A Medical Image Segmentation Method with Intra-Class Region Dynamic Decoupling
Through the combination of intra-class region decoupling and dynamic threshold modules, the problem of difficulty in accurately identifying and distinguishing different regions of vascular images in the prior art is solved, and the accurate segmentation of vascular images and the relief of intra-class imbalance problems are achieved.
Patent Information
- Application Number
- CN202510363056.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-26
- Publication Date
- 2025-06-24
- Estimated Expiration
- 2045-03-26
AI Technical Summary
The existing dynamic decoupling method of partially-class regions is difficult to accurately identify and distinguish different regions when dealing with complex vascular structures, resulting in poor segmentation effect.
Through intra-class area decoupling, the vascular image is divided into difficult sub-regions and easy sub-regions, and the dynamic threshold module is used for precise decoupling, and the model is trained through a semi-supervised learning method to alleviate the problem of intra-class imbalance.
The accurate segmentation of blood vessel images is achieved, the problem of intra-class imbalance is alleviated, and the segmentation effect and model generalization ability are improved.
Smart Images
Figure CN119887808B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of digital image processing, and particularly to a medical image segmentation method for intra-class region dynamic decoupling. Background Art
[0002] Precisely segmenting blood vessels from computed tomography (CT) is an important prerequisite for diagnosing and treating various diseases. Different from other organs with simple and single shapes, blood vessels have a complex tree-like structure, so it is difficult to obtain complete and accurate blood vessel labels. As a method with low requirements for labeled data, semi-supervised medical image segmentation greatly alleviates the dependence of neural networks on labeled data. However, there are significant intra-class differences in the number of voxels, the number of branches, and the morphological structure between the main branches and peripheral microvessels of blood vessels. This severe intra-class distribution imbalance problem hinders the practical application of semi-supervised methods in blood vessel segmentation. Therefore, it is worth researching and designing a general semi-supervised medical image segmentation method.
[0003] Some existing intra-class region dynamic decoupling methods use traditional image processing techniques to simply segment images based on the gray values of the images. When dealing with complex blood vessel structures, since the structures of blood vessels are often complex and variable and there are individual differences, traditional image processing techniques may have difficulty accurately identifying and distinguishing different regions.
[0004] In summary, some existing intra-class region dynamic decoupling methods use traditional image processing techniques, which may have difficulty accurately identifying and distinguishing different regions. Therefore, it is necessary to propose a medical image segmentation method for intra-class region dynamic decoupling. Summary of the Invention
[0005] To solve the above problems, the present invention provides a medical image segmentation method for intra-class region dynamic decoupling. By intra-class region decoupling, blood vessel images are divided into difficult sub-regions and easy sub-regions. By adopting different optimization methods for different regions to guide the model to learn different intra-class features, the intra-class imbalance problem is alleviated. The dynamic threshold module accurately decouples different regions within the class and generates learnable thresholds for images during training.
[0006] To achieve the above object, the technical solution of the present invention is as follows: A medical image segmentation method for intra-class region dynamic decoupling, comprising the following steps:
[0007] S1, dataset preprocessing: Collect dataset images through the Internet, truncate and normalize the voxel intensities of all dataset images, perform data augmentation on them to obtain augmented data, and at the same time divide the images containing blood vessels to be segmented into bounding boxes.
[0008] S2. Establish a three-branch baseline segmentation network: Use a U-shaped network as the baseline network for segmentation. Establish a number of student models and teacher models, and combine the student models and teacher models into student-student models and teacher-student models. All student models and teacher models are 3D U-Net networks with the same architecture but different initial perturbations.
