Medical image classification-oriented passive domain adaptation method and system, electronic device
By introducing a Gaussian mixture model to filter high-confidence samples and combining it with contrastive learning to optimize the loss function in medical image classification, the problem of low pseudo-label quality in passive domain adaptation is solved, improving the adaptive performance and prediction accuracy of the model, making it suitable for medical image classification tasks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2024-07-05
- Publication Date
- 2026-07-07
AI Technical Summary
Existing medical image classification models suffer from a distribution mismatch between training and test samples in practical applications, resulting in insufficient generalization ability. Furthermore, the low quality and high noise of pseudo-label generation in passive domain adaptive methods negatively impact prediction performance.
A passive domain adaptive strategy is adopted, using a Gaussian mixture model to select high-confidence samples to generate pseudo-labels, and improving the model's adaptability through contrastive learning and multi-loss function optimization.
It avoids cross-center data privacy and security issues, improves the model's adaptive performance and prediction accuracy, and enhances the model's generalization ability in the target domain.
Smart Images
Figure CN118628836B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of medical image classification technology, and in particular to a passive domain adaptive method and system, and electronic equipment for medical image classification. Background Technology
[0002] Medical image classification is crucial in computer-aided clinical diagnostic applications, including disease diagnosis, preoperative assessment, and prognostic prediction. In recent years, deep learning has demonstrated superior performance in medical image classification tasks. However, most traditional deep learning models are based on the assumption of independent and identically distributed (i.i.d.) data, meaning that training and test samples have the same or similar data distributions. In practical applications, many factors, such as equipment vendors, imaging protocols, and image modalities, can cause distribution shifts (domain shifts) between training and test samples. Improving the adaptability and generalization ability of deep learning models to achieve clinical usability is a pressing issue that needs to be addressed.
[0003] Currently, Domain Adaptation (DA) is a mainstream technique that improves model generalization performance by utilizing labeled source domain data and partially unlabeled target domain data for data augmentation or feature alignment. However, due to privacy concerns and transmission limitations of medical data, it is difficult to directly and simultaneously obtain multi-center source domain data in practical applications. Source-Free Domain Adaptation (SFDA) aims to improve generalization performance through model transfer while avoiding data privacy issues and transmission limitations.
[0004] In general, existing SFDA methods can be mainly divided into: data-based methods and model-based methods.
[0005] The former utilizes source domain information from the embedding model to simulate source domain data or explore inherent data structures or clustering information hidden in unlabeled target domain data; while data-based methods use source domain information from the embedding model to simulate source domain data or explore inherent data structures or clustering information hidden in unlabeled target domain data. However, clustering results may be affected by data distribution, leading to instability. Especially in high-dimensional data spaces, the performance of clustering algorithms may be hampered, making it difficult to accurately capture the inherent structure of the data.
[0006] The latter approach divides the model into multiple sub-modules and uses strategies such as self-learning and contrastive learning to adjust the parameters of some modules to adapt to the target domain. Model-based methods mainly employ semi-supervised learning to align the source and target domains for better domain adaptation. Examples include pseudo-labeling and contrastive learning. Pseudo-labeling is the most widely used, but it has the following problems: a. Current methods mainly focus on pseudo-label generation strategies, and label confidence screening is generally done using simple threshold screening, resulting in low-quality pseudo-generated labels. b. Most current methods directly use all data for pseudo-label supervision, introducing a lot of noise and affecting the model's predictive performance. Summary of the Invention
[0007] The purpose of this application is to provide a passive domain adaptive method, system, and electronic device for medical image classification. By adopting a passive domain adaptive strategy, it avoids cross-center data privacy and data security issues.
[0008] According to a first aspect of the embodiments of this application, a passive domain adaptive method for medical image classification is provided, comprising:
[0009] Supervised training of a medical image classification model is performed on source domain medical image data to initialize a medical image classification model in the target domain, wherein the medical image classification model consists of an encoder and a classifier.
