Multi-site unsupervised domain adaptation method for medical image classification

Abstract

The application discloses a multi-site unsupervised domain self-adaptive medical image classification method. First, multi-site medical image samples are acquired, and the samples are preprocessed to obtain feature vectors of the samples. Then, samples of one site are selected as target domain data, and samples of the rest sites are all regarded as source domain data. Feature vectors of the samples of the source domains are stacked to construct a source domain feature matrix, and the source domain feature matrix and the target domain feature matrix are screened to extract key feature information. Then, a classifier is constructed, and a target function is constructed by taking alignment of feature distributions of multiple source domains and a single target domain as an optimization target. Finally, the classifier is iteratively optimized until the target function converges, and optimal classifier parameters are acquired. The method aligns feature distributions between multiple source domains and single target domain data by using a domain self-adaptive method, so as to eliminate distribution differences between the source domain data and the target domain data. In the classifier optimization process, the contribution degrees of the source domains in the migration learning process of the source domain data to the target domain data are calculated, and the method also has high classification precision under the condition of limited samples.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image classification, and in particular to a multi-site unsupervised domain adaptive medical image classification method. BACKGROUND

[0002] In recent years, with the continuous development and progress of medical image technology, medical image analysis has become an indispensable means in medical research, clinical disease diagnosis and treatment. Traditional medical image analysis mainly relies on manual processing by doctors to extract features, which is not only low in efficiency but also prone to judgment errors.

[0003] With the rapid development of deep learning technology, it provides a means for medical image classification, recognition and segmentation. Patent application with publication number CN114359213A discloses a classification method for brain resting state functional magnetic resonance imaging, which uses a combination of traditional machine learning and deep learning methods to analyze the user's neuroimaging results, solving the problem that the analysis method of resting state functional magnetic resonance imaging cannot have high classification performance and low training data volume. The method provides a solution to the problem of low training data volume, but its limitation is still that using only a small amount of data cannot meet the sample data diversity requirements of machine learning and deep learning, so the classification performance improvement is limited. Patent application with publication number CN115221946A belongs to the field of brain neural signal processing technology, especially relates to a brain neural signal classification optimization method and system based on functional magnetic resonance imaging, which uses a signal acquisition device to obtain functional magnetic resonance imaging data of a subject, and performs data format conversion, deletes the front time point, time layer correction, head motion correction, spatial standardization, smoothing, linear drift, white matter and cerebrospinal fluid signal regression, filtering and other preprocessing operations on the collected data; for the preprocessed data, a functional connectivity matrix is generated by using an information entropy algorithm to construct a brain functional network, and then the topological properties of the brain neural network are obtained, and further, a model is built using a machine learning algorithm to obtain the health grade of the subject's brain neural signal, to optimize the classification of the subject and the health control group. The health grade of the subject's brain neural signal obtained can be used as intermediate data to assist doctors in diagnosing brain neural diseases of patients, which is convenient for actual scene application. The invention does not consider the scarcity of functional magnetic resonance imaging data in actual use, and it is difficult to fit the classification model under the condition of small amount of labeled data, and the obtained classification model may not work well. Patent application with publication number CN115116607A discloses a brain disease prediction system based on resting state magnetic resonance transfer learning, which respectively acquires healthy adult resting state magnetic resonance big data set and corresponding patient resting state magnetic resonance data of brain disease, and registers to standard brain space after preprocessing; extract the time signal of each brain region by matching to the brain atlas of standard brain space; construct a deep learning model using graph convolution network and gated recurrent network, and pre-train based on the health adult big data set; based on the pre-trained model and patient resting state functional magnetic resonance data, model fine-tuning and disease prediction are performed. The invention extracts the spatio-temporal features in the resting state magnetic resonance signal, pre-trains the deep learning model using the health adult resting state magnetic resonance big data set, fully excavates the inherent spatio-temporal feature patterns in the health adult resting state magnetic resonance, and migrates the prior spatio-temporal feature patterns learned from the health adult data to brain disease prediction, effectively improving the prediction performance of the model.The application proposes a solution using transfer learning for the case that brain disease samples are difficult to obtain, first, a large number of healthy samples are used to fit a pre-trained model, but the acquisition of healthy sample data still has a high cost, and fine-tuning on such a pre-trained model is difficult to guarantee the accommodation of disease sample diversity.

