Semantic transferable medical image segmentation method based on semi-supervised learning

Abstract

The invention discloses a semantic transferable medical image segmentation method based on semi-supervised learning. The semantic transferable medical image segmentation method comprises the following steps: step 1, constructing a labeled source domain data set and a non-labeled target domain data set; step 2, according to the data set, pre-training a student network and obtaining a source domain data representation and a target domain data representation with a pseudo tag; step 3, constructing sampling distribution according to the statistics of the two representations, then performing TSA on the source domain data representation, and constructing and optimizing a segmentation loss function; inputting the two representations into a GraphCL framework to obtain a loss function; 4, constructing a model overall loss function to optimize parameters of the student network after pre-training; the parameters of the teacher network are updated by using EMA according to the optimized parameters of the student network; and a trained student network is obtained through multiple iterations, and medical image segmentation is realized. According to the method, the problems of domain offset of anatomical structure characterization and excessive dependence on source domain data are solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image segmentation, and particularly relates to a medical image segmentation method based on semi-supervised learning and semantic transferability. BACKGROUND

[0002] Medical image segmentation plays a key role in modern medicine, enabling precise diagnosis. Although deep learning technology has completely revolutionized this field, its success is highly dependent on large-scale labeled data sets, and in medical images, expert labeling is often costly and time-consuming, which poses a major challenge. Semi-supervised learning methods, as a promising solution, can help improve the generalization ability of the model by utilizing limited labeled data and a large amount of unlabeled data, and fully exploiting the structure and context information in the unlabeled data, thereby alleviating this challenge.

[0003] In recent years, semi-supervised medical image segmentation (SSMIS) has made much progress, exploring various strategies, including consistency regularization, pseudo-labeling, and adversarial training. These methods encourage the model to learn robust feature representations from unlabeled data while maintaining good performance with limited labeled samples. The literature “Wenjia Bai, Ozan Oktay, Matthew Sinclair, Hideaki Suzuki, Martin Rajchl, Giacomo Tarroni, Ben Glocker, Andrew King, Paul M Matthews, and Daniel Rueckert. Semi-supervised learning for network-based cardiac mr image segmentation. In International Conference on Medical Image Computing and Computer-assisted Intervention, pages 253–260, 2017.” proposes a self-training method for cardiac magnetic resonance image segmentation, proposing the SemiFCN model, which jointly utilizes labeled and unlabeled data for training. The literature “Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation. In International Conference on Medical Image Computing and Computer-assisted Intervention, pages 605–613, 2019.” proposes an uncertainty-guided Mean Teacher framework combined with a transformation consistency strategy, proposing the UA-MT model to improve model performance.

[0004] However, current semi-supervised learning methods face two fundamental limitations: one is the domain shift problem of anatomical structure representation, and the other is the excessive reliance on source domain labeled data for feature adaptation, both of which limit their practical application in clinical practice. These challenges are particularly prominent in medical images, as differences in scanning protocols, patient populations, and anatomical structures can lead to significant inconsistencies in the distribution between labeled and unlabeled data.

[0005] The first limitation is the domain shift of anatomical structure representation, which is manifested in that a model trained on data from a certain clinical institution or a specific type of scanner is difficult to generalize to a new environment. For example, a segmentation model trained on enhanced CT images from hospital A may not perform well on non-enhanced CT images from hospital B due to differences in image contrast and noise characteristics.

[0006] The second limitation is due to the excessive dependence on source domain labeled data in feature learning. Most current semi-supervised learning methods rely on generating pseudo labels or imposing consistency constraints on the model trained from limited labeled data. This approach is prone to a "self-reinforcing cycle", in which the model's bias towards a small-scale labeled set is further transmitted to unlabeled data. In medical images, this bias is particularly severe in scenarios with extreme class imbalance (e.g., small lesions relative to large background regions), often leading to systematic under-segmentation of rare but clinically critical target regions. SUMMARY

[0007] To overcome the deficiencies of the prior art, the technical problem to be solved by the present application is to provide a semi-supervised learning-based semantic transferable medical image segmentation method.

