A semi-supervised medical image segmentation method and system based on multi-view joint training

Abstract

The present invention discloses a method and system for semi-supervised medical image segmentation with multi-view joint training. The method comprises: acquiring and segmenting medical image data; constructing a segmentation network model based on networks of different categories; defining a teacher model and a student model; calculating an overall loss; updating the student model based on the overall loss; screening the trained student models; and inputting the image to be segmented into the final segmentation model. The system comprises: a dataset construction module, a model construction module, a model definition module, a loss calculation module, a model update module, a model screening module, and a model application module. By using the present invention, the understanding and prediction capabilities of semi-supervised medical image segmentation can be enhanced. The present invention can be widely applied in the field of image processing.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to a multi-view joint training semi-supervised medical image segmentation method and system. Background Art

[0002] Medical image segmentation refers to the extraction of pixels identifying anatomical or pathological structures from the background in biomedical images. However, unlike ordinary natural images, the scarcity of high-quality annotated training data has always been a major challenge in the field of medical image segmentation. Most advanced segmentation models are typically networks with a large number of model parameters. Their strong generalization ability is highly dependent on large-scale and high-quality pixel-level annotated data, which is often difficult to obtain in clinical and medical tasks. Manually annotating medical images at the pixel level is an expensive and time-consuming process that requires the knowledge of experienced clinical experts. The scarcity of annotated medical image data can be further exacerbated by data differences between patient populations, acquisition parameters and protocols, series, vendors, and centers, which can lead to significant statistical bias. Therefore, due to the heterogeneity of medical imaging data and the strict requirements for patient privacy, even collecting sufficient training data is a challenging task.

[0003] Semi-supervised learning can fully utilize a small amount of labeled data and a large amount of unlabeled data to train models. However, the models in current semi-supervised learning methods are self-reliant and prone to cognitive bias, resulting in poor quality of medical image segmentation in data-scarce medical images. Summary of the Invention

[0004] In view of this, in order to solve the technical problem that semi-supervised learning in existing medical image segmentation methods is prone to cognitive bias, which in turn leads to low segmentation quality, firstly, the present invention proposes a semi-supervised medical image segmentation method with multi-view joint training, which includes the following steps:

[0005] Obtain medical image data and divide it into labeled datasets and unlabeled datasets;

[0006] Building a segmentation network model based on different types of networks to obtain a first segmentation network and a second segmentation network;

[0007] The first segmentation network is provided with an adversarial noise generation module;

[0008] Build a model with the same structure as the first segmentation network as a teacher model, and use the first segmentation network as a student model;

[0009] Inputting the labeled data sets into the student model respectively and calculating the losses to obtain a first supervised loss, a second supervised loss and a first adversarial noise loss;

[0010] Inputting the unlabeled data set into the teacher model and the student model respectively to obtain a first unsupervised loss, a second unsupervised loss, and a second adversarial noise loss;

[0011] Obtain an overall loss by performing a weighted summation based on the first supervised loss, the second supervised loss, the first adversarial noise loss, the first unsupervised loss, the second unsupervised loss, and the second adversarial noise loss;

[0012] Update the student model based on the overall loss to obtain a trained student model;

[0013] Screening the trained student models to obtain a final segmentation model;

[0014] The image to be segmented is input into the final segmentation model, and the segmentation result is output.

[0015] In some embodiments, the first segmentation network is constructed based on an adversarial noise generation module and a convolutional neural network, and the second segmentation network is constructed based on a transformer.

[0016] In some embodiments, the parameter update of the teacher model takes into account historical information and the update of the student network parameters, which is expressed as follows:

[0017] θ' r =αθ' r-1 +(1-α)θ r

[0018] Among them, α represents the decay rate of the exponential moving average, r refers to the training steps of the model, and θ r Refers to the weight of the student model at the training step r, θ' r-1 Refers to the weight of the teacher model at the previous step.

[0019] In some embodiments, the step of inputting the labeled data set into the student model respectively and calculating the loss to obtain the supervised loss and the first anti-noise loss specifically includes:

[0020] Input the labeled data sets in the same batch into two student models respectively;

[0021] For the first segmentation network, adding noise to the initial image in the labeled dataset based on the adversarial noise generation module to generate noisy original data;

[0022] Inputting the noisy raw data into a convolutional neural network to perform initial prediction to obtain a predicted value;

[0023] Performing a dice loss calculation based on the predicted value and the label value in the labeled data set, and calculating a gradient value of the dice loss with respect to noise;

[0024] Adding the gradient value to the corresponding initial image to obtain an adversarial noise sample;

[0025] Inputting the adversarial noise sample into a convolutional neural network, and comparing the result with the label to obtain a first adversarial noise loss;

[0026] Inputting the labeled data sets into the student model respectively to generate prediction values;

[0027] A first supervised loss and a second supervised loss are calculated based on the predicted values.