[0010] Unlabeled target domain medical image data is input into the encoder for feature extraction, and then the extracted features are input into the classifier to obtain the classification probability.
[0011] The classification probability is used as input to the Gaussian mixture model to optimize the fitting of the Gaussian mixture model, and the value of the first distribution of the target domain medical image data in the optimized fitting Gaussian mixture model is used to divide the target domain medical image data into source domain class and target domain specific class.
[0012] Based on the characteristics of the source domain class, generate the class center for each class;
[0013] The target domain medical image is enhanced to obtain an enhanced image. Then, the original image and the enhanced image are input into the encoder for feature extraction to obtain the original image features and the enhanced image features.
[0014] The original image features and enhanced image features are input into the classifier to obtain the predicted classification probability. The predicted classification probability and intermediate features of the medical image data are used to calculate the various losses and obtain the total loss.
[0015] The medical image classification model is fine-tuned by minimizing the total loss until the total loss reaches the preset convergence condition, at which point training stops.
[0016] The finely tuned medical image classification model is used to predict the target domain medical image data to obtain the final classification result.
[0017] According to a second aspect of the embodiments of this application, a passive domain adaptive system for medical image classification is provided, comprising:
[0018] An initialization module is used to perform supervised training on a medical image classification model on source domain medical image data, thereby initializing a medical image classification model in the target domain, wherein the medical image classification model consists of an encoder and a classifier.
[0019] The feature extraction and probability prediction module is used to input unlabeled target domain medical image data into the encoder for feature extraction, and then input the extracted features into the classifier to obtain the classification probability.
[0020] The data domain partitioning module is used to take the classification probability as input to the Gaussian mixture model to optimize the fitting of the Gaussian mixture model, and to use the value of the first distribution of the target domain medical image data in the optimized fitting Gaussian mixture model to partition the target domain medical image data into a source domain class and a target domain specific class.
[0021] The class center generation module is used to generate the class center of each class based on the characteristics of the class source domain class;
[0022] The image enhancement and extraction module is used to enhance the target domain medical image to obtain the enhanced image, and then input the original image and the enhanced image into the encoder for feature extraction to obtain the original image features and the enhanced image features;
[0023] The loss calculation module is used to input the original image features and enhanced image features into the classifier to obtain the predicted classification probability, and to calculate various losses using the predicted classification probability and intermediate features of the medical image data to obtain the total loss;
[0024] The fine-tuning training module is used to fine-tune the medical image classification model by minimizing the total loss until the total loss reaches the preset convergence condition and training stops.
[0025] The prediction module is used to predict the target domain medical image data using the fine-tuned medical image classification model to obtain the final classification result.
[0026] According to a third aspect of the embodiments of this application, an electronic device is provided, comprising:
[0027] One or more processors;
[0028] Memory, used to store one or more programs;
[0029] When the one or more programs are executed by the one or more processors, the one or more processors perform the method as described in the first aspect.
[0030] According to a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the method as described in the first aspect.
[0031] The technical solutions provided by the embodiments of this application may include the following beneficial effects:
[0032] As can be seen from the above embodiments, this application adopts a passive domain adaptive strategy, avoiding cross-center data privacy and data security issues, and also avoiding the high cost of medical image data annotation. This invention introduces the Gaussian Mixture Model (GMM) method to screen high-confidence samples, thereby improving the confidence of the generated pseudo-labels and thus better improving the adaptive performance of the model. In addition, in the pseudo-label supervision process, this invention only uses high-confidence samples for supervised training, uses all samples for regularization optimization, promotes the model to continuously optimize towards the target domain during training without causing excessive deviation, and uses a contrastive learning method to bring the high-confidence source domain and target domain data domains closer in the feature space, and combines multiple loss functions for optimization to further improve model performance.
[0033] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description
[0034] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.
[0035] Figure 1 This is a flowchart illustrating a passive domain adaptive method for medical image classification according to an exemplary embodiment.
[0036] Figure 2 This is a structural block diagram illustrating a passive domain adaptive method for medical image classification according to an exemplary embodiment.