[0008] The technical solution of the present application to solve the technical problem is to provide a semi-supervised learning-based semantic transferable medical image segmentation method, characterized in that the method comprises the following steps:

[0009] Step 1, obtaining labeled source domain data and unlabeled target domain data where x i s represents the i-th source domain data, x i t represents the i-th target domain data, y i s represents the label class of the i-th source domain data, N s represents the number of source domain data, N t represents the number of target domain data.

[0010] Step 2, using the source domain data D s of step 1 to pre-train the student network, obtaining the pre-trained student network f stu (·); then input the target domain data D t of step 1 into the pre-trained student network f stu (·) to make a prediction, obtaining pseudo labels; input the source domain data D s into the pre-trained student network f stu (·), obtaining source domain data representation; input the target domain data into the teacher network ftea In the (·), the target domain data representation is obtained; and the pseudo label is corresponded to the target domain data representation to form the target domain data representation with the pseudo label;

[0011] Step 3, the statistics of the source domain data representation obtained in step 2 and the target domain data representation with the pseudo label are calculated respectively, then the sampling distribution is constructed according to the statistics of the source domain data representation and the target domain data representation with the pseudo label, and then the TSA is performed on the source domain data representation in step 2 based on the sampling distribution, to obtain the augmented feature set Then, the augmented feature set is used to construct the segmentation loss function The segmentation loss function is analyzed and optimized to obtain the optimized segmentation loss function

[0012] Meanwhile, the source domain data representation obtained in step 2 and the target domain data representation with the pseudo label are input into the GraphCL framework to obtain the loss function under the framework

[0013] Step 4, based on the loss function and the optimized segmentation loss function the model overall loss function is obtained Then, the model overall loss function is used to optimize the parameters of the pre-trained student network f stu in step 2; meanwhile, the parameters of the teacher network f tea in step 2 are updated in the form of exponential moving average according to the parameters of the optimized student network; after multiple iterations, the trained student network and the trained teacher network are obtained;

[0014] Step 5, the test data is input into the trained student network to realize medical image segmentation.

[0015] Compared with the prior art, the beneficial effects of the present application are as follows:

[0016] (1) The present application solves the problems of domain shift and excessive dependence on source domain data of anatomical structure representation through the teacher-student framework and semantic transferable augmentation, and promotes robust feature learning in various clinical fields and imaging modalities.

[0017] (2) The present application designs a cross-domain feature alignment mechanism, adopts a teacher-student framework, and the student network is responsible for processing the labeled data of the source domain, extracting domain-independent general features, and explicitly modeling the cross-domain difference through the statistics of specific label categories ; meanwhile, the teacher network models the cross-domain difference through the feature statistics Perform EMA updates to gradually adapt to target domain characteristics.

[0018] (3) The application proposes a semantic transferable augmented TSA, which transforms features based on anatomical feature perception, specifically by sampling from a constructed fine multivariate normal distribution Where the cross-domain mean difference is Δμ c Captures the systematic domain shift between the source domain and the target domain, and the target domain data covariance Indicates the distribution change within the class, and preserves the anatomical structure relationship under the labeled class condition. It can be used as a plug-in module and integrated into other existing semi-supervised medical image segmentation methods.

[0019] (4) The application reduces the dependence on high-cost labeled data, and shows significant potential in improving the accessibility and clinical scalability of medical image segmentation.