[0028] In some embodiments, the first supervised loss and the second supervised loss each include a cross entropy loss and a dice loss, wherein:

[0029] The calculation formula for cross entropy loss is as follows:

[0030]

[0031] Among them, Y represents the class label value, P represents the prediction result, N represents the number of pixels, and L i Represents the calculation formula of the loss between pixels, M represents the number of categories, represents the true probability that the i-th pixel belongs to category c, Indicates the predicted probability that the i-th pixel belongs to category c;

[0032] The formula for calculating dice loss is as follows:

[0033]

[0034] In some embodiments, the step of inputting the unlabeled dataset into the teacher model and the student model respectively to obtain an unsupervised loss and a second anti-noise loss specifically includes:

[0035] Inputting the unlabeled data set into the teacher model to generate teacher model pseudo labels;

[0036] For the first segmentation network, calculating a second adversarial noise loss based on the unlabeled dataset and the teacher model pseudo labels;

[0037] Input the unlabeled data sets into the student model respectively to obtain corresponding prediction values;

[0038] Perform pseudo-label fusion according to the predicted value to obtain a pseudo-label;

[0039] An unsupervised loss is calculated based on the unlabeled data and the pseudo labels.

[0040] In some embodiments, the pseudo-label fusion rules are as follows:

[0041] Among all the predicted values, for one pixel position, if more than half of them believe that it belongs to a certain category and the probability value of any output is greater than the preset threshold, then the pixel position is considered to be a confident position, and if it is lower than the threshold, it is considered to be an uncertain position. After calculating all the pixel points, a pseudo label with high confidence is finally generated.

[0042] In some embodiments, the step of calculating the unsupervised loss based on the unlabeled data and the pseudo-labels specifically includes:

[0043] inputting the unlabeled data into the student model;

[0044] Constructing a confidence template based on the pseudo-label;

[0045] The consistency loss is calculated based on the confidence template and the output of the student model to obtain an unsupervised loss.

[0046] In a second aspect, the present invention further proposes a multi-view joint training semi-supervised medical image segmentation system, the system comprising:

[0047] The dataset construction module acquires medical image data and divides it into labeled and unlabeled datasets;

[0048] A model construction module constructs a segmentation network model based on networks of different categories to obtain a first segmentation network and a second segmentation network; the first segmentation network is provided with an adversarial noise generation module;

[0049] A model definition module is used to build a model with the same structure as the first segmentation network as a teacher model, and use the first segmentation network and the second segmentation network as student models;

[0050] a loss calculation module, configured to input the labeled data set into the student model respectively and calculate the loss to obtain a first supervised loss, a second supervised loss, and a first adversarial noise loss; input the unlabeled data set into the teacher model and the student model respectively to obtain a first unsupervised loss, a second unsupervised loss, and a second adversarial noise loss; and perform a weighted sum based on the first supervised loss, the second supervised loss, the first adversarial noise loss, the first unsupervised loss, the second unsupervised loss, and the second adversarial noise loss to obtain an overall loss;

[0051] A model updating module updates the student model based on the overall loss to obtain a trained student model;

[0052] A model screening module is used to screen the trained student models to obtain a final segmentation model;

[0053] The model application module is used to input the image to be segmented into the final segmentation model and output the segmentation result.

[0054] The present invention also proposes a semi-supervised medical image segmentation device for multi-view joint training, comprising:

[0055] at least one processor;

[0056] at least one memory for storing at least one program;

[0057] When the at least one program is executed by the at least one processor, the at least one processor implements the above-mentioned multi-view joint training semi-supervised medical image segmentation method.

