Semi-supervised broad-learning-based breast cancer diagnosis method based on integrated self-training

Abstract

The present invention relates to the technical fields of artificial intelligence and medical image processing. Disclosed is a semi-supervised broad-learning-based breast cancer diagnosis method based on integrated self-training. The semi-supervised broad-learning-based breast cancer diagnosis method comprises: initializing an integrated diagnosis model including a plurality of sub-diagnosis models; using the integrated diagnosis model to predict unlabeled breast cancer imaging data, so as to generate candidate pseudo-labeled samples; determining detection priorities on the basis of a preset scheduling strategy, and performing perturbation impact detection on the candidate samples, so as to obtain stability impact detection results; on the basis of the detection results, performing screening to obtain high-quality pseudo-labeled samples; and using the high-quality pseudo-labeled samples to execute an incremental update on the integrated diagnosis model, so as to complete one iteration. By screening pseudo labels by means of perturbation impact detection, the present invention effectively solves the problem of error accumulation; and incremental update characteristics are incorporated to reduce the computational overheads, thereby improving the stability and practicability of the method.

Description

A semi-supervised width breast cancer diagnosis method based on integrated self-training TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and medical image processing, in particular to a semi-supervised width breast cancer diagnosis method based on integrated self-training. BACKGROUND

[0002] Breast cancer is one of the most common malignant tumors in women worldwide, and its early and accurate diagnosis is crucial to improving patient survival rates. Medical image analysis based on artificial intelligence is an effective way to achieve this goal, and its core lies in building a high-precision computer-aided diagnosis model. However, the training of such models relies on large-scale and professionally annotated image datasets, which are costly and time-consuming to obtain, limiting the rapid development and clinical deployment of related technologies.

[0003] To address the problem of insufficient labeled data, semi-supervised learning techniques are introduced, with self-training being one of the commonly used methods. Self-training methods use existing models to predict unlabeled data, and use high-confidence prediction results as pseudo-labels, which are added to the training set along with the corresponding data for subsequent model iteration. Some technologies use ensemble learning strategies to improve the accuracy of pseudo-labels, that is, to reduce the impact of single model prediction bias by making comprehensive judgments based on the prediction results of multiple models.

[0004] Existing integrated self-training methods still have technical limitations in application. Their pseudo-label screening mechanism usually only relies on prediction confidence, which cannot guarantee the correctness of the pseudo-labels, thus may cause error accumulation and damage the stability of the model, and does not consider the differentiated value of different candidate samples for model optimization, resulting in low training efficiency. In addition, such methods completely retrain the ensemble model at each iteration, which is computationally expensive and limits the overall practicality of the method.

[0005] Therefore, the present application proposes a semi-supervised width breast cancer diagnosis method based on integrated self-training to solve the problems of the prior art. SUMMARY

[0006] To address the deficiencies of the prior art, the present application provides a semi-supervised width breast cancer diagnosis method based on integrated self-training, which solves the technical problems of error accumulation caused by relying solely on prediction confidence to screen pseudo-labels and the high computational overhead caused by complete retraining of the model at each iteration.

[0007] To achieve the above purpose, the present application is implemented by the following technical solutions:

[0008] The first aspect of the present application provides a semi-supervised wide breast cancer diagnosis method based on integrated self-training, comprising the following steps:

[0009] S1, initializing an integrated diagnosis model based on labeled breast cancer image data samples, wherein the integrated diagnosis model comprises a plurality of sub-diagnosis models;

[0010] S2, using the integrated diagnosis model to predict unlabeled breast cancer image data samples to generate candidate pseudo-label samples;

[0011] S3, determining the detection priority of the candidate pseudo-label samples according to a preset scheduling strategy to form a priority detection queue;

[0012] S4, according to the order of the priority detection queue, the candidate pseudo-label samples are subjected to disturbance influence detection to obtain a stability influence detection result corresponding to each of the candidate pseudo-label samples;

[0013] S5, based on the stability influence detection result, high-quality pseudo-label samples are selected from the candidate pseudo-label samples;

[0014] S6, using the high-quality pseudo-label samples, updating the integrated diagnosis model to complete an integrated self-training iteration.