[0020] (5) The application has the ability to generalize across multiple imaging modalities and anatomical structures, which helps to improve the accuracy of diagnosis and treatment planning, and is particularly suitable for resource-poor medical environments. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is the overall flowchart of the application;

[0022] Figure 2 is the cardiac structure segmentation effect comparison chart of the application example 1 on ACDC with only 5% data labeling;

[0023] Figure 3 is the cardiac structure segmentation effect comparison chart of the application example 1 on ACDC with only 10% data labeling;

[0024] Figure 4 is the pancreatic segmentation effect comparison chart of the application example 1 on Pancreas-NIH dataset;

[0025] Figure 5 is the feature embedding result chart of the application example 1 using t-SNE visualization;

[0026] Figure 6 is the segmentation loss function of the application example 1 without using optimization is the feature embedding result chart of the application example 1 using t-SNE visualization;

[0027] Figure 7 is the parameter sensitivity analysis result chart of the application example 1 with only 5% data labeling on ACDC with weight hyperparameter β;

[0028] Figure 8The parameter sensitivity analysis result figure of the weight hyperparameter β of embodiment 1 of the present application under 10% data labeling on ACDC. DETAILED DESCRIPTION

[0029] The specific embodiments of the present application are given below. The specific embodiments are only used to further illustrate the present application and do not limit the protection scope of the present application.

[0030] The present application provides a medical image segmentation method based on semi-supervised learning (referred to as method for short), characterized in that the method comprises the following steps:

[0031] Step 1, obtaining a small amount of labeled source domain data (i.e. source domain medical image data) based on medical equipment and a large amount of unlabeled target domain data (i.e. target domain medical image data) for model training; wherein x i s represents the i th source domain data, x i t represents the i th target domain data, y i s represents the label category of the i th source domain data, N s represents the number of source domain data, N t represents the number of target domain data.

[0032] Preferably, in step 1, the medical equipment is a CT scanner, a magnetic resonance imaging system and an ultrasonic diagnostic instrument, and the corresponding data are CT image data, nuclear magnetic resonance image data and ultrasonic image data respectively.

[0033] Preferably, in step 1, the number ratio of source domain data to target domain data is 1-4:16-19, i.e. N s :N t =1-4:16-19.

[0034] Step 2, pre-training the student network using the source domain data D s of step 1 to obtain the pre-trained student network f stu (·); then input the target domain data D t of step 1 into the pre-trained student network f stu (·) to make a prediction and obtain pseudo labels; input the source domain data D s into the pre-trained student network f stu (·) to obtain source domain data representation; input the target domain data into the teacher network f tea (·) to obtain target domain data representation; and then one-to-one correspondence between the pseudo labels and the target domain data representation forms the target domain data representation with pseudo labels;

[0035] Step 3, respectively calculate the statistics of the source domain data representation and the target domain data representation with pseudo labels obtained in step 2, then construct a sampling distribution according to the statistics of the source domain data representation and the target domain data representation with pseudo labels, and then perform TSA (Transferable Semantic Augmentation) on the source domain data representation in step 2 based on the sampling distribution to obtain an augmented feature set Then, according to the augmented feature set , a segmentation loss function is constructed Then analyze and optimize the segmentation loss function Improve the model training performance to obtain an optimized segmentation loss function

[0036] At the same time, input the source domain data representation and the target domain data representation with pseudo labels obtained in step 2 into the GraphCL framework to obtain the loss function under the framework

[0037] Preferably, in step 3, the statistics are calculated as follows:

[0038] For each label category c, the statistics of the source domain data representation are calculated based on the source domain data representation with labels, including the source domain data mean And the source domain data covariance

[0039]

[0040] In formula (1), represents the source domain data belonging to the source domain data set with label category c, x s represents the sample in the source domain data set with label category c, and Cov(·) represents the calculation of covariance.

[0041] For each label category c, the statistics of the target domain data representation with pseudo labels are calculated based on the target domain data representation with pseudo labels, including the target domain data mean And the target domain data covariance

[0042]

[0043] In formula (2), represents the target domain pixel point belonging to the target domain data set with label category c, x t represents the sample in the target domain data set with label category c, and Cov(·) represents the calculation of covariance.

