Double-standard active learning method and system for medical image segmentation
Through the dual-standard active learning method, combined with the potential spatial distribution representativeness and uncertainty evaluation, high information-quantity samples were screened for annotation, solving the problems of high labeling costs and limited model performance in the existing technology, and achieving efficient medical image segmentation.
Patent Information
- Application Number
- CN202510567418.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-30
- Publication Date
- 2025-08-08
AI Technical Summary
The existing active learning methods have problems such as single-dimensional sample selection dimensions and insufficient multi-dimensional fusion performance in medical image segmentation, resulting in high labeling costs and limited model performance.
The dual-standard active learning method is adopted to capture the global distribution characteristics of medical image data through a variational autoencoder, and combined with the potential spatial distribution representativeness and uncertainty evaluation, samples with global representativeness and high information volume are screened for annotation.
It significantly reduces the amount of labeling data, improves the segmentation accuracy and generalization ability of medical image segmentation models, and reduces the labeling cost.
Smart Images
Figure CN120451558A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image processing, and in particular to a dual-standard active learning method and system for medical image segmentation. Background Art
[0002] In recent years, deep learning-based medical image segmentation models have demonstrated outstanding performance across a wide range of tasks. However, their training relies heavily on large amounts of labeled data, which presents significant challenges in practical applications. Medical image annotation is not only complex but also requires specialized medical expertise, resulting in high annotation costs. As data volumes expand, overall annotation costs are increasing exponentially.
[0003] To address this challenge, active learning (AL) methods intelligently select the most informative examples for annotation, effectively reducing the need for annotation. Research has shown that models trained on high-quality, informative examples can perform comparable to models trained using fully annotated data. However, existing active learning methods still face two key bottlenecks:
[0004] (1) The problem of single dimension in sample selection: Most current algorithms focus only on a single dimension, uncertainty or representativeness. Pure uncertainty strategies tend to select "difficult samples" near the decision boundary. Although this helps the model refine the classification boundary, it ignores the global distribution characteristics of the data. Pure representative strategies, while covering the entire data distribution, weaken the learning of key boundary samples, affecting segmentation accuracy.
[0005] (2) Insufficient multi-dimensional fusion efficiency: Some algorithms attempt to fuse uncertainty and representativeness, but they are limited in the following aspects: First, the measurement standards of the two types of indicators are fundamentally different, making direct coordination difficult; second, the existing weighting strategies lack adaptive capabilities, often resulting in fusion effects that are inferior to random selection benchmarks. This inefficient fusion not only fails to take advantage of the combination, but may even lead to performance degradation.
[0006] The study "Abdomen Atlas-8K: Annotating 8,000 CTVolumes for Multi-Organ Segmentation in Three Weeks" has significant limitations in its sample selection strategy: it selects samples solely based on model prediction uncertainty, completely ignoring the representativeness of the data distribution. This single-dimensional selection criterion can result in the final annotated sample set deviating from the true data distribution, thus restricting further improvement in model performance.
[0007] The method in "One-shot Active Learning for Image Segmentation via Contrastive Learning and Diversity-based Sampling" exhibits another extreme tendency in its sample selection strategy: it over-relies on diverse sampling based on contrastive learning (i.e., representativeness metric) while completely ignoring the critical dimension of model prediction uncertainty. This selection bias makes it difficult for the model to effectively identify key samples near the decision boundary, ultimately leading to poor generalization performance in complex segmentation tasks.
[0008] Although "D2ADA: Dynamic Density-aware Active Domain Adaptation" innovatively proposes a dynamic density-aware sample selection mechanism, its method has limitations in specific application scenarios: the dynamic weighting method designed by this strategy mainly targets the significant domain shift problem in domain adaptation tasks. In scenarios such as medical image segmentation where the target domain distribution is relatively stable, its complex fusion mechanism may cause insufficient adaptability, making it difficult to achieve optimal sample selection effects. Summary of the Invention
[0009] The purpose of the present invention is to propose a dual-standard active learning method and system for medical image segmentation, which significantly reduces the amount of required annotation data, thereby effectively reducing the overall annotation cost of medical image analysis.
[0010] According to a first aspect of an embodiment of the present disclosure, a dual-standard active learning method for medical image segmentation is provided, comprising the following steps:
[0011] In Medical Image PoolX pool We train a variational autoencoder for representational active learning to map medical image data into a latent space that captures the global distributional properties of the learning data.