[0015] The second aspect of the present application provides a computer device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to realize the method according to any one of claims 1-9.

[0016] The present application provides a semi-supervised wide breast cancer diagnosis method based on integrated self-training.

[0017] 1、The present application constructs an integrated diagnosis model comprising a plurality of sub-diagnosis models, and designs a high-quality sample screening mechanism based on disturbance influence detection, effectively solving the error accumulation problem in the traditional semi-supervised self-training method. Before accepting a high-confidence candidate pseudo-label sample, the method will detect the influence of its addition on the stability of the whole model decision and internal representation. Only when the influence is within a controllable range, the sample is determined as a high-quality sample. This cautious screening strategy can filter out potential noise and incorrect labels from the source, ensuring the quality of pseudo-labels for model updating, and thus ensuring the steady and reliable improvement of the performance of the semi-supervised breast cancer diagnosis model.

[0018] 2、The application realizes efficient and stable utilization of unlabeled data by introducing a priority scheduling strategy based on information entropy gain and combining subsequent stability detection. The scheduling strategy pays more attention to candidate samples with large prediction differences in the integrated diagnostic model, because these samples often carry the most valuable information for optimizing the model decision boundary. After locating these high-information samples, stability detection is used to ensure that they do not harm the model. This "exploration first, verification later" mechanism enables the model to learn the most challenging and informative knowledge first, while avoiding violent oscillation in the learning process, significantly enhancing the robustness of the semi-supervised breast cancer diagnosis process.

[0019] 3、The application solves the problem of huge computational cost in repeated iterative training in semi-supervised learning by using a width learning system as a sub-diagnostic model and designing an efficient incremental update strategy based on its mathematical properties. In each integrated self-training iteration, after selecting high-quality pseudo-label samples, the method does not need to perform complete retraining on the entire integrated diagnostic model. Instead, it directly updates the output weights of each sub-diagnostic model through analytical solution. This feature significantly reduces the time cost of training, making the entire diagnostic method highly efficient and practical in handling large-scale breast cancer image data sets, especially suitable for clinical auxiliary diagnosis scenarios that require rapid iteration. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a flowchart of a semi-supervised width breast cancer diagnosis method based on integrated self-training according to the present application;

[0021] Figure 2 is a schematic diagram of the structure of a computer device according to the present application.

[0022] Among them, 100, computer device; 101, processor; 102, memory; 103, bus. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings of the present application specification. Obviously, the described embodiments are only a part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0024] Please refer to Figure 1, which is a flowchart of a semi-supervised width breast cancer diagnosis method based on integrated self-training according to the present application. The present application provides a semi-supervised width breast cancer diagnosis method based on integrated self-training, which can include the following steps:

[0025] S1, initializing an integrated diagnosis model based on the labeled breast cancer image data samples, the integrated diagnosis model comprising a plurality of sub-diagnosis models;

[0026] S2, predicting the unlabeled breast cancer image data samples by using the integrated diagnosis model to generate candidate pseudo-label samples;

[0027] S3, determining a detection priority of the candidate pseudo-label samples according to a preset scheduling strategy to form a priority detection queue;

[0028] S4, performing perturbation influence detection on the candidate pseudo-label samples in the order of the priority detection queue to obtain a stability influence detection result corresponding to each candidate pseudo-label sample;

[0029] S5, screening high-quality pseudo-label samples from the candidate pseudo-label samples based on the stability influence detection result;

[0030] S6, updating the integrated diagnosis model by using the high-quality pseudo-label samples to complete an integrated self-training iteration.

[0031] Referring to FIG. 1, in step S1, the method first initializes an integrated diagnosis model by using the labeled breast cancer image data samples. It should be noted that the integrated diagnosis model is a composite model, which does not directly make a single end-to-end prediction, but makes a final diagnosis decision by combining the prediction results of a plurality of independent sub-diagnosis models. The sub-diagnosis model is the basic unit of the integrated diagnosis model, and each sub-diagnosis model itself is a complete model with independent prediction capability.