[0009] S3. Decouple the intra-class regions: Divide the augmented data into labeled data and unlabeled data, and decouple the blood vessel categories into difficult sub-regions and easy sub-regions, and apply different constraints to the difficult sub-regions and easy sub-regions. When performing the decoupling operation, send the predicted probability map of one student model and the noisy pseudo-labels of another student model into the dynamic threshold module to obtain the first dynamic threshold; when the degree of consistency of the voxels of the probability map generated by another student model is greater than the first dynamic threshold, it is considered that this student model reaches consistency in this region, and at this time the region where this student model is located is the easy sub-region; send the probability map of the teacher model and the average value of the noisy pseudo-labels of all student models into the dynamic threshold module to obtain the second dynamic threshold. If the degree of consistency of the voxels of the teacher model is less than the second dynamic threshold, it is considered that there is ambiguity here, which is the difficult sub-region.
[0010] S4. Construct a dynamic threshold module: Construct a dynamic threshold module, and accurately decouple different difficult sub-regions and easy sub-regions through the dynamic threshold module.
[0011] S5. Establish a constraint optimization network: All student models perform representation learning on different difficult sub-regions and easy sub-regions, and the teacher model performs consistency constraints on all student models.
[0012] S6. Train the semi-supervised blood vessel segmentation network: Utilize the labeled data and unlabeled data through the loss function to train the semi-supervised blood vessel segmentation network.
[0013] S7. Segmentation target: Input the test image, and the semi-supervised blood vessel segmentation network tests the test image to complete the segmentation of the blood vessel region.
[0014] Furthermore, in S1, use the smallest bounding box containing the blood vessel to be segmented as the range of the region of interest of the student model.
[0015] Furthermore, in S1, the data augmentation method adopts one or more of random horizontal flipping, vertical flipping, and random rotation from -90 degrees to 90 degrees.
[0016] Furthermore, in S2, the teacher model is the self-ensemble of the average value of the student models, obtained by exponential moving average.
[0017] Furthermore, in S2, the expression of the student model is as follows:
[0018] (1)
[0019] (2)
[0020] The expression of the teacher model is as follows:
[0021] (3)
[0022] Among them, are model parameters, and are both student models, is the teacher model.
[0023] Furthermore, in S3, the unlabeled data includes the probability map generated by the student model and the noisy pseudo-labels.
[0024] The easy sub-regions of the vascular image are obtained by using the prediction consistency of the student-student model, and the information of the easy sub-regions is stored in the confidence map. The specific formula for this process is as follows:
[0025] (4)
[0026] (5)
[0027] Among them, is the indicator function, is the dynamic threshold module, and are the probability map and the noisy pseudo-labels generated by one of the student models, and are the probability map and the noisy pseudo-labels generated by another student model, is the first dynamic threshold for each category, is the regional confidence map, is the degree of consistency between the student models.
[0028] Furthermore, in S3, when performing the decoupling operation, the difficult sub-regions of the vascular image are obtained by using the prediction uncertainty of the teacher-student model, and the information of the difficult sub-regions is stored in the uncertainty map. The specific formula for this process is as follows:
[0029] (6)
[0030] (7)
[0031] (8)
[0032] Among them, is the probability map of the teacher model, is The average value of , is the second dynamic threshold for each category, is the uncertainty map, and
[0033] Further, in S4, when performing the precise decoupling operation, first cover the pseudo-label on the probability map and find the voxels of each category in the probability map corresponding to the category in the pseudo-label. Finally, calculate the threshold for each category respectively. The specific formula is as follows:
[0034] (9)
[0035] (10)
[0036] Among them, is the indicator function, M represents the mask, represents the noisy pseudo-label, represents the probability map, represents the threshold.
[0037] Further, in S5, for the easy sub-region, pseudo-label cross-supervision is performed between different student models. The definition formula of the pseudo-label cross-supervision loss is as follows:
[0038] (11)
[0039] (12)
[0040] (13)
[0041] Among them, is the dice loss, and are respectively and the pseudo-labels generated by and are respectively and the probability maps generated by is the total number of data in the dataset, is the number of labeled data, is the student model the confidence map of the easy region of the noisy pseudo-label.
