Semi-supervised medical image automatic segmentation method based on pseudo label optimization

Abstract

The invention discloses a semi-supervised medical image automatic segmentation method based on pseudo label optimization. The method comprises the following steps: step 1, preprocessing and dividing a data set; 2, constructing a semi-supervised training framework which comprises a pseudo label refining module, a triple loss module and a mutual correction framework; refining processing is carried out through an SLIC superpixel segmentation method and an information entropy voting mechanism; a triple loss function is controlled to be minimized; a mutual correction loss function is controlled to be minimized; step 3, constructing a supervision loss function based on the result of the supervision test; the total loss function convergence is controlled; and step 4, segmenting the verification set, and evaluating segmentation results based on two indexes of a Dice coefficient and an ASD average surface distance. According to the method, a semi-supervised training framework is constructed based on a pseudo-label refining module, a triple loss module and a mutual correction framework, the number of pseudo-labels is increased, the spatial consistency and accuracy of the pseudo-labels are improved, and the capture capability and segmentation precision of boundary information are optimized.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image segmentation, and specifically to a semi-supervised medical image automatic segmentation method based on pseudo-label optimization. Background Art

[0002] Medical image segmentation is one of the key technologies in computer-aided diagnosis systems, and its goal is to accurately extract target regions from medical images, such as organs, lesions, etc. Traditional medical image segmentation methods mainly rely on contour-based algorithms and traditional machine learning techniques, and these methods often perform limitedly when dealing with complex anatomical structures. With the development of deep learning technologies, especially the wide application of convolutional neural networks such as CNN, the accuracy and efficiency of medical image segmentation have been significantly improved. UNet and its variants have become the mainstream architectures for medical image segmentation because their encoder-decoder structures and skip connections can effectively capture multi-scale features.

[0003] However, one of the main challenges faced by medical image segmentation is the scarcity of labeled data. The annotation of medical images requires the participation of professional doctors, which is costly and time-consuming. Therefore, semi-supervised learning has become an important solution to improve the generalization ability of the model by using a large amount of unlabeled data. Existing semi-supervised medical image segmentation methods usually use pseudo-labels to utilize unlabeled data, but these methods often only focus on high-confidence prediction results in practice and ignore the value of low-confidence pseudo-labels. However, low-confidence pseudo-labels can depict the boundaries of anatomical structures to a certain extent, indicating their importance in the segmentation task. These low-confidence pseudo-labels often contain valuable boundary information, which can help the model better understand complex anatomical structures.

[0004] In existing research on enhancing the utilization rate of pseudo-labels, most methods focus on optimizing the distribution of pseudo-labels by predicting confidence or measuring uncertainty. However, these methods mainly operate only at the feature level and fail to effectively increase the quantity or quality of pseudo-labels, resulting in unsatisfactory segmentation results.

[0005] That is to say, in the existing technologies for medical image segmentation, there are at least the following disadvantages: 1) Insufficient utilization of low-confidence pseudo-labels, and the overall data cannot be fully utilized; 2) Lack of optimization of the usage method of pseudo-labels, especially the lack of optimization of the quantity and quality of pseudo-labels; 3) The effect of segmented images is not good. Summary of the Invention

[0006] To solve the above three problems, the purpose of the present invention is to provide a semi-supervised medical image automatic segmentation method based on pseudo-label optimization. Through the superpixel segmentation and information entropy voting mechanism of the pseudo-label refinement module, the pseudo-label boundary is dynamically adjusted, significantly reducing boundary noise and increasing the number of pseudo-labels, improving the spatial consistency and accuracy of the pseudo-labels; also in the triplet loss module, through hierarchical pseudo-labels and feature extraction, the model's ability to capture boundary information is optimized, enabling the model to more accurately segment complex anatomical structures; also in the mutual correction framework, through the output differences of two parallel sub-networks, the errors in the pseudo-labels are iteratively corrected, significantly improving the robustness of the pseudo-labels and the segmentation accuracy of the model; in addition, through the synergistic effect of pseudo-label refinement, triplet loss calculation and mutual correction framework in the overall architecture, the performance of semi-supervised medical image segmentation is significantly improved, making full use of the overall data and performing effective optimization.