[0032] In the specific embodiments of the present application, the integrated diagnosis model is implemented as an integrated width learning system. Accordingly, the integrated width learning system is a diagnosis system set composed of a plurality of independent width learning systems, wherein each width learning system operates as a sub-diagnosis model.

[0033] The initialization process of a single width learning system as a sub-diagnosis model is as follows: the labeled breast cancer image data sample is taken as input data X, and the corresponding label is taken as Y. The input data X is first processed by the feature node layer. The feature node layer comprises n groups of feature mappings, the i-th group of feature mappings is transformed by a group of randomly generated weights W ei and bias β ei , the input data X is transformed into the i-th group of feature nodes Z i .

[0034] All n groups of feature nodes are combined to form a feature node matrix Z n =[Z1,Z2,…,Z n ]. The feature node matrix Zn Further input to the enhanced node layer. The enhanced node layer contains m sets of nonlinear transformation, the jth set of transformation is performed by a set of randomly generated weights W hj and bias β hj , the feature node matrix Z n is transformed into the jth set of enhanced node H j .

[0035] All m sets of enhanced nodes are combined and form the enhanced node matrix H m = [H1, H2, …, H m ]. Then, the feature node matrix Z n and the enhanced node matrix H m are concatenated along the column direction to form the augmented matrix A = [Z n | H m ] of the width learning system.

[0036] The output of the width learning system is the product of the augmented matrix A and an output weight matrix W. The core of the initialization process is to solve the output weight matrix W using the labeled input data X and its corresponding label data Y. Specifically, the output weight matrix W can be obtained by solving the analytical solution of the following ridge regression problem: W = (A T A + λI) -1 A T Y;

[0037] where W is the output weight matrix to be solved; A is the augmented matrix composed of feature nodes and enhanced nodes; Y is the label matrix corresponding to the labeled data; λ is a constant, which is a regularization coefficient used to control the complexity of the model; I is the identity matrix; A T is the transpose of the matrix A.

[0038] In order to construct an integrated width learning system containing multiple sub-diagnostic models, the method generates K width learning systems by performing the above initialization process K times. In each generation process, a set of completely new and independently generated random weights (i.e., W ei , β ei , W hj , β hj ) are used to construct feature nodes and enhanced nodes. Due to the difference in random weights, each width learning system uses the same labeled data for training, but its internal structure and the output weight matrix W obtained by final calculation are different. Thus, the initialization of the integrated diagnostic model containing K width learning systems as sub-diagnostic models is completed. Each sub-diagnostic model in the integrated diagnostic model has completed training and has the ability to predict new data, thereby laying the foundation for the subsequent semi-supervised learning step.

[0039] Please refer to Figure 1. After the initialization of the integrated diagnostic model is completed, step S2 uses the integrated diagnostic model to predict all unlabeled breast cancer image data samples. The purpose is to generate a set of candidate pseudo-label samples with high confidence, providing a data basis for subsequent self-training iterations.

[0040] In the specific implementation of this step, for any unlabeled breast cancer imaging data sample x in the dataset u This is then input into the initialized ensemble diagnostic model. Since this ensemble diagnostic model contains K independent sub-diagnostic models, therefore x u Each of these K sub-models will produce K independent prediction probability vectors p. k (x u ), where k = 1, 2, ..., K. Then, by taking the arithmetic mean of these K predicted probability vectors, the x-value is calculated. u Integrated prediction probability vector

[0041] in, For x u The ensemble prediction probability vector; K is the total number of sub-diagnostic models in the ensemble diagnostic model; p k (x u ) represents the k-th sub-diagnostic model for sample x u The output is the predicted probability vector.

[0042] Based on this integrated prediction probability vector That can determine the relationship with x u The corresponding prediction result. This prediction result contains two parts: the predicted category y′ u and prediction confidence level c u Predict category y′ u The class is determined by the element with the maximum value in the ensemble prediction probability vector. Prediction confidence c u Then, the maximum probability value is directly used. Applying this process to all unlabeled samples yields a set of prediction results that correspond one-to-one with the unlabeled breast cancer imaging data samples and include prediction confidence.