[0004] The above method classifies using resting state magnetic resonance images, but the current classification of medical images still faces the following challenges: 1) the accuracy of the machine learning model is heavily dependent on a large amount of labeled data, but most small medical institutions cannot have their own large-scale labeled data set; 2) the current public large data set is often divided into multiple sites, the data of different sites are scanned and obtained by different devices of different medical institutions, there is often heterogeneity between the data, and each site only has a small number of labeled samples, and only using the data of a single site cannot meet the requirement of training the model using large-scale data. SUMMARY

[0005] In view of the deficiencies of the prior art, the technical problem to be solved by the present application is to provide a multi-site unsupervised domain adaptive medical image classification method.

[0006] The technical solution adopted by the present application to solve the technical problem is as follows:

[0007] A multi-site unsupervised domain adaptive medical image classification method, characterized in that the method comprises the following steps:

[0008] First, obtain resting state functional magnetic resonance images of the same brain tissue from multiple sites, each site containing multiple samples; pre-process the samples to construct a functional connectivity matrix; process the functional connectivity matrix to obtain a feature vector of the sample;

[0009] Second, construct a source domain feature matrix and a target domain feature matrix, and select the source domain features and the target domain features;

[0010] Step 2.1, group all the pre-processed samples according to the site, randomly select the samples of a site as target domain data, and all the samples of the remaining sites as source domain data, each site corresponds to a group of source domain data, each group of source domain data contains multiple source domain samples, and the feature vectors of all source domain samples in the same group of source domain data are stacked into a source domain feature matrix according to the sample dimension; similarly, the feature vectors of all target domain samples are stacked into a target domain feature matrix according to the sample dimension;

[0011] Step 2.2, select the source domain and target domain features to extract key feature information, and obtain the source domain feature matrix and the target domain feature matrix after feature selection;

[0012] Thirdly, constructing the classifier; constructing the objective function with the alignment of the feature distribution of multiple source domains and a single target domain as the optimization goal, the expression of the objective function is:

[0013]

[0014] In the formula, W D represents the Wasserstein distance, which is related to the transfer cost matrix and the transfer plan matrix; represents the target domain feature distribution estimation, x t represents the target domain feature set, y t represents the target domain pseudo label set generated by the classifier f; alpha j represents the weight of the jth source domain; represents the jth source domain feature distribution estimation, respectively represents the feature set and the label set of the jth source domain; theta represents the classifier parameter; c j , gamma j respectively represent the transfer cost matrix and the transfer plan matrix between the target domain and the jth source domain, beta j represents the jth component of the distance normalized vector beta, and J represents the number of source domains;

[0015] The classifier is initialized, and the pseudo label of the target domain is generated by using the initialized classifier; the transfer cost matrix and the transfer plan matrix between the target domain and each source domain data are calculated; the initial value of the weight of each source domain is set as 1 / J, and the classifier is initialized according to the source domain weight; the target domain feature matrix after feature selection is input into the initialized classifier to generate the target domain pseudo label set; the distance matrix between the target domain label space and the jth source domain label space is calculated according to the target domain pseudo label set and the jth source domain label set, and then the Euclidean distance matrix between the target domain feature set and the jth source domain feature set is calculated; the two kinds of distance matrices are added to obtain the transfer cost matrix between the target domain and the jth source domain, and the expression is:

[0016]

[0017] Among them, represents the Euclidean distance matrix between the target domain feature set x t and the jth source domain feature set , represents the distance matrix between the target domain label space and the jth source domain label space, and lambda t represents a hyperparameter;

[0018] The transfer cost matrix is normalized, the maximum value is selected from each normalized transfer cost matrix, and these maximum values are combined to form a distance normalized vector beta element of R J×1,β j Let be the component of β, representing the maximum value in the j-th normalized transmission cost matrix;

[0019] Mapping each sample of the source domain data to each sample of the target domain data corresponds to a transfer vector; therefore, the source domain contains n. s n source domain samples were obtained in total. s There are J transmission vectors, which form a transmission plan matrix; a transmission plan matrix is ​​generated between each source domain and the destination domain, for a total of J transmission plan matrices. The transmission plan matrix between the destination domain and the j-th source domain is denoted as . The initial transfer plan is to uniformly map each sample in the source domain data to each sample in the target domain data. Therefore, the transfer plan matrix γ... j Each element in the set takes the value 1 / n. s ×n t ;

[0020] The fourth step involves training the classifier using the source domain feature matrix after feature selection, and then optimizing the classifier based on the objective function to obtain the optimized classifier. The target domain data is then input into the optimized classifier to output the classification result.

