Image classification method and system based on source domain independent field adaptation and active learning

Abstract

The application belongs to the technical field of image processing, and provides an image classification method and system based on source domain independent field adaptation and active learning. The method comprises the following steps: based on unlabeled target domain medical image samples, a source domain model is used to select samples at one time; a feature extractor is used to extract the one-time selected samples to obtain the features of all samples; clustering operation is performed on the features of all samples to obtain pseudo labels of all samples; the value of neighbor uncertainty of each sample is calculated, when the value of neighbor uncertainty meets a certain range, the sample corresponding to the value of neighbor uncertainty is selected for relaxation operation, and the candidate label in the pseudo label is selected to perform subsequent labeling to obtain all labeled samples; according to the unlabeled target domain medical image samples and the medical image samples queried by active labeling, all labeled samples are combined, a target domain model is used, a loss function is calculated, and a trained target domain multi-label classification model is obtained for image classification.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, and in particular relates to an image classification method and system based on source domain-independent domain adaptation and active learning. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Automated initial screening and classification of large-scale medical images (such as lung CT images) is a crucial problem. In real-world scenarios, doctors need to classify numerous lung CT images into categories such as normal, pneumonia, and COVID-19. However, manual annotation on such large datasets is time-consuming, labor-intensive, and extremely difficult. Given time and resource constraints, we can leverage existing annotated medical images from other hospitals, utilizing the similarity between the annotated images and the current medical image distribution to guide the classification task. For example, if Hospital B needs to annotate a large number of unannotated medical images, we can collect and utilize a large number of similar, already annotated medical images from Hospital A to help Hospital B classify the images. The distribution differences between the images from the two hospitals stem from variations in equipment and imaging clarity. Therefore, a common domain adaptation method can be used to align images from different hospitals, thereby enabling the classification of Hospital B's images. The existing annotated images are called the source domain, and the images to be classified are called the target domain. The data distributions of the source and target domains are similar but different. When there are distributional differences between the source and target domains, the goal of domain adaptation is to label the unlabeled data in the target domain by using the existing labeled data in the source domain.

[0004] With increasing public concern for privacy and data security, it's often difficult to obtain complete, labeled image data from Hospital A. Instead, Hospital A can provide a pre-trained source domain model on that data. The scenario of classifying similar medical images from another hospital using a source domain model from one hospital is called source-domain-independent adaptation for medical image classification. However, due to the lack of source domain data, traditional domain adaptation methods for aligning the two domains all fail. Furthermore, because the source domain model contains insufficient source domain information, positive transfer is difficult to guarantee, leading to poor performance of existing source-domain-independent adaptation methods: that is, even with various strategies, the performance of existing methods is difficult to improve significantly.

[0005] Currently, most methods use the confidence level or information entropy of the model output to select samples with high uncertainty as samples with high information content to be added to the model training, hoping to significantly improve the model's performance.

[0006] The key to active learning-based medical image classification methods that are source-domain independent and domain-adaptive lies in selecting a small number of valuable samples. Existing active learning methods for selecting valuable samples can be mainly divided into the following two categories:

[0007] Uncertainty-based methods: Due to model overconfidence and domain variability, the standards of model output, such as entropy and confidence, are biased, making it difficult to accurately measure the uncertainty of medical images. Therefore, existing methods do not select the most valuable active samples, resulting in limited performance.

[0008] Diversity-based methods may select some well-adapted samples biased towards the source domain. These samples have already been adapted, so their contribution to model training is limited.

[0009] Therefore, current methods are unable to accurately select a small number of valuable samples. Summary of the Invention

[0010] To address the technical problems mentioned above, this invention provides an image classification method and system based on source domain-independent domain adaptation and active learning. It designs an algorithm to select a small number of samples that are crucial for medical image classification, allowing experts to annotate them, thereby significantly improving the performance of a given source domain model with minimal data cost.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] The first aspect of the present invention provides an image classification method based on source domain-independent domain adaptation and active learning.