[0044] Preferably, in step 3, the sampling distribution is constructed specifically as:

[0045] According to the difference Δμ of the mean of the target domain data and the mean of the source domain data c and the covariance of the target domain data The normal distribution conditioned on the label category c is constructed, defined as:

[0046]

[0047] In formula (3), the cross-domain mean difference captures the systematic bias between the source domain and the target domain, and the target domain data covariance represents the distribution change within the class.

[0048] Preferably, in step 2, the TSA is specifically:

[0049] In order to obtain more direct and reasonable structure of explicit augmentation effect, the present application further reconstructs the feature augmentation process, which maintains the consistency of anatomical structure while having good computational efficiency. The source domain pixel-level representation belonging to the label category c represents the pixel-level representation of the i-th source domain data belonging to the label category c, which is augmented M times while keeping the original label unchanged, thereby generating the augmented feature set

[0050]

[0051] In formula (4), δ m represents the perturbation vector sampled from the sampling distribution in the m-th augmentation process, y s,i represents the label category of the i-th source domain data; n s represents the total number of pixels of the source domain data.

[0052] Preferably, in step 3, the segmentation loss function is constructed as:

[0053]

[0054] In formula (5), n s represents the total number of pixels of the source domain data; M represents the number of augmentations; ω c ∈R d ,b c ∈R d respectively represent the classifier weight vector and the bias term of the label category c; respectively represent the label category y s,i ​The classifier weight vector and bias term; exp(·) represents taking the exponential function; C represents the number of label categories.

[0055] Preferably, in step 3, the segmentation loss function is analyzed and optimized. The specific steps are as follows:

[0056] Step 3.1: Analyze the segmentation loss function Traditional explicit augmentation methods typically require multiple perturbation samplings of each pixel feature to achieve stable convergence, but this incurs high memory overhead. Unlike these strategies, this invention treats the augmentation process as a continuous transformation space, where each pixel feature is enriched through an implicit distribution. This allows the network to effectively approximate an augmented view of an infinite set. As the augmentation iteration M→∞, the augmented feature set... According to the asymptotic analysis of the distribution, based on the law of large numbers, the empirical risk almost converges to its mathematical expectation, thus yielding the segmentation loss function. The upper bound loss function is shown in equation (6):

[0057]

[0058] In equation (6), It is in equation (5) abbreviation, Represent the expected function;

[0059] Step 3.2: Optimize the segmentation loss function Directly optimizing formula (6) is computationally infeasible because it involves expectations for infinitely augmented samples. To simplify the optimization process, this invention applies Jensen's inequality to the concave function log(·). Based on this inequality, the optimized segmentation loss function is derived. The derivation process of the upper bound:

[0060]

[0061]

[0062] In equation (7), This represents the expected function.

[0063] Step 4: Based on the loss function and the optimized segmentation loss function Obtain the overall loss function of the model Then, based on the overall loss function of the model To optimize the pre-trained student network f in step 2 stu The parameters of (·); at the same time, the teacher network f in step 2 teaThe parameters of (·) are updated according to the parameters of the optimized student network using an exponential moving average (EMA) method; after multiple iterations, a trained student network and a trained teacher network are obtained;

[0064] Preferably, in step 4, the model overall loss function is:

[0065]

[0066] In formula (8), β is a weight hyperparameter for controlling the contribution of the optimized segmentation loss function to the model overall loss function .

[0067] Preferably, in step 4, the specific operation of the exponential moving average is:

[0068] θ tea ← λθ tea + (1-λ)θ stu (9)

[0069] In formula (9), θ tea and θ stu are the parameters of the teacher network and the student network respectively, and λ is a weight hyperparameter.

[0070] Step 5, test phase, input test data into the trained student network to realize medical image segmentation.

[0071] Example 1:

[0072] The present application evaluates the performance of the present application and the current most advanced semi-supervised medical segmentation method on three medical image segmentation benchmark datasets. The experimental design aims to test the ability of the present application to fully utilize unlabelled data in the presence of significant domain shifts, covering different image modalities and anatomical structures.