[0012] From the medical image pool X pool Randomly extract M samples from the image set to construct the initial labeled image set X an ; In this initial labeled image set X an By minimizing the cross entropy loss function Train a neural network model θ1 for medical image segmentation;
[0013] After T rounds of active learning, the final medical image segmentation model θ is obtained T+1 .
[0014] In one embodiment, in each round t of active learning, the following steps are specifically performed:
[0015] Using representative active learning variational autoencoders based on latent space distribution, we extract images from a pool of medical images X. pool Select R t globally representative candidate samples;
[0016] In R t Among the candidate samples, the uncertainty active learning algorithm based on the entropy loss gradient is used to further screen out S t samples;
[0017] To S t The samples are labeled to obtain their true labels, and the unlabeled samples are X non ;
[0018] The newly marked S t samples are added to the initial labeled image set X an In the above example, update the labeled image set X an ;
[0019] After updating the labeled image set X an Retrain the neural network model θ1 for medical image segmentation.
[0020] In one embodiment, the uncertainty active learning algorithm based on entropy loss gradient is implemented as follows:
[0021] Specifically, a deep neural network generates a predicted label for the sample x Get the loss function in The gradient vector g relative to the parameters of the last layer of the network x , based on the gradient vector g x The norm of θ1 is used to judge the parameter change of the neural network model for medical image segmentation.
[0022] In one embodiment, by obtaining the gradient vectors of all unlabeled samples, a subset of samples with high information content in the gradient space is selected for labeling.
[0023] In one embodiment, the representative active learning variational autoencoder is implemented as follows: enc : Make X from medical image pool pool The observed value x i Mapped to In each round of active learning, the unlabeled sample X non Select query samples in batches, which represent the medical image pool X pool Distribution statistics of X an Covering insufficient area space.
[0024] In one embodiment, since the mode of the latent space Z encodes the medical image pool X pool The most frequent attribute, so the ideal sample x * Query by the following formula:
[0025]
[0026] Through iteration, based on the query sample x * This will make the posterior distribution p(z|X an ) and p(z|X pool ) are aligned so that X an The observed value of X pool The breadth and mode of , to achieve the expected goal; to calculate formula (1) using Bayesian inference:
[0027]
[0028] The equation on the right contains an equivalent expression of the posterior distribution p(z|X), which is simplified to equation (3):
[0029]
[0030] To approximate the mapping function f enc , trained an infoVAE-based model using MMD regularization, with a loss function L infoVAE =L AE +L MMD , where L AE is the reconstruction error:
[0031]
[0032] Among them, q φ (z|x) represents the approximate posterior distribution learned by the encoder, mapping the input data x to the distribution of latent variables; p θ (x|z) represents the conditional generative distribution learned by the decoder, which represents the probability distribution of generating x given the latent variable z;
[0033] L MMD is the maximum mean difference (MMD) loss:
[0034]
[0035] where p is the prior distribution, q is the posterior inference in the latent space by the encoder, and k(z,z′) is the distance metric in the kernel space; we choose p(z) as the standard multivariate normal distribution and use a Gaussian as the kernel mapping k(z,z′) = exp(-||zz′|| / 2σ) 2), where σ = 1; the expectation in (5) is based on the expected value from the sample pairs z and z′.
[0036] In one embodiment, the true posterior inference p(z|x) is approximated by q by the learned parameter set φ of the encoder φ (z|x), thus formula (1) is approximated as formula (6):
[0037]
[0038] Formula (3) is approximately q φ (x i |X,z)q φ (z|X); therefore, (i)X an and (ii) X pool The samples in are projected into the latent space of the infoVAE-based model; in order to obtain q φ (z|X), respectively, fit these two projections into a multivariate diagonal Gaussian distribution; use the error function To estimate q φ (z|x i ,X an ) and q φ (z|x i ,X pool ) is as shown in formula (7):
[0039]
[0040] where μ X and σ X are the parameters of the fitted Gaussian distribution.
[0041] According to a second aspect of an embodiment of the present disclosure, a dual-standard active learning system for medical image segmentation is provided, comprising:
[0042] Active learning training module, in medical image pool X pool We train a variational autoencoder for representational active learning to map medical image data into a latent space that captures the global distributional properties of the learning data.
