponential moving average strategy, t is the number of training steps;

Multiply the student segmentation probability result map S(X _u ) and the teacher segmentation probability result map R(X _u ) with the original unlabeled MRI image X _u pixel by pixel respectively to obtain the corresponding student segmentation area

Divide the area with the teacher

Divide students into areas

Divide the area with the teacher

It is passed to the discriminator A for similarity comparison, and the multi-scale features of students and the multi-scale features of teachers are extracted respectively, and the consistency loss is calculated according to the multi-scale features of students and the multi-scale features of teachers.

Among them, according to formula (2), the consistency loss is calculated

in,

is the voxel-by-voxel multiplication of the two images; f( ) is the hierarchical feature map extracted from the corresponding segmented area; _δmae is

K is the number of network layers in the discriminator A; f(x _i ) is the feature vector output by the i-th layer.

It should be noted that the entire training set can be represented as the set S={X _n ,Y _l }, including all images X _n and the gold standard Y _l of labeled images, where X _n ={X _l ,X _u }= {x ₁ ,…,x _L ,x _L+1 ,…,x _L+U }∈R ^H×W×D×N , Y _l ={y ₁ ,…,y _L }∈R ^{H×W×D ×C×L} , the size of each image is H×W×D, the number of label categories in each segmentation task is C, the number of images with ground truth label maps is L, and the number of images in the training set is set to N.

When the original unlabeled MRI image X _u was input into the student model S, in order to obtain similar samples required for consistent training, Gaussian distribution-based noise was also added to the same original unlabeled MRI image X _u , Similar inputs were generated for the teacher model R. Based on the assumption of a consensus mechanism, the two networks are expected to produce similar segmentation results, using an exponential moving average to update the weights θ' of the teacher model according to the weights θ of the student model at each training step t in the training process.

Meanwhile, different from previous mean teacher methods based on simple consistency, the discriminator A for adversarial learning is used as another important component in the framework, and a consistency loss calculated based on multi-scale features is designed. Specifically, from the student model S and the teacher model R, output the original unlabeled MRI image Xu and its corresponding noise unlabeled MRI image corresponding to the student segmentation probability result map S(X _u ₎ and the teacher segmentation probability result map R respectively. After (X _u ), it is overlaid on the original unlabeled MRI image X _u to obtain two sets of segmented regions in MRI, which are generated by pixel-by-pixel multiplication of the input MRI and the segmentation probability map , that is, the student segmentation area

Divide the area with the teacher

In consistency training, the two obtained segmentation regions are encouraged to be similar, instead of only considering the consistency of the segmentation probability map as in the traditional mean teacher model.

Since CNNs can effectively learn image features at multiple scales, in order to better measure the consistency of segmented regions, the hierarchical features of segmented regions are extracted from CNN-based discriminator A and concatenated together, and compared the student segmented regions

Divide the area with the teacher

Corresponding multi-scale features, which are considered as student segmentation regions

Divide the area with the teacher

difference between.

The specific process is, according to formula (3), calculate the total segmentation loss

_λcon is a weighting coefficient used to balance the relative importance of the designed loss function.

Then, use the total segmentation loss

Optimizing a semi-supervised image segmentation framework.

In the embodiment of the present invention, in addition to generating the above-mentioned multi-scale features for calculating consistency loss, the discriminator A also outputs a confidence map for self-training. This confidence map can be used to guide and constrain the target region so that the learned distribution is closer to the true distribution. By thresholding the confidence map, reliable confidence regions can be obtained to select high-confidence segmentation results and convert them to pseudo-labels for self-training. Therefore, a subset of valid segmentation results from _unlabeled MRI images Xu can be directly viewed as labels, which can be added to the training set to further enrich the dataset.

Self-training loss for discriminator A

As shown in formula (4):

in,

The concatenation of the segmentation probability result map and the corresponding segmentation area for students, where || represents the cascade operation of the two images; A( ) is the

is the one-hot encoding of the ground truth generated from argmax _c S(X _u );

The gold standard corresponding to the segmentation probability result map for students, only when the corresponding voxel value of the confidence map output by discriminator A is greater than the user-defined threshold _μself .