[0073] Table 1

[0074]

[0075] Table 2

[0076]

[0077] Table 1 and Table 2 respectively show the comparison results of the present application and the current most advanced semi-supervised medical image segmentation method on LA(5%) and LA(10%) datasets. On the LA(10%) dataset, only 8 samples (about 10% of the labeled data) are used, and the present application achieves the best performance in Dice coefficient (89.62%), Jaccard index (81.31%) and 95% Hausdorff distance (6.68 voxel), slightly better than the GraphCL benchmark, and significantly superior in performance to methods such as UA-MT and the fully supervised V-Net method. Even on the LA(5%) dataset with lower labeling rate, the present application still shows excellent performance, achieving the best results in Dice (87.65%) and Jaccard (78.24%) indicators, thanks to the TSA proposed by the present application.

[0078] Table 3

[0079]

[0080] Table 4

[0081]

[0082]

[0083] Table 3 and Table 4 respectively show the comparison results of the present application and the current most advanced semi-supervised medical image segmentation method on ACDC(5%) and ACDC(10%) datasets. The present application again demonstrates excellent performance. On the ACDC(10%) dataset, only 7 samples (10% of the labeled data) are used, and the best results are achieved in all evaluation indicators: Dice (89.96%), Jaccard (82.31%), 95HD (1.61 voxel) and ASD (0.64 voxel), and the present application achieves significant performance improvement compared with GraphCL. In Figure 2 and Figure 3 In Table 3 and Table 4, the qualitative segmentation results on the ACDC dataset under the semi-supervised setting of 5% and 10% labeled data are shown. Each row represents a method: BCP (first row), GraphCL (second row) and the present application. The red outline in the figure represents the true label, and the outlines of other colors represent the prediction segmentation results of the corresponding method. From the visualization results, it can be seen that the segmentation results generated by the present application are closer to the true label in various cardiac structures, and are more accurate than BCP and GraphCL. For example, under the labeling settings of 5% and 10%, the present application can more accurately divide the cardiac chambers and boundaries, and produce fewer errors and false segmentations. While BCP and GraphCL often have inconsistencies and segmentation ambiguities in the transition areas of boundaries and structures.

[0084] Table 5

[0085]

[0086] Table 5 shows the comparison results of the present application and the current most advanced semi-supervised medical image segmentation method on the Pancreas-NIH dataset. On Pancreas-NIH, using 12 samples (20% of the labeled data), the present application still maintains the leading position, achieving the best results in Dice (83.06%), Jaccard (71.26%), 95HD (5.67 voxel) and ASD (1.51 voxel) indicators, and comprehensively surpassing all comparison methods. The pancreas segmentation task is more challenging, because the pancreas has large anatomical differences, blurred boundaries and is difficult to identify in CT images. The excellent performance of the present application in boundary-related indicators highlights the effectiveness of the covariance perception enhancement mechanism in maintaining the anatomical consistency of the target domain CT data. In Figure 4 , the qualitative segmentation comparison results on the Pancreas-NIH dataset are shown, comparing the differences between BCP, GraphCL and the present application and the real label. The red contour in the figure is the real label boundary, and the green contour represents the predicted segmentation result. In the first and second columns of the figure, the present application can better depict the fine structure of the pancreas and align with the GT boundary, and is obviously better than BCP and GraphCL in structure detail retention and boundary position. For example, in the first column, BCP and GraphCL have slight under-segmentation of the head of the pancreas, while the present application can effectively alleviate this problem. The third column shows a typical difficult case with irregular structure morphology and low image contrast, and the present application still shows more robust structure boundary recovery capability in this case.

[0087] In summary, the present application exhibits the most advanced segmentation performance on three different datasets, which cover different image modalities (magnetic resonance image data and CT image data), anatomical regions and labeling proportions. The performance improvement on each dataset fully verifies the important contribution of the present application in improving cross-domain migration ability and segmentation accuracy.