[0013] Image classification methods based on source-domain-independent domain adaptation and active learning include:

[0014] Unlabeled target domain medical images are acquired, and a trained target domain multi-label classification model is used to obtain annotations for the target domain medical images.

[0015] The training process of the target domain multi-label classification model includes: based on unlabeled target domain medical image samples, initializing the feature extractor and classifier of the target domain model using the source domain model, and selecting samples all at once; extracting the features of all samples using the feature extractor; performing clustering operations on the features of all samples to obtain pseudo-labels for all samples; calculating the neighbor uncertainty value of each sample, and when the neighbor uncertainty value meets a certain range, selecting the sample corresponding to the neighbor uncertainty value for relaxation operations, and selecting candidate labels from the pseudo-labels for subsequent labeling to obtain all labeled samples; based on the unlabeled target domain medical image samples and the actively labeled query medical image samples, combined with all labeled samples, using the target domain model, calculating the loss function, and obtaining the trained target domain multi-label classification model when the output results achieve the expected effect.

[0016] Furthermore, the value of the neighbor uncertainty is calculated using the following formula:

[0017] NAU(x) = NP(x) * NA(x)

[0018]

[0019]

[0020] Where NP(x) represents neighbor purity, NA(x) represents neighbor affinity, K represents the total number of categories, and p k S represents the probability that sample x is labeled as class k in the label space of its surrounding neighbors. Nq It describes the cosine similarity between a sample and its q-th neighbor.

[0021] Furthermore, the neighbor purity describes the degree of confusion among the neighbor labels around the sample.

[0022] Furthermore, the neighbor affinity describes the closeness between the sample and its surrounding neighbors.

[0023] Furthermore, the loss function is:

[0024]

[0025] in, L represents the cross-entropy loss function. ent L represents the information entropy loss function. div This represents the KL divergence loss function.

[0026] Furthermore, the clustering operation is performed using the k-menas clustering algorithm.

[0027] Furthermore, the process of selecting the sample corresponding to the neighbor uncertainty value for relaxation operation when the neighbor uncertainty value meets a certain range specifically includes: for candidate samples whose neighbor uncertainty value meets a certain range, evaluating the nearest neighbor sample of the candidate sample; if the nearest neighbor sample has already been selected, skipping the selection of candidate samples.

[0028] A second aspect of the present invention provides an image classification system based on source domain-independent domain adaptation and active learning.

[0029] Image classification systems based on source-domain-independent domain adaptation and active learning include:

[0030] The labeling and classification module is configured to: acquire unlabeled target domain medical images, and use a trained target domain multi-label classification model to obtain labels for the target domain medical images;

[0031] The target domain multi-label classification model training module is configured as follows: Based on unlabeled target domain medical image samples, the feature extractor and classifier of the target domain model are initialized using the source domain model, and samples are selected all at once; the feature extractor extracts the features of the selected samples to obtain the features of all samples; clustering is performed on the features of all samples to obtain pseudo-labels for all samples; the uncertainty value of each sample's neighbors is calculated, and when the uncertainty value of the neighbors meets a certain range, the sample corresponding to the uncertainty value of the neighbors is selected for relaxation operation, and candidate labels are selected from the pseudo-labels for subsequent labeling to obtain all labeled samples; based on the unlabeled target domain medical image samples and the actively labeled query medical image samples, combined with all labeled samples, the target domain model is used to calculate the loss function, and when the output results achieve the expected effect, the trained target domain multi-label classification model is obtained.

[0032] A third aspect of the present invention provides a computer-readable storage medium.

[0033] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the image classification method based on source-domain-independent domain adaptation and active learning as described in the first aspect above.

[0034] A fourth aspect of the present invention provides a computer device.

[0035] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the image classification method based on source domain-independent domain adaptation and active learning as described in the first aspect above.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] This invention is the first to introduce active learning into source-domain-independent domain-adaptive medical image classification, successfully alleviating the poor performance problem of source-domain-independent domain adaptation. Under conditions of privacy protection and the existence of domain differences, it achieves the classification of all target domain images using only a small number of labeled target domain images.

