Ultrasound image multi-center plaque classification noise label correction method based on double classifier prediction

Abstract

A noise label correction method for ultrasonic image multi-center plaque classification based on double-classifier prediction is characterized by comprising the steps that an ultrasonic image is preprocessed and then input into a feature extractor, a multi-center feature adaptation module is additionally arranged behind the feature extractor, a modeling and processing module for constructing noise labels is additionally arranged behind the feature extractor, and classification of multi-center plaque echoes is achieved; the multi-center feature adaptation module introduces a domain adaptation layer composed of two full-connection layers, uses the maximum average difference as the quantitative measurement of the difference, constructs two independent classifiers, and takes the output average of the two classifiers as the input of modeling of the next noise label. Aiming at inherent difficulties existing in multi-center plaque ultrasonic data, a multi-center feature adaptation module and a noise tag modeling and correction module are provided, while feature differences are further reduced, pseudo tag modeling is carried out on a target center, and better modeling estimation and processing are carried out on generated noise tags.

Description

Technical Field

[0001] The present invention relates to the field of computers, and in particular to a medical image classification method. Background Art

[0002] Cardiovascular and cerebrovascular diseases, such as stroke, have gradually become one of the diseases with the highest morbidity and mortality rates in the world. The World Health Organization has declared it a pandemic. One of the main causes of CVD is the continuous development of carotid atherosclerosis. Therefore, the concept of vulnerable plaques was proposed, and researchers prevented cardiovascular diseases by identifying atherosclerotic plaques with higher risks. In clinical practice, due to the shallow location of the carotid artery, B-ultrasound, as a non-invasive detection instrument, is currently one of the main means of preoperative diagnosis of carotid plaque properties in clinical practice. However, its imaging is greatly affected by the operator's technique and instrument. Different operators and different instruments have a certain impact on the final imaging effect. At the same time, due to the lack of a gold standard for diagnosis, the accuracy of diagnosis depends largely on the experience and level of clinicians, and the universality is poor. Therefore, it is very necessary to develop a computer-assisted diagnosis and treatment system based on multi-center plaque data.

[0003] At present, deep learning has been gradually applied to plaque-related fields as an efficient auxiliary diagnosis method, especially for plaque data from a single center, and has achieved good results. However, one of its prerequisites is that it requires a large amount of labeled data. However, the collection of medical data itself is different from natural images. The annotation of carotid plaque data also requires doctors with rich experience. These problems have limited the development of multi-center plaque risk auxiliary diagnosis systems to a certain extent. Therefore, unsupervised learning methods have been proposed to solve this type of problem. Unsupervised learning aims to train a domain-invariant classifier by bringing the data features of two different fields closer. This also means that unsupervised learning does not require the label of the target center data to classify it. In order to achieve this goal in the multi-center risk prediction system for carotid plaques, there are still some difficult problems that need to be solved.

[0004] For multi-center ultrasound of carotid plaques, the main difficulties are: 1) the patient's swallowing, the proficiency of the operating doctor, etc. will affect the final imaging. At the same time, due to the imaging characteristics of ultrasound, it will produce artifacts, and the resolution still needs to be improved. These problems will cause the characteristics of some data to be difficult to identify and classify, making it easy for the classifier to give noise labels. In addition, it will also affect the knowledge transfer between fields. 2) The consistency of ultrasound imaging features depends on the same ultrasound probe and the same doctor's operating techniques. That is to say, plaque ultrasound data from different centers will form a large imaging difference. Such differences cause feature differences between data from different centers, increasing the difficulty of aligning the features of data from different centers. It is worth noting that an efficient multi-center auxiliary diagnosis and treatment system relies on the alignment of data features from different sources and the accuracy of the classifier. Therefore, it is very necessary to overcome the above two difficulties. In order to solve the above difficulties, the present invention proposes a multi-center plaque risk prediction algorithm based on dual classifier noise label correction. Summary of the invention