[0058] Based on the above scheme, the present invention provides a semi-supervised medical image segmentation method and system using multi-view joint training. This method utilizes the characteristics of convolutional networks and transformer networks, combined with the concept of joint training in consistency learning, to improve the quality of medical image segmentation. By adding a teacher model, the low accuracy of the transformer network in the early stages of training, which can harm the network, can be mitigated. Furthermore, pseudo-label fusion generates high-quality pseudo-labels, allowing the characteristics of both types of networks to fully utilize the principles of consistency learning and leverage the strengths of both. Finally, adversarial noise is added to the network to further enhance the training effect. This, in turn, improves the quality of semi-supervised medical image segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a flowchart of the steps of a multi-view joint training semi-supervised medical image segmentation method of the present invention;

[0060] Figure 2 1 is a data flow diagram of label-free data processing according to a specific embodiment of the present invention;

[0061] Figure 3 Schematic diagram of adversarial sample generation used in the present invention;

[0062] Figure 4 This is a structural block diagram of a multi-view joint training semi-supervised medical image segmentation system of the present invention. DETAILED DESCRIPTION

[0063] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0064] It should be noted that, for ease of description, only the parts related to the invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of this application can be combined with each other.

[0065] It should be understood that the terms "system," "device," "unit," and / or "module" used in this application are a method for distinguishing different components, elements, parts, portions, or assemblies at different levels. However, if other terms can achieve the same purpose, the terms may be replaced by other expressions.

[0066] Unless the context clearly indicates an exception, the terms "a," "an," "an," and / or "the" are not intended to refer to the singular and may include the plural. Generally speaking, the terms "comprise" and "include" only indicate the inclusion of the steps and elements specifically identified, and these steps and elements do not constitute an exclusive list; a method or apparatus may also include other steps or elements. The phrase "comprises a..." does not preclude the presence of additional identical elements in the process, method, product, or apparatus that includes the elements.

[0067] In the description of the embodiments of this application, "plurality" refers to two or more than two. The terms "first" and "second" below are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, features specified as "first" or "second" may explicitly or implicitly include one or more of the features.

[0068] In addition, flow charts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, the steps may be processed in reverse order or simultaneously. Furthermore, other operations may be added to these processes, or one or more operations may be removed from these processes.

[0069] Reference Figure 1 , which is a flow chart of an optional example of the semi-supervised medical image segmentation method for multi-view joint training proposed in the present invention. The method can be applied to a computer device. The medical image segmentation method proposed in this embodiment may include but is not limited to the following steps:

[0070] Step S1: Obtain medical image data and divide it into labeled data sets and unlabeled data sets;

[0071] Among them, the labeled data set is denoted as The unlabeled dataset is denoted as Among them, X in i represents the input image data, and Y represents the corresponding pixel-level annotation mask.

[0072] Pixel-level annotations are sets of categories for each pixel in image data, with a size equal to the image itself. In the annotations, each pixel's number represents its category. Labeled datasets contain images with corresponding annotation masks and are used for the supervised portion of model training. Unlabeled datasets contain only images without corresponding annotation masks and are used for the unsupervised portion of model training.

[0073] Step S2: constructing a segmentation network model based on networks of different categories to obtain a first segmentation network and a second segmentation network; the first segmentation network is provided with an anti-noise generation module;

[0074] Among them, the main frameworks of the two segmentation network models include an encoder and a decoder; the encoder encodes the input image to obtain a feature vector; the decoder processes the feature vector to obtain the segmentation result.

[0075] Step S3: building a model with the same structure as the first segmentation network as a teacher model, and using the first segmentation network and the second segmentation network as student models;

[0076] Step S4: input the labeled data sets into the student model respectively, and calculate the losses to obtain a first supervised loss, a second supervised loss, and a first adversarial noise loss;

[0077] Step S5: input the unlabeled data set into the teacher model and the student model respectively to obtain a first unsupervised loss, a second unsupervised loss, and a second adversarial noise loss;

[0078] Step S6: performing weighted summation based on the supervised loss, the unsupervised loss, the first anti-noise loss, and the second anti-noise loss to obtain an overall loss;

[0079] Step S7: updating the student model based on the overall loss to obtain a trained student model;

[0080] Step S8: Screen the trained student model to obtain a final segmentation model;

[0081] Step S9: input the image to be segmented into the final segmentation model and output the segmentation result.

[0082] In some feasible embodiments, step S1 specifically includes:

[0083] The original dataset is first divided into a training dataset and a validation dataset in a ratio of 7:3. The dataset is used to train the model, and the validation dataset is used to test the model's accuracy. The training dataset is then divided into a supervised dataset and an unsupervised dataset in a ratio of 1:9.

[0084] In some feasible embodiments, step S2 specifically includes:

[0085] S2.1. Build a convolutional neural network as the first segmentation network. This experiment uses the commonly used unet neural network

[0086] S2.2. Build a transformer network as the second segmentation network. In this experiment, we choose the swin-unet network, which has a similar shape to the unet but a different kernel, and initialize it with the ImageNet pre-trained parameters.