For adversarial learning, discriminator A is also used to define the adversarial loss

It can further enhance the ability of the student model to fool the discriminator, as shown in Equation (5):

Among them, the adversarial loss

Can be applied to all training samples, as it depends only on the adversarial network, regardless of whether there are labels.

During the adversarial training of the framework, the student model S and the teacher model R are forced to generate consistent segmentation probability maps to fool the discriminator A, while the discriminator A is trained to enhance the ability to distinguish the student segmentation probability maps from the teacher segmentation probability maps. Hence, the spatial cross-entropy loss of discriminator A is defined. As shown in formula (6):

Among them, E _n =0 means that the segmentation probability map input to the discriminator A is generated by the student model S. _En = 1 indicates that the sample is from the teacher model R;

is the concatenation of the teacher segmentation probability result map and the teacher segmentation region, which is another input to the discriminator A.

It can be seen that the self-training loss of discriminator A can be

and its adversarial loss

with supervised segmentation loss

and consistency loss

Combine, update the total segmentation loss

Therefore, the method further includes:

According to the student segmentation probability result map and its corresponding gold standard, the self-training loss of the discriminator is calculated, and the adversarial loss of the discriminator is obtained, and the self-training loss of the discriminator and its adversarial loss are further combined with the supervised segmentation loss and Consistency losses are combined to update the total segmentation loss and optimize the semi-supervised image segmentation framework based on the updated total segmentation loss.

That is, according to Equation (7), update the total segmentation loss

where λ _con , λ _self and λ _adv are the corresponding weighting coefficients to balance the relative importance of the designed loss function.

As shown in Figure 2, it is an application scene diagram of brain MRI segmentation jointly trained by mean teacher model and adversarial network in a method for constructing a semi-supervised image segmentation framework according to an embodiment of the present invention.

As shown in FIG. 3 , in an embodiment of the present invention, a system for constructing a semi-supervised image segmentation framework is provided, including an image segmentation framework construction unit 110, a supervised segmentation loss calculation unit 120, a consistency loss calculation unit 130, and an image segmentation frame construction unit 110. Segmentation frame optimization unit 140; wherein,

an image segmentation frame construction unit 110, configured to construct a semi-supervised image segmentation frame comprising a student model, a teacher model and a discriminator;

The supervised segmentation loss calculation unit 120 is used to obtain the marked MRI image and its corresponding gold standard, and import the marked MRI image as the first training set image into the student model for training to obtain the segmentation probability graph, and further combined with the gold standard to calculate a supervised segmentation loss;

The consistency loss calculation unit 130 is configured to obtain the original unlabeled MRI image and the noise unlabeled MRI image after it is combined with the noise of the preset Gaussian distribution to obtain a second training set image, and the second training set image is obtained. The set images are imported into the student model and the teacher model for training respectively, and the corresponding student segmentation probability result map and teacher segmentation probability result map are obtained. Further, the student segmentation probability result map and the teacher segmentation probability result map are respectively After being overlaid on the original unlabeled MRI image, the corresponding student segmentation area and teacher segmentation area are generated and passed to the discriminator for similarity comparison to calculate the consistency loss; wherein, the teacher model is in During the training process, the model parameters are updated based on the weight of the student model using an exponential moving average strategy;

The image segmentation framework optimization unit 140 is configured to obtain a total segmentation loss according to the supervised segmentation loss and the consistency loss, and optimize the semi-supervised image segmentation framework according to the total segmentation loss.

Among them, it also includes:

The image segmentation framework re-optimization unit 150 is used to calculate the self-training loss of the discriminator according to the student segmentation probability result map and the gold standard set correspondingly, and obtain the confrontation loss of the discriminator, and further The discriminator's self-training loss and its adversarial loss are combined with the supervised segmentation loss and the consistency loss to update the total segmentation loss, and segment the semi-supervised image according to the updated total segmentation loss The framework is optimized.