[0042] For the difficult sub-region, the teacher model is introduced to calculate the consistency loss. The parameters of the teacher model are updated by exponential moving average. The specific formula is as follows:
[0043] (14)
[0044] where \(t\) is the current iteration number of training, and are hyperparameters for controlling the exponential moving average update.
[0045] The formula for the consistency loss is specifically as follows:
[0046] (15)
[0047] where is the squared loss function, is the unlabeled data, and are both model parameters of the student model, is the model parameter of the teacher model.
[0048] Furthermore, in S6, the loss function includes a supervised loss and an unsupervised loss, and the specific formula for the loss function is as follows:
[0049] (16)
[0050] where represents the supervised loss, represents the unsupervised loss, is a balance parameter for balancing the supervised loss and the unsupervised loss.
[0051] The specific formula for the supervised loss is as follows:
[0052] (17)
[0053] where is a linear combination of the cross-entropy loss and the dice loss, is the labeled data, is the label of the labeled data.
[0054] The unsupervised loss includes the pseudo-label cross-supervision loss between student models and the consistency loss between the teacher-student models, and the specific formula is as follows:
[0055] (18)
[0056] where represents the pseudo-label cross-supervision loss, represents the consistency loss.
[0057] Adopting the above scheme has the following beneficial effects:
[0058] 1. This solution, as a general medical image segmentation method, can achieve accurate segmentation of various medical images. Through intra-class region decoupling, vascular images are separated into difficult sub-regions and easy sub-regions. By adopting different optimization methods for different regions to guide the model to learn different intra-class features, the serious intra-class imbalance problem is alleviated. Through a dynamic threshold module based on confidence learning, different regions within the class are precisely decoupled, and the dynamic threshold module generates an adaptive and learnable threshold for the image during the training process.
[0059] 2. In this solution, by limiting the range of the region of interest of the student model to the smallest bounding box containing the blood vessels to be segmented, the student model will focus on processing this part of the image data, thus significantly reducing the amount of data that needs to be processed with emphasis, reducing the computational complexity, and accelerating the image processing speed.
[0060] 3. In this solution, by randomly rotating and flipping the image, different viewing situations at different angles and directions can be simulated, thus increasing the data diversity. This diversity helps the teacher model and the student model learn more comprehensive features and improve the generalization ability of the teacher model and the student model. At the same time, operations such as random rotation and flipping can force the teacher model and the student model to learn more robust feature representations, and these features can still maintain consistency under different transformations, which helps improve the performance of the teacher model and the student model on the test set.
[0061] 4. In this solution, by integrating the parameters of the student model in multiple training steps, the teacher model can capture a more extensive feature representation. Since these feature representations have good generalization ability on different datasets and test sets, the teacher model can better handle unseen data and improve the generalization performance of the teacher model.
[0062] 5. In this solution, by combining the student model and the teacher model and using the dynamic threshold module to decouple the difficult sub-regions and easy sub-regions of the blood vessel category, this method can more accurately identify and process complex regions in the image. This fine-grained processing strategy makes the segmentation result more accurate.
[0063] Additional aspects and advantages of the present invention will be given in part in the following description, become apparent in part from the following description, or be learned through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS
[0064] Figure 1 It is a schematic flow chart of the medical image segmentation method for dynamic decoupling of intra-class regions of the present invention.
[0065] Figure 2 It is a schematic flow chart of establishing the student model and the teacher model in the medical image segmentation method for dynamic decoupling of intra-class regions of the present invention.
[0066] Figure 3 It is a classification schematic diagram of blood vessel categories in the medical image segmentation method for intra-class region dynamic decoupling of the present invention. Specific implementation manners
[0067] The following is a further detailed description through specific implementation manners:
[0068] Example 1:
[0069] As shown in Figure 1 , Figure 2 and Figure 3 : A medical image segmentation method for intra-class region dynamic decoupling includes the following steps:
[0070] S1, Dataset preprocessing: Collect dataset images through the Internet, truncate and normalize the voxel intensities of all dataset images, perform data augmentation on them to obtain augmented data, and at the same time divide the images containing the blood vessels to be segmented into bounding boxes.