[0038] This invention defines three properties that key image samples for adaptive image classification in an active source-domain-independent domain must satisfy: neighborhood disorder, individual variability, and bias towards the target domain. In this way, information point samples are found, and three losses are introduced to ensure the learning of these information point samples and the remaining samples. Attached Figure Description

[0039] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0040] Figure 1 This is a flowchart illustrating the image classification method based on source domain-independent domain adaptation and active learning as presented in this invention. Detailed Implementation

[0041] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0042] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0043] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0044] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0045] Example 1

[0046] This embodiment provides an image classification method based on source domain-independent domain adaptation and active learning. This embodiment uses the application of this method to a server as an example for illustration. It is understood that this method can also be applied to terminals, and can also be applied to systems including terminals, servers, and other components, and implemented through interaction between the terminal and the server. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network servers, cloud communication, middleware services, domain name services, CDN security services, and big data and artificial intelligence platforms. The terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. The terminal and server can be directly or indirectly connected via wired or wireless communication, which is not limited herein. In this embodiment, the method includes the following steps:

[0047] Unlabeled target domain medical images are acquired, and a trained target domain multi-label classification model is used to obtain annotations for the target domain medical images (the model assigns specific labels: normal, pneumonia, COVID-19).

[0048] The training process of the target domain multi-label classification model includes: based on unlabeled target domain medical image samples, initializing the feature extractor and classifier of the target domain model using the source domain model, and selecting samples all at once; extracting the features of all samples using the feature extractor; performing clustering operations on the features of all samples to obtain pseudo-labels for all samples; calculating the neighbor uncertainty value of each sample, and when the neighbor uncertainty value meets a certain range, selecting the sample corresponding to the neighbor uncertainty value for relaxation operations, and selecting candidate labels from the pseudo-labels for subsequent labeling to obtain all labeled samples; based on the unlabeled target domain medical image samples and the actively labeled query medical image samples, combined with all labeled samples, using the target domain model, calculating the loss function, and obtaining the trained target domain multi-label classification model when the output results achieve the expected effect.

[0049] The following is in conjunction with the appendix Figure 1 This embodiment will be described in detail:

[0050] Step 1: Input the unlabeled target domain medical image to be processed, and use the source domain model to initialize the feature extractor and classifier of the target domain model.

[0051] Step 2: Input the image into the feature extractor of the source domain model and output the features of all samples.

[0052] Step 3 (Selecting Active Samples): Cluster all features using k-menas to obtain pseudo-labels for all samples. Calculate the neighbor uncertainty value NAU(x) for each sample, select candidate labels using neighbor relaxation, and then have them labeled by experts. When the budget is exhausted, obtain the set D of all labeled samples. L .

[0053] Step 4: Input all unlabeled target domain medical images and all actively labeled query medical images into the target domain model, and learn from different samples using different losses.

[0054] Step 5: Repeat step 4 until the desired effect is achieved and a well-trained target domain multi-label classification model is obtained.

[0055] Samples that satisfy the following three characteristics are more valuable for active learning of medical image classification with source domain independence: (1) neighbor confusion, (2) individual differences, and (3) target domain bias.

[0056] To explore and mine these information points, we designed three strategies: neighborhood environment uncertainty, neighborhood diversity relaxation, and one-time visits. When utilizing these information points, we use cross-entropy loss for training. Simultaneously, we introduce information entropy loss and KL divergence to update the model, allowing it to focus more on the valuable samples we selected. Next, we will discuss the three characteristics, analyze the three proposed strategies, and outline the training loss function.

[0057] 1. Valuable sample characteristic analysis

[0058] (1-1) Neighbor Confusion: If a sample's neighbors have confusing labels, not only does the sample itself have greater uncertainty, but its neighbors also have greater uncertainty. Correcting the label of this sample can achieve self-correction while also correcting the labels of its neighbors, resulting in greater performance gains. In addition, samples with confusing neighbor labels are more likely to be at the boundary of classification decisions, and learning from these samples is beneficial for learning clearer decision boundaries.