[0043] Segmentation model training module, from medical image pool X pool Randomly extract M samples from the image set to construct the initial labeled image set X an ; In this initial labeled image set X an By minimizing the cross entropy loss function Train a neural network model θ1 for medical image segmentation;
[0044] Active learning module, after T rounds of active learning, obtains the final medical image segmentation model θ T+1 .
[0045] According to a third aspect of an embodiment of the present disclosure, an electronic device is provided, comprising a memory, a processor, and a computer program stored and running on the memory, wherein when the processor executes the program, the dual-standard active learning method for medical image segmentation is implemented.
[0046] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, the dual-standard active learning method for medical image segmentation is implemented.
[0047] Compared with the existing technology, the above technical solutions adopted by the present invention have the following advantages: 1. The present invention proposes a dual-standard active learning screening mechanism, which realizes efficient sample selection by organically combining the two methods of gradient-based uncertainty assessment and latent space distribution representativeness measurement. In terms of uncertainty assessment, the entropy loss function is used to replace the traditional loss function, which significantly reduces the dependence on pseudo-labels; in terms of representativeness measurement, by analyzing the characteristics of the latent space distribution, the samples that best represent the global distribution of the data and are significantly different from the labeled samples are preferentially selected, effectively avoiding the problem of sample redundancy. This dual selection mechanism ensures both the amount of information in the sample and the comprehensiveness of the data distribution.
[0048] 2. This paper proposes a phased fusion active learning sample screening strategy: first, based on a representativeness metric, a candidate sample set with global distribution characteristics is screened from the original dataset. Uncertainty assessment methods are then applied to this candidate set to select the most informative final samples. This two-stage fusion mechanism offers two advantages: first, it effectively overcomes the distribution bias problem caused by traditional methods' excessive focus on difficult samples during the initial training phase; second, by limiting uncertainty calculations to the representative candidate set, it significantly reduces the algorithm's computational complexity and significantly improves active learning efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] The drawings in the specification, which constitute a part of this application, are used to provide further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute improper limitations on this application.
[0050] Figure 1 Flowchart of the dual-criteria active learning method for medical image segmentation;
[0051] Figure 2 Flowchart of the uncertainty active learning algorithm based on entropy loss gradient. DETAILED DESCRIPTION
[0052] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.
[0053] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.
[0054] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.
[0055] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the possible implementation architecture, functions and operations of the methods and systems according to the various embodiments of the present disclosure. It should be noted that each box in the flowchart or block diagram can represent a module, program segment, or a part of code, and the module, program segment, or a part of code can include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, or they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the flowchart and / or block diagram, and the combination of the boxes in the flowchart and / or block diagram, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or can be implemented using a combination of dedicated hardware and computer instructions.
[0056] In response to the limitations of single-dimensional sample selection strategies, the present invention deeply analyzes the inherent defects of different selection criteria: although the uncertainty-based method can effectively identify difficult samples near the decision boundary, it will lead to insufficient understanding of the model's overall data distribution; and although the representativeness-based method can ensure comprehensive coverage of the data distribution, it is difficult to improve the model's processing capabilities for complex boundary samples. This either-or selection strategy seriously restricts the overall performance of the model. Secondly, in response to the effectiveness of multi-dimensional indicator fusion, the present invention reveals the deep contradictions in existing methods when combining uncertainty and representativeness: on the one hand, there are essential differences in the measurement standards of the two indicators, and it is difficult to establish a unified evaluation system; on the other hand, simple linear weighting strategies often lead to the advantages of the indicators offsetting each other. These problems not only fail to achieve the fusion effect, but may even cause the algorithm performance to degenerate to a level below the baseline level of random sampling, greatly limiting the practical application value of active learning technology.
[0057] Example 1:
[0058] This embodiment provides a dual-standard active learning method for medical image segmentation, comprising the following steps:
[0059] S1. In the medical image pool X pool We train a variational autoencoder for representational active learning to map medical image data into a latent space that captures the global distributional properties of the learning data.
[0060] S2. From the medical image pool X pool Randomly extract M samples from the image set to construct the initial labeled image set X an ; In this initial labeled image set X an By minimizing the cross entropy loss function Train the neural network model θ1 for medical image segmentation, providing a basis for the subsequent active learning process;
[0061] S3. Obtain the final medical image segmentation model θ after T rounds of active learning T+1 In each round t of active learning, the following steps are performed:
[0062] S31. Using representative active learning variational autoencoders based on latent space distribution, from the medical image pool X pool Select R t globally representative candidate samples;
[0063] Specifically, the mapping function f enc : Make from X pool The observed value x i Mapped to This is a continuously defined latent space with a desired probability distribution. non Select query samples in batches, which represent the medical image pool X pool Distribution statistics of X an Covering insufficient area space.