[0071] S2, Establish a three-branch baseline segmentation network: Use the U-shaped network as the baseline network for segmentation, establish several student models and teacher models, and combine the student models and teacher models into student-student models and teacher-student models. All student models and teacher models are 3D U-Net networks with the same architecture but different initial perturbations.
[0072] The expression of the student model is as follows:
[0073] (1)
[0074] (2)
[0075] The expression of the teacher model is as follows:
[0076] (3)
[0077] Wherein, are model parameters, and are both student models, is the teacher model.
[0078] S3, Decouple the intra-class regions: Divide the augmented data into labeled data and unlabeled data, and decouple the blood vessel categories into difficult sub-regions and easy sub-regions, and apply different constraints to the difficult sub-regions and easy sub-regions. Among them, the unlabeled data includes the probability maps generated by the student models and the noisy pseudo-labels.
[0079] When performing the decoupling operation, first send the predicted probability map of one student model and the noisy pseudo-labels of another student model into the dynamic threshold module to obtain the first dynamic threshold; when the degree of consistency of the voxels in the probability map generated by the other student model is greater than the first dynamic threshold, it is considered that the student model reaches consistency in this area, and the area where the student model is located at this time is the easy sub-region.
[0080] After that, send the probability map of the teacher model and the average value of the noisy pseudo-labels of all student models into the dynamic threshold module to obtain the second dynamic threshold. If the degree of consistency of the voxels of the teacher model is less than the second dynamic threshold, it is considered that there is ambiguity here, which is the difficult sub-region.
[0081] S4. Construct the dynamic threshold module: Construct the dynamic threshold module and accurately decouple different difficult sub-regions and easy sub-regions through the dynamic threshold module.
[0082] First, obtain the easy sub-region of the vascular image by using the prediction consistency of the student-student model, and store the information of the easy sub-region in the confidence map. The specific formula for this process is as follows:
[0083] (4)
[0084] (5)
[0085] Among them, is the indicator function, is the dynamic threshold module, and are the probability map and noisy pseudo-labels generated by one of the student models, and are the probability map and noisy pseudo-labels generated by another student model, is the first dynamic threshold for each category, is the regional confidence map, is the degree of consistency between student models.
[0086] Subsequently, obtain the difficult sub-region of the vascular image by using the prediction uncertainty of the teacher-student model, and store the information of the difficult sub-region in the uncertainty map. The specific formula for this process is as follows:
[0087] (6)
[0088] (7)
[0089] (8)
[0090] Among them, is the probability map of the teacher model, is the average value of , is the second dynamic threshold for each category, is the uncertainty map, is the degree of consistency of the teacher model.
[0091] When performing the precise decoupling operation, first cover the pseudo-label on the probability map and find the voxels of each category in the probability map corresponding to the category in the pseudo-label. Finally, calculate the threshold for each category respectively. The specific formula is as follows:
[0092] (9)
[0093] (10)
[0094] where is the indicator function, M represents the mask, represents the noisy pseudo-label, represents the probability map, represents the threshold.
[0095] S5. Establish a constrained optimization network: All student models perform representation learning on different difficult sub-regions and easy sub-regions, and the teacher model performs consistency constraints on all student models.
[0096] For the easy sub-region, cross-supervision of pseudo-labels is performed between different student models. The formula for defining the cross-supervision loss of pseudo-labels is as follows:
[0097] (11)
[0098] (12)
[0099] (13)
[0100] where is the dice loss, and are respectively and the generated pseudo-labels, and are respectively and the generated probability maps, is the total number of data in the dataset, is the number of labeled data, is the student model confidence map of the easy region of the noisy pseudo-label.