[0005] The purpose of the present invention is that the traditional plaque risk prediction model requires experienced diagnostic doctors to label a large amount of plaque data, which is time-consuming and labor-intensive. Constructing a plaque risk prediction model suitable for multiple centers can reduce related dependencies to a certain extent. It still faces two major challenges: 1) Different doctors and operating instruments will cause characteristic differences between plaque data of different centers. 2) The quality of ultrasound imaging is seriously affected by noise such as artifacts and low resolution. These two major challenges have led to a decrease in the generalization performance of the model, especially when the single-center model is extended to multi-center data, its accuracy will be greatly reduced. By constructing a plaque risk prediction network based on dual classifier noise label correction, while shortening the feature distance of multiple center data, attention is paid to plaque data that is difficult to transfer and classify due to noise. The impact of plaque noise on prediction is handled by end-to-end modeling and correction of the data of the target center.

[0006] To achieve the above objectives, the concept of the present invention is: on the basis of bringing the multi-center data features closer, the collective wisdom of the dual classifiers is used to continuously correct the labels to transfer knowledge between different centers. Two independent classifiers can increase the complexity of the model, thereby reducing the risk of overfitting. This is particularly beneficial for the ultrasound multi-center problem, because when there is less target domain data, overfitting may lead to performance degradation. In more detail, we constructed a multi-layer domain adaptation layer for feature sharing to bring the feature differences of plaque data from different centers closer. After passing the adaptation layer, independent dual classifiers are used to work together to give pseudo-labels to the data of the target center. Based on the pseudo-labels, the labels are iteratively updated by learning the features and calculating the loss function, using a generative adversarial approach.

[0007] The technical solution of the present invention:

[0008] A noise label correction method for multi-center plaque classification in ultrasound images based on dual classifier prediction: First, the multi-center plaque image is preprocessed, and the feature distance between centers is shortened through the multi-center feature adaptation module. On this basis, the noise label is modeled and corrected, and the model prediction affected by noise is corrected and processed to achieve the classification of plaque echoes. Specific steps:

[0009] A. First, complete data preprocessing through image enhancement operations such as normalization and resolution adjustment.

[0010] The multi-center feature adaptation module introduces a domain adaptation layer consisting of two fully connected layers, and uses the maximum mean difference (MMD) as a quantitative measure of the difference. In addition, two independent classifiers are constructed to increase the differences between classifiers, which can help the model better capture the correct signal and reduce the impact of noise labels. Finally, the output of the two classifiers is averaged and used as the input for the next step.

[0011] This part involves two loss functions, and the calculation process includes:

[0012] B1. For a given source center image x s and its corresponding label y s , there will be a set of data sets from the source center {X s , Y s}. For the unlabeled image x at the target center t , there will be a set of data sets {X t}. Train a feature extractor G and two classifiers C1andC2. Ultrasound image data x of the source center and the target center s and x t Input the feature extraction network G respectively to obtain the corresponding feature representation.

[0013] B2, MMD loss, its calculation formula is as follows:

[0014]

[0015] B3. Difference loss uses the absolute value difference of the output of the classifier for the target center data to perform quantitative measurement. Its calculation formula is:

[0016]

[0017] B4. By maximizing L diversity and minimize L MMD , the model will get two classifiers that maximize the difference, which can increase the model's ability to discriminate data at the target center. The overall goal of this part is:

[0018] min(γL MMD -L diversity )(4)

[0019] B5. Through multi-center domain adaptation, the model can obtain the prediction results f1(x, θ) and f2(x, θ) from the two classifiers respectively, and average them to obtain a prediction distribution that is convenient for modeling, which is defined as follows:

[0020]

[0021] C. After obtaining the prediction for the label, it is further processed through the noise label modeling and processing module.

[0022] The modeling calculation process of the label is as follows:

[0023] C1. To model labels, the model will use two sets of label distributions. and To describe the data from the two centers. In the specific model learning process, each training and will be input into the model as the pseudo-true label. For label initialization, the source center will use its own true label y s For the patch data at the target center, there is no initial true label for initialization, so DCLCN uses the model pre-trained by the source center to give the target center a set of estimates of the distribution of noisy pseudo labels. The initial distribution estimate of the pseudo label at the target center includes not only the influence of image noise but also the noise label.