It is worth noting that, in the above system embodiment, each system unit included is only divided according to functional logic, but is not limited to the above-mentioned division, as long as the corresponding function can be realized; in addition, the specific The names are only for the convenience of distinguishing from each other, and are not used to limit the protection scope of the present invention.

Those skilled in the art can understand that all or part of the steps in the methods of the above embodiments can be implemented by instructing relevant hardware through a program, and the program can be stored in a computer-readable storage medium, and the storage Media such as ROM/RAM, magnetic disk, optical disk, etc.

The above disclosure is only a preferred embodiment of the present invention, and of course, it cannot limit the scope of the rights of the present invention. Therefore, the equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

A method for constructing a semi-supervised image segmentation framework, comprising the following steps:

Step S1, constructing a semi-supervised image segmentation framework including a student model, a teacher model and a discriminator;

Step S2, obtaining a marked MRI image and its corresponding gold standard, and importing the marked MRI image as a first training set image into the student model for training, obtaining a segmentation probability map, and further combining the Gold standard to calculate supervised segmentation loss;

Step S3, obtain the original unlabeled MRI image and the noise unlabeled MRI image after it is combined with the noise of the preset Gaussian distribution, obtain a second training set image, and import the second training set image into the student The model and the teacher model are trained respectively to obtain the corresponding student segmentation probability result map and the teacher segmentation probability result map. Further, the student segmentation probability result map and the teacher segmentation probability result map are respectively overlaid on the original After the marked MRI image, the corresponding student segmentation area and teacher segmentation area are generated and passed to the discriminator for similarity comparison to calculate the consistency loss; wherein, the teacher model is based on the The weights of the student model use an exponential moving average strategy to update the model parameters;

Step S4: Obtain a total segmentation loss according to the supervised segmentation loss and the consistency loss, and optimize the semi-supervised image segmentation framework according to the total segmentation loss.
The method for constructing a semi-supervised image segmentation framework according to claim 1, wherein the step S3 specifically comprises:

obtaining the original unlabeled MRI image and the noise unlabeled MRI image obtained by combining the original unlabeled MRI image with the preset Gaussian distribution noise, to obtain a second training set image;

Import the original unlabeled MRI image of the second training set image into the student model for training to obtain a corresponding student segmentation probability result map, and import the noise unlabeled MRI image of the second training set image into the student model. Training is performed in the teacher model, and the teacher model uses the exponential moving average strategy to update the model parameters based on the weight of the student model in the training process to obtain a teacher segmentation probability result map;

The student segmentation probability result map and the teacher segmentation probability result map and the original unlabeled MRI image are respectively multiplied pixel by pixel to obtain the corresponding student segmentation area and teacher segmentation area;

The student segmentation area and the teacher segmentation area are passed to the discriminator for similarity comparison, and the student multi-scale features and teacher multi-scale features are extracted respectively, and according to the student multi-scale features and the teacher multi-scale features Scale features, and calculate the consistency loss.
The method for constructing a semi-supervised image segmentation framework according to claim 2, wherein the model parameter updated by the teacher model is a weight, which is calculated by the formula θ' t =αθ' t-1 +(1-α)θ t to achieve; wherein, θ' is the weight of the teacher model, θ is the weight of the student model, α is a hyperparameter that controls the decay of the exponential moving average strategy, and t is the number of training steps.
The method for constructing a semi-supervised image segmentation framework according to claim 2, wherein the calculation formula of the consistency loss is:
in,