[0101] For difficult sub-regions, a teacher model is introduced to calculate the consistency loss, and the parameters of the teacher model are updated by exponential moving average. The specific formula is as follows:
[0102] (14)
[0103] where t is the current training iteration number, and are hyperparameters that control the exponential moving average update.
[0104] The formula for the consistency loss is specifically as follows:
[0105] (15)
[0106] where, is the squared loss function, is the unlabeled data, and are both the model parameters of the student model, is the model parameter of the teacher model.
[0107] S6. Train the semi-supervised vascular segmentation network: Utilize the labeled data and unlabeled data through the loss function to train the semi-supervised vascular segmentation network.
[0108] The loss function includes the supervised loss and the unsupervised loss. The specific formula of the loss function is as follows:
[0109] (16)
[0110] where, represents the supervised loss, represents the unsupervised loss, is the balance parameter used to balance the supervised loss and the unsupervised loss.
[0111] The specific formula of the supervised loss is as follows:
[0112] (17)
[0113] where, is a linear combination of the cross-entropy loss and the dice loss, is the labeled data, is the label of the labeled data.
[0114] The unsupervised loss includes the pseudo-label cross-supervision loss between student models and the consistency loss between the teacher-student models. The specific formula is as follows:
[0115] (18)
[0116] where, represents the pseudo-label cross-supervision loss, represents the consistency loss.
[0117] S7, segmentation objective: Input the image to be tested, and the semi-supervised vascular segmentation network tests the image to be tested to complete the segmentation of the vascular region.
[0118] The specific implementation process is as follows: In this embodiment, the axial dimension of all slices of the image is 512×512, and the spatial resolution is 0.5mm - 1.0mm. The image labels are obtained by carefully annotating the images by 3 radiologists with more than 5 years of clinical experience.
[0119] First, collect dataset images through the Internet, truncate the voxel intensity of all dataset images within the window of [-1000, 600] Hounsfield units, normalize it to [0, 1], and perform data augmentation processing on it to obtain augmented data. At the same time, divide the images containing the blood vessels to be segmented into bounding boxes. Use the U-shaped network as the baseline network for segmentation, establish several student models and teacher models, and combine the student models and teacher models into student-student models and teacher-student models. All the established student models and teacher models are 3D U-Net networks with the same architecture but different initial perturbations.
[0120] In this embodiment, the number of student models is two, and the expressions of the student models are as follows:
[0121] (1)
[0122] (2)
[0123] The expressions of the teacher models are as follows:
[0124] (3)
[0125] Among them, are model parameters, and are both student models, is the teacher model.
[0126] After completing the establishment of the student models and teacher models, divide the augmented data into labeled data and unlabeled data. At the same time, decouple the vascular categories into difficult sub-regions and easy sub-regions, and apply different constraints to the difficult sub-regions and easy sub-regions.
[0127] After completing the decoupling of the intra-class regions, construct a dynamic threshold module, and use the dynamic threshold module to accurately decouple different difficult sub-regions and easy sub-regions.
[0128] When decoupling the vascular category into difficult sub-regions and easy sub-regions, first, the easy sub-regions of the vascular image are obtained by using the prediction consistency of the student-student model, and the information of the easy sub-regions is stored in the confidence map. The predicted probability map of one student model and the noisy pseudo-label of another student model are fed into the dynamic threshold module to obtain the first dynamic threshold; when the degree of consistency of the voxels of the probability map generated by the other student model is greater than the first dynamic threshold, it is considered that the student model reaches consistency in this region, and the region where the student model is located at this time is the easy sub-region.
[0129] The specific formula for this process is as follows:
[0130] (4)
[0131] (5)
[0132] Among them, is the indicator function, is the dynamic threshold module, and are the probability map and the noisy pseudo-label generated by one of the student models, and are the probability map and the noisy pseudo-label generated by another student model, is the first dynamic threshold for each category, is the regional confidence map, is the degree of consistency between the student models.
