Semi-supervised medical image segmentation method based on dynamic ambiguity perception

Abstract

The invention discloses a semi-supervised medical image segmentation method based on dynamic ambiguity perception, and designs a dynamic ambiguity perception framework which comprises a heterogeneous feature extraction module, a dynamic ambiguity bootstrap module and an ambiguity perception module. Firstly, an original medical image enters a heterogeneous feature extraction module, and two competitive subnets extract high-level semantic features and generate different pseudo tags. The pseudo labels generated by the two competitive subnets enter a dynamic ambiguity bootstrap module, and the competitive subnets are divided into a strong network and a weak network by calculating the quality scores of the two pseudo labels. After a strong network and a weak network are divided, prediction generated by the weak network focuses on mutual ambiguity perception which biases to targeted learning of pseudo label knowledge, and prediction generated by the strong network focuses on self-ambiguity perception of self-ambiguity examination. The medical image segmentation accuracy is improved, the diagnosis time of doctors can be effectively shortened, and the efficiency is improved. Meanwhile, missed diagnosis or misdiagnosis caused by different diagnosis levels or fatigue of doctors is reduced.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional image processing technology, specifically relating to a semi-supervised medical image segmentation method based on dynamic ambiguity perception. Background Technology

[0002] Deep learning has demonstrated significant potential in medical image segmentation, with diagnostic accuracy even surpassing that of human experts, particularly in the identification of complex structures such as tumors and organs. However, the realization of high-performance models typically relies on large-scale, high-quality labeled data, and the labeling process for medical images requires the deep involvement of professional physicians, resulting in high costs and lengthy processes. With the rapid popularization of medical imaging technology, the amount of unlabeled image data is growing exponentially. How to efficiently utilize limited labeled data and massive amounts of unlabeled data has become a core bottleneck restricting the practical application of this technology.

[0003] Currently, semi-supervised learning is considered a key approach to resolving this contradiction. These methods help improve the model's generalization ability by uncovering potential patterns in unlabeled data. However, existing methods face several challenges in practice: First, the pseudo-label generation mechanism easily introduces noise, and with training iterations, errors accumulate and propagate, ultimately leading to model performance degradation. Second, most methods employ a fixed teacher-student network architecture, lacking dynamic adjustment capabilities and struggling to adapt to the learning needs of pseudo-labels of varying quality. Furthermore, traditional methods often focus on consistency constraints in the label space, while paying insufficient attention to the separability of the feature space, limiting the model's ability to capture subtle structures.

[0004] It is worth noting that target regions in medical images often exhibit varied morphologies and small proportions (such as vascular lesions and micro-tumors), making it easy for existing methods to lose crucial details using voxel-level averaging strategies. Although recent studies have attempted to alleviate these problems through uncertainty quantification or multi-model collaboration, shortcomings remain, such as insufficient information utilization and coarse error suppression mechanisms. Therefore, developing a semi-supervised learning framework capable of dynamically perceiving ambiguous regions, adaptively filtering high-quality knowledge, and implementing multi-dimensional constraints has become an important direction for overcoming current technological bottlenecks. Summary of the Invention

[0005] The present invention aims to at least partially solve one of the technical problems existing in the related art.

[0006] The purpose of this invention is to provide a semi-supervised medical image segmentation method based on dynamic ambiguity perception, which improves the accuracy and robustness of medical image segmentation.

[0007] To achieve the above objectives, this invention provides a semi-supervised medical image segmentation method based on dynamic ambiguity perception, comprising the following steps:

[0008] S1. Obtain a medical image dataset, which includes labeled medical images and unlabeled medical images;

[0009] S2. Construct a semi-supervised medical image segmentation model based on dynamic ambiguity perception. This semi-supervised medical image segmentation model includes a heterogeneous feature extraction module, a dynamic ambiguity bootstrapping module, and an ambiguity perception module. Then, use this semi-supervised medical image segmentation model to process medical images, specifically:

[0010] The medical image is input into the heterogeneous feature extraction module, which uses two competing subnets with different structures to extract features and generate pseudo-labels.

[0011] The pseudo-labels are then input into the dynamic ambiguity bootstrapping module to calculate the quality score of the pseudo-labels generated by each competing subnet. Based on the quality score, the competing subnets are dynamically divided into strong networks and weak networks. The pseudo-labels generated by the strong network have a higher quality than those generated by the weak network.

