A semi-supervised width-based breast cancer diagnostic method based on ensemble self-training

Abstract

This invention discloses a semi-supervised width-based breast cancer diagnosis method based on ensemble self-training, comprising: S1, acquiring, labeling, and preprocessing breast cancer image data; S2, training multiple homogeneous labeled BLSs; S3, all homogeneous labeled BLSs voting to generate pre-labels; S4, calculating the purity of classification labels based on the votes, selecting a set of unlabeled breast cancer image data, assigning pseudo-labels to them, and using them as auxiliary training data; S5, adjusting the structure of all labeled BLSs using a dynamic node mechanism, adding the auxiliary training data to the training set, and training these labeled BLSs; S6, if the termination condition is met, training the final decision BLS using the currently labeled breast cancer image data and applying it to actual breast cancer data to output the diagnostic result. This invention combines ensemble learning and self-training mechanisms to solve the labeling and imbalance problems in the field of breast cancer diagnosis, and uses width learning as a base classifier to avoid the problems of high cost and difficult deployment of deep learning-based breast cancer diagnosis methods.

Description

Technical Field

[0001] This invention belongs to the field of medical diagnostic technology, specifically relating to a semi-supervised width breast cancer diagnostic method based on integrated self-training. Background Technology

[0002] Currently, breast cancer diagnostic technologies are mainly divided into those based on traditional medical methods and those based on deep learning. Traditional medical methods utilize reagents to detect breast cancer or observe medical images. These methods have a high recognition rate, but require the cooperation of professional medical personnel, and the diagnostic process is time-consuming. Compared with traditional medical methods, intelligent breast cancer diagnostic technology based on deep learning has unparalleled advantages: it eliminates the need to wait for specialists, provides fast diagnosis, and is applicable to various medical images. However, deep learning-based methods are costly, have large parameter sets, are difficult to deploy in a lightweight manner, and cannot meet the needs of online real-time processing. On the other hand, due to factors such as privacy, high cancer treatment costs, and difficulties in long-term tracking, breast cancer data suffers from problems such as limited data and a lack of diagnostic or prognostic information (i.e., labels), which poses a significant challenge to training large-scale deep learning models. In addition, the number of cancer patients is relatively small, and breast cancer data exhibits severe imbalance. Therefore, designing a breast cancer diagnostic technology with high data tolerance (low data requirements, weak label restrictions), low time consumption, high robustness, and real-time response is crucial. Summary of the Invention

[0003] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and propose a semi-supervised width breast cancer diagnosis method based on integrated self-training.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A semi-supervised width-based breast cancer diagnosis method based on ensemble self-training includes the following steps:

[0006] S1. Collect and label breast cancer image data, and preprocess the image data;

[0007] S2. Multiple homogeneous label width learning models are trained in parallel using labeled breast cancer image data.

[0008] S3. All homogeneous label width learning models vote on unlabeled breast cancer image data to generate pre-labels;

[0009] S4. Calculate the classification label purity based on the voting of the homogeneous label width learning model, select a suitable set of unlabeled breast cancer image data, assign pseudo-labels to them, and use them as auxiliary training data.

[0010] S5. Use a data-driven dynamic node mechanism to adjust the structure of all label width learning models and add auxiliary training data to the training set to train these label width learning models.

[0011] S6. Determine whether the preset termination condition has been met. If yes, train the final decision width learning model using the current labeled breast cancer image data and apply it to the actual breast cancer data, and output the diagnosis result. If no, proceed to S2 and perform subsequent steps until the preset termination condition is met.

[0012] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0013] 1. In response to the problems of high resource consumption and long training time in current mainstream deep learning-based breast cancer diagnosis methods, this invention is based on width learning, which uses fewer parameters to achieve high recognition accuracy and response speed, and facilitates rapid model development and technology deployment.

[0014] 2. This invention introduces a self-training framework that utilizes unlabeled data to effectively solve the problem of missing prognostic information for breast cancer caused by privacy concerns, tracking interruptions, and other reasons. It fully utilizes breast cancer feature information and improves the recognition performance of intelligent methods.