[0133] Subsequently, the difficult sub-regions of the vascular image are obtained by using the prediction uncertainty of the teacher-student model, and the information of the difficult sub-regions is stored in the uncertainty map. The probability map of the teacher model and the average value of the noisy pseudo-labels of all student models are fed into the dynamic threshold module to obtain the second dynamic threshold. If the degree of consistency of the voxels of the teacher model is less than the second dynamic threshold, it is considered that there is ambiguity here, which is the difficult sub-region.
[0134] The specific formula for this process is as follows:
[0135] (6)
[0136] (7)
[0137] (8)
[0138] Among them, is the probability map of the teacher model, is and the average value of, is the second dynamic threshold for each category, is the uncertainty graph, is the consistency degree of the teacher model.
[0139] When performing the precise decoupling operation, first cover the pseudo-label on the probability graph and find the voxels of each category in the probability graph corresponding to the categories in the pseudo-label. Finally, calculate the threshold of each category respectively. The specific formula is as follows:
[0140] (9)
[0141] (10)
[0142] Among them, is the indicator function, M represents the mask, represents the noisy pseudo-label, represents the probability graph, represents the threshold.
[0143] Subsequently, establish a constrained optimization network to enable all student models to further perform representation learning on different difficult sub-regions and easy sub-regions, and perform consistency constraints on all student models through the teacher model.
[0144] For the easy sub-region, perform pseudo-label cross-supervision between different student models. The definition formula of the pseudo-label cross-supervision loss is as follows:
[0145] (11)
[0146] (12)
[0147] (13)
[0148] Among them, is the dice loss, and are respectively and the generated pseudo-labels, and are respectively and the generated probability graphs, is the total number of data in the dataset, is the number of labeled data, is the student model the confidence map of the easy region of the noisy pseudo-label.
[0149] For the difficult sub-region, introduce the teacher model to calculate the consistency loss. The parameters of the teacher model are updated by exponential moving average. The specific formula is as follows:
[0150] (14)
[0151] Among them, t is the iteration number of the current training, and are hyperparameters for controlling the exponential moving average update.
[0152] The formula for the consistency loss is specifically as follows:
[0153] (15)
[0154] Among them, is the squared loss function, is the unlabeled data, and are both model parameters of the student model, is the model parameter of the teacher model.
[0155] Finally, the labeled data and the unlabeled data are utilized through the loss function to train the semi-supervised vascular segmentation network. After the staff inputs the test image into the semi-supervised vascular segmentation network, the semi-supervised vascular segmentation network tests the image and completes the segmentation of the vascular region.
[0156] The loss function includes the supervised loss and the unsupervised loss. The specific formula of the loss function is as follows:
[0157] (16)
[0158] Among them, represents the supervised loss, represents the unsupervised loss, is the balance parameter for balancing the supervised loss and the unsupervised loss.
[0159] The specific formula of the supervised loss is as follows:
[0160] (17)
[0161] Among them, is a linear combination of the cross-entropy loss and the dice loss, is the labeled data, is the label of the labeled data.
[0162] The unsupervised loss includes the pseudo-label cross-supervision loss between student models and the consistency loss between the teacher-student models. The specific formula is as follows:
[0163] (18)
[0164] Among them, represents the pseudo-label cross-supervision loss, Represents the consistency loss. Embodiment
[0165] As shown in the appendix Figure 1 As shown, the difference from Embodiment 1 is that in S1, the smallest bounding box containing the blood vessel to be segmented is used as the range of the region of interest of the student model.
[0166] The specific implementation process is as follows: When processing a smaller dataset, limiting the range of the region of interest of the student model to the blood vessel region can reduce the risk of overfitting of the model. At the same time, the student model can focus more on learning the characteristics of the blood vessel region rather than various features in the entire image, enabling the student model to more accurately identify the blood vessel boundary and improve the segmentation accuracy.