[0064] Considering that the mode of the latent space Z encodes X pool The most frequent attribute, ideal sample x * You can query it through formula (1):
[0065]
[0066] Through iteration, based on x * The query sample will make the posterior distribution p(z|X an ) and p(z|X pool ) are aligned so that X an The observed value of X pool To calculate formula (1), Bayesian inference is used, as shown in formula (2):
[0067]
[0068] The equation on the right contains an equivalent expression of the posterior distribution p(z|X), which can be simplified to equation (3):
[0069]
[0070] To approximate the mapping function f enc , an infoVAE-based model was trained using MMD regularization with a loss function of L infoVAE =L AE +L MMD Among them, L AE To reconstruct the error, we need to ensure that the sample reconstructed by the latent variable z is as close to the input data as possible; the loss function L AE As shown in formula (4):
[0071]
[0072] q φ (z|x) represents the approximate posterior distribution learned by the encoder, mapping the input data x to the distribution of latent variables. θ (x|z) represents the conditional generative distribution learned by the decoder, which represents the probability distribution of generating x given the latent variable z.
[0073] L MMDis the maximum mean difference (MMD) loss, which is used to measure the latent variable distribution q φ (z) and the distance between the prior distribution p(z), ensuring that the distribution in the latent space can match the prior distribution p(z), the loss function L MMD As shown in formula (5):
[0074]
[0075] Where p is the prior distribution, q is the posterior inference in the latent space by the encoder, and k(z,z′) is the distance metric in the kernel space. We choose p(z) as the standard multivariate normal distribution and use Gaussian as the kernel mapping k(z,z′) = exp(-||zz′|| / 2σ) 2 ), where σ = 1. The expectation in Equation (5) is based on the expected value from the sample pair z and z′. Then, through the learned parameter set φ of the encoder, the true posterior inference p(z|x) can be approximated as q φ (z|x), so that formula (1) can be approximated as formula (6):
[0076]
[0077] Formula (3) is approximately q φ (x i |X,z)q φ (z|X). Therefore, (i)X an and (ii) X pool The samples in are projected into the latent space of infoVAE. Next, in order to obtain q φ (z|X), respectively fit a multivariate diagonal Gaussian distribution to these two projections. Finally, use the error function To estimate q φ (z|x i ,X an ) and q φ (z|x i ,X pool ) is as shown in formula (7):
[0078]
[0079] where μ X and σ X are the parameters of the fitted Gaussian distribution. In other words, use the first half of the cumulative distribution function of the fitted Gaussian distribution as it moves around its expected value μ X It is symmetrical.
[0080] S32.In R t Among the candidate samples, the uncertainty active learning algorithm based on the entropy loss gradient is used to further screen out St samples;
[0081] Specifically, a deep neural network is used to generate a predicted label for the sample x Get the loss function in The gradient vector g relative to the parameters of the last layer of the network x If the model has high confidence in its predicted label, the gradient vector g x The norm of will be relatively small, indicating that the label only produces a slight update to the model parameters. In contrast, if the model shows a large uncertainty in the predicted label, the gradient vector g x The norm of will increase significantly, indicating that this label may cause significant changes in model parameters. To reduce the reliance of traditional gradient-based active learning methods on pseudo-labels, the loss function is replaced with an entropy loss. By using entropy loss, gradient-based active learning methods can better capture the inherent uncertainty of samples without over-reliance on pseudo-labels, thereby improving the model's generalization ability. By obtaining the gradient vectors of all unlabeled samples, a subset of samples with high information content in the gradient space can be selected for labeling, effectively improving model performance.
[0082] S33.To S t The samples are labeled to obtain their true labels, and the unlabeled samples are X non ;
[0083] S34. The newly marked S t samples are added to the initial labeled image set X an In the above example, update the labeled image set X an ;
[0084] S35. In the updated labeled image set X an Retrain the neural network model θ1 for medical image segmentation.