[0007] It is achieved through the following technical solutions: A semi-supervised medical image automatic segmentation method based on pseudo-label optimization, using a deep network model for segmentation, includes the following steps: Step 1: Preprocess the dataset of medical images. The preprocessing includes at least image normalization and data augmentation; generate pseudo-labels for the preprocessed dataset and divide the dataset into a training set, a test set, and a validation set used by the deep network model; Step 2: Construct a semi-supervised training framework for the training set. The semi-supervised training framework includes a pseudo-label refinement module, a triplet loss module, and a mutual correction framework. The pseudo-label refinement module and the triplet loss module are two parallel sub-networks; Step 2.1: In the pseudo-label refinement module, refine the training set in Step 1 through the SLIC superpixel segmentation method and the information entropy voting mechanism, and dynamically adjust the pseudo-label boundary; Step 2.2: In the triplet loss module, determine the anchor sample, positive sample, and negative sample of the triplet loss function from the training set in Step 1, construct the triplet loss function, and control the minimization of the triplet loss function; Step 2.3: In the mutual correction framework, based on the output differences of the two parallel sub-networks, construct a mutual correction loss function and control the minimization of the mutual correction loss function; Step 3: After the semi-supervised training framework in Step 2 completes the training of the training set, select multiple labeled data from the test set for supervised testing in the deep network model, and construct a supervised loss function based on the results of the supervised testing; For the triplet loss function, the mutual correction loss function, and the supervised loss function, set different weights and construct a total loss function, and use the gradient descent method to update the model parameters and adjust each weight to control the convergence of the total loss function; Step 4. After the total damage function converges in Step 3, use the deep network model to segment the validation set, and evaluate the segmentation results based on two metrics: the Dice coefficient and the average surface distance (ASD).

[0008] Through the semi-supervised training framework, the model can be effectively trained based on the training set, thereby optimizing the number of pseudo-labels and the judgment of the boundaries, and obtaining an efficient and highly accurate deep network model. Additionally, multiple functions are constructed respectively, and different weights are set for updating, which can comprehensively consider the overall data. Finally, by combining the two metrics of the Dice coefficient and the average surface distance (ASD) for evaluation, the effect of image segmentation can be effectively judged.

[0009] Preferably, the data augmentation in Step 1 includes random flipping, rotation, adding Gaussian noise, and blurring. Through operations such as random flipping and rotation, the data can be effectively optimized, thereby improving the accuracy and effectiveness of subsequent calculations.

[0010] Preferably, when dividing the data in Step 1, the data volume in the training set is denoted as E, the data volume in the test set is denoted as R, and the data volume in the validation set is denoted as T. Then E > 1.4R > 4T ≥ 0. Strictly restricting the quantity ratio of the training set, test set, and validation set can effectively improve the training results of the model.

[0011] Preferably, the refinement process in Step 2.1 specifically includes the following steps: I) First, generate multiple superpixel region sets through the SLIC superpixel segmentation method, calculate the entropy value of each superpixel region, compare each entropy value with a preset threshold, and adjust the boundaries of the superpixel regions corresponding to each entropy value that exceeds the preset threshold; II) Then, use the information entropy voting mechanism to optimize the consistency of the pseudo-labels of any adjacent superpixel regions, dynamically adjust the pseudo-label boundaries of each superpixel region, and after the adjustment is completed, return to Step I) for iterative loop until the upper limit of the number of iterations is reached or the adjustment convergence index is satisfied; where the adjustment convergence index is used to represent the ratio of the superpixel blocks that need to be adjusted to the superpixel blocks that do not need to be adjusted. The refinement process can effectively increase the number of pseudo-labels and reduce boundary noise.

[0012] Preferably, the specific method for constructing the triplet loss function in Step 2.2 is as follows: First, divide the pseudo-labels in the training set into multiple confidence levels through hierarchical processing; then, use the feature extractor to calculate the cosine similarity between the corresponding boundary pixels and internal pixels of different pseudo-labels divided into multiple confidence levels respectively; based on multiple confidence levels and multiple cosine similarities, determine the anchor sample, positive sample, and negative sample of the triplet loss function, and construct the triplet loss function. By constraining the distances between the anchor, positive, and negative samples, the separation of the feature space can be driven, thereby improving the model's ability to distinguish complex boundaries.

[0013] Preferably, in step 2.2, the expression of the triplet loss function is as follows: L triplet = max(0, d(a, p) - d(a, n) + α); where a is the anchor sample, p is the positive sample, n is the negative sample, d represents the distance metric in the feature space, and α is a preset margin parameter, and α is used to control the distance difference between the positive and negative samples.