[0088] Ablation experiment analysis. In order to quantitatively evaluate the effect of the proposed optimized segmentation loss function module, the present application performs ablation experiments on the ACDC and Pancreas-NIH datasets for different labeled / unlabeled data divisions. As shown in Table 6, when the optimized segmentation loss function is removed, the performance of each indicator decreases, highlighting the important role of this module in improving segmentation quality under limited supervision.

[0089] Table 6 Ablation experiment

[0090]

[0091] From Table 6, it can be found that, with only 5% labeled data in the ACDC dataset, adding the optimized segmentation loss function can improve the Dice coefficient from 87.96% to 88.14%, the Jaccard index from 79.26% to 79.48%, the 95HD from 4.10 to 3.42, and the ASD from 1.28 to 0.97. When the proportion of labeled data increases to 10%, the performance improves more significantly. Similar trends also appear on the Pancreas-NIH dataset. These results collectively verify the effectiveness of the present application.

[0092] In addition, in order to qualitatively evaluate the effect of the optimized segmentation loss function , the present application visualizes the learned feature embedding using t-SNE: when the optimized segmentation loss function is used, it can be observed from Figure 5 that the features form well-separated and compact cluster structures, and there is a clear alignment relationship between the source domain and the target domain; in contrast, when the optimized segmentation loss function is removed, it can be observed from Figure 6 that the features between different domains exhibit obvious overlap and entanglement. These results further verify the important role of the optimized segmentation loss function in reducing the domain difference and improving semantic consistency.

[0093] Sensitivity analysis. The present application performs sensitivity analysis on the weight hyperparameter β of the optimized segmentation loss function on the ACDC dataset using 5% and 10% labeled data, respectively, as shown in Figure 7 and Figure 8 . The experimental results show that when β is in the range of 0.2-1.0, the segmentation indicators perform relatively stably, and the best performance is achieved when β = 0.4. Therefore, in all experiments, the present application sets β to 0.4 to match the optimal result.

[0094] The present application does not cover the prior art.

Claims

1. A method for semantic transferable medical image segmentation based on semi-supervised learning, characterized in that, The method comprises the following steps: Step 1, obtaining labeled source domain data and unlabeled target domain data wherein, x i s represents the i-th source domain data, x i t represents the i-th target domain data, y i s represents the label category of the i-th source domain data, N s represents the number of source domain data, N t represents the number of target domain data; Step 2: Use the source domain data D from Step 1 s The student network is pre-trained to obtain the pre-trained student network f. stu (·); then the target domain data D from step 1 is... t Input into the pre-trained student network f stu Prediction is performed in (·) to obtain pseudo-labels; the source domain data D is then used for prediction. s Input into the pre-trained student network f stu In (·), the source domain data representation is obtained; the target domain data is then input into the teacher network f. tea In (·), the target domain data representation is obtained; then the pseudo-labels are matched one-to-one with the target domain data representation to form the target domain data representation with pseudo-labels. Step 3, respectively calculate the statistics of the source domain data representation and the target domain data representation with pseudo labels obtained in step 2, then construct a sampling distribution according to the statistics of the source domain data representation and the target domain data representation with pseudo labels, and then perform TSA on the source domain data representation in step 2 based on the sampling distribution to obtain an augmented feature set Then, the augmented feature set is used to construct a segmentation loss function Then, the segmentation loss function is analyzed and optimized to obtain an optimized segmentation loss function Meanwhile, the source domain data representation obtained in step 2 and the target domain data representation with pseudo labels are input into the GraphCL framework to obtain a loss function under the framework Step 4, optimizing the segmentation loss function based on the loss function and the optimized segmentation loss function to obtain a model overall loss function According to the model overall loss function , the pre-trained student network f of step 2 is optimized stu ; at the same time, the parameters of the teacher network f of step 2 tea are updated using an exponential moving average method according to the parameters of the optimized student network; after multiple iterations, a trained student network and a trained teacher network are obtained. Step 5, input the test data into the trained student network to realize medical image segmentation.