[0059] (1-2) Individual differences reflect the diversity of the sample. By ensuring individual differences, we can prevent the repeated selection of very similar samples and avoid sample redundancy.

[0060] (1-3) The property of target domain bias indicates that the sample selection is biased towards the target domain distribution. According to the covariance assumption, all target domain data can be divided into data that are easily adapted but biased towards the source domain and data that are not easily adapted but biased towards the target domain. Selecting data biased towards the target domain can ensure the model's generalization in the target domain.

[0061] 2. Three selection strategies

[0062] (2-1) Uncertainty of the neighboring environment:

[0063] like Figure 1 To address this uncertainty, we designed a neighborhood uncertainty mechanism. Intuitively, samples with highly inconsistent neighbor labels, coupled with close proximity to their neighbors, appear to violate clustering assumptions and thus have higher uncertainty. Based on this intuition, we designed neighbor purity and neighbor affinity to measure the uncertainty surrounding the neighbors.

[0064] First, let's define what a neighbor is:

[0065] At the feature space level, we define the q samples with high cosine similarity to sample x as the neighbors of sample x:

[0066]

[0067] Where q represents the number of surrounding neighbors, S N Denotes the sample space of the neighbors around x. Y represents the pseudo-label space of the samples surrounding x's neighbors. The pseudo-labels are obtained based on clustering. Nq This represents the pseudo-label of the q-th neighbor.

[0068] Neighbor purity (NP): describes the degree of confusion among the labels of a sample's neighbors. To calculate the neighbor purity NP(x), we first determine the probability distribution space of the neighbor classes of sample x:

[0069]

[0070] Where, p k This represents the probability that sample x is labeled as class k in the label space of its surrounding neighbors. Figure 1 For example, the probability distribution spaces of the neighbor classes of a and b are as follows:

[0071] The neighbor purity NP here is measured by the following formula:

[0072]

[0073] Where log represents the log function, K represents the total number of all classes, and the larger the neighbor purity NP is, the more chaotic the labels around the sample are, and the higher the uncertainty of the sample.

[0074] Neighbor affinity (NA): describes how closely a sample is connected to its surrounding neighbors. To measure affinity, we first define a neighbor similarity space for each sample.

[0075]

[0076] Among them, S N1 The cosine similarity between a sample and its first neighbor is described. Furthermore, neighbor affinity is measured by the average similarity between this sample and all its surrounding neighbors:

[0077]

[0078] The lower the neighbor affinity, the lower the similarity between the sample and its surrounding neighbors.

[0079] Multiplying neighbor purity and neighbor affinity together forms the standard NAU of uncertainty surrounding our neighborhood:

[0080] NAU(x)=NP(x)*NA(x) (6)

[0081] The closer a sample's neighbors are, and the more chaotic the neighbor labels, the higher their uncertainty. Compared to existing methods for addressing uncertainty in active learning, we shift the focus from individual uncertainty to measuring the uncertainty of the sample's environment. Similar to the advantages of ensemble learning, this improves the tolerance for sample selection.

[0082] (2-2) Neighbor diversity relaxation:

[0083] We propose a neighborhood diversity relaxation strategy to ensure individual differences among samples. Existing methods for ensuring diversity are mainly based on clustering, but they may select some worthless samples because they ignore the existence of domain differences. In contrast, the neighborhood diversity relaxation method ensures sample diversity by performing a relaxation operation on candidate samples with high neighbor uncertainty. Specifically, for a candidate sample x with a high NAU value, we evaluate its nearest neighbor samples N. x If N x If a sample has already been selected, it means that samples similar to x have already been selected, so we skip selecting x. Through the neighbor relaxation strategy, neighbor samples with similar characteristics have no chance of being selected, thus ensuring individual differences in the selected sample set.