[0014] Preferably, in step 2.3, the expression of the mutual correction loss function is: ; where P1(x i ) and P2(x i ) respectively represent the predicted outputs of two parallel sub-networks for the input x i of the training set; represents the L2 norm in the form of, which is used to measure the consistency of the prediction results of the two sub-networks; Consistency(P1, P2) represents the spatial consistency constraint term of the prediction results of the two sub-networks; λ is a balance weight parameter, which is used to adjust the strength of the consistency constraint.

[0015] Preferably, when adjusting each weight in step 3, the weights of the triplet loss function and the mutual correction loss function are updated according to the exponentially increasing strategy. When the model parameters have not yet converged, the generated pseudo-labels have a large amount of noise. If the weights of the triplet loss and the mutual correction loss are too high at this time, it is equivalent to the weight of the semi-supervised loss being too high, and the model is easily misled by the wrong labels, resulting in unstable training; while in the later stage of training, as the model gradually stabilizes, the quality of the pseudo-labels improves, and it is necessary to increase the weight of the semi-supervised loss to fully exploit the value of the unlabeled data. This exponentially increasing strategy naturally matches this progressive learning process through the design of low initial weights and exponential growth in the later stage.

[0016] Preferably, in step 3, the expression of the supervised damage function is: ; where M represents the number of labeled samples, c represents the number of classes, y i，c is the one-hot encoding of the true label in the labeled data, and P i，c is the probability distribution predicted by the model.

[0017] Preferably, the expression of the total loss function in step 3 is: ; where μ, β, and γ are the weight coefficients corresponding to the triplet loss function, the mutual correction loss function, and the supervised loss function respectively.

[0018] Compared with the prior art, the significant advantages of the present invention are: Through the superpixel segmentation and information entropy voting mechanism of the pseudo-label refinement module, the present invention dynamically adjusts the pseudo-label boundaries, significantly reducing boundary noise and increasing the number of pseudo-labels, enhancing the spatial consistency and accuracy of the pseudo-labels; in the triplet loss module, through hierarchical pseudo-labels and feature extraction, it optimizes the model's ability to capture boundary information, enabling the model to more accurately segment complex anatomical structures; in the mutual correction framework, through the output differences of two parallel sub-networks, it iteratively corrects the errors in the pseudo-labels, significantly enhancing the robustness of the pseudo-labels and the segmentation accuracy of the model; in addition, through the synergistic effect of pseudo-label refinement, triplet loss calculation, and mutual correction framework, the overall architecture significantly improves the performance of semi-supervised medical image segmentation, making full use of the overall data and performing effective optimization. Brief Description of the Drawings

[0019] Figure 1 is a flowchart of a semi-supervised medical image automatic segmentation method based on pseudo-label optimization; Figure 2 is an overall process and internal detail display of a semi-supervised medical image automatic segmentation method based on pseudo-label optimization; Figure 3 is a comparison chart of the actual test data of the public dataset Synapse with other semi-supervised methods using the method of the present application; Figure 4 is a comparison chart of the actual test data of the public dataset AMOS with other semi-supervised methods using the method of the present application. Detailed Embodiment

[0020] Next, the technical solutions in the embodiments of the present invention will be described in detail in conjunction with the attached Figures 1 to 4 drawings in the embodiments of the present invention.

[0021] As Figure 1 shown, it is a flowchart of a semi-supervised medical image automatic segmentation method based on pseudo-label optimization; as Figure 2 shown, it is an overall process and internal detail display of a semi-supervised medical image automatic segmentation method based on pseudo-label optimization; in combination with Figure 1 and Figure 2 shown, when the present invention uses a deep network model for segmentation, the total medical image dataset contains a small amount of pre-annotated real data and a large amount of unannotated data. The semi-supervised loss is calculated using a semi-supervised framework. After training is completed, the test set is used for testing and optimization is carried out by setting the weights of the total loss function. Finally, a small amount of validation set is used for validation and a comprehensive effect evaluation is carried out while taking into account both the overall and local aspects. The specific steps are as follows: Step 1: Preprocess the medical image dataset. The preprocessing includes at least image normalization and data augmentation. Data augmentation includes operations such as random flipping, rotation, adding Gaussian noise, and blurring, which can optimize the data in the dataset and improve the accuracy and effectiveness of subsequent calculations. Then, generate pseudo-labels for the preprocessed dataset and divide the dataset into a training set, a test set, and a validation set for use in the deep network model.