[0037] Figure 3 This is a block diagram illustrating a passive domain adaptive system for medical image classification according to an exemplary embodiment. Detailed Implementation
[0038] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.
[0039] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.
[0040] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."
[0041] Figure 1 This is a flowchart illustrating a passive domain adaptive method for medical image classification according to an exemplary embodiment. Figure 2 This is a structural block diagram illustrating a passive domain adaptive method for medical image classification according to an exemplary embodiment, such as... Figure 1 and Figure 2 As shown, the method may include the following steps:
[0042] S1: Supervised training of a medical image classification model is performed on the source domain medical image data to initialize the target domain medical image classification model, which consists of an encoder and a classifier; this step includes the following sub-steps:
[0043] S11: Perform uniform size and normalization processing on labeled source domain medical image data;
[0044] Specifically, bilinear interpolation is used to scale all source domain medical images to the same size, thus conforming to the input specifications of the medical image classification model.
[0045] S13: Input the processed source domain medical image data into the medical image classification model and output the predicted classification probability;
[0046] Specifically, the medical image classification model consists of an encoder and a classifier. The encoder can be any type of network feature extractor, such as ResNet or ViT, while the classifier can be a common MLP with a softmax function.
[0047] The encoder extracts intermediate features from the processed source domain medical image, and the classifier then uses these intermediate features as input to obtain unnormalized predicted classification probabilities. Then, the normalized predicted classification probability is obtained using the Softmax activation function. .
[0048] S14: Calculate the classification loss based on the predicted classification probability and the one-hot encoding of the true label;
[0049] Specifically, based on the predicted classification probabilities and the one-hot encoding of the true labels, the classification loss of the medical image classification model is calculated using cross-entropy loss. It is used to evaluate the degree of difference between the predicted values and the true values of a medical image classification model.
[0050] S15: Perform backpropagation based on the classification loss to update the parameters of the medical image classification model.
[0051] Specifically, the network of the medical image classification model automatically calculates the gradient of the classification loss and uses the stochastic gradient descent (SGD) method to update the parameters of the medical image classification model, thereby improving the classification accuracy of the auxiliary medical image classification model.
[0052] S16: After the medical image classification model is trained, the medical image classification model trained on the source domain medical image data is used to initialize the medical image classification model in the target domain.
[0053] Specifically, the medical image classification model in the target domain uses the parameters trained on the medical image data in the source domain as initial parameters.
[0054] S2: Input the unlabeled target domain medical image data into the encoder for feature extraction to obtain intermediate features, and then input the extracted features into the classifier to obtain the classification probability. This step includes the following sub-steps:
[0055] S21: Perform the same size and normalization processing on the unlabeled target domain medical image data as on the source domain medical image data to obtain the processed target domain medical image.
[0056] Specifically, bilinear interpolation is used to scale all source domain medical images to the same size, thus conforming to the input specifications of the medical image classification model.
[0057] S22: Input the processed target domain medical image into the encoder for feature extraction and output the intermediate features of the image;
[0058] Specifically, the processed target domain medical image is input into the encoder to obtain the intermediate features of the image.
[0059] S23: Input the extracted features into the classifier to obtain the classification probability of the image;
[0060] Specifically, the intermediate features of the image The unnormalized predicted classification probability of the image is obtained by inputting it into the classifier. Then, the normalized predicted classification probability is obtained using the Softmax activation function. .
[0061] S3: The classification probability is used as input to the Gaussian mixture model to optimize the fitting of the Gaussian mixture model, and the target domain medical image data is divided into source domain class and target domain specific class by using the value of the first distribution of the target domain medical image data in the optimized fitting Gaussian mixture model. This step includes the following sub-steps:
[0062] S31: Fit and optimize the Gaussian mixture model based on the classification probability of the target domain medical image;
[0063] Specifically, a binary Gaussian mixture model (GMM) is initialized, and parameters such as the EM iteration stopping threshold and the maximum number of iterations are set. Then, the classification probability is input into the binary Gaussian mixture model for optimization fitting to obtain the optimized Gaussian mixture model.