2. The method of claim 1, wherein the semi-supervised learning-based semantically transferable medical image segmentation method is characterized by, In step 1, the data are CT image data, nuclear magnetic resonance image data and ultrasonic image data. 3.The semi-supervised learning based semantically transferable medical image segmentation method of claim 1, wherein, In step 1, the number ratio of source domain data to target domain data is 1-4:16-19.

4. The semi-supervised learning based semantically transferable medical image segmentation method according to claim 1, characterized in that, In step 3, the statistical quantity is specifically calculated as: For each label class c, the statistics of the source domain data representation are computed based on the labeled source domain data representation, including the source domain data mean and the source domain data covariance In formula (1), denotes the source domain data belongs to the source domain data set of label category c, x s denotes the sample in the source domain data set of the source domain data label category c, Cov(·) denotes the calculation of the covariance; For each label category c, compute the statistics of the target domain data representation with pseudo labels, which are computed by the target domain data representation with pseudo labels, including the target domain data mean and the target domain data covariance In formula (2), x represents the target domain data set to which the target domain pixel point belongs to the label category c t x represents the sample in the target domain data set whose target domain data label category is c, and Cov(·) represents calculating the covariance.

5. The method of claim 1, wherein the semi-supervised learning-based semantically transferable medical image segmentation method is characterized by, In step 3, the sampling distribution is specifically constructed as: According to the difference Δμ between the target domain data mean and the source domain data mean c and the target domain data covariance a normal distribution conditioned on the label class c is constructed, defined as:​ In formula (3), the cross-domain mean difference captures the systematic shift between source and target domains, the target domain data covariance represents the distribution change within the class.

6. The semi-supervised learning based semantically transferable medical image segmentation method according to claim 1, characterized in that, In step 2, the TSA is specifically: pixel-level representation of the source domain belonging to label class c pixel-level representation of the i-th source domain data belonging to label class c, augmented M times, keeping the original label unchanged, thus generating the augmented feature set In formula (4), δ m denotes the sampling distribution the perturbation vector of the sampling, y s,i denotes the label category of the i-th source domain data; n s denotes the total number of pixel points of the source domain data.

7. The semi-supervised learning based semantically transferable medical image segmentation method according to claim 1, characterized in that, In step 3, the segmentation loss function is constructed is: In formula (5), n s represents the total number of pixel points of the source domain data; M represents the number of augmentations; ω c ∈R d ,b c ∈R d respectively represent the classifier weight vector and the bias term of the label category c. respectively represent the classifier weight vector and the bias term of the label category y s,i of the i-th source domain data; exp(·) represents the exponential function; and C represents the number of label categories.

8. The semi-supervised learning based semantically transferable medical image segmentation method according to claim 1, characterized in that, In step 3, the segmentation loss function is analyzed and optimized The specific steps are as follows: Step 3.1, Analyzing the split loss function When the number of augmentations M→∞, by asymptotic analysis of the distribution of the augmented feature set , according to the law of large numbers, the empirical risk converges to its mathematical expectation, and the split loss function is obtained. The upper bound loss function is shown in equation (6): In formula (6), is an abbreviation for in formula (5), denotes the desired function; Step 3.2, optimizing the segmentation loss function Applying Jensen's inequality to the concave function log(·) Based on this inequality, the derivation process of the upper bound of the optimized segmentation loss function is obtained: In formula (7), represents a desired function. 9.The semi-supervised learning based semantically transferable medical image segmentation method of claim 1, wherein, In step 4, the model overall loss function is: In Equation (8), β is a weight hyper-parameter that controls the contribution of the optimized segmentation loss function to the overall loss function of the model .

10. The method of claim 1, wherein the semi-supervised learning-based semantically transferable medical image segmentation method is characterized by, In step 4, the specific operation of the exponential moving average is: θ tea ← λθ tea + (1 - λ)θ stu (9) In formula (9), θ tea and θ stu are parameters of the teacher network and the student network, respectively, and λ is a weight hyperparameter.

Images