is the consistency loss;
is the voxel-by-voxel multiplication of the two images;
a student segmentation region obtained by multiplying the original unlabeled MRI image by the student segmentation probability result map;
is the teacher segmentation area obtained by multiplying the original unlabeled MRI image and the teacher segmentation probability result map; X u is the original unlabeled MRI image; S(X u ) is the student segmentation probability result Figure; R(X u ) is the teacher segmentation probability result map; f( ) is the hierarchical feature map extracted from the corresponding segmentation area; h, w, d are the height, width and length of each image; δ mae for
K is the number of network layers in the discriminator; f(x i ) is the feature vector output by the ith layer.
The method for constructing a semi-supervised image segmentation framework according to claim 1, wherein the calculation formula of the supervised segmentation loss is:
in,

is the supervised segmentation loss; Y l is the gold standard for labeled images; h, w, d are the height, width, and length of each image; C is the number of label categories; c is a certain number of label categories in C a class; X l is the labeled MRI image; S(X l ) is the segmentation probability map.
The method for constructing a semi-supervised image segmentation framework according to claim 4 or 5, wherein the method further comprises:

According to the student segmentation probability result map and its corresponding gold standard, the self-training loss of the discriminator is calculated, the adversarial loss of the discriminator is obtained, and the self-training loss of the discriminator and its corresponding loss are obtained. The adversarial loss is combined with the supervised segmentation loss and the consistency loss, the total segmentation loss is updated, and the semi-supervised image segmentation framework is optimized according to the updated total segmentation loss.
The method for constructing a semi-supervised image segmentation framework according to claim 6, wherein the calculation formula of the self-training loss of the discriminator is:

in,

is the self-training loss of the discriminator;
is the concatenation of the student segmentation probability result map and the corresponding segmentation region, where || represents the cascade operation of the two images; A( ) is the

The corresponding confidence map generated; μ self is the confidence threshold;
is the one-hot encoding of the ground truth generated from argmax c S(X u );
The gold standard set for the student segmentation probability result map correspondingly.
The method for constructing a semi-supervised image segmentation framework according to claim 6, wherein the formula for calculating the adversarial loss of the discriminator is:
in,

is the adversarial loss of the discriminator; X n is the image set formed by the labeled MRI image X l and the original unlabeled MRI image Xu , X n =(X l , Xu }.
A system for constructing a semi-supervised image segmentation framework is characterized by comprising an image segmentation framework construction unit, a supervised segmentation loss calculation unit, a consistency loss calculation unit and an image segmentation framework optimization unit; wherein,

The image segmentation framework building unit is used to construct a semi-supervised image segmentation framework including a student model, a teacher model and a discriminator;

A supervised segmentation loss calculation unit is used to obtain the labeled MRI image and its corresponding gold standard, and import the labeled MRI image as the first training set image into the student model for training to obtain a segmentation probability map , and further combined with the gold standard to calculate the supervised segmentation loss;

The consistency loss calculation unit is used to obtain the original unlabeled MRI image and the noise unlabeled MRI image after it is combined with the noise of the preset Gaussian distribution to obtain a second training set image, and the second training set The image is imported into the student model and the teacher model for training respectively, and the corresponding student segmentation probability result map and teacher segmentation probability result map are obtained. Further, the student segmentation probability result map and the teacher segmentation probability result map are respectively covered On the original unlabeled MRI image, the corresponding student segmentation area and teacher segmentation area are generated and passed to the discriminator for similarity comparison, so as to calculate the consistency loss; wherein, the teacher model is trained during training. In the process, the model parameters are updated using an exponential moving average strategy based on the weight of the student model;

An image segmentation framework optimization unit, configured to obtain a total segmentation loss according to the supervised segmentation loss and the consistency loss, and optimize the semi-supervised image segmentation framework according to the total segmentation loss.
The system for constructing a semi-supervised image segmentation framework according to claim 9, further comprising:

The image segmentation framework re-optimization unit is used to calculate the self-training loss of the discriminator according to the student segmentation probability result map and the gold standard set correspondingly, and obtain the confrontation loss of the discriminator, and further The self-training loss of the discriminator and its adversarial loss are combined with the supervised segmentation loss and the consistency loss to update the total segmentation loss, and according to the updated total segmentation loss, the semi-supervised image segmentation framework optimize.