[0012] Finally, the pseudo-labels are input into the ambiguity perception module, which includes a mutual ambiguity perception unit and a self-ambiguity perception unit. In the mutual ambiguity perception unit, the weak network learns the pseudo-label knowledge of the strong network in the ambiguity region. In the self-ambiguity perception unit, the input of the strong network is perturbed and self-ambiguity perception is performed based on prediction consistency.

[0013] S3. Train the semi-supervised medical image segmentation model based on the supervised loss of labeled medical images and the self-supervised loss of unlabeled medical images generated by pseudo-labels and ambiguity perception.

[0014] S4. Use the trained semi-supervised medical image segmentation model to predict the medical image to be segmented and generate the segmentation result.

[0015] A further preferred technical solution of the present invention is that the medical image dataset in step S1 includes a large number of unlabeled brain CTA scan images, left atrial MRI scan images and pancreatic CT scan images, as well as a small number of high-precision labeled scan images of the above three types of medical images.

[0016] Preferably, step S2 involves inputting the medical image into the heterogeneous feature extraction module, extracting features using two structurally dissimilar competing subnets, and generating pseudo-labels; specifically:

[0017] A heterogeneous feature extraction module is constructed by using two deep learning networks with different structures as two competing subnetworks in parallel.

[0018] The medical images are input into two competing subnets, which extract high-level semantic features from the medical images. The extracted features are then processed by cross-entropy to generate pseudo-labels for the two competing subnets.

[0019] Preferably, step S2 involves inputting the pseudo-labels into the dynamic ambiguity bootstrapping module, calculating the quality score of the pseudo-labels generated by each competing subnet, and dynamically dividing the competing subnets into strong and weak networks based on the quality scores; specifically:

[0020] For the voxel location of the pseudo-label Its prediction probability is denoted as This indicates that the voxel location belongs to category The predicted probability; where , , , These represent the height, width, and depth dimensions of the image, respectively; the number of categories. ;

[0021] For the voxel location of the pseudo-label The degree of ambiguity in the voxel prediction at a given location is denoted as:

[0022] ;

[0023] The quality score of the pseudo-label is calculated using the following formula:

[0024] ;

[0025] Next, the quality scores of the pseudo-labels obtained from the two competing subnets are compared. The competing subnet with the higher quality score is designated as a strong network, and the competing subnet with the lower quality score is designated as a weak network.

[0026] Preferably, in step S2, in the mutually ambiguous sensing unit, the weak network learns the pseudo-label knowledge of the strong network in the ambiguous region; specifically:

[0027] Set an ambiguity threshold T, and generate an auxiliary ambiguity map based on the comparison between the voxel ambiguity predicted by the weak network and the preset threshold T, as follows:

[0028] ;

[0029] Then, a mutually ambiguous graph is generated based on the prediction differences between the strong and weak networks, represented as:

[0030] ;

[0031] Multiplying the auxiliary ambiguous graph by the mutually ambiguous graph yields the final ambiguous graph, represented as:

[0032] ;

[0033] Based on this final ambiguity graph, the loss is calculated to determine the regions and weights where the weak network learns from the strong network.

[0034] Preferably, step S2 involves applying a perturbation to the input of the strong network and performing self-ambiguity sensing based on prediction consistency in the self-ambiguity sensing unit; specifically:

[0035] A slight perturbation is applied to the original image input to the strong network to obtain a perturbed image; the perturbed image is then input into the strong network to obtain the perturbation prediction result;

[0036] Based on the prediction of the original image by the strong network and the prediction result of the perturbation, a self-ambiguous joint confidence matrix is ​​constructed. Then, the self-ambiguous joint distribution matrix is ​​further calculated to identify unstable voxels in the prediction of the strong network and to apply consistency regularization constraints to the prediction of unstable voxels.

[0037] As a preferred option, the self-ambiguous joint confidence matrix C is constructed as follows:

[0038] ;

[0039] The self-ambiguous joint distribution matrix Q is constructed as follows:

[0040] ;

[0041] in, This represents the input image. Represents voxels; This is represented as a noisy prediction for a strong network. Voxel representation belong The probability, As a potential real label, Represented as The number of voxels; Indicates the number of categories. This represents the average confidence level.

[0042] Preferably, the supervised loss in step S3 adopts the cross-entropy loss function, and the self-supervised loss includes mutual ambiguity perception loss and self-ambiguity perception loss. The mutual ambiguity perception loss is weighted based on the final ambiguity graph, and the self-ambiguity perception loss is calculated based on the consistency deviation of unstable voxels.