[0043] This method selects predictions with a confidence level higher than a preset confidence threshold τ from the obtained prediction result set. conf (For example, τ) conf The prediction result can be set to 0.95. This step involves calculating the prediction confidence level c for each sample. u Compared with the preset confidence threshold τ conf This is achieved through comparison, retaining only those that satisfy condition c. u >τ conf Those samples and their prediction results.

[0044] Each filtered prediction result is combined with the original unlabeled breast cancer image data sample from which the prediction result is generated to form a candidate pseudo-label sample. Specifically, a candidate pseudo-label sample is formed by a data pair (x u , y′ u ), where x u is the unlabeled breast cancer image data sample whose confidence meets the requirement, and y′ u is the predicted class determined by the ensemble diagnostic model. The set of all these data pairs, i.e., the candidate pseudo-label sample set generated in this step.

[0045] Referring to FIG. 1, step S3 processes the candidate pseudo-label sample set generated in step S2, aiming to determine a probe priority for each candidate pseudo-label sample according to a preset scheduling strategy, and finally form an ordered priority probe queue. In this embodiment, the preset scheduling strategy is a priority ordering strategy based on information entropy gain, the core of which is to quantify the degree of disagreement of the ensemble diagnostic model on the prediction result of each candidate pseudo-label sample.

[0046] The implementation process of the scheduling strategy is carried out for each candidate pseudo-label sample. It is necessary to determine the ensemble prediction class corresponding to the candidate pseudo-label sample. The ensemble prediction class is defined as the class with the highest average prediction probability of the unlabeled breast cancer image data sample constituting the candidate pseudo-label sample by multiple sub-diagnostic models. In specific calculation, for x u , the ensemble prediction probability vector is obtained by calculating the arithmetic mean of the prediction probability vectors output by all K sub-diagnostic models. The class corresponding to the maximum element in the vector is the ensemble prediction class c * .

[0047] where c * is the ensemble prediction class of x u ; and is the value of the jth element in the ensemble prediction probability vector , representing the average prediction probability of class j; arg max j represents the class index j that makes the value of the expression maximum.

[0048] After determining the ensemble prediction class c * , the information entropy gain G(x u ) corresponding to the candidate pseudo-label sample is calculated. The information entropy gain is specifically calculated by calculating the information entropy gain of the ensemble prediction class c *The variance of the prediction probability of the candidate pseudo-label sample is obtained. The calculated variance is determined as the information entropy gain corresponding to the candidate pseudo-label sample. The calculation formula is as follows:

[0049] wherein G(x u ) is the information entropy gain corresponding to x u ; K is the total number of sub-diagnosis models; p k (c * |x u ) is the prediction probability value corresponding to the integrated prediction class c * in the prediction result output by the kth sub-diagnosis model for x u . is the average value of the prediction probability of all sub-diagnosis models for the integrated prediction class c * , that is, the element value of the integrated prediction class c * in the integrated prediction probability vector .

[0050] This calculation process is applied to all candidate pseudo-label samples, so that a quantitative information entropy gain value is obtained for each sample.

[0051] According to the values of the information entropy gain G(x u ) corresponding to all candidate pseudo-label samples, the candidate pseudo-label samples are sorted in descending order. The ordered sample queue formed after the sorting is the priority detection queue finally output in this step. In this way, the candidate pseudo-label samples that cause greater differences between the sub-diagnosis models (i.e., the information entropy gain values are higher) will be placed at the front of the queue, so that they will be preferentially detected in the subsequent steps.

[0052] Referring to FIG. 1, steps S4 and S5 are sequentially performed, and the core purpose is to carefully investigate and screen the candidate pseudo-label samples with high priority, so as to ensure that only the samples beneficial to the model stability are finally adopted.

[0053] This process performs perturbation influence detection on the candidate pseudo-label samples in the queue one by one according to the order of the priority detection queue formed in step S3. For a candidate pseudo-label sample (x u , y′ u ) selected from the queue to be detected.