[0015] 3. This invention solves the problem of performance instability in imbalanced data scenarios by using voting by multiple labeled BLS classifiers and label purity to determine the pseudo-label of unlabeled data, thereby improving the robustness of the model and making it suitable for imbalanced and data drift scenarios. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention;

[0017] Figure 2 This is a flowchart of the training process of the width learning model in the method of this invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0019] Example

[0020] like Figure 1 As shown, the present invention, a semi-supervised width breast cancer diagnosis method based on ensemble self-training, includes the following steps:

[0021] S1. Collect and label breast cancer image data, and preprocess the image data;

[0022] In this method, breast cancer image data includes three types, specifically:

[0023] Raw labeled breast cancer image data Unlabeled breast cancer image data Auxiliary training data Where r, s, and t represent the amount of original labeled breast cancer image data, unlabeled breast cancer image data, and auxiliary training data, respectively.

[0024] In step S1, the preprocessing specifically involves dividing the breast cancer image data into labeled breast cancer image data and unlabeled breast cancer image data.

[0025] S2. Multiple homogeneous label width learning models (BLS) are trained in parallel using labeled breast cancer image data;

[0026] Among them, training multiple homogeneous label width learning models adopts multi-threaded parallel computing, such as... Figure 2 As shown, the specific steps include:

[0027] Initialize the parameters of the width learning model, including the number of mapped feature groups, the number of features within each group, and the number of augmentation nodes;

[0028] For breast cancer data X, first calculate its feature mapping feature Z. n ≡[Z1, Z2, ..., Z n The process is a linear mapping, where the i-th group of mapping features Z i The calculation is as shown in formula (1), and the connection weight W ei and β ei It is generated randomly. For linear mappings, each set of features contains k mapping features;

[0029]

[0030] By mapping node Z n Further calculation of the enhancement feature H corresponding to breast cancer m ≡[H1, H2, ..., H m This process is a non-linear mapping, where the enhanced feature H... j The calculation is as shown in formula (2), and the connection weight W hj and β hj It is randomly generated, ζ(Z, W) hj ,β hj () is a non-linear activation function;

[0031] H j ≡ζ(Z n W hj +β hj ), j = 1, 2, ..., m(2)

[0032] Mapping nodes represent the linear features of breast cancer image data, while enhancement nodes represent the nonlinear features of breast cancer image data.

[0033] Linear and nonlinear features are combined into a feature layer A = [Z|H], which is then connected to the model output layer Y.

[0034] Since the information of the model's output layer Y is known, namely the diagnosis corresponding to the breast cancer image data, it is only necessary to calculate the connection weights W between the feature layer and the output layer. The calculation formula is as follows:

[0035] W = A -1 Y (3)

[0036] In the calculation, ridge regression is used to solve for the weight matrix, that is, W is solved by the optimization problem of formula (3):

[0037]

[0038] Where σ1=σ2=v=u=2, we can solve for:

[0039] W=(λI+AA T ) -1 A T Y (5).

[0040] S3. All homogeneous label width learning models vote on unlabeled breast cancer image data to generate pre-labels; the specific pre-labels generated by the voting of each label width learning model are obtained according to formula (6):

[0041] y′ p =argmax(y) (6)

[0042] Where y is the output of the marking width learning model, expressed as formula (7):

[0043]

[0044] S4. Calculate the classification label purity based on the voting of the homogeneous label width learning model, select a suitable set of unlabeled breast cancer image data, assign pseudo-labels to them, and use them as auxiliary training data.

[0045] The formula for calculating the purity of the classification label is as follows:

[0046]

[0047] Among them, f m It is the frequency of the mode class in the voting results; #(f i ! = 0) indicates that f i The number of non-zero numbers; N bls N represents the number of BLS labels in the labeling classifier. CIndicates the number of categories in the dataset; and These represent the concentration and dispersion of the voting results of multiple labeling classifiers, respectively. Typically, only instances with high label purity, i.e., high concentration and low dispersion, can be labeled by the voting results of most classifiers.

[0048] Select a suitable set of unlabeled breast cancer image data, assign pseudo-labels to it, and use it as auxiliary training data, specifically:

[0049] Pseudo-labels are assigned only to image data whose classification label purity is greater than a threshold, i.e., unlabeled data where LP >> ∈. The remaining image data are unlabeled. This process is represented by formula (9):

[0050]

[0051] Wherein, ∈ is the preset label purity threshold, and its value is between [0, 1].