[0043] Preferably, in step S4, before predicting the medical image to be segmented, a test is performed using an labeled medical image, specifically as follows:

[0044] Medical image segmentation is performed by averaging the weights of two trained competing subnet models.

[0045] Then, the accuracy of the segmentation results and model performance are verified based on the segmentation results and annotations, including:

[0046] The accuracy of the entire segmentation is calculated by the proportion of correctly predicted voxels to the total number of voxels; the segmentation intersection-union ratio is calculated by comparing the ratio of the number of correctly predicted voxels to the actual number of voxels in the medical image and the union of the predicted results, in order to evaluate the model performance.

[0047] Beneficial Effects: The semi-supervised medical image segmentation method based on dynamic ambiguity perception of this invention achieves efficient utilization of limited labeled data and a large amount of unlabeled data by constructing a collaborative learning framework comprising three core modules: a heterogeneous feature extraction module, a dynamic ambiguity bootstrapping module, and an ambiguity perception module. The heterogeneous feature extraction module employs two competing sub-networks with different structures but comparable performance to learn features from the input data, ensuring the extraction of complementary and robust feature representations. These features are fed into the dynamic ambiguity bootstrapping module, which quantifies the overall quality of pseudo-labels generated by each sub-network through uncertainty analysis and dynamically divides the sub-networks into strong networks that generate high-quality pseudo-labels and weak networks that generate low-quality pseudo-labels based on quality scores. The ambiguity perception module is a key innovation of this invention, consisting of a cross-ambiguity perception block and a self-ambiguity perception block. The cross-ambiguity perception block handles the relationship between strong and weak networks, accurately selecting reliable voxel predictions from the strong network that need to be learned from the weak network by constructing auxiliary ambiguity graphs and interactive ambiguity graphs, while avoiding interference from potential erroneous predictions in the strong network, thereby achieving safe and targeted knowledge transfer. The self-ambiguity sensing block focuses on optimizing the strong network itself. By introducing input perturbations and constructing a joint confidence matrix, it identifies and suppresses potential local ambiguities or errors in its high-quality pseudo-labels, achieving a self-cleaning effect.

[0048] During model training, the parameters of the two competing sub-networks are updated independently, and their roles may change as training progress under the evaluation of the dynamic ambiguity bootstrapping module. This dynamic mechanism effectively prevents error solidification caused by fixed teacher-student roles in the traditional mean-teacher method. The final segmentation result is generated by averaging the prediction weights of the two sub-networks, further enhancing the model's stability and robustness.

[0049] In summary, this invention successfully improves the accuracy and robustness of semi-supervised medical image segmentation by dynamically evaluating pseudo-label quality, differentiating reliability predictions, and implementing a dual ambiguity perception mechanism. This method not only alleviates the problem of strong dependence on labeled data but also provides an effective technical approach for achieving more accurate and reliable segmentation in complex medical images, possessing significant practical value for assisting clinical diagnosis. Attached Figure Description

[0050] Figure 1 This is an overall flowchart of the semi-supervised medical image segmentation method based on dynamic ambiguity perception of the present invention.

[0051] Figure 2 This is a flowchart of the data processing of the dynamic ambiguity bootstrapping module in this invention.

[0052] Figure 3 This is a flowchart of the data processing of the mutual ambiguity sensing unit in the ambiguity sensing module of this invention.

[0053] Figure 4 This is a flowchart of the data processing of the self-ambiguity sensing unit of the ambiguity sensing module in this invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0055] The following is combined Figures 1-4 This invention describes a semi-supervised medical image segmentation method based on dynamic ambiguity perception.

[0056] Example: This example provides a semi-supervised medical image segmentation method based on dynamic ambiguity perception, such as... Figure 1 As shown, it includes the following steps:

[0057] S1. Obtain the medical image dataset.

[0058] A large number of unlabeled brain CTA scan images, left atrial MRI scan images, and pancreatic CT scan images were collected, along with a small number of high-precision labeled scan images of these three medical images. The collected image data was preprocessed to remove redundant information, and various enhanced views of the initial input were generated to construct the medical image dataset.

[0059] This dataset includes both labeled and unlabeled medical images. The small number of labeled medical images serves three purposes: first, it can be used to calculate supervised loss during semi-supervised model training; second, it can be used for offline calibration of the quality score threshold mentioned later; and third, it serves as a test set during the testing phase.