[0064] The specific concepts of Gaussian mixture models are as follows:
[0065] A Gaussian Mixed Model (GMM) is a linear combination of multiple Gaussian distribution functions. Theoretically, a GMM can fit any type of distribution. GMMs are typically used to solve cases where data from the same dataset contains multiple distinct distributions.
[0066] S32: Obtain the probability values of the first distribution of all target domain medical image samples in the Gaussian mixture model;
[0067] Specifically, the classification probabilities of the medical image samples in the target domain are input again into the fitted GMM model to obtain the probability of the samples in the first distribution within the GMM model. .
[0068] S33: Set the data partitioning threshold;
[0069] Specifically, thresholds are manually set according to task requirements. .
[0070] S34: Based on the probability value and the data partitioning threshold, the target domain data is divided into source domain classes and target domain specific classes;
[0071] Specifically, based on the probability of the sample in the first distribution of the fitted GMM network. ,Will Above a given threshold The samples are classified into the source domain class, while other samples are classified into the target domain-specific class. The formula is as follows:
[0072]
[0073] S4: Generate the class center for each class based on the characteristics of the source domain class; this step includes the following sub-steps:
[0074] S41: Based on the classification probability of the target domain medical image samples, the source domain data is divided into K classes, where K is the number of classes;
[0075] Specifically, the class with the highest classification probability value of the medical image sample in the target domain is taken as the category of the sample.
[0076] S42: Average the features of the medical image samples of the same category source domain to obtain the class center of each class;
[0077] Specifically, class center for each category is generated using class source domain class data. The specific formula is as follows:
[0078]
[0079] in Indicates sample The initial tag, This represents the summation of data samples from the source domain. Indicates sample for The value is 1 if the condition is met, otherwise it is 0. Indicates sample Its characteristics.
[0080] S5: Enhance the target domain medical image to obtain an enhanced image. Then, input the original image and the enhanced image into the encoder for feature extraction to obtain the original image features and the enhanced image features. This step includes the following sub-steps:
[0081] S51: Perform image enhancement on the unlabeled target domain medical image data;
[0082] Specifically, bilinear interpolation is used to process all input images. All images were scaled to a fixed size, and then the images were processed to the same size. Photometric enhancement is achieved by employing methods such as illumination, color, and Gaussian noise reduction to obtain enhanced data. To simulate the domain offset of medical images;
[0083] S52: Use the encoder to extract features from the original image and the enhanced image, and perform instance normalization on the features;
[0084] Specifically, two encoders with shared parameters are used to process the original image. and image enhancement Feature extraction is performed to obtain the original image features. and enhance image features The two features are instance-normalized to maintain the independence between each image instance, thereby accelerating model convergence.
[0085] S6: Input the original image features and enhanced image features into the classifier to obtain the predicted classification probability, and calculate the various losses using the predicted classification probability and intermediate features of the medical image data to obtain the total loss; this step includes the following sub-steps:
[0086] S61: Input the features extracted from the original image into the classifier to obtain the predicted classification probability of each instance image;
[0087] Specifically, the features extracted from the original image The unnormalized predicted classification probability of the original image is obtained by inputting the data into the classifier. Then, the normalized predicted classification probability is obtained using the Softmax activation function. .
[0088] S62: Calculate the cosine similarity between the features extracted from the original image and the class center to obtain the pseudo-label of the image;
[0089] Specifically, by calculating features and various class centers The cosine similarity between them is used to obtain pseudo-labels with high confidence. The specific formula is as follows:
[0090]
[0091] Where cos is the cosine similarity calculation function. Represents a collection of class source domain class data.