[0024] The label correction is calculated as follows:

[0025] C1. and There are two sets of labels that model the impact of noise. The model gradually corrects this part of the impact during continuous learning. To achieve this goal, the network parameters θ will be continuously updated to predict the label distribution of the source center and the target center. and will also be added to the training process and will be continuously updated during the training process, using Y d To represent their collection. Three types of tags will be involved, namely y. is the prediction result from the previous classifier, which is the probability distribution of a category; y is the true label of the data set (note that in this article, the data label of the target center will not be involved in the training); The pre-trained classifier gives the target center data a noise pseudo-label to initialize the target center; is a set of hard labels that are continuously updated by back propagation. The goals of this part are as follows:

[0026]

[0027] C2. For The source center will use y, which is the true label, and the target center will use the initial noise pseudo label To complete:

[0028]

[0029] Where K is a constant.

[0030] C3, After a softmax function, we get a set of continuously updated probability distributions y required by the model. d .

[0031]

[0032] C4, cross entropy loss: For the source center data, it has a true label. We use cross entropy loss to improve the model's classification ability for the source center:

[0033]

[0034] C5, classification loss: In order to reduce the impact of patch noise, the model needs to continuously update the label y d and parameter θ. The update of this part mainly depends on y d And the prediction results of the dual classifier in the previous part The DCLCN uses KL loss to measure the deviation.

[0035]

[0036] C6, compatibility loss: Although the pseudo-labels predicted by the model include the predictions of those patches that are not easily identifiable, in each subsequent round of training, the results obtained by the model should still be mostly the same as the initial pseudo-labels. This view has also been confirmed in previous studies. Therefore, the model adds another compatibility loss, which is defined as follows:

[0037]

[0038] C7, by minimizing the overall loss of label correction, can further correct the impact of noise on prediction. Calculated as follows:

[0039]

[0040] The final overall training steps are as follows:

[0041] D1. In order to obtain a preliminary model that is more suitable for multi-center plaque classification, DCLCN uses the plaque data from the source center to train the model. The training goal of this step is to minimize the cross entropy loss, and the goals are as follows:

[0042]

[0043] D2, add data from the target center for training. For a clearer algorithm flow, please refer to Algorithm 1. First, use the multi-center domain adaptation layer to shorten the distance between the target center and the source center. The corresponding calculation loss. At the same time, by maximizing Expand the difference between the two classifiers. After these two steps, the label modeling and updating parts are calculated respectively. Three losses are used to update network parameters and pseudo labels. The specific parameters of each loss function are described in the experimental section.

[0044]

[0045] D3, train the previous feature extraction part and the multi-center domain adaptation layer to minimize the two parts For purpose.

[0046]

[0047] Compared with the prior art, the present invention has the following outstanding substantive features and significant advantages:

[0048] The existing work on plaque echo classification for multi-center carotid ultrasound images is still in its early stages. Due to the interference of noise such as artifacts in carotid images, small plaques and complex morphological changes, the existing multi-center classification model cannot effectively handle the influence of multi-center feature differences and noise. The newly proposed model not only shortens the feature distance of multiple center data, but also focuses on plaque data that is difficult to transfer and classify due to noise. The above problems can be solved by end-to-end modeling and correction of the data of the target center.

[0049] Different from previous methods, DCLCN focuses more on the differences between domains and data samples that are difficult to be identified by classifiers in view of the inherent difficulties of multi-center plaque ultrasound data. In response to these two problems, a multi-center feature adaptation module and a noise label modeling and correction module are proposed respectively. While further reducing the feature differences, a pseudo-label modeling is performed on the target center, which provides better modeling estimation and processing for the generated "noise" labels. At the same time, this method has also conducted a large number of effective experiments on the three collected plaque ultrasound datasets to further verify the rationality and effectiveness of the proposed method. For models that have not undergone domain adaptation, the results have been greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is an example of the imaging differences of multi-center ultrasound data of plaques: These three pictures are from three centers. (1) From a single picture: (a) The yellow box is the ultrasound artifact, which covers up some useful information and produces noise interference. (b) It can be clearly seen that the resolution of this ultrasound image is very low. (2) Combining the data of the three data sets: (a) (b) (c) The three pictures have obvious imaging differences, which will cause certain difficulties for subsequent research.