[0060] S2. Construct a semi-supervised medical image segmentation model based on dynamic ambiguity perception.

[0061] The semi-supervised medical image segmentation model in this embodiment includes a heterogeneous feature extraction module, a dynamic ambiguity bootstrapping module, and an ambiguity perception module. The medical image is processed using this semi-supervised medical image segmentation model as follows:

[0062] S21. Input the medical image into the heterogeneous feature extraction module, and extract features and generate pseudo-labels using two competing subnets with different structures.

[0063] Two deep learning networks with different structures but comparable performance are used as two competing subnetworks, such as one subnetwork based on a convolutional neural network and the other based on a Transformer architecture, to form a heterogeneous feature extraction module;

[0064] The medical images are input into two competing subnets, which extract high-level semantic features from the medical images. The extracted features are then processed by cross-entropy to generate pseudo-labels for the two competing subnets.

[0065] S22. Input the pseudo-labels into the dynamic ambiguity bootstrapping module, calculate the quality score of the pseudo-labels generated by each competing subnet, and dynamically divide the competing subnets into strong networks and weak networks according to the quality score. The quality of the pseudo-labels generated by the strong network is higher than that of the pseudo-labels generated by the weak network.

[0066] like Figure 2 As shown, for the voxel position of the pseudo-label Its prediction probability is denoted as This indicates that the voxel location belongs to category The predicted probability; where , , , These represent the height, width, and depth dimensions of the image, respectively; the number of categories. ;

[0067] For the voxel location of the pseudo-label The degree of ambiguity in the voxel prediction at a given location is denoted as:

[0068] ;

[0069] The quality score of the pseudo-label is calculated using the following formula:

[0070] ;

[0071] Next, the quality scores of the pseudo-labels obtained from the two competing subnets are compared. The competing subnet with the higher quality score is designated as a strong network, and the competing subnet with the lower quality score is designated as a weak network.

[0072] S23. Finally, the pseudo-labels are input into the ambiguity perception module to process the pseudo-labels from the dynamic ambiguity bootstrapping module in a targeted manner.

[0073] The ambiguity sensing module includes a mutually ambiguous sensing unit and a self-ambiguous sensing unit. In the mutually ambiguous sensing unit, the weak network learns the pseudo-label knowledge of the strong network in ambiguous regions; for example... Figure 3 As shown, specifically:

[0074] Set an ambiguity threshold T, and generate an auxiliary ambiguity map based on the comparison between the voxel ambiguity predicted by the weak network and the preset threshold T, as follows:

[0075] ;

[0076] Then, a mutually ambiguous graph is generated based on the prediction differences between the strong and weak networks, represented as:

[0077] ;

[0078] Multiplying the auxiliary ambiguous graph by the mutually ambiguous graph yields the final ambiguous graph, represented as:

[0079] ;

[0080] Based on this final ambiguity graph, the loss is calculated to determine the regions and weights where the weak network learns from the strong network.

[0081] In the self-ambiguity sensing unit, a perturbation is applied to the input of the strong network, and self-ambiguity sensing is performed based on prediction consistency; such as Figure 4 As shown, specifically:

[0082] A slight perturbation is applied to the original image input to the strong network to obtain a perturbed image; the perturbed image is then input into the strong network to obtain the perturbation prediction result;

[0083] Based on the prediction of the original image by the strong network and the prediction result of the perturbation, a self-ambiguous joint confidence matrix is ​​constructed. Then, the self-ambiguous joint distribution matrix is ​​further calculated to identify unstable voxels in the prediction of the strong network and to apply consistency regularization constraints to the prediction of unstable voxels.

[0084] The self-ambiguous joint confidence matrix C is constructed as follows:

[0085] ;

[0086] The self-ambiguous joint distribution matrix Q is constructed as follows:

[0087] ;

[0088] in, This represents the input image. Represents voxels; This is represented as a noisy prediction for a strong network. Voxel representation belong The probability, As a potential real label, Represented as The number of voxels; Indicates the number of categories. This represents the average confidence level.

[0089] S3. Train the semi-supervised medical image segmentation model based on the supervised loss of labeled medical images and the self-supervised loss of unlabeled medical images generated by pseudo-labels and ambiguity perception.