[0054] A temporary candidate model, that is, an instantaneous diagnosis model, which is functionally identical to the current integrated diagnosis model, is created in the memory of the computing device. This avoids direct modification of the original integrated diagnosis model.

[0055] The instant diagnosis model is updated once using the candidate pseudo-label sample to be detected. Since the sub-diagnosis model in the embodiment adopts a width learning system, the output weight can be quickly reconstructed by an analytical solution, so the incremental updating process does not need complete retraining, and the calculation efficiency is very high.

[0056] After the update is completed, the stability change of the instant diagnosis model compared with the original integrated diagnosis model is evaluated based on a preset verification subset D v ( which remains fixed during the entire execution of the method), and the quantitative evaluation result of the stability change is taken as the stability influence detection result corresponding to the candidate pseudo-label sample to be detected.

[0057] Specifically, the stability influence detection result is obtained by calculating two key indicators.

[0058] The first indicator is a decision stability factor S D . The decision stability factor is used to measure the consistency degree of the prediction decisions of the instant diagnosis model and the integrated diagnosis model on the verification subset. The calculation method is as follows:

[0059] Wherein, |D v | is the total number of samples in the verification subset D v ; x v is a sample in the verification subset; f(x v ) is the predicted class of the sample x v by the original integrated diagnosis model; f ′ (x v ) is the predicted class of the sample x v by the instant diagnosis model; and I(·) is an indicator function, which is 1 when the internal condition is true, and 0 otherwise.

[0060] The second indicator is a representation consistency factor S R . The representation consistency factor is used to measure the consistency degree of the internal feature representation generated by the instant diagnosis model and the integrated diagnosis model for the verification subset. The calculation method is as follows:

[0061] Wherein, A(x v ) is the internal feature representation generated by the original integrated diagnosis model for the sample x v , which is specifically the augmented matrix of the width learning system in the embodiment; and A ′ (x v ) is the internal feature representation generated by the instant diagnosis model for the sample x vgenerated internal feature representation; · denotes vector dot product, ||·|| denotes L2 norm of vector. The fraction is the cosine similarity between two feature representation vectors.

[0062] After obtaining the decision stability factor S D and the representation consistency factor S R , step S5 is to select high-quality pseudo-label samples from the candidate pseudo-label samples based on the stability influence detection result.

[0063] The implementation of the selection is to determine whether the decision stability factor S D corresponding to a certain candidate pseudo-label sample is not less than a preset decision stability threshold τ D (For example, τ D may be set to 0.995), and whether the corresponding representation consistency factor S R is not less than a preset representation consistency threshold τ R (For example, τ R may be set to 0.999).

[0064] If the determination result is yes, that is, both S D ≥τ D and S R ≥τ R are satisfied, the candidate pseudo-label sample is determined as a high-quality pseudo-label sample and is added to a high-quality sample set. Otherwise, if any condition is not satisfied, it is considered that the candidate pseudo-label sample will cause damage to the model stability, and it is directly discarded. This process is sequentially performed on the samples in the priority detection queue until the queue is processed.

[0065] Referring to FIG. 1, step S6 is the last link of completing one integrated self-training iteration. In this step, all high-quality pseudo-label samples selected in step S5 are used to perform a formal update on the current integrated diagnostic model.

[0066] The update process is specifically implemented as an incremental update on the integrated diagnostic model. Since the integrated diagnostic model in the embodiment is an integrated wide learning system including K wide learning systems as sub-diagnostic models, the incremental update is applied to each independent sub-diagnostic model respectively.

[0067] For the kth sub-diagnostic model (k = 1, …, K), the incremental update process is as follows:

[0068] Let D qual be the set of high-quality pseudo-label samples selected in the current iteration, the corresponding breast cancer image data sample be X qual , and the pseudo-label be Y qual .

[0069] X samples of breast cancer imaging data qual The data is input into the k-th sub-diagnostic model, and its inherent random weight parameters are used to generate an augmented matrix A corresponding to this batch of new data. inc,k .