[0092] S63: Calculate the pseudo-label supervision loss using the predicted classification probability of the source domain class data and the pseudo-labels. ;
[0093] Specifically, the loss of supervision by false labels The pseudo-label obtained by S62 As a supervisory signal, it is used in conjunction with the output probability of the classifier. The cross-entropy loss is calculated to measure the difference between the source domain sample output and the high-confidence pseudo-label, and the cross-entropy loss is used as the loss function, as shown in the following formula:
[0094]
[0095] in Represents the cross-entropy loss function. and Let S represent the predicted probability distribution of the image and the pseudo-labels generated in S62. Indicates the index of the image. Representative class source domain data domain The amount of data. Minimizing the pseudo-label supervision loss. It can optimize model performance and improve the classification accuracy of the model.
[0096] S64: Calculate the consistency loss using the original and enhanced image input features of all target domain medical image data. ;
[0097] Specifically, consistency loss The original image features are constrained using the squared L2 norm. and enhance image features The consistency between them prompts the Encoder to learn invariant representations, as shown in the following formula:
[0098]
[0099] in Let represent the square of the L2 norm. This is achieved by minimizing the consistency loss. This allows the model to focus on domain-shared semantic features while ignoring domain-specific style features, thereby learning domain-invariant representations and improving the network's out-of-domain generalization performance.
[0100] S65: Calculate the entropy minimization loss using the predicted classification probability of the target domain medical image data. ;
[0101] Specifically, entropy minimization loss Minimize the predicted probability entropy of all target domain samples to enhance prediction confidence, clarify the classification boundary, and thus improve the model's robustness and resistance to interference. The specific formula is as follows:
[0102]
[0103] in .
[0104] S66: Calculate the domain alignment loss using the original image features and class centers of the target domain medical image data;
[0105] Specifically, domain alignment loss Contrastive learning is used to align the features of source domain class data with the features of target domain specific class data, thereby bringing the target domain closer to the source domain in the feature space and improving the accuracy of target domain specific data. Specifically, if If a sample is a source domain sample, then its positive sample is the class center of its class. Negative samples are the average of the features of that class in the target domain's specific class data domain; if For data belonging to a specific target domain, the opposite applies. The specific formula is shown below:
[0106]
[0107] in Represents expectations, This is the temperature coefficient, a common parameter in contrastive learning, primarily used to control the model's ability to distinguish between negative samples. Indicates sample Features Indicates a positive sample. This indicates a negative sample.
[0108] S67: Combine all losses into the final total loss. ;
[0109] Specifically, the total loss The pseudo-label supervision loss Consistency loss Domain alignment loss and entropy minimization loss A linear combination of [variables]. The total loss is shown in the following formula.
[0110]
[0111] in, and This is a hyperparameter used to balance the impact of the two types of losses on the total loss.
[0112] S7: Fine-tune the medical image classification model by minimizing the total loss until the total loss reaches the preset convergence condition, then stop training.
[0113] Specifically, after calculating the total loss, the backpropagation algorithm is used to update the model's parameters until the total loss function reaches the preset convergence condition, at which point training stops.
[0114] S8: Use the fine-tuned medical image classification model to predict the target domain medical image data and obtain the final classification result;
[0115] Specifically, bilinear interpolation is used to scale all target domain medical images to the same size. The processed images are then input into the trained model to obtain unnormalized predicted classification probabilities. The normalized predicted classification probabilities are obtained by processing with the Softmax function, and the category corresponding to the maximum probability is taken as the category of the target domain medical image.
[0116] As described in the above embodiments, this application introduces a Gaussian Mixture Model (GMM) method to screen high-confidence samples, thereby improving the confidence of the generated pseudo-labels and thus better enhancing the adaptive performance of the model. Furthermore, in the pseudo-label supervision process, this application only uses high-confidence samples for supervised training, using all samples for regularization optimization. This promotes the model's continuous optimization towards the target domain during training without causing excessive deviation. A contrastive learning method is used to bring the high-confidence source domain and target domain data domains closer together in the feature space, and multiple loss functions are combined for optimization, further improving model performance. This application has strong universality and can be applied to various adaptive classification tasks in passive domains.
[0117] Corresponding to the aforementioned embodiments of the passive domain adaptive method for medical image classification, this application also provides embodiments of a passive domain adaptive device for medical image classification.