[0090] The supervised loss adopts the cross-entropy loss function, and the self-supervised loss includes mutual ambiguity perception loss and self-ambiguity perception loss. The mutual ambiguity perception loss is weighted based on the final ambiguity graph, and the self-ambiguity perception loss is calculated based on the consistency deviation of unstable voxels.

[0091] S4. Test the trained semi-supervised medical image segmentation model using labeled medical images.

[0092] Medical image segmentation is performed by averaging the weights of two trained competing subnet models.

[0093] Then, the accuracy of the segmentation results and model performance are verified based on the segmentation results and annotations, including:

[0094] The accuracy of the entire segmentation is calculated by the proportion of correctly predicted voxels to the total number of voxels; the segmentation intersection-union ratio is calculated by comparing the ratio of the number of correctly predicted voxels to the actual number of voxels in the medical image and the union of the predicted results, in order to evaluate the model performance.

[0095] S5. Use the trained semi-supervised medical image segmentation model to predict the medical image to be segmented and generate the segmentation result.

[0096] This embodiment proposes a semi-supervised medical image segmentation method based on dynamic ambiguity perception. A dynamic ambiguity perception framework is designed, and a semi-supervised medical image segmentation model is constructed. The framework mainly consists of three modules: a heterogeneous feature extraction module, a dynamic ambiguity bootstrapping module, and an ambiguity perception module. These three modules work together to effectively complete the semi-supervised medical image segmentation task. First, the original medical image enters the heterogeneous feature extraction module, where two competing sub-networks extract high-level semantic features and generate different pseudo-labels. Furthermore, the pseudo-labels generated by the two competing sub-networks enter the dynamic ambiguity bootstrapping module, which further divides the competing sub-networks into strong and weak networks by calculating the quality scores of the two pseudo-labels. Simultaneously, after dividing the strong and weak networks (based on pseudo-label quality), targeted processing is performed. The predictions generated by the weak network focus on mutual ambiguity perception, which is biased towards learning pseudo-label knowledge, while the predictions generated by the strong network emphasize self-ambiguity perception, which focuses on self-ambiguity review. This method can improve the accuracy of semi-supervised medical image segmentation, effectively reducing the time doctors spend diagnosing and improving efficiency. It also reduces missed diagnoses or misdiagnoses caused by varying diagnostic skills or fatigue among doctors.

[0097] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A semi-supervised medical image segmentation method based on dynamic ambiguity perception, characterized in that, Includes the following steps: S1. Obtain a medical image dataset, which includes labeled medical images and unlabeled medical images; S2. Construct a semi-supervised medical image segmentation model based on dynamic ambiguity perception. This semi-supervised medical image segmentation model includes a heterogeneous feature extraction module, a dynamic ambiguity bootstrapping module, and an ambiguity perception module. Then, use this semi-supervised medical image segmentation model to process medical images, specifically: The medical image is input into the heterogeneous feature extraction module, which uses two competing subnets with different structures to extract features and generate pseudo-labels. The pseudo-labels are then input into the dynamic ambiguity bootstrapping module to calculate the quality score of the pseudo-labels generated by each competing subnet. Based on the quality score, the competing subnets are dynamically divided into strong networks and weak networks. The pseudo-labels generated by the strong network have a higher quality than those generated by the weak network. Finally, the pseudo-labels are input into the ambiguity perception module, which includes a mutual ambiguity perception unit and a self-ambiguity perception unit. In the mutual ambiguity perception unit, the weak network learns the pseudo-label knowledge of the strong network in the ambiguity region. In the self-ambiguity perception unit, the input of the strong network is perturbed and self-ambiguity perception is performed based on prediction consistency. S3. Train the semi-supervised medical image segmentation model based on the supervised loss of labeled medical images and the self-supervised loss of unlabeled medical images generated by pseudo-labels and ambiguity perception. S4. Use the trained semi-supervised medical image segmentation model to predict the medical image to be segmented and generate the segmentation result.

2. The semi-supervised medical image segmentation method based on dynamic ambiguity perception according to claim 1, characterized in that, The medical image dataset in step S1 includes a large number of unlabeled brain CTA scan images, left atrial MRI scan images, and pancreatic CT scan images, as well as a small number of high-precision labeled scan images of the above three types of medical images.