[0021] Furthermore, for classifier optimization, the objective function includes two parameters to be optimized: the source domain weight vector α and the classifier parameters θ. First, the two parameters are randomly initialized, and the step size of the source domain weight vector is set to μ. α The step size of the classifier parameter θ is μ. θ ;

[0022] Next, with the source domain weight vector α fixed, the classifier parameters θ are updated through gradient descent. The optimization process is expressed as formula (7):

[0023]

[0024] in, express Partial derivatives with respect to θ;

[0025] Then, with the classifier parameters θ fixed, the source domain weight vector α is updated through gradient descent. The optimization process is expressed as formula (8):

[0026]

[0027] in, This means calculating the corresponding weight components in each of the J source domains. express Partial derivatives with respect to α;

[0028] Repeat the above update process until the objective function converges, and obtain the optimized classifier.

[0029] Further, the classifier comprises four fully connected layers, and a Relu activation function is followed after the first three fully connected layers, and a Softmax activation function is followed after the last fully connected layer; the expression of the classifier is:

[0030] F1=Relu(Linear(X)) (1)

[0031] F2=Relu(Linear(F1)) (2)

[0032] F3=Relu(Linear(F2)) (3)

[0033] output=Softmax(Linear(F3)) (4)

[0034] Wherein, Linear(.) represents a fully connected layer, Relu(.) represents a Relu activation function, Softmax(.) represents a Softmax activation function, F1-F3 represent the outputs of the first three fully connected layers respectively, And output represents an output matrix of the classifier.

[0035] Compared with the prior art, the outstanding substantial features and significant progress of the present application are as follows:

[0036] 1. The present application adopts a domain adaptive method to align the feature distribution between multiple source domains and a single target domain data in view of the heterogeneity of different site data, so as to reduce the distribution difference between the source domain and the target domain data. The traditional domain adaptive method only reduces the heterogeneity between the source domain and the single target domain data. The present application minimizes the transmission cost between the source domain and the target domain data, and simultaneously considers the contribution degree of each source domain data in the migration learning process of the source domain data to the target domain data, constructs a target function, trains the classifier, and makes full use of the limited samples of the multi-site data as much as possible under the premise of considering the site difference, so that the method can achieve high classification accuracy on the unlabeled target domain, and small institutions lacking exclusive data sets can train models using public data sets and predict on their own small-scale unlabeled samples.

[0037] 2. The present application adopts an end-to-end form, which is different from the traditional domain adaptive model which first converts the features of the source domain and the target domain samples, and then trains the classifier to perform target domain sample classification prediction. The present application only uses one classifier, reduces the complexity, and has good usability. DETAILED DESCRIPTION

[0038] Figure 1 is the overall flowchart of the present application. DETAILED DESCRIPTION

[0039] The technical solutions of the present application will be described in detail below in combination with the drawings and specific embodiments, but the protection scope of the present application is not limited thereto.

[0040] The present application is a multi-site unsupervised domain adaptation medical image classification method, and the specific process is shown in Figure 1 , comprising the following steps:

[0041] First, obtain medical images of the same brain tissue from multiple sites, each site containing multiple samples; pre-process the samples to obtain the feature vector of each sample;

[0042] The medical images are resting-state functional magnetic resonance images, and all samples are pre-processed to construct a functional connection matrix; since the functional connection matrix has symmetry, the meanings and values of the upper triangular part and the lower triangular part are the same, and the values of the elements on the diagonal line only represent the correlation of each region of interest with itself; to retain only meaningful features, the elements of the diagonal line and the lower triangular part of the functional connection matrix are removed, and then the functional connection matrix is converted into a vector as the feature vector of the sample;

[0043] Second, source domain and target domain data preparation and feature selection;

[0044] Step 2.1, construct the source domain feature matrix, the target domain feature matrix, and the source domain label vector;