[0087] Furthermore, another UNET network is built as its teacher model, and the parameter update mechanism between the teacher model and the student model adopts the exponential moving average method to ensure that the experience of the teacher model can be effectively transferred to the student model. Specifically, the parameter update of the teacher model takes into account historical information and the update of the student network parameters to improve the robustness and generalization ability of the model. The specific update formula is:

[0088] θ' r =αθ' r-1 +(1-α)θ r

[0089] Among them, α represents the decay rate of the exponential moving average, r refers to the training steps of the model, and θ r Refers to the weight of the student model at the training step r, θ' r-1 Refers to the weight of the teacher model at the previous step.

[0090] In some feasible embodiments, in step S3, referring to Figure 3 , supervised training specifically includes:

[0091] Inputting the supervised data into an adversarial generation module;

[0092] Apply random initial noise to the data and then input it into the network for initial prediction;

[0093] The predicted value output by the network and the input label value are calculated using the dice loss, and then the gradient of the loss with respect to the noise value is calculated.

[0094] The calculated gradient value is added to the original image to obtain the adversarial noise sample, which is then input into the network and the result is compared with the label to obtain the adversarial noise loss;

[0095] The data is input into two types of models respectively to obtain multiple prediction values, and then the supervised loss is calculated, specifically the cross entropy loss L mse , dice loss L dice , expressed as

[0096]

[0097]

[0098] The supervised loss is expressed as follows:

[0099] L seg =aL mse +(1-a)L dice

[0100] a is used to control L mse and L dice The ratio ranges from 0 to 1.

[0101] The supervised loss is calculated using a weighted sum of cross-entropy and dice loss. By simultaneously feeding labeled datasets into two student network models and leveraging the combination of cross-entropy and dice loss, the model's accurate grasp of medical image features improves, enabling better segmentation.

[0102] In some feasible embodiments, in step S4, referring to Figure 2 , its unsupervised training specifically includes:

[0103] Inputting the unsupervised data into the teacher model to generate teacher model pseudo labels;

[0104] Input the unsupervised data and the teacher model pseudo-label into the adversarial noise module and perform the same steps as above to calculate the adversarial loss.

[0105] The unsupervised data is input into two types of models respectively to obtain multiple prediction values, which are then input into the pseudo-label fusion module;

[0106] The pseudo-label fusion rule is as follows: in the student and teacher models of the convolutional network and the transformer network, if more than half of the prediction results for the position are considered to belong to a certain class C, and the probability value of any result is higher than a certain accuracy rate, the position is considered a confident position (if it is lower than a certain accuracy rate, it is considered an uncertain position. After calculating all pixels, a high-confidence pseudo-label is finally generated). After confirming each pixel, a confidence template that matches the size of the image is obtained;

[0107] The unsupervised loss is calculated by combining multiple predicted values ​​with the pseudo labels generated by the pseudo label fusion module.

[0108] In this embodiment, the unsupervised loss is the consistency loss. This calculation is done by inputting the unlabeled dataset into the student and teacher networks, generating a confidence template by voting, and then calculating the consistency loss L. u .

[0109] After obtaining the confidence template, the student model output is compared with the template one by one using the dice loss consistency loss.

[0110] In some feasible embodiments, the total loss L total The specific expressions are as follows:

[0111] L total =L seg1 +L seg2 +L lds +βL u

[0112] Among them L seg1 L seg2 is the supervised loss of the convolutional model and the transformer model, β is the parameter of the unsupervised loss ratio, which is used to suppress the early unsupervised damage network and make full use of the unsupervised characteristics in the later stage. u represents the consistency loss, L lds represents the noise resistance loss.

[0113] In some feasible embodiments, since there are two types of models that exist independently, a model with higher accuracy can be selected as the final model after calculation on the validation set.

[0114] The performance of the present invention, a semi-supervised medical image segmentation method with multi-view joint training, will be evaluated below: It should be noted that the dataset used in this section is the mainstream open source cardiac dataset ACDC, which contains sample data of 100 patients. After dividing the dataset, the data of 7 patients are selected as the labeled dataset and the data of 63 patients are selected as the unlabeled dataset, and then slice training is performed on them.