[0084] (2-3) One-time access:

[0085] To explore samples biased towards the target domain, we employ a one-time access strategy, selecting samples only once based on the original source domain model. There are two reasons for this. First, in the absence of source data, we can measure domain dissimilarity by observing the target data's performance on the untrained source model. Examples easily classified by the source model can be considered source domain-similar samples, primarily occurring in clusters with clean neighbor labels. Otherwise, samples with mixed neighbor labels are severely misclassified and biased towards the target domain; these samples can be effectively explored using our proposed criteria. Second, as the trained model gradually tilts towards the target domain, existing strategies struggle to determine whether samples are biased towards the source or target domain.

[0086] We define the sample set selected using the above three strategies as D. L We use standard cross-entropy for training these samples:

[0087]

[0088] Where, represents the softmax function; k It is a one-hot vector, where a value of 1 represents the true label.

[0089] In addition, we added information entropy loss and KL divergence to ensure clear and balanced categories. These two losses are commonly used in self-supervised learning and domain adaptation, and their formulas are as follows:

[0090]

[0091]

[0092] Among them, L ent L represents the information entropy loss. div Denotes the KL divergence, D t q represents all target domain data. k Indicates uniform distribution. This represents the mean output of the model for all samples in the k-th class.

[0093] The final training objective of the model is:

[0094]

[0095] To address the performance issues, we propose a method for medical image classification based on source-domain-independent domain adaptation through active learning. This method involves designing an algorithm to select a small number of samples crucial for medical image classification, which are then labeled by experts. This significantly improves the performance of a given source-domain model with minimal data cost. In practical applications, finding domain experts to label images incurs considerable costs, making this approach particularly relevant for medical image classification scenarios.

[0096] The core of applying existing active learning to source-domain-independent domain adaptation methods is to use the model output criteria to measure the uncertainty of samples, thereby selecting valuable medical image samples as active samples to participate in the model training process, thus enhancing the model's performance.

[0097] Example 2

[0098] This embodiment provides an image classification system based on source domain-independent domain adaptation and active learning.

[0099] Image classification systems based on source-domain-independent domain adaptation and active learning include:

[0100] The labeling and classification module is configured to: acquire unlabeled target domain medical images, and use a trained target domain multi-label classification model to obtain labels for the target domain medical images;

[0101] The target domain multi-label classification model training module is configured as follows: Based on unlabeled target domain medical image samples, the feature extractor and classifier of the target domain model are initialized using the source domain model, and samples are selected all at once; the feature extractor extracts the features of the selected samples to obtain the features of all samples; clustering is performed on the features of all samples to obtain pseudo-labels for all samples; the uncertainty value of each sample's neighbors is calculated, and when the uncertainty value of the neighbors meets a certain range, the sample corresponding to the uncertainty value of the neighbors is selected for relaxation operation, and candidate labels are selected from the pseudo-labels for subsequent labeling to obtain all labeled samples; based on the unlabeled target domain medical image samples and the actively labeled query medical image samples, combined with all labeled samples, the target domain model is used to calculate the loss function, and when the output results achieve the expected effect, the trained target domain multi-label classification model is obtained.

[0102] It should be noted that the above-described labeling and classification module and target domain multi-label classification model training module are the same examples and application scenarios implemented in Embodiment 1, but are not limited to the content disclosed in Embodiment 1. It should also be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0103] Example 3

[0104] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the image classification method based on source domain-independent domain adaptation and active learning as described in Embodiment 1 above.

[0105] Example 4

[0106] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the image classification method based on source domain-independent domain adaptation and active learning as described in Embodiment 1 above.