[0085] In recent years, how to achieve an organic combination of uncertainty active learning (AL) and representative active learning has been a research difficulty in this field. Inappropriate fusion strategies often cause the algorithm to overemphasize one side, thereby weakening the overall performance. Representative AL has significant advantages in the early stages of training: in the first few rounds of iterations, due to the selection of samples that widely cover the data distribution, the model performance can be rapidly improved, and the overall segmentation quality is significantly improved. However, as the training deepens, the problem of insufficient refinement of difficult-to-segment areas such as boundary areas by representative AL gradually emerges. Using uncertainty AL alone has an obvious risk of local optimization: excessive focus on samples near the decision boundary, resulting in a concentrated sampling area and a lack of diversity. Especially in the early stages of training, when the basic capabilities of the model have not been fully established, this selection strategy is very likely to cause overfitting problems.
[0086] The combination method proposed in the present invention is: first use the representative metric to select R t candidate samples, and then use uncertainty measurement from R t The final S is selected from the candidate samples t This combination method first ensures that the selected samples have global coverage, and then uses uncertainty to ensure that the difficult samples on the decision boundary are selected. It avoids the phenomenon that the uncertainty AL method pays too much attention to the difficult samples on the decision boundary in the early stage of active learning. At the same time, in the combination method proposed by the present invention, it is not necessary to measure uncertainty for all samples. In each round of active learning, only the R samples selected by representativeness need to be measured. t The uncertainty measurement is performed on candidate samples, which greatly reduces the amount of calculation.
[0087] Example 2:
[0088] This embodiment provides a dual-standard active learning system for medical image segmentation, including:
[0089] Active learning training module, in medical image pool X pool We train a variational autoencoder for representational active learning to map medical image data into a latent space that captures the global distributional properties of the learning data.
[0090] Segmentation model training module, from medical image pool X pool Randomly extract M samples from the image set to construct the initial labeled image set X an ; In this initial labeled image set X an By minimizing the cross entropy loss function Train a neural network model θ1 for medical image segmentation;
[0091] Active learning module, after T rounds of active learning, obtains the final medical image segmentation model θ T+1 .
[0092] Example 3:
[0093] An electronic device includes a memory, a processor, and a computer program stored and running on the memory, wherein when the processor executes the program, it implements the above-mentioned dual-standard active learning method for medical image segmentation, including:
[0094] In Medical Image PoolX pool We train a variational autoencoder for representational active learning to map medical image data into a latent space that captures the global distributional properties of the learning data.
[0095] From the medical image pool X pool Randomly extract M samples from the image set to construct the initial labeled image set Xan ; In this initial labeled image set X an By minimizing the cross entropy loss function Train a neural network model θ1 for medical image segmentation;
[0096] After T rounds of active learning, the final medical image segmentation model θ is obtained T+1 .
[0097] Example 4:
[0098] A computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the program implements the above-mentioned dual-standard active learning method for medical image segmentation, comprising:
[0099] In Medical Image PoolX pool We train a variational autoencoder for representational active learning to map medical image data into a latent space that captures the global distributional properties of the learning data.
[0100] From the medical image pool X pool Randomly extract M samples from the image set to construct the initial labeled image set X an ; In this initial labeled image set X an By minimizing the cross entropy loss function Train a neural network model θ1 for medical image segmentation;
[0101] After T rounds of active learning, the final medical image segmentation model θ is obtained T+1 .
[0102] Those skilled in the art will appreciate that the modules or steps of the present disclosure described above can be implemented using a general-purpose computer device. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. The present disclosure is not limited to any specific combination of hardware and software.
[0103] The above description is merely a preferred embodiment of the present application and is not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.
[0104] Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Those skilled in the art should understand that on the basis of the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.
Claims
1. A dual-standard active learning method for medical image segmentation, characterized in that: The following steps are involved: In Medical Image PoolX pool We train a variational autoencoder for representational active learning to map medical image data into a latent space that captures the global distributional properties of the learning data. From the medical image pool X pool Randomly extract M samples from the image set to construct the initial labeled image set X an ; In this initial labeled image set X an By minimizing the cross entropy loss function Train a neural network model θ1 for medical image segmentation; After T rounds of active learning, the final medical image segmentation model θ is obtained T+1 .