[0022] In this embodiment, when dividing the data, the data volume in the training set is denoted as E, the data volume in the test set is denoted as R, and the data volume in the validation set is denoted as T. Then E > 1.4R > 4T ≥ 0, and E, R, and T are all integers. Strictly restricting the quantity ratio of the training set, the test set, and the validation set can effectively improve the training results of the model. The dataset is usually divided into a training set, a validation set, and a test set according to a ratio of 7:1:2. In the medical image segmentation task of this embodiment, using the dataset ratio method of E > 1.4R > 4T ≥ 0 can give priority to ensuring the reliability of the test set, avoid overfitting misjudgment, and while balancing the scales of the training set and the test set, ensure the flexibility of the validation set. This division method is more suitable for our image segmentation task.

[0023] Step 2: Construct a semi-supervised training framework for the training set. The semi-supervised training framework includes a pseudo-label refinement module, a triplet loss module, and a mutual correction framework. The pseudo-label refinement module and the triplet loss module are two parallel subnets.

[0024] Step 2.1: In the pseudo-label refinement module, refine the training set in Step 1 through the SLIC superpixel segmentation method and the information entropy voting mechanism, and dynamically adjust the pseudo-label boundaries. Specifically, the refinement process includes the following steps: I) First, generate multiple superpixel region sets through the SLIC superpixel segmentation method, calculate the entropy value of each superpixel region, compare each entropy value with a preset threshold, and adjust the boundaries of the superpixel regions corresponding to each entropy value that exceeds the preset threshold; II) Then, use the information entropy voting mechanism to optimize the consistency of the pseudo-labels of any adjacent superpixel regions, dynamically adjust the pseudo-label boundaries of each superpixel region, and after the adjustment is completed, return to Step I) for iterative loop until the upper limit of the iteration times is reached or the adjustment convergence index is satisfied; where the adjustment convergence index is used to characterize the ratio of the superpixel blocks that need to be adjusted to the superpixel blocks that do not need to be adjusted. The refinement process can effectively increase the number of pseudo-labels and reduce boundary noise.

[0025] This dynamically adjusted and iteratively cycled method, based on the ability to identify high-confidence regions, can also perform an exclusive OR operation according to the result differences in each identification, conduct Pair combination, and then incorporate some easily overlooked low-confidence regions into the calculation, thereby reducing boundary noise and increasing the number of pseudo-labels.

[0026] It should be noted that the preset threshold is set by the user in advance. For example, a threshold can be set in the range of 0.6 - 0.8 in advance, and then adjusted according to the actual type of the image. For high-contrast organs such as the liver, the threshold is increased, and for low-contrast organs such as the pancreas, the threshold is decreased.

[0027] Step 2.2: In the triplet loss module, determine the anchor sample, positive sample, and negative sample of the triplet loss function from the training set in Step 1, construct the triplet loss function, and control the minimization of the triplet loss function.

[0028] In this embodiment, the specific method for constructing the triplet loss function is as follows: First, divide the pseudo-labels in the training set into multiple confidence levels through hierarchical processing; then, through the feature extractor, calculate the cosine similarity between the corresponding boundary pixels and internal pixels of different pseudo-labels divided into multiple confidence levels; based on multiple confidence levels and multiple cosine similarities, determine the anchor sample, positive sample, and negative sample of the triplet loss function, and construct the triplet loss function. Through the distance constraints of the anchor, positive, and negative samples, drive the separation of the feature space in the Euclidean space, thereby improving the model's ability to distinguish complex boundaries.

[0029] In this embodiment, in Step 2.2, the expression of the triplet loss function is: L triplet = max(0, d(a, p) - d(a, n) + α); where a is the anchor sample, p is the positive sample, n is the negative sample, the operation form of d() represents the distance metric in the feature space, and α is the preset boundary interval parameter, and α is used to control the distance difference between the positive and negative samples.

[0030] Step 2.3: In the mutual correction framework, construct a mutual correction loss function based on the output differences of two parallel subnets, and control the minimization of the mutual correction loss function.