[0118] Figure 3 This is a block diagram of a passive domain adaptive system for medical image classification, illustrated according to an exemplary embodiment. (Refer to...) Figure 3 The system includes:
[0119] Initialization module 1 is used to perform supervised training on a medical image classification model on source domain medical image data, thereby initializing a medical image classification model in the target domain. The medical image classification model consists of an encoder and a classifier.
[0120] The feature extraction and probability prediction module 2 is used to input unlabeled target domain medical image data into the encoder for feature extraction, and then input the extracted features into the classifier to obtain the classification probability.
[0121] The data domain partitioning module 3 is used to take the classification probability as input to the Gaussian mixture model to optimize the fitting of the Gaussian mixture model, and to use the value of the first distribution of the target domain medical image data in the optimized fitting Gaussian mixture model to partition the target domain medical image data into a source domain class and a target domain specific class.
[0122] The class center generation module 4 is used to generate the class center of each class based on the characteristics of the class source domain class;
[0123] The image enhancement and extraction module 5 is used to enhance the target domain medical image to obtain an enhanced image, and then input the original image and the enhanced image into the encoder for feature extraction to obtain the original image features and the enhanced image features.
[0124] The loss calculation module 6 is used to input the original image features and enhanced image features into the classifier to obtain the predicted classification probability, and to calculate various losses using the predicted classification probability and intermediate features of the medical image data to obtain the total loss.
[0125] The fine-tuning training module 7 is used to fine-tune the medical image classification model by minimizing the total loss until the total loss reaches the preset convergence condition and training stops.
[0126] Prediction module 8 is used to predict the target domain medical image data using the fine-tuned medical image classification model to obtain the final classification result.
[0127] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.
[0128] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.
[0129] Accordingly, this application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; and when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the passive domain adaptive method for medical image classification as described above.
[0130] Accordingly, this application also provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the passive domain adaptive method for medical image classification as described above.
[0131] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.
[0132] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.
Claims
1. A passive domain adaptive method for medical image classification, characterized in that, include: Supervised training of a medical image classification model is performed on source domain medical image data to initialize a medical image classification model in the target domain, wherein the medical image classification model consists of an encoder and a classifier. Unlabeled target domain medical image data is input into the encoder for feature extraction, and then the extracted features are input into the classifier to obtain the classification probability. The classification probability is used as input to the Gaussian mixture model to optimize the fitting of the Gaussian mixture model, and the value of the first distribution of the target domain medical image data in the optimized fitting Gaussian mixture model is used to divide the target domain medical image data into source domain class and target domain specific class. Based on the characteristics of the source domain class, generate the class center for each class; The target domain medical image is enhanced to obtain an enhanced image. Then, the original image and the enhanced image are input into the encoder for feature extraction to obtain the original image features and the enhanced image features. The original image features and enhanced image features are input into the classifier to obtain the predicted classification probability. The predicted classification probability and intermediate features of the medical image data are used to calculate the various losses and obtain the total loss. The medical image classification model is fine-tuned by minimizing the total loss until the total loss reaches the preset convergence condition, at which point training stops. The finely tuned medical image classification model is used to predict the target domain medical image data to obtain the final classification result.
2. The method according to claim 1, characterized in that, Supervised training of a medical image classification model on source domain medical image data is performed to initialize a medical image classification model on the target domain, including the following sub-steps: Labeled source domain medical image data are subjected to uniform size and normalization processing; Construct a medical image classification model consisting of an encoder (Encoder) and a classifier (Classifier); The processed source domain medical image data is input into the medical image classification model, and the predicted classification probability is output. The classification loss is calculated based on the predicted classification probability and the one-hot encoding of the true label; Backward gradient propagation is performed based on the classification loss to update the parameters of the medical image classification model; After the medical image classification model is trained, the medical image classification model trained on the source domain medical image data is used to initialize the medical image classification model in the target domain.