[0052] S5. Use a data-driven dynamic node mechanism to adjust the structure of all label width learning models and add auxiliary training data to the training set to train these label width learning models.

[0053] The dynamic node mechanism specifically includes:

[0054] The number of augmentation nodes in the label width learning model is adjusted according to the number of labeled breast cancer image data l(t) currently used in the label width learning model and the number of auxiliary training data im(t) added. The new number of nodes m′ is calculated by formula (10):

[0055]

[0056] Where m(t) is the number of augmentation nodes in the current label width learning model, im(t) is the amount of auxiliary training data, and l(t) is the amount of labeled breast cancer image data used by the current label width model. It is an up-rounding operation.

[0057] S6. Determine whether the preset termination condition has been met. If yes, train the final decision width learning model using the current labeled breast cancer image data and apply it to the actual breast cancer data to output the diagnosis result. If no, proceed to S2 and continue with the subsequent steps until the preset termination condition is met.

[0058] The preset termination condition is either that the number of unlabeled samples is 0 or the maximum number of iterations is reached. Generally, in this method, most of the unlabeled data is added to the training data in the first 20 generations, and the improvement in model performance is not significant in the later stages. Therefore, the termination condition is usually set to the maximum number of iterations.

[0059] To further describe the technical effects of the method of the present invention, the publicly available BreastMNIST dataset was used for testing. This dataset contains 780 breast ultrasound images, divided into three categories: normal, benign, and malignant. In this embodiment, the task is simplified to binary classification, combining normal and benign images as positive and classifying malignant images as negative. The dataset is divided into training, validation, and test sets in a 7:1:2 ratio. The training set contains 54 labeled data entries and 492 unlabeled data entries.

[0060] Three models were compared. Traditional statistical methods included S3VM, TSVM, LapSVM, PseudoLabel, SemiBoost, and STSVM. Deep learning methods included SSVAE, DBN, Ladder Network, ParsNet, SkipE-RNN, π-Model, and Temporal Ensembling. For shallow neural networks, SS-ELM, SS-HELM, SS-BLS, and EBLS3 were used. Among them, SemiBoost, π-Model, Temporal Ensembling, and EBLS3 all employed ensemble learning techniques. Table 1 below shows a performance comparison of different methods on the breast cancer diagnosis task.

[0061]

[0062]

[0063] Table 1

[0064] As shown in Table 1, deep learning methods achieved a maximum accuracy of 70.76% in breast cancer diagnosis, but their time consumption was also relatively high, approximately 530 seconds. In contrast, the method of this invention achieved the highest recognition performance, with an accuracy of 73.72%, while consuming only 0.21 seconds. The three manifold regularization-based methods, SS-ELM, SS-HELM, and SS-BLS, exhibited lower performance because they could not construct an effective connected graph when the data volume was small, thus affecting the recognition effect. Furthermore, it can be observed that many deep learning methods have relatively low diagnostic accuracy. This is because deep learning methods rely on large-scale data, while in breast cancer diagnosis, the data volume often cannot meet the data requirements of most deep learning methods, thus leading to performance degradation. The above results demonstrate that the method of this invention effectively solves the problems of current deep learning-based methods and can improve model performance using unlabeled data.

[0065] It should also be noted that, in this specification, terms such as "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0066] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A semi-supervised width-based breast cancer diagnosis method based on ensemble self-training, characterized in that, Includes the following steps: S1. Collect and label breast cancer image data, and preprocess the image data; S2. Multiple homogeneous label width learning models are trained in parallel using labeled breast cancer image data. S3. All homogeneous label width learning models vote on unlabeled breast cancer image data to generate pre-labels; S4. Calculate the classification label purity based on the voting of the homogeneous label width learning model, select a suitable set of unlabeled breast cancer image data, assign pseudo-labels to them, and use them as auxiliary training data. S5. Use a data-driven dynamic node mechanism to adjust the structure of all label width learning models and add auxiliary training data to the training set to train these label width learning models. S6. Determine whether the preset termination condition has been met. If yes, train the final decision width learning model using the current labeled breast cancer image data and apply it to the actual breast cancer data to output the diagnosis result. If no, proceed to S2 and continue with the subsequent steps until the preset termination condition is met. Step S3 is as follows: The pre-labels generated by the voting of the learning model for each label width are obtained according to formula (6): and' p =argmax(y)(6) Where y is the output of the marking width learning model, expressed as formula (7): In step S4, the formula for calculating the purity of the classification label is: Among them, f m It is the frequency of the mode class in the voting results; #(f i ! = 0) indicates that f i The number of non-zero numbers; N bls N represents the number of BLS labels in the labeling classifier. C Indicates the number of categories in the dataset; and These represent the degree of concentration and dispersion of the voting results from multiple labeling classifiers, respectively.