[0051] Figure 2 This is the main idea of ​​the present invention: Left: The initial distribution of data features of the two centers is very messy and cannot be classified by the classifier. b: After further shortening the feature distance, some data is still difficult to be classified and identified by the classifier. This is because some ultrasound images contain noise, which makes it difficult to transfer features between multiple centers. c: The final effect that the model hopes to achieve.

[0052] Figure 3 It is a framework diagram of the present invention, which is mainly composed of two parts: multi-center feature adaptation and label modeling and correction. The label distribution continuously updated by the model and the images of the two centers are used as the input of the model.

[0053] Figure 4 The results of this method are compared with those of similar classical methods and more advanced methods.

[0054] Figure 5 The experimental results of six source domain and target domain combinations from three centers are listed.

[0055] Figure 6 The effects of different numbers of multi-center domain adaptation layers on the model are listed.

[0056] Figure 7 The rationality of the loss function selected by the proof method is listed, and the results of ablation experiments on several losses are listed. DETAILED DESCRIPTION

[0057] The preferred embodiments of the present invention are further described in detail below with reference to the accompanying drawings.

[0058] Embodiment 1:

[0059] Classification methods include:

[0060] A. Data preprocessing: First, data preprocessing is completed through image enhancement operations such as normalization and resolution adjustment.

[0061] B. Multi-center feature adaptation module, introduces a domain adaptation layer consisting of two fully connected layers, and uses the maximum mean difference (MMD) as a quantitative measure of the difference. In addition, two independent classifiers are constructed to increase the difference between classifiers.

[0062] C. Noise label modeling and processing module: By modeling and correcting the labels, the model prediction affected by noise is processed. The model corrects the ultrasound plaque labels affected by noise such as artifacts and low resolution by maintaining and updating the end-to-end label distribution from the two central datasets and minimizing the cross entropy loss, classification loss, and joint loss.

[0063] Example 2: In this example, the experiment is based on Resnet50 and pre-trained parameters from ImageNet. The learning rate of the model is 0.002, α=0.3, β=0.3. Denoising, flattening and horizontal random flipping are performed as data preprocessing and enhancement. SGD is used with a momentum of 0.9, a weight decay of 10-4, and a batch size of 16. The epoch is set to 150, of which 20 are source center pre-training in the first stage, 100 are dual classifier pseudo-label domain adaptation in the second stage, and the remaining epochs are used for model fine-tuning. . In terms of performance evaluation, the accuracy (ACC), recall (Recall), and precision (PRE) are used to evaluate the classification performance of the method. The larger the values ​​of these five evaluation indicators, the better the prediction performance of the model.

[0064] The data sets used in the present invention come from three cooperative hospitals. Among them, the data set of hospital A contains 244 plaques from 156 patients, with a total of 759 images. There are 506 stable plaque images, 253 unstable plaque images, and 759 ultrasound plaque images. The data set of hospital B contains 310 plaque ultrasound images, including 132 stable plaque images and 178 unstable plaque images. The data set of hospital C contains 124 stable plaques and 193 unstable plaques, with a total of 317 ultrasound plaque images. All three data sets are annotated by radiologists with many years of clinical experience.

[0065] The main flow chart of the present invention is as follows Figure 3 As shown:

[0066] A. Data preprocessing and establishing classification framework: Figure 3 As shown in the figure, the input video is first divided into 30 segments equally. Then, considering that in the previous study, the three selection methods of "randomly extracting a frame", "fixedly selecting the first frame" and "fixedly selecting the last frame" were tested, and it was found that the experimental effect of "fixedly selecting the first frame" was the best, so the first frame in each segment was fixedly selected to obtain a 30-frame sequence as the subsequent input. After that, each frame was resized to 224×224 and normalized for subsequent model training.