[0045] Group all samples pre-processed in the first step according to the site, randomly select the samples of one site as the target domain data, and all samples of the remaining sites as the source domain data, each site corresponds to a group of source domain data, each group of source domain data contains multiple source domain samples, and the feature vectors of all source domain samples in the same group of source domain data are stacked according to the sample dimension to form a source domain feature matrix Source domain feature matrix X s contains the feature information of all source domain samples, n s represents the total number of source domain samples, n represents the length of the feature vector, and R represents the matrix space; the target domain data contains multiple target domain samples, and the feature vectors of all target domain samples are stacked according to the sample dimension to form a target domain feature matrix Target domain feature matrix X t contains the feature information of all target domain samples, n t represents the total number of target domain samples;

[0046] The label information of the source domain data will be used in the subsequent feature selection process, so the labels of each group of source domain data are stacked according to the sample dimension to form a source domain label vector Source domain label vector Y sLabel information of all source domain samples;

[0047] In step 2.2, source domain and target domain features are selected to extract key feature information, and source domain feature matrix and target domain feature matrix after feature selection are obtained;

[0048] Single variable selection method is used as the feature extraction method, F distribution is used as the basis, variance analysis (ANOVA) is calculated by sum of squares and degrees of freedom, F = σ between / σ inner , wherein σ between represents inter-group mean square, σ inner represents intra-group mean square, and the features contained in the feature vector are screened according to the F value; this method is used to select the features of the source domain and the target domain respectively, to reduce the length of the feature vector, and to obtain the source domain feature matrix and target domain feature matrix after feature selection, wherein n' represents the length of the feature vector after feature selection, and n' < n;

[0049] In step 3, a classifier is constructed, and an objective function for iterative optimization of the classifier is designed;

[0050] In step 3.1, a classifier is constructed;

[0051] The classifier includes four fully connected layers, and a Relu activation function is followed after the first three fully connected layers, and a Softmax activation function is followed after the last fully connected layer; the expression of the classifier is:

[0052] F1 = Relu(Linear(X)) (1)

[0053] F2 = Relu(Linear(F1)) (2)

[0054] F3 = Relu(Linear(F2)) (3)

[0055] output = Softmax(Linear(F3)) (4)

[0056] Wherein, Linear(.) represents a fully connected layer, Relu(.) represents a Relu activation function, Softmax(.) represents a Softmax activation function, F1-F3 represent the outputs of the first three fully connected layers respectively, and output represents the output matrix of the classifier;

[0057] In step 3.2, an objective function for iterative optimization of the classifier is designed;

[0058] The domain adaptation task of the application is to realize the alignment of the feature distribution between the source domain and the target domain data, so as to convert the domain adaptation problem into an optimization problem, assuming that there are J source domains in total, and the target of iterative optimization is to align the feature distribution of the J source domains and a single target domain, and the target function is shown in formula (5), and the actual meaning of the target function is to minimize the transmission cost of the source domain and the target domain data;

[0059]

[0060] In the formula, W D represents the Wasserstein distance, and is related to the transmission cost matrix and the transmission plan matrix; represents the target domain feature distribution estimation, x t represents the target domain feature set, y t represents the target domain pseudo label set generated by the classifier f; alpha j represents the weight of the jth source domain, and represents the contribution degree of the source domain in the migration process of the source domain data to the target domain data. represents the jth source domain feature distribution estimation, represents the feature set and the label set of the jth source domain respectively; theta represents the classifier parameter; c j , gamma j represent the transmission cost matrix and the transmission plan matrix between the target domain and the jth source domain respectively, beta j represents the jth component of the distance normalized vector beta;

[0061] In step 3.3, the classifier is initialized, and the pseudo label of the target domain is generated by using the initialized classifier; the transmission cost matrix and the transmission plan matrix between the target domain and each source domain data are calculated.

[0062] The initial value of each source domain weight is set to 1 / J, and the classifier is initialized according to the source domain weight; the target domain feature matrix X' t selected by the feature is input into the initialized classifier, and the pseudo label set y t of the target domain is generated.