[0167] Embodiment 3:
[0168] As shown in the appendix Figure 1 As shown, the difference from Embodiment 2 is that in S1, the data augmentation method uses one or more of random horizontal flipping, vertical flipping, and random rotation from -90 degrees to 90 degrees. In this embodiment, a combination of horizontal flipping and vertical flipping is used.
[0169] The specific implementation process is as follows: Operations such as random horizontal flipping, vertical flipping, and random rotation can be used flexibly in combination to generate more diverse training samples. This flexibility enables the data augmentation method to be customized and optimized according to the specific task and the characteristics of the dataset.
[0170] Embodiment 4:
[0171] As shown in the appendix Figure 2 As shown, the difference from Embodiment 3 is that in S2, the teacher model is the self-ensemble of the average value of the student model, obtained by exponential moving average.
[0172] The specific implementation process is as follows: When processing complex data, such as noisy data, the teacher model can mitigate the impact of data complexity on the model performance through its stability and robustness. At the same time, the teacher model can also capture the potential patterns and rules in the data, thereby helping the student model better understand and process this data.
[0173] Obviously, the above embodiments are merely examples given for clear illustration and are not limitations on the implementation methods. For those of ordinary skill in the art, other different forms of changes or variations can be made based on the above description. It is not necessary and impossible to enumerate all the implementation methods here. And the obvious changes or variations derived therefrom are still within the protection scope of the present invention.
Claims
1. A medical image segmentation method with dynamic decoupling of intra-class regions, characterized in that: The following steps are involved: S1, dataset preprocessing: dataset images are collected through the Internet, the voxel intensities of all dataset images are truncated and normalized, and data enhancement is performed on them to obtain enhanced data. At the same time, the images containing the blood vessels to be segmented are divided into bounding boxes; S2, establish a three-branch baseline segmentation network: use the U-net as the baseline network for segmentation, establish several student models and teacher models, and combine the student model and the teacher model into a student-student model and a teacher-student model. All student models and teacher models are 3D U-Net networks with the same architecture but different initialization perturbations. S3, decoupling intra-class regions: the enhanced data is divided into labeled data and unlabeled data, and the blood vessel category is decoupled into difficult sub-regions and easy sub-regions, and different constraints are imposed on the difficult sub-regions and easy sub-regions; when performing the decoupling operation, the predicted probability map of one student model and the noisy pseudo-label of another student model are sent to the dynamic threshold module to obtain the first dynamic threshold; when the consistency of the voxels of the probability map generated by the other student model is greater than the first dynamic threshold, it is considered that the student model has reached consistency in this area, and the area where the student model is located is the easy sub-region; the probability map of the teacher model and the average of the noisy pseudo-labels of all student models are sent to the dynamic threshold module to obtain the second dynamic threshold. If the consistency of the voxels of the teacher model is less than the second dynamic threshold, it is considered that there is ambiguity here, which is a difficult sub-region; S4, constructing a dynamic threshold module: constructing a dynamic threshold module, and accurately decoupling different difficult sub-regions and easy sub-regions through the dynamic threshold module; S5, establish a constrained optimization network: all student models learn to represent different difficult sub-regions and easy sub-regions, and the teacher model constrains all student models to be consistent; S6, training a semi-supervised blood vessel segmentation network: using the loss function to utilize labeled data and unlabeled data to train a semi-supervised blood vessel segmentation network; S7, segmentation target: input the image to be tested, and the semi-supervised blood vessel segmentation network tests the image to be tested to complete the segmentation of the blood vessel area.
2. The medical image segmentation method with dynamic decoupling of intra-class regions according to claim 1, characterized in that: In S1, the minimum bounding box containing the blood vessel to be segmented is used as the range of the region of interest of the student model.
3. The medical image segmentation method with dynamic decoupling of intra-class regions according to claim 1, characterized in that: In S1, the data augmentation method uses one or more of random horizontal flipping, vertical flipping, and random -90 degree to 90 degree rotation.