2. A dual-standard active learning method for medical image segmentation according to claim 1, characterized in that: In each round of active learning, the following steps are performed: Using representative active learning variational autoencoders based on latent space distribution, we extract images from a pool of medical images X. pool Select R t globally representative candidate samples; In R t Among the candidate samples, the uncertainty active learning algorithm based on the entropy loss gradient is used to further screen out S t samples; To S t The samples are labeled to obtain their true labels, and the unlabeled samples are X non ; The newly marked S t samples are added to the initial labeled image set X an In the above example, update the labeled image set X an ; After updating the labeled image set X an Retrain the neural network model θ1 for medical image segmentation.
3. A dual-standard active learning method for medical image segmentation according to claim 2, characterized in that: The uncertainty active learning algorithm based on entropy loss gradient is implemented as follows: a prediction label is generated for the sample x through a deep neural network Get the loss function in The gradient vector g relative to the parameters of the last layer of the network x , based on the gradient vector g x The norm of θ1 is used to judge the parameter change of the neural network model for medical image segmentation.
4. A dual-standard active learning method for medical image segmentation according to claim 3, characterized in that: By obtaining the gradient vectors of all unlabeled samples, a subset of samples with high information content in the gradient space is selected for labeling.
5. The dual-standard active learning method for medical image segmentation according to claim 2, characterized in that: The variational autoencoder for representative active learning is implemented as follows: Make X from medical image pool pool The observed value x i Mapped to In each round of active learning, the unlabeled sample X non Select query samples in batches, which represent the medical image pool X pool Distribution statistics of X an Covering insufficient area space.
6. A dual-standard active learning method for medical image segmentation according to claim 5, characterized in that: Since the mode of the latent space Z encodes the medical image pool X pool The most frequent attribute, so the ideal sample x * Query by the following formula: Through iteration, based on the query sample x * This will make the posterior distribution p(z|X an ) and p(z|X pool ) are aligned so that X an The observed value of X pool The breadth and mode of , to achieve the expected goal; to calculate formula (1) using Bayesian inference: The equation on the right contains an equivalent expression of the posterior distribution p(z|X), which is simplified to equation (3): To approximate the mapping function f enc , trained an infoVAE-based model using MMD regularization, with a loss function L infoVAE =L AE +L MMD , where L AE is the reconstruction error: Among them, q φ (z|x) represents the approximate posterior distribution learned by the encoder, mapping the input data x to the distribution of latent variables; p θ (x|z) represents the conditional generative distribution learned by the decoder, which represents the probability distribution of generating x given the latent variable z; L MMD is the maximum mean difference (MMD) loss: where p is the prior distribution, q is the posterior inference in the latent space by the encoder, and k(z,z′) is the distance metric in the kernel space; we choose p(z) as the standard multivariate normal distribution and use a Gaussian as the kernel mapping k(z,z′) = exp(-||zz′|| / 2σ) 2 ), where σ = 1; the expectation in (5) is based on the expected value from the sample pairs z and z′.
7. A dual-standard active learning method for medical image segmentation according to claim 6, characterized in that: Through the encoder’s learning parameter set φ, the true posterior inference p(z|x) is approximated as q φ (z|x), thus formula (1) is approximated as formula (6): Formula (3) is approximately q φ (x i |X,z)q φ (z|X); therefore, (i)X an and (ii) X pool The samples in are projected into the latent space of the infoVAE-based model; in order to obtain q φ (z|X), respectively, fit these two projections into a multivariate diagonal Gaussian distribution; use the error function To estimate q φ (z|x i ,X an ) and q φ (z|x i ,X pool ) is as shown in formula (7): where μ X and σ X are the parameters of the fitted Gaussian distribution.
8. A dual-standard active learning system for medical image segmentation, characterized in that: include: Active learning training module, in medical image pool X pool We train a variational autoencoder for representational active learning to map medical image data into a latent space that captures the global distributional properties of the learning data. Segmentation model training module, from medical image pool X pool Randomly extract M samples from the image set to construct the initial labeled image set X an ; In this initial labeled image set X an By minimizing the cross entropy loss function Train a neural network model θ1 for medical image segmentation; Active learning module, after T rounds of active learning, obtains the final medical image segmentation model θ T+1 .
9. An electronic device comprising a memory, a processor, and a computer program stored and running on the memory, characterized in that: When the processor executes the program, the dual-standard active learning method for medical image segmentation according to any one of claims 1 to 7 is implemented.
10. A computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the dual-standard active learning method for medical image segmentation according to any one of claims 1 to 7 is implemented.