3. The semi-supervised medical image segmentation method based on dynamic ambiguity perception according to claim 1, characterized in that, Step S2 involves inputting the medical image into the heterogeneous feature extraction module, whereby features are extracted and pseudo-labels are generated using two structurally dissimilar competing subnets; specifically: A heterogeneous feature extraction module is constructed by using two deep learning networks with different structures as two competing subnetworks in parallel. The medical images are input into two competing subnets, which extract high-level semantic features from the medical images. The extracted features are then processed by cross-entropy to generate pseudo-labels for the two competing subnets.

4. The semi-supervised medical image segmentation method based on dynamic ambiguity perception according to claim 1, characterized in that, Step S2 involves inputting the pseudo-labels into the dynamic ambiguity bootstrapping module, calculating the quality score of the pseudo-labels generated by each competing subnet, and dynamically dividing the competing subnets into strong and weak networks based on the quality scores; specifically: For the voxel location of the pseudo-label Its prediction probability is denoted as This indicates that the voxel location belongs to category The predicted probability; where , , , These represent the height, width, and depth dimensions of the image, respectively; the number of categories. ; For the voxel location of the pseudo-label The degree of ambiguity in the voxel prediction at a given location is denoted as: ； The quality score of the pseudo-label is calculated using the following formula: ； Next, the quality scores of the pseudo-labels obtained from the two competing subnets are compared. The competing subnet with the higher quality score is designated as a strong network, and the competing subnet with the lower quality score is designated as a weak network.

5. The semi-supervised medical image segmentation method based on dynamic ambiguity perception according to claim 4, characterized in that, Step S2 describes the process in which the weak network learns the pseudo-label knowledge of the strong network in the ambiguous region within the mutually ambiguous sensing unit; specifically: Set an ambiguity threshold T, and generate an auxiliary ambiguity map based on the comparison between the voxel ambiguity predicted by the weak network and the preset threshold T, as follows: ； Then, a mutually ambiguous graph is generated based on the prediction differences between the strong and weak networks, represented as: ； Multiplying the auxiliary ambiguous graph by the mutually ambiguous graph yields the final ambiguous graph, represented as: ； Based on this final ambiguity graph, the loss is calculated to determine the regions and weights where the weak network learns from the strong network.

6. The semi-supervised medical image segmentation method based on dynamic ambiguity perception according to claim 5, characterized in that, Step S2 describes applying a perturbation to the input of the strong network in the self-ambiguity sensing unit and performing self-ambiguity sensing based on prediction consistency; specifically: A slight perturbation is applied to the original image input to the strong network to obtain a perturbed image; the perturbed image is then input into the strong network to obtain the perturbation prediction result; Based on the prediction of the original image by the strong network and the prediction result of the perturbation, a self-ambiguous joint confidence matrix is ​​constructed. Then, the self-ambiguous joint distribution matrix is ​​further calculated to identify unstable voxels in the prediction of the strong network and to apply consistency regularization constraints to the prediction of unstable voxels.

7. The semi-supervised medical image segmentation method based on dynamic ambiguity perception according to claim 6, characterized in that, The self-ambiguous joint confidence matrix C is constructed as follows: ； The self-ambiguous joint distribution matrix Q is constructed as follows: ； in, This represents the input image. Represents voxels; This is represented as a noisy prediction for a strong network. Voxel representation belong The probability, As a potential real label, Represented as The number of voxels; Indicates the number of categories. This represents the average confidence level.

8. The semi-supervised medical image segmentation method based on dynamic ambiguity perception according to claim 6, characterized in that, The supervised loss in step S3 adopts the cross-entropy loss function, and the self-supervised loss includes mutual ambiguity sensing loss and self-ambiguity sensing loss. The mutual ambiguity sensing loss is weighted based on the final ambiguity graph, and the self-ambiguity sensing loss is calculated based on the consistency deviation of unstable voxels.

9. The semi-supervised medical image segmentation method based on dynamic ambiguity perception according to claim 1, characterized in that, In step S4, before predicting the segmented medical image, a test is performed using an labeled medical image. Specifically: Medical image segmentation is performed by averaging the weights of two trained competing subnet models. Then, the accuracy of the segmentation results and model performance are verified based on the segmentation results and annotations, including: The accuracy of the entire segmentation is calculated by the proportion of correctly predicted voxels to the total number of voxels. The segmentation intersection-union ratio (SUIR) is calculated by comparing the ratio of the correctly predicted number of voxels to the union of the actual number of voxels in the medical image and the predicted results, in order to evaluate the model performance.

Images