4. The medical image segmentation method with dynamic decoupling of intra-class regions according to claim 1, characterized in that: In S2, the teacher model is the self-integration of the average of the student models, obtained by exponential moving average.
5. The medical image segmentation method with dynamic decoupling of intra-class regions according to claim 1, characterized in that: In S2, the expression of the student model is as follows: （1）； （2）； The expression of the teacher model is as follows: （3）； in, are model parameters, and All are student models. Model for teachers.
6. The medical image segmentation method with dynamic decoupling of intra-class regions according to claim 1, characterized in that: In S3, the unlabeled data includes the probability map generated by the student model and the noisy pseudo labels; The prediction consistency of the student-student model is used to obtain the easy sub-region of the vascular image, and the information of the easy sub-region is stored in the confidence map. The specific formula of the process is as follows: （4）； （5）； in, is the indicator function, is the dynamic threshold module, and is a probability map and noisy pseudo-label generated by one of the student models, and is a probability map and noisy pseudo-label generated by another student model, is the first dynamic threshold for each category, is the regional confidence map, is the degree of consistency between student models.
7. The medical image segmentation method with dynamic decoupling of intra-class regions according to claim 6, characterized in that: In S3, when performing the decoupling operation, the prediction uncertainty of the teacher-student model is used to obtain the difficult sub-region of the vascular image, and the information of the difficult sub-region is stored in the uncertainty map. The specific formula of this process is as follows: （6）； （7）； （8）； in, is the probability graph of the teacher model, for and The average value of is the second dynamic threshold for each category, is the uncertainty diagram, is the consistency of the teacher model.
8. The medical image segmentation method with dynamic decoupling of intra-class regions according to claim 7, characterized in that: In S4, when performing precise decoupling operations, firstly, the pseudo labels are overlaid on the probability map and the voxels of the corresponding categories in the pseudo labels of each category in the probability map are found. Finally, the threshold of each category is calculated separately. The specific formula is as follows: （9）； （10）； in, is the indicator function, M represents mask, represents a noisy pseudo label, represents a probability graph, Indicates the threshold value.
9. The medical image segmentation method with dynamic decoupling of intra-class regions according to claim 8, characterized in that: In S5, for the easy sub-region, pseudo-label cross-supervision is performed between different student models, and the pseudo-label cross-supervision loss is defined as follows: （11）； （12）； （13）； in, is the dice loss, and They are and The pseudo labels generated are and They are and The generated probability map, is the total number of data in the dataset, is the number of labeled data, Model for students Easy region confidence map with noisy pseudo labels; For difficult sub-regions, a teacher model is introduced to calculate the consistency loss. The parameters of the teacher model are updated by exponential moving average. The specific formula is as follows: （14）； in, is the number of iterations of the current training, and Hyperparameters for controlling exponential moving average updates; The formula for consistency loss is as follows: （15）； in, is the squared loss function, For unlabeled data, and are the model parameters of the student model, are the model parameters of the teacher model.
10. The medical image segmentation method with dynamic decoupling of intra-class regions according to claim 9, characterized in that: In S6, the loss function includes supervised loss and unsupervised loss. The specific formula of the loss function is as follows: （16）； in, represents the supervision loss, represents the unsupervised loss, is a balance parameter used to balance supervised loss and unsupervised loss; The specific formula of supervision loss is as follows: （17）； in, is a linear combination of cross entropy loss and dice loss, For labeled data, is the label of the labeled data; The unsupervised loss includes the pseudo-label cross-supervision loss between student models and the consistency loss between teacher and student models. The specific formula is as follows: （18）； in, represents the pseudo-label cross-supervision loss, Indicates consistency loss.
Citation Information
Patent Citations
3D medical image segmentation model establishment method based on mask modeling and application thereof
CN116664588A
Semi-supervised medical image segmentation method, system, equipment and medium
CN117095014A