[0067] B. For a given source-centered image x s and its corresponding label y s , there will be a set of data sets from the source center {X s ,Y s}. For the unlabeled image x at the target center t , there will be a set of data sets {X t}. Train a feature extractor G and two classifiers C1andC2. Ultrasound image data x of the source center and the target center s and x t Input the feature extraction network G respectively to obtain the corresponding feature representation.

[0068] MMD loss is calculated as follows:

[0069]

[0070] The difference loss uses the absolute value difference of the output of the classifier for the target center data to perform quantitative measurement. Its calculation formula is:

[0071]

[0072] By maximizing L diversity and minimize L MMD , the model will get two classifiers that maximize the difference, which can increase the model's ability to discriminate data at the target center. The overall goal of this part is:

[0073] min(γL MMD -L diversity )(4)

[0074] Through multi-center domain adaptation, the model can obtain the prediction results f1(x, θ) and f2(x, θ) from the two classifiers respectively, and average them to obtain a prediction distribution that is convenient for modeling, which is defined as follows:

[0075]

[0076] After obtaining the prediction for the label, it is further processed through the noise label modeling and processing module.

[0077] C. To model the labels, the model will use two sets of label distributions and To describe the data from the two centers. In the specific model learning process, each training and will be input into the model as the pseudo-true label. For label initialization, the source center will use its own true label y s For the patch data at the target center, there is no initial true label for initialization, so this method uses the model pre-trained by the source center to give the target center a set of estimates of the distribution of noisy pseudo labels. The initial distribution estimate of the pseudo label at the target center includes not only the influence of image noise but also the noise label.

[0078] C1. and There are two sets of labels that model the impact of noise. The model gradually corrects this part of the impact during continuous learning. To achieve this goal, the network parameters θ will be continuously updated to predict the label distribution of the source center and the target center. and will also be added to the training process and will be continuously updated during the training process, using Y d To represent their collection. Three types of tags will be involved, namely y. is the prediction result from the previous classifier, which is the probability distribution of a category; y is the true label of the data set (note that in this article, the data label of the target center will not be involved in the training); The pre-trained classifier gives the target center data a noise pseudo-label to initialize the target center; is a set of hard labels that are continuously updated by back propagation. The goals of this part are as follows:

[0079]

[0080] C2. For The source center will use y, which is the true label, and the target center will use the initial noise pseudo label To complete:

[0081]

[0082] Where K is a constant.

[0083] C3. After a softmax function, we get a set of continuously updated probability distributions y required by the model. d .

[0084]

[0085] C4. Cross entropy loss: For the source center data, it has a true label. We use cross entropy loss to improve the model's classification ability for the source center:

[0086]

[0087] C5. Classification loss: In order to reduce the impact of patch noise, the model needs to continuously update the label y d and parameter θ. The update of this part mainly depends on y d And the prediction results of the dual classifier in the previous part The DCLCN uses KL loss to measure the deviation.

[0088]

[0089] C6. Compatibility Loss: Although the pseudo-labels predicted by the model include the patch predictions that are not easily identifiable, in each subsequent round of training, the results obtained by the model should still be mostly the same as the initial pseudo-labels. This view has also been confirmed in previous studies. Therefore, the model adds another item, CompatibilityLoss, which is defined as follows:

[0090]

[0091] The effect of noise on prediction can be further corrected by minimizing the overall loss of label correction. The calculation is as follows:

[0092]

[0093] The final overall training steps are as follows:

[0094] D. In order to obtain a preliminary model that is more suitable for multi-center plaque classification, DCLCN uses the plaque data from the source center to train the model. The training goal of this step is to minimize the cross entropy loss, and the goals are as follows:

[0095]

[0096] The data from the target center is added for training. First, a multi-center domain adaptation layer is used to shorten the distance between the target center and the source center. The corresponding calculation loss. At the same time, by maximizing Expand the difference between the two classifiers. After these two steps, the label modeling and updating parts are calculated respectively. Three losses are used to update network parameters and pseudo labels. The specific parameters of each loss function are described in the experimental section.