[0115] In this example, the present invention performs segmentation prediction on the ACDC dataset, and the performance of the present invention is evaluated based on the prediction results. The experiment uses the Dice similarity coefficient (Dice loss) and the Hausdorff distance 95% (HD95) score as performance evaluation indicators, and compares them with two currently advanced benchmark models: MC-Net and SS-Net. The model implementation and execution follow the steps in the open source code for the model. The experimental code is consistent with the implementation method provided by the present invention.

[0116] Table 1 Performance test results of the proposed method on the ACDC dataset

[0117]

[0118]

[0119] Table 1 shows our proposed semi-supervised medical image segmentation method with multi-view joint training. In our experiments, this method achieved excellent performance, with a final Dice coefficient of 88.6% and an HD95 score of 4.8. It is worth emphasizing that our final model prediction only uses the final output of the unet model for calculation.

[0120] like Figure 4 As shown in FIG, a semi-supervised medical image segmentation system for multi-view joint training includes:

[0121] The dataset construction module acquires medical image data and divides it into labeled and unlabeled datasets;

[0122] A model construction module constructs a segmentation network model based on networks of different categories to obtain a first segmentation network and a second segmentation network; the first segmentation network is provided with an adversarial noise generation module;

[0123] A model definition module is used to build a model with the same structure as the first segmentation network as a teacher model, and use the first segmentation network and the second segmentation network as student models;

[0124] a loss calculation module, configured to input the labeled data set into the student model respectively and calculate the loss to obtain a first supervised loss, a second supervised loss, and a first adversarial noise loss; input the unlabeled data set into the teacher model and the student model respectively to obtain a first unsupervised loss, a second unsupervised loss, and a second adversarial noise loss; and perform a weighted sum based on the first supervised loss, the second supervised loss, the first adversarial noise loss, the first unsupervised loss, the second unsupervised loss, and the second adversarial noise loss to obtain an overall loss;

[0125] A model updating module updates the student model based on the overall loss to obtain a trained student model;

[0126] A model screening module is used to screen the trained student models to obtain a final segmentation model;

[0127] The model application module is used to input the image to be segmented into the final segmentation model and output the segmentation result.

[0128] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0129] A semi-supervised medical image segmentation device with multi-view joint training:

[0130] at least one processor;

[0131] at least one memory for storing at least one program;

[0132] When the at least one program is executed by the at least one processor, the at least one processor implements the above-mentioned multi-view joint training semi-supervised medical image segmentation method.

[0133] The contents of the above method embodiments are all applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0134] A storage medium stores processor-executable instructions, which, when executed by a processor, are used to implement the above-mentioned multi-view joint training semi-supervised medical image segmentation method.

[0135] The contents of the above method embodiments are all applicable to the present storage medium embodiment. The functions specifically implemented by the present storage medium embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0136] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A semi-supervised medical image segmentation method based on multi-view joint training, characterized in that: The following steps are involved: Obtain medical image data and divide it into labeled datasets and unlabeled datasets; Building a segmentation network model based on different types of networks to obtain a first segmentation network and a second segmentation network; The first segmentation network is provided with an adversarial noise generation module; Build a model with the same structure as the first segmentation network as a teacher model, and use the first segmentation network and the second segmentation network as student models; Inputting the labeled data set into the first segmentation network and the second segmentation network of the student model respectively, and calculating the loss to obtain a first supervised loss, a second supervised loss and a first adversarial noise loss; Inputting the unlabeled data set into the teacher model and the student model respectively to obtain a first unsupervised loss, a second unsupervised loss, and a second adversarial noise loss; Obtain an overall loss by performing a weighted summation based on the first supervised loss, the second supervised loss, the first adversarial noise loss, the first unsupervised loss, the second unsupervised loss, and the second adversarial noise loss; Update the student model based on the overall loss to obtain a trained student model; Screening the trained student models to obtain a final segmentation model; The image to be segmented is input into the final segmentation model, and the segmentation result is output.

2. The method for semi-supervised medical image segmentation based on multi-view joint training according to claim 1, characterized in that: The first segmentation network is constructed based on an adversarial noise generation module and a convolutional neural network, and the second segmentation network is constructed based on a transformer.

3. The method for semi-supervised medical image segmentation based on multi-view joint training according to claim 1, characterized in that: The parameter update of the teacher model takes into account the historical information and the update of the student network parameters, and its formula is expressed as follows: θ′ r =θ′ r-1 +(1-α)θ r Among them, α represents the decay rate of the exponential moving average, r refers to the training steps of the model, and θ r Refers to the weight of the student model at the training step r, θ′ r-1 Refers to the weight of the teacher model at the previous step.