2. The semi-supervised width breast cancer diagnosis method according to claim 1, characterized in that, The method includes three types of breast cancer image data, specifically: Raw labeled breast cancer image data Unlabeled breast cancer image data Auxiliary training data Where r, s, and t represent the amount of original labeled breast cancer image data, unlabeled breast cancer image data, and auxiliary training data, respectively.

3. The semi-supervised width breast cancer diagnosis method according to claim 1, characterized in that, In step S1, the preprocessing specifically involves dividing the breast cancer image data into labeled breast cancer image data and unlabeled breast cancer image data.

4. The semi-supervised width breast cancer diagnosis method according to claim 1, characterized in that, In step S2, training multiple homogeneous label width learning models employs multi-threaded parallel computation, specifically including the following steps: Initialize the parameters of the width learning model, including the number of mapped feature groups, the number of features within each group, and the number of augmentation nodes; For breast cancer data X, first calculate its feature mapping feature Z. n ≡[Z1,Z2,…,Z n The process is a linear mapping, where the i-th group of mapping features Z i The calculation is as shown in formula (1), and the connection weight W ei and β ei It is generated randomly. It is a linear mapping, and each set of features contains k mapping features; By mapping node Z n Further calculation of the enhancement feature H corresponding to breast cancer m ≡[H1,H2,…,H m This process is a non-linear mapping, where the enhanced feature H... j The calculation is as shown in formula (2), and the connection weight W hj and β hj It is randomly generated, ζ(Z,W) hj ,β hj () is a non-linear activation function; H j ≡ζ(Z n W hj +β hj ),j=1,2,…,m (2) Mapping nodes represent the linear features of breast cancer image data, while enhancement nodes represent the nonlinear features of breast cancer image data. Linear and nonlinear features are combined into a feature layer A = [Z|H], which is then connected to the model output layer Y. Since the information of the model's output layer Y is known, namely the diagnosis corresponding to the breast cancer image data, it is only necessary to calculate the connection weights W between the feature layer and the output layer. The calculation formula is as follows: W=A -1 Y (3) In the calculation, ridge regression is used to solve for the weight matrix, that is, W is solved by the optimization problem of formula (3): Where σ1=σ2=v=u=2, we can solve for: W=(λI+AA T ) -1 A T Y (5)。 5. The semi-supervised width breast cancer diagnosis method according to claim 1, characterized in that, In step S4, a suitable set of unlabeled breast cancer image data is selected, assigned pseudo-labels, and used as auxiliary training data, specifically as follows: Pseudo-labels are assigned only to image data whose classification label purity is greater than a threshold, i.e., unlabeled data where LP >> ∈. The remaining image data are unlabeled. This process is represented by formula (9): Wherein, ∈ is the preset label purity threshold, and its value is between [0,1].

6. The semi-supervised width breast cancer diagnosis method according to claim 1, characterized in that, In step S5, the dynamic node mechanism is specifically as follows: The number of augmentation nodes in the label width learning model is adjusted based on the amount of labeled breast cancer image data l(t) currently used in the model and the amount of auxiliary training data im(t) added. The new number of nodes m is then determined. ′ Calculated using formula (10): Where m(t) is the number of augmentation nodes in the current label width learning model, im(t) is the amount of auxiliary training data, and l(t) is the amount of labeled breast cancer image data used by the current label width model. It is an up-rounding operation.

7. The semi-supervised width breast cancer diagnosis method according to claim 1, characterized in that, In step S5, the preset termination condition is that the number of unlabeled samples is 0 or the maximum number of iterations is reached.

Images