[0097]

[0098] Train the previous feature extraction part and the multi-center domain adaptation layer to minimize the For purpose.

[0099]

[0100] Reference Figure 4 , showing the comparison between this method and the classic and current mainstream methods. This method has achieved relatively good results in all three indicators. Although its accuracy is not as high as DAN, the accuracy is not much different (0.5%), and there is a certain improvement in precision and recall. This also shows that this method has a relatively good resolution for positive samples, that is, unstable plaques and other models. In addition, DMCD and CoUDA are two models that consider noise, and DCLCN has a certain degree of improvement in all three indicators. CoUDA considers the impact of noise labels on knowledge transfer, while DMCD adds high-confidence samples to the training set for denoising, and neither considers the impact of ultrasonic noise itself on multi-center knowledge transfer. In addition, compared with Resnet-50 (SourceOnly) which has no multi-center adaptation at all, this method has made very significant progress, with accuracy, precision and recall increased by 28.32%, 39.48% and 32.22% respectively, proving that the multi-center adaptation method is effective. However, compared with the fully supervised method, there is still a certain gap.

[0101] Reference Figure 5 , showing the comparison results of 6 tasks divided from the collected plaque data from three centers. The experimental results show that for the 6 tasks, the accuracy, specificity and recall rate are maintained at a level of about 70%, among which the three indicators of the tasks of centers A and B are generally higher than those of the tasks related to center C, with accuracy (4.18), precision (10.52), and recall (3.67). This may be due to the lack of cross-section data of blood vessels in the data of center C.

[0102] Reference Figure 6, Since different numbers of multi-center domain adaptation layers have different effects on the model, we explored the most appropriate number of layers. Experiments have shown that multi-center domain adaptation layers do contribute to better domains. Although the multi-center domain adaptation layer with one layer has the highest accuracy, it is not much different from the two-layer one (0.41%), but the two-layer domain adaptation layer has better results in both precision and recall. This may be because compared with the single-layer multi-center domain adaptation layer, the two-layer structure can provide more modeling capabilities and help better capture the subtle differences between data from different hospitals. As the number of layers is superimposed, the depth of the model deepens, which may cause overfitting of the model.

[0103] Reference Figure 7 , shows the ablation experiment on the effect of different loss functions. The diversity of classifiers, multi-center domain adaptation, and the two losses for label correction all contribute to the final classification accuracy, so combining the three methods to solve the problem of plaque multi-center risk prediction is effective.

[0104] The description herein is combined with the drawings and specific embodiments to help understand the method and core idea of ​​the present invention. The method described in the present invention is not limited to the embodiments described in the specific embodiments. Other embodiments derived by those skilled in the art based on the method and ideas of the present invention also belong to the technical innovation scope of the present invention. The content of this specification should not be understood as limiting the present invention.

Claims

1. A noise label correction method for multi-center plaque classification of ultrasound images based on dual classifier prediction, characterized in that The method includes: inputting the ultrasound image into the feature extractor after preprocessing, adding a multi-center feature adaptation module after the feature extractor, constructing a modeling and processing module for noise labels, and realizing the classification of multi-center plaque echoes; the multi-center feature adaptation module introduces a domain adaptation layer composed of two fully connected layers, uses the maximum average difference as a quantitative measure of the difference, constructs two independent classifiers and takes the average of the outputs of the two classifiers as the input for the next step of noise label modeling.

2. The noise label correction method for multi-center plaque classification of ultrasound images based on dual classifier prediction according to claim 1 is characterized in that Ultrasound image preprocessing includes: cropping the area containing patient information and completing data preprocessing through image enhancement operations such as normalization and resolution adjustment.