4. The method for semi-supervised medical image segmentation based on multi-view joint training according to claim 2, characterized in that: The step of inputting the labeled data set into the first segmentation network and the second segmentation network of the student model respectively, and calculating the loss to obtain the first supervised loss, the second supervised loss and the first anti-noise loss specifically includes: Input the labeled data sets in the same batch into the first segmentation network and the second segmentation network of the student model respectively; For the first segmentation network, adding noise to the initial image in the labeled dataset based on the adversarial noise generation module to generate noisy original data; Inputting the noisy raw data into a convolutional neural network to perform initial prediction to obtain a predicted value; Performing a dice loss calculation based on the predicted value and the label value in the labeled data set, and calculating a gradient value of the dice loss with respect to noise; Adding the gradient value to the corresponding initial image to obtain an adversarial noise sample; Inputting the adversarial noise sample into a convolutional neural network, and comparing the result with the label to obtain a first adversarial noise loss; Inputting the labeled data set into the first segmentation network and the second segmentation network of the student model respectively to generate a prediction value; A first supervised loss and a second supervised loss are calculated based on the predicted values.

5. The method for semi-supervised medical image segmentation based on multi-view joint training according to claim 4, characterized in that: The first supervised loss and the second supervised loss both include cross entropy loss and dice loss, where: The calculation formula for cross entropy loss is as follows: Among them, Y represents the class label value, P represents the prediction result, N represents the number of pixels, and L i Represents the calculation formula of the loss between pixels, M represents the number of categories, represents the true probability that the i-th pixel belongs to category c, Indicates the predicted probability that the i-th pixel belongs to category c; The formula for calculating dice loss is as follows:

6. The method for semi-supervised medical image segmentation based on multi-view joint training according to claim 1, characterized in that: The step of inputting the unlabeled dataset into the teacher model and the student model respectively to obtain a first unsupervised loss, a second unsupervised loss, and a second anti-noise loss specifically includes: Inputting the unlabeled data set into the teacher model to generate teacher model pseudo labels; For the first segmentation network, calculating a second adversarial noise loss based on the unlabeled dataset and the teacher model pseudo labels; Inputting the unlabeled data set into the first segmentation network and the second segmentation network of the student model respectively to obtain corresponding prediction values; Perform pseudo-label fusion according to the predicted value to obtain a pseudo-label; An unsupervised loss is calculated based on the unlabeled data and the pseudo labels.

7. The method for semi-supervised medical image segmentation based on multi-view joint training according to claim 6, characterized in that: The rules for pseudo-label fusion are as follows: Among all the predicted values, for one pixel position, if more than half are considered to belong to a certain category and any output probability value is greater than a preset threshold, then the pixel position is considered to be a confident position.

8. The method for semi-supervised medical image segmentation based on multi-view joint training according to claim 6, characterized in that: The step of calculating the unsupervised loss based on the unlabeled data and the pseudo-labels specifically includes: inputting the unlabeled data into the student model; Constructing a confidence template based on the pseudo-label; The consistency loss is calculated based on the confidence template and the output of the student model to obtain an unsupervised loss.

9. A multi-view joint training semi-supervised medical image segmentation system, characterized by: include: The dataset construction module acquires medical image data and divides it into labeled and unlabeled datasets; A model construction module constructs a segmentation network model based on networks of different categories to obtain a first segmentation network and a second segmentation network; the first segmentation network is provided with an adversarial noise generation module; A model definition module is used to build a model with the same structure as the first segmentation network as a teacher model, and use the first segmentation network and the second segmentation network as student models; a loss calculation module, configured to input the labeled data set into the first segmentation network and the second segmentation network of the student model, respectively, and calculate the losses to obtain a first supervised loss, a second supervised loss, and a first adversarial noise loss; input the unlabeled data set into the teacher model and the student model, respectively, to obtain a first unsupervised loss, a second unsupervised loss, and a second adversarial noise loss; and perform weighted summation based on the first supervised loss, the second supervised loss, the first adversarial noise loss, the first unsupervised loss, the second unsupervised loss, and the second adversarial noise loss to obtain an overall loss; A model updating module updates the student model based on the overall loss to obtain a trained student model; A model screening module is used to screen the trained student models to obtain a final segmentation model; The model application module is used to input the image to be segmented into the final segmentation model and output the segmentation result.

10. A semi-supervised medical image segmentation device with multi-view joint training, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the multi-view joint training semi-supervised medical image segmentation method as described in any one of claims 1 to 8.

Images