3. The method according to claim 1, characterized in that, Unlabeled target domain medical image data is input into the encoder for feature extraction to obtain intermediate features. The extracted features are then input into the classifier to obtain the classification probability, including the following sub-steps: The unlabeled target domain medical image data is subjected to the same uniform size and normalization process as the source domain medical image data to obtain the processed target domain medical image. The processed target domain medical image is input into the encoder for feature extraction, and the intermediate features of the image are output. The intermediate features are input into the classifier Classifer to obtain the classification probability of the image.
4. The method according to claim 1, characterized in that, The classification probability is used as input to a Gaussian mixture model to optimize the fit of the Gaussian mixture model. The target domain medical image data is then divided into source domain classes and target domain-specific classes using the value of the first distribution in the optimized Gaussian mixture model. This process includes the following sub-steps: The Gaussian mixture model is fitted and optimized based on the classification probability of the target domain medical image. Obtain the probability values of the first distribution of all medical image samples in the target domain in the Gaussian mixture model; Set data partitioning thresholds; Based on the probability value and the data segmentation threshold, the target domain data is divided into source domain classes and target domain specific classes.
5. The method according to claim 1, characterized in that, Based on the characteristics of the source domain class, the class center of each class is generated, including the following sub-steps: Based on the classification probability of the target domain medical image samples, the source domain data is divided into K classes, where K is the number of classes; The features of medical image samples of the same category in the target domain are averaged to obtain the class center of each class.
6. The method according to claim 1, characterized in that, The original image features and enhanced image features are input into a classifier to obtain predicted classification probabilities. The predicted classification probabilities and intermediate features of the medical image data are used to calculate various losses to obtain the total loss. This includes the following steps: The features extracted from the original image are input into the classifier to obtain the predicted classification probability of each instance image; The cosine similarity between the features extracted from the original image and the class centers is calculated to obtain the pseudo-label of the image; The pseudo-label supervision loss is calculated using the predicted classification probability of the source domain class data and the pseudo-labels. The consistency loss is calculated using the original and enhanced image input features of all target domain medical image data. The entropy minimization loss is calculated using the predicted classification probability of the target domain medical image data; The domain alignment loss is calculated using the original image features and class centers of the target domain medical image data; Combine all losses into a total loss.
7. The method according to claim 1, characterized in that, The trained model is used to predict the classification results of medical images in the target domain, and the final classification results of the samples are obtained. This includes the following sub-steps: Obtain all target domain images The target domain medical image data is input into the trained model to obtain the predicted output probability; Set the category corresponding to the maximum probability as the category of the target domain medical image.
8. A passive domain adaptive system for medical image classification, characterized in that, include: An initialization module is used to perform supervised training on a medical image classification model on source domain medical image data, thereby initializing a medical image classification model in the target domain, wherein the medical image classification model consists of an encoder and a classifier. The feature extraction and probability prediction module is used to input unlabeled target domain medical image data into the encoder for feature extraction, and then input the extracted features into the classifier to obtain the classification probability. The data domain partitioning module is used to take the classification probability as input to the Gaussian mixture model to optimize the fitting of the Gaussian mixture model, and to use the value of the first distribution of the target domain medical image data in the optimized fitting Gaussian mixture model to partition the target domain medical image data into a source domain class and a target domain specific class. The class center generation module is used to generate the class center of each class based on the characteristics of the class source domain class; The image enhancement and extraction module is used to enhance the target domain medical image to obtain the enhanced image, and then input the original image and the enhanced image into the encoder for feature extraction to obtain the original image features and the enhanced image features; The loss calculation module is used to input the original image features and enhanced image features into the classifier to obtain the predicted classification probability, and to calculate various losses using the predicted classification probability and intermediate features of the medical image data to obtain the total loss; The fine-tuning training module is used to fine-tune the medical image classification model by minimizing the total loss until the total loss reaches the preset convergence condition and training stops. The prediction module is used to predict the target domain medical image data using the fine-tuned medical image classification model to obtain the final classification result.
9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-7.
10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-7.