3. The noise label correction method for multi-center plaque classification of ultrasound images based on dual classifier prediction according to claim 1 is characterized in that The calculation process of the multi-center feature adaptation module includes: B1. For a given source center image x s and its corresponding label y s , there will be a set of data sets from the source center {X s , Y s }. For the unlabeled image x at the target center t , there will be a set of data sets {X t }, train a feature extractor G and two classifiers C1 and C2, the ultrasound image data x of the source center and the target center s and x t Input the feature extraction network G respectively to obtain the corresponding feature representation; B2. Multi-center feature adaptation. In the specific model implementation, this part consists of two fully connected layers. The maximum mean difference (MMD) projects the features from the kernel function to the high-dimensional space, and calculates the expected difference between the random variables of the two feature distributions after mapping. The maximum value is the standard for measuring distance. By minimizing the MMD loss, the invariant features between multiple centers can be found. The calculation formula is: B3. DiversityLoss is introduced to further measure the difference between the two classifiers. The absolute value difference of the output of the classifier for the target center data is used for quantitative measurement. The calculation formula is: B4. By maximizing L diversity and minimize L MMD , the model will get two classifiers that maximize the difference, which can increase the model's ability to discriminate data at the target center. The overall goal of this part is: min(γL MMD- L diversity ) (4) B5. Through multi-center domain adaptation, the model can obtain the prediction results f1(x, θ) and f2(x, θ) from the two classifiers respectively, and average them to obtain a prediction distribution that is convenient for modeling, which is defined as follows:

4. The noise label correction method for multi-center plaque classification of ultrasound images based on dual classifier prediction according to claim 1 is characterized in that The calculation process of the noise label modeling and processing module is as follows: C1. To model labels, the model will use two sets of label distributions. and To describe the data from the two centers, and in the specific model learning process, each training and will be input into the model as a pseudo-true label. For label initialization, the source center will use its own true label y s For the patch data at the target center, a model pre-trained by the source center is used to give the target center a set of estimates of the distribution of noise pseudo-labels. The initial distribution estimate of the pseudo-labels at the target center includes the influence of image noise and noise labels. C2, and There are two sets of label models that include the impact of noise. The model gradually corrects this part of the impact during continuous learning. To achieve this goal, the network parameters θ will be continuously updated to predict the label distribution of the source center and the target center. and will also be added to the training process and will be continuously updated during the training process, using Y d To represent their collection, three types of tags will be involved, namely y, where is the prediction result from the previous classifier, which is the probability distribution of a category; y is the true label of the data set; The pre-trained classifier gives the target center data a noise pseudo-label to initialize the target center. is a set of hard labels that are continuously updated by back propagation. The goals of this part are as follows: C3. For The source center will use the true label y, and the target center will use the initial noise pseudo label To complete: Where K is a constant; C4, After a softmax function, we get a set of continuously updated probability distributions y required by the model. d : C5, cross entropy loss: For the source center data, it has a true label. We use cross entropy loss to improve the model's classification ability for the source center: C6, classification loss: In order to reduce the impact of patch noise, the model needs to continuously update the label y d The update of this part mainly depends on the prediction results of yd and the double classifier in the previous part. Guided by the deviation, DCLCN uses KL loss to measure the deviation. C7, compatibility loss: The model adds another compatibility loss, which is defined as follows: C8, by minimizing the overall loss of label correction, the impact of noise on prediction can be further corrected, calculated as follows:

5. The noise label correction method for multi-center plaque classification of ultrasound images based on dual classifier prediction according to any one of claims 1 to 4, characterized in that The training steps of the model are as follows: D1. In order to obtain a preliminary model that is more suitable for multi-center plaque classification, DCLCN uses plaque data from the source center to train the model. The training goal is to minimize the cross entropy loss. The goals are as follows: D2, adding data from the target center for training, first using the multi-center domain adaptation layer to shorten the distance between the target center and the source center, the corresponding calculation loss, and at the same time, by maximizing To expand the difference between the two classifiers, after these two steps, we need to calculate the modeling and updating parts of the labels respectively. Three losses to update network parameters and pseudo labels: D3, train the previous feature extraction part and the multi-center domain adaptation layer to minimize the two parts For the purpose:

Images