[0070] The core of incremental updates in a wide-range learning system lies in efficiently updating the key matrix required to solve for the output weights without repeatedly calculating them on all historical data. Specifically, let P be the inverse matrix used to solve for the output weights of the k-th sub-diagnostic model in its state before the current update. k The inverse matrix is ​​calculated as follows: Where A k This is an augmented matrix representing all the data used by the sub-diagnostic model. When a new breast cancer imaging data sample X is introduced... qual Then, the new inverse matrix P′ k The following formula can be used to obtain P k Obtained directly through update:

[0071] Among them, P′ k P is the inverse matrix after the k-th sub-diagnostic model is updated; k A is the inverse matrix before the k-th sub-diagnostic model is updated; inc,k I is the augmented matrix generated from the high-quality pseudo-label sample data in the kth sub-diagnostic model; I is the identity matrix. Let matrix A i The transpose of nc,k.

[0072] After obtaining the updated inverse matrix P′ k Then, the output weight matrix W of the k-th sub-diagnostic model k It is also updated accordingly. The new output weight matrix W′ k The calculation formula is:

[0073] Among them, W′ k W is the updated output weight matrix of the k-th sub-diagnostic model; k Y is the output weight matrix before the k-th sub-diagnostic model is updated; qual This is the pseudo-label matrix corresponding to the high-quality pseudo-label samples.

[0074] Applying the above incremental update process to all K sub-diagnostic models in the ensemble diagnostic model completes the update of the entire ensemble diagnostic model. This concludes one complete ensemble self-training iteration. Afterward, the method can return to step S2, using the updated ensemble diagnostic model to perform a new round of prediction and filtering on the remaining unlabeled data, repeating this process until the preset termination condition is met.

[0075] Referring to FIG. 2, FIG. 2 is a schematic diagram of a computer device structure according to an embodiment of the present application. The embodiment of the present application also provides a computer device, which can be a server, a workstation, a personal computer or any other electronic device with comparable computing capability.

[0076] The computer device 100 can include at least one processor 101, a memory 102 and a bus 103 for connecting the above components in hardware. The memory 102 stores a set of computer program instructions, which, when executed by the processor 101, can implement the semi-supervised breast cancer diagnosis method based on integrated self-training described in the foregoing embodiments. In functional logic, the instructions can be divided into a plurality of functional modules, each of which is responsible for executing a specific step of the method.

[0077] Specifically, the computer device 100 includes:

[0078] a model initialization module for performing the S1 step;

[0079] a candidate sample generation module for performing the S2 step;

[0080] a priority scheduling module for performing the S3 step;

[0081] a perturbation influence detection module for performing the S4 step;

[0082] a high-quality sample screening module for performing the S5 step;

[0083] a main model updating module for performing the S6 step.

[0084] During the execution of the entire method, these modules are invoked by the processor 101 and work cooperatively. Data, such as the initial labeled and unlabeled data set, the intermediate generated candidate pseudo-label sample, the priority detection queue and the model parameters in the updating process, are stored and read and written in the memory 102. By cyclically executing the functions of the candidate sample generation, priority scheduling, perturbation detection, sample screening and main model updating modules, the computer device can continuously and stably use unlabeled data to improve the performance of the diagnosis model.

[0085] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A semi-supervised breast cancer diagnosis method based on integrated self-training, characterized in that, The method comprises the following steps: S1, initializing an integrated diagnosis model based on labeled breast cancer image data samples, the integrated diagnosis model comprising a plurality of sub-diagnosis models; S2, predicting unlabeled breast cancer image data samples using the integrated diagnosis model to generate candidate pseudo-label samples; S3, determining the detection priority of the candidate pseudo-label samples according to a preset scheduling strategy to form a priority detection queue; S4, performing perturbation influence detection on the candidate pseudo-label samples in the order of the priority detection queue to obtain a stability influence detection result corresponding to each candidate pseudo-label sample; S5, screening high-quality pseudo-label samples from the candidate pseudo-label samples based on the stability influence detection result; S6, updating the integrated diagnosis model using the high-quality pseudo-label samples to complete an integrated self-training iteration.