[0107] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0108] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0109] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0110] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0111] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0112] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An image classification method based on source domain independent field adaptation and active learning, characterized in that, The method comprises the following steps: obtaining unlabeled target domain medical images, and using a trained target domain multi-label classification model to obtain labels of the target domain medical images; the process of training the target domain multi-label classification model comprises: based on the unlabeled target domain medical image samples, using a source domain model to initialize a feature extractor and a classifier of a target domain model, and selecting samples at one time; using the feature extractor to extract the samples selected at one time to obtain features of all the samples; performing clustering operation on the features of all the samples to obtain pseudo-labels of all the samples; calculating a neighbor uncertainty value of each sample, and when the neighbor uncertainty value meets a certain range, selecting the sample corresponding to the neighbor uncertainty value to perform relaxation operation, and selecting a candidate label in the pseudo-labels to perform subsequent labeling to obtain all the labeled samples; the neighbor uncertainty value is calculated by the following formula: where NP(x) denotes the neighbor purity, NA(x) denotes the neighbor affinity, and K denotes the total number of all classes, denotes the sample labeled as class in the surrounding neighbor label space, describes the cosine similarity between the sample and its qth neighbor; when the neighbor uncertainty value meets the certain range, the process of selecting the sample corresponding to the neighbor uncertainty value to perform relaxation operation specifically comprises: for the candidate sample whose neighbor uncertainty value meets the certain range, evaluating a nearest neighbor sample of the candidate sample, and if the nearest neighbor sample has been selected, skipping the selection of the candidate sample; based on the unlabeled target domain medical image samples and the medical image samples queried by active labeling, combining all the labeled samples, using the target domain model to calculate a loss function, and when the output result meets an expected effect, obtaining the trained target domain multi-label classification model. 2.The image classification method based on source domain independent field adaptation and active learning according to claim 1, characterized in that, The neighbor purity describes a chaotic degree of neighbor labels around the sample. 3.The image classification method based on source domain independent field adaptation and active learning according to claim 1, characterized in that, The neighbor affinity describes a close degree of the sample and the neighbor around the sample. 4.The image classification method based on source domain independent field adaptation and active learning according to claim 1, characterized in that, the loss function is: wherein, represents a cross-entropy loss function, represents an information entropy loss function, represents a KL divergence loss function. 5.The image classification method based on source domain independent field adaptation and active learning according to claim 1, characterized in that, the clustering operation is performed by using a k-menas clustering algorithm.

6. An image classification system based on source domain independent adaptation and active learning, characterized in that, The method comprises the following steps: a labeling classification module configured to: obtain unlabeled target domain medical images, and use a trained target domain multi-label classification model to obtain labels of the target domain medical images; a target domain multi-label classification model training module configured to: based on the unlabeled target domain medical image samples, use a source domain model to initialize a feature extractor and a classifier of a target domain model, and select samples at one time; use the feature extractor to extract the samples selected at one time to obtain features of all the samples; perform clustering operation on the features of all the samples to obtain pseudo-labels of all the samples; calculate a neighbor uncertainty value of each sample, and when the neighbor uncertainty value meets a certain range, select the sample corresponding to the neighbor uncertainty value to perform relaxation operation, and select a candidate label in the pseudo-labels to perform subsequent labeling to obtain all the labeled samples; the neighbor uncertainty value is calculated by the following formula: where NP(x) denotes the neighbor purity, NA(x) denotes the neighbor affinity, and K denotes the total number of all classes, denotes the sample labeled as class in the surrounding neighbor label space, describes the cosine similarity between the sample and its qth neighbor; when the neighbor uncertainty value meets the certain range, the process of selecting the sample corresponding to the neighbor uncertainty value to perform relaxation operation specifically comprises: for the candidate sample whose neighbor uncertainty value meets the certain range, evaluating a nearest neighbor sample of the candidate sample, and if the nearest neighbor sample has been selected, skipping the selection of the candidate sample; According to the unlabeled target domain medical image samples and the actively labeled query medical image samples, all the labeled samples are combined, a target domain model is used to calculate a loss function, and when an output result reaches an expected effect, a trained target domain multi-label classification model is obtained.

7. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by a processor, implements the steps in the image classification method based on source domain independent field adaptation and active learning of any one of claims 1-5.

8. 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 implements the steps in the image classification method based on source domain independent field adaptation and active learning of any one of claims 1-5 when executing the program.

Images