[0031] In this embodiment, in Step 2.3, the expression of the mutual correction loss function is: ; where P1(x i ) and P2(x i ) respectively represent the predicted outputs of two parallel sub-networks for the input x i of the training set; The L2 norm is represented in the form of, which is used to measure the consistency of the prediction results of two sub-networks; Consistency(P1, P2) represents the spatial consistency constraint term of the prediction results of two sub-networks; λ is the balance weight parameter, which is used to adjust the strength of the consistency constraint.

[0032] Step 3: After the semi-supervised training framework in Step 2 completes the training of the training set, select multiple labeled data from the test set for supervised testing in the deep network model. Based on the results of the supervised testing, use the true value of the labeled data as the comparison boundary standard Lsup, so as to compare the test results of the supervised testing with the true value and construct a supervised loss function.

[0033] Then, for the triplet loss function, the mutual correction loss function, and the supervised loss function, set different weights and construct the total loss function. Use the gradient descent method to update the model parameters, that is, update the sizes of different weights, and control the convergence of the total loss function by adjusting each weight.

[0034] In this embodiment, when adjusting each weight in Step 3, the weights of the triplet loss function and the mutual correction loss function are updated according to the exponential increase strategy. When the model parameters have not converged, the generated pseudo-labels have a large amount of noise. If the weights of the triplet loss and the mutual correction loss are too high at this time, it is equivalent to that the weight of the semi-supervised loss is too high, and the model is easily misled by the wrong labels, resulting in unstable training; while in the later stage of training, as the model gradually stabilizes and the quality of the pseudo-labels improves, it is necessary to increase the weight of the semi-supervised loss to fully exploit the value of the unlabeled data. This exponential increase strategy naturally matches this progressive learning process through the design of low initial weights and exponential growth in the later stage.

[0035] In this embodiment, in Step 3, the expression of the supervised damage function is: ; where M represents the number of labeled samples, c represents the number of categories, y i，c is the one-hot encoding of the true label in the labeled data, and P i，c is the probability distribution predicted by the model. One-hot encoding is a method of converting categorical data into a numerical format that can be processed by machine learning algorithms.

[0036] In this embodiment, the expression of the total loss function in Step 3 is: ; where μ, β, and γ are the weight coefficients corresponding to the triplet loss function, the mutual correction loss function, and the supervised loss function respectively. After the triplet loss function and the mutual correction loss function are weighted and added, they form what is equivalent to the semi-supervised loss function. Therefore, the total loss function can also be said to be composed of the semi-supervised damage function and the supervised damage function.

[0037] Step 4. After the total damage function converges in Step 3, use the deep network model to segment the validation set, and evaluate the segmentation results based on two metrics: the Dice coefficient and the average surface distance (ASD).

[0038] It should be noted that the full name of Dice is Dice Similarity Coefficient. The Dice coefficient is sensitive to the change in the size of the region and can quantify the degree of regional overlap between the predicted segmentation result and the true label, thus reflecting the overall accuracy of the segmentation. It is very suitable for evaluating the image segmentation effect. The full name of ASD is Average Surface Distance, that is, the average surface distance, which can quantify the boundary error and is used to measure the average spatial distance between the predicted segmentation boundary and the true boundary, reflecting whether the edge segmentation is accurate. The specific standard used needs to be converted according to the actual influencing resolution, such as using millimeters or pixels as the unit. Therefore, the Dice coefficient focuses on the overall regional overlap degree, and the ASD average surface distance focuses on the local boundary accuracy. Combining the results of these two metrics can achieve a comprehensive evaluation of the segmentation effect during validation, both globally and locally.

[0039] As Figure 3 shown, it is a comparison chart of the actual test data of the public dataset Synapse and the results of other semi-supervised methods using the method of this application; as Figure 4 shown, it is a comparison chart of the actual test data of the public dataset AMOS and the results of other semi-supervised methods using the method of this application. Combining Figure 3 and Figure 4 in the content, here is a specific application example: Taking abdominal CT images as the input, the datasets used are the Synapse dataset and the AMOS abdominal CT image dataset. The Synapse dataset is a dataset specifically for the abdominal CT image segmentation task, and the AMOS dataset is an abdominal multi-organ segmentation dataset. The Synapse dataset contains 30 CT scans, of which 18 are for training and 12 are for testing, with a total of 3779 slice data, covering 8 abdominal organs, including artery 1, pancreas, gallbladder, stomach, adrenal gland, liver, vein, and artery 2. The AMOS dataset contains 360 CT scans, of which 216 are for training, 24 are for validation, and 120 are for testing, and it contains 13 foreground classes, that is, 13 labeled data. Data augmentation operations are performed on the images, including random flipping, rotation, adding Gaussian noise, and blurring, to improve the generalization ability of the model.