2. The semi-supervised breast cancer diagnosis method based on integrated self-training according to claim 1, wherein, In step S2, the step of predicting unlabeled breast cancer image data samples using the integrated diagnosis model to generate candidate pseudo-label samples comprises: predicting the unlabeled breast cancer image data samples using the integrated diagnosis model to obtain a group of prediction results corresponding one-to-one to the unlabeled breast cancer image data samples and containing prediction confidence; screening prediction results with prediction confidence higher than a preset confidence threshold from the group of prediction results, and combining each screened prediction result with the unlabeled breast cancer image data sample from which the prediction result is generated to constitute the candidate pseudo-label sample.

3. The semi-supervised breast cancer diagnosis method based on integrated self-training according to claim 1, wherein, In step S3, the preset scheduling strategy is a priority sorting strategy based on information entropy gain, and the step of determining the detection priority of the candidate pseudo-label samples according to the preset scheduling strategy to form a priority detection queue comprises: determining the integrated prediction category corresponding to each candidate pseudo-label sample, the integrated prediction category being the category with the highest average prediction probability of the plurality of sub-diagnosis models for the unlabeled breast cancer image data sample constituting the candidate pseudo-label sample; calculating the variance of the prediction probabilities of the integrated prediction category assigned by the plurality of sub-diagnosis models, and determining the variance as the information entropy gain corresponding to the candidate pseudo-label sample; sequentially sorting the candidate pseudo-label samples in descending order according to the values of the information entropy gains to form the priority detection queue.

4. The semi-supervised breast cancer diagnosis method based on integrated self-training according to claim 1, wherein, In step S4, the step of performing perturbation influence detection on the candidate pseudo-label samples in the order of the priority detection queue to obtain a stability influence detection result corresponding to each candidate pseudo-label sample comprises: selecting a candidate pseudo-label sample to be detected from the priority detection queue; creating a transient diagnosis model based on the current integrated diagnosis model; incrementally updating the transient diagnosis model using the candidate pseudo-label sample to be detected; Based on a preset verification subset divided from the labeled breast cancer image data samples, a stability change of the instantaneous diagnosis model relative to the ensemble diagnosis model is evaluated as a stability influence detection result corresponding to the candidate pseudo-label samples to be detected.

5. The semi-supervised breast cancer diagnosis method based on integrated self-training according to claim 4, wherein, The step of evaluating the stability change of the instantaneous diagnosis model relative to the ensemble diagnosis model comprises: calculating a decision stability factor for measuring a consistency degree of prediction decisions of the instantaneous diagnosis model and the ensemble diagnosis model on the verification subset.

6. The semi-supervised breast cancer diagnosis method based on integrated self-training according to claim 5, wherein, The step of evaluating the stability change of the instantaneous diagnosis model relative to the ensemble diagnosis model further comprises: calculating a representation consistency factor for measuring a consistency degree of internal feature representations generated by the instantaneous diagnosis model and the ensemble diagnosis model for the verification subset.

7. The semi-supervised breast cancer diagnosis method based on integrated self-training according to claim 6, wherein, In step S5, the step of screening high-quality pseudo-label samples from the candidate pseudo-label samples based on the stability influence detection result comprises: determining whether the decision stability factor corresponding to a certain candidate pseudo-label sample is not less than a preset decision stability threshold and whether the representation consistency factor is not less than a preset representation consistency threshold; if the determination result is yes, the candidate pseudo-label sample is determined as a high-quality pseudo-label sample.

8. The semi-supervised breast cancer diagnosis method based on integrated self-training of claim 1, wherein, The ensemble diagnosis model is an ensemble width learning system containing multiple width learning systems as sub-diagnosis models.

9. The semi-supervised breast cancer diagnosis method based on integrated self-training of claim 1, wherein, The step of S6 specifically comprises: using the high-quality pseudo-label samples to realize incremental update of the ensemble diagnosis model by updating an inverse matrix of each width learning system for solving an output weight and calculating a new output weight matrix based on the updated inverse matrix.

10. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the method of any one of claims 1-9. The processor executes the computer program to realize the method of any one of claims 1-9.

Images