[0040] Using the aforementioned semi-supervised structure for model training, the pseudo-label refinement module generates a set of superpixel regions through the SLIC algorithm, calculates the entropy value of each superpixel region, and dynamically adjusts the pseudo-label boundary according to a preset threshold; the triplet loss module calculates the cosine similarity between boundary pixels and internal pixels through hierarchical pseudo-labels and a feature extractor to optimize the model's ability to capture boundary information; the mutual correction framework uses the output differences of two subnets, Subnet A and Subnet B, to iteratively correct errors in the pseudo-labels and improve the robustness and segmentation accuracy of the pseudo-labels.

[0041] Then, based on the triplet loss, mutual correction loss, and supervised loss, a total loss function is constructed, and the model parameters are optimized by the gradient descent method. During the training process, an exponential increase strategy is adopted to update the coefficients of the triplet loss and mutual correction loss to balance the contributions of each loss until the model converges.

[0042] After repeating the loop iteration, the Dice coefficient and ASD average surface distance are used as indicators to evaluate the quality of the model. After the indicators of the Dice coefficient and ASD average surface distance meet the user's requirements, the model is used for verification testing. The actual effect diagram after testing and the results of several existing methods, UA-MT, URPC, CPS, SS-Net, DST, DePL, Adsh, CReST, SimiS, Basak et al., CLD, DHC, are compared. The comparison results are referred to Figure 3 and Figure 4 . For Figure 3 when segmenting arteries with severe artifacts, unclear boundaries, and low contrast in Figure 4 , this method completely retains small arteries and is close to the true result, while CPS shows more vascular segmentations. And in

[0043] the segmentation of adrenal CT, this method completely segments the tissue, while the URPC segmentation method has obvious displacement.

[0044] In summary, through the superpixel segmentation and information entropy voting mechanism of the pseudo-label refinement module, the present invention dynamically adjusts the boundaries of pseudo-labels, significantly reducing boundary noise and increasing the number of pseudo-labels, enhancing the spatial consistency and accuracy of pseudo-labels; also in the triplet loss module, through hierarchical pseudo-labels and feature extraction, it optimizes the model's ability to capture boundary information, enabling the model to more accurately segment complex anatomical structures; and in the mutual correction framework, through the output differences of two parallel sub-networks, it iteratively corrects errors in pseudo-labels, significantly enhancing the robustness of pseudo-labels and the segmentation accuracy of the model; in addition, through the synergistic effect of pseudo-label refinement, triplet loss calculation, and the mutual correction framework, the overall architecture significantly improves the performance of semi-supervised medical image segmentation, fully utilizes the overall data and conducts effective optimization, showing significant progressiveness.

[0045] The above embodiments are only used to illustrate the technical idea of the present invention, and the protection scope of the present invention cannot be limited thereby. Any changes made on the basis of the technical solution according to the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A semi-supervised medical image automatic segmentation method based on pseudo-label optimization, using a deep network model for segmentation, characterized in that: The following steps are involved: Step 1: Preprocess the medical image data set, which includes at least image normalization and data enhancement; Generate pseudo labels for the preprocessed dataset and divide the dataset into training set, test set and validation set used by the deep network model; Step 2: construct a semi-supervised training framework for the training set, the semi-supervised training framework includes a pseudo-label refinement module, a triplet loss module and a mutual correction framework, the pseudo-label refinement module and the triplet loss module are two parallel subnetworks; Step 2.1: In the pseudo-label refinement module, the training set in step 1 is refined by the SLIC superpixel segmentation method and the information entropy voting mechanism, and the pseudo-label boundary is dynamically adjusted; Step 2.2, in the triplet loss module, determine the anchor samples, positive samples and negative samples of the triplet loss function from the training set in step 1, construct the triplet loss function, and control the triplet loss function to be minimized; Step 2.3, in the mutual correction framework, based on the output difference of the two parallel sub-networks, a mutual correction loss function is constructed, and the mutual correction loss function is minimized; Step 3: After the semi-supervised training framework in step 2 completes the training of the training set, multiple labeled data are selected from the test set to perform supervised testing in the deep network model, and a supervised loss function is constructed based on the results of the supervised testing; For the triplet loss function, mutual correction loss function and supervision loss function, different weights are set and the total loss function is constructed. The model parameters are updated using the gradient descent method, and each weight is adjusted to control the convergence of the total loss function. Step 4: After the total damage function converges in step 3, the deep network model is used to segment the validation set, and the segmentation results are evaluated based on the two indicators of Dice coefficient and ASD average surface distance.

2. According to claim 1, a semi-supervised medical image automatic segmentation method based on pseudo-label optimization is characterized in that: The data augmentation in step 1 includes random flipping, rotation, adding Gaussian noise, and blurring.

3. The method for automatic semi-supervised medical image segmentation based on pseudo-label optimization according to claim 1, characterized in that: When dividing the data in step 1, the amount of data in the training set is recorded as E, the amount of data in the test set is recorded as R, and the amount of data in the validation set is recorded as T, then E>1.4R>4T≥0.

4. The method for automatic semi-supervised medical image segmentation based on pseudo-label optimization according to claim 1, characterized in that: The refining process in step 2.1 specifically includes the following steps: I) First, multiple superpixel region sets are generated by the SLIC superpixel segmentation method, the entropy value of each superpixel region is calculated, each entropy value is compared with a preset threshold, and the boundary of the superpixel region corresponding to each entropy value exceeding the preset threshold is adjusted; II) Then use the information entropy voting mechanism to optimize the consistency of the pseudo-labels of any adjacent superpixel areas, dynamically adjust the pseudo-label boundaries of each superpixel area, and after the adjustment is completed, return to step I) to iterate until the upper limit of the number of iterations is reached or the adjustment convergence index is met; among which, the adjustment convergence index is used to characterize the ratio of superpixel blocks that need to be adjusted to superpixel blocks that do not need to be adjusted.

5. The method for automatic semi-supervised medical image segmentation based on pseudo-label optimization according to claim 1, characterized in that: The specific method for constructing the triplet loss function in step 2.2 is: First, the pseudo-labels in the training set are divided into multiple confidence levels through hierarchical processing. Then, the cosine similarities between the corresponding boundary pixels and internal pixels of different pseudo-labels divided into multiple confidence levels are calculated through the feature extractor. Based on multiple confidence levels and multiple cosine similarities, the anchor samples, positive samples and negative samples of the triplet loss function are determined, and the triplet loss function is constructed.

6. The method for automatic semi-supervised medical image segmentation based on pseudo-label optimization according to claim 1, characterized in that: In step 2.2, the triple loss function is expressed as: Ltriplet=max(0, d(a, p)-d(a, n)+α); where a is the anchor sample, p is the positive sample, n is the negative sample, d represents the distance metric in the feature space, and α is the preset boundary interval parameter, which is used to control the distance difference between positive and negative samples.

7. The method for automatic semi-supervised medical image segmentation based on pseudo-label optimization according to claim 6, characterized in that: In step 2.3, the expression of the mutual correction loss function is: ; Where P1(xi) and P2(xi) represent the predicted outputs of two parallel sub-networks for the input xi of the training set; The form of is used to represent the L2 norm, which is used to measure the consistency of the prediction results of the two sub-networks; Consistency (P1, P2) represents the spatial consistency constraint term of the prediction results of the two sub-networks; λ is the balancing weight parameter, which is used to adjust the strength of the consistency constraint.

8. The method for automatic semi-supervised medical image segmentation based on pseudo-label optimization according to claim 1, characterized in that: When adjusting each weight in step 3, the weights of the triplet loss function and the mutual correction loss function are updated according to the exponential increase strategy.

9. The method for semi-supervised automatic segmentation of medical images based on pseudo-label optimization according to claim 7, characterized in that: In step 3, the expression of the supervised damage function is: ; Where M represents the number of labeled samples, c represents the number of categories, yi,c is the one-hot encoding of the true label in the labeled data, and Pi,c is the probability distribution predicted by the model.

10. The method for semi-supervised automatic segmentation of medical images based on pseudo-label optimization according to claim 9, characterized in that: The expression of the total loss function in step 3 is: ; Among them, μ, β, and γ are the weight coefficients corresponding to the triplet loss function, mutual correction loss function, and supervision loss function, respectively.

Images