[0063] According to the pseudo label set y t of the target domain and the label set of the jth source domain, the distance matrix between the target domain label space and the jth source domain label space is calculated, and then the Euclidean distance matrix between the target domain feature set x t and the jth source domain feature set is calculated; the two kinds of distance matrices are added to obtain the transmission cost matrix c j, see equation (6); each source domain and target domain generates a transfer cost matrix, and J transfer cost matrices are generated in total;

[0064]

[0065] wherein, denotes the Euclidean distance matrix between the target domain feature set x t and the jth source domain feature set , denotes the distance matrix between the target domain label space and the jth source domain label space, λ t denotes a hyperparameter;

[0066] The transfer cost matrix is normalized, the maximum value is selected from each normalized transfer cost matrix, and the maximum values are combined to form a distance normalization vector β∈R J×1 , β j is a component of β, and represents the maximum value in the jth normalized transfer cost matrix;

[0067] The transfer between the source domain and the target domain data is essentially mapping each sample of the source domain data to each sample of the target domain data. Each mapping of a source domain sample to a target domain sample corresponds to a transfer vector. Therefore, the source domain contains n s source domain samples, and n s transfer vectors are obtained in total, which form a transfer plan matrix. Each source domain and target domain generates a transfer plan matrix, and J transfer plan matrices are generated in total. The transfer plan matrix between the target domain and the jth source domain is denoted as The initial transfer plan is to uniformly map each sample in the source domain data to each sample of the target domain data. Therefore, the value of each element in the transfer plan matrix γ j is 1 / n s ×n t ;

[0068] In the fourth step, the source domain feature matrix after feature selection is used to train the classifier, and the classifier is optimized according to the target function to obtain an optimized classifier. The target domain data is input into the optimized classifier to output a classification result.

[0069] The optimization of the classifier is also the optimization of the target function. There are two parameters to be optimized in the target function, which are the source domain weight vector α∈R J×1 and the classifier parameter θ. The source domain weight vector α is composed of the weights of each source domain. If the two parameters are optimized simultaneously, one of them will quickly fall into a local minimum value, and the final classification performance will be poor. Therefore, the alternating projection gradient descent method is used to optimize the two parameters.

[0070] First, two parameters are randomly initialized, the step size of source domain weight vector a is μ α , and the step size of classifier parameter θ is μ θ ;

[0071] Next, fixing the source domain weight vector a, the only variable in the objective function is the classifier parameter θ, which is updated by gradient descent method, and the optimization process is described as formula (7):

[0072]

[0073] Where μ θ represents the step size of the classifier parameter, represents the partial derivative of θ;

[0074] Then, fixing the classifier parameter θ, the only variable in the objective function is the source domain weight vector a, which is updated by gradient descent, and the optimization process is described as formula (8):

[0075]

[0076] Where, represents the calculation of the corresponding weight component on the J source domains, represents the partial derivative of a;

[0077] Repeat the above update process until the objective function converges, and obtain the optimized classifier, and also obtain the optimal source domain weight vector, which reflects the optimal contribution degree of each source domain in the transfer learning process from source domain data to target domain data.

[0078] The training data is: the source domain feature matrix after feature selection and the source domain label vector Y s , The test data is: the target domain feature matrix after feature selection The output of the classifier is a binary classification result matrix

[0079] ​​Experimental verification: the experimental running hardware environment is that the display card is NVIDIA GeForce GTX 1080Ti, the software environment is Windows 10 system, CUDA 10.2 and Pytorch deep learning framework; the experimental data used is ABIDE (Autism Brain Imaging Data Exchange) data set, which combines structural and functional MRI data of 1112 subjects from 17 international collection sites and extensive phenotype information, and the public preprocessed version of the data set is provided by the preprocessed connectome project (PCP), the subjects are re-grouped into 20 sites according to the location, so it is a highly heterogeneous database; the resting state functional magnetic resonance image (rs-fMRI) data is preprocessed by the brain connectome analysis program (CPAC), the preprocessing process includes time layer correction, head motion correction and spatial standardization, and the signal fluctuation caused by scanner error and subject breathing and heart beating is removed by interference regression, then the anatomical automatic labeling (AAL) atlas containing 116 predefined regions of interest is aligned to each image, and the average response time sequence in each region of interest is extracted; the original data used in the experiment of the application is the resting state functional magnetic resonance image composed of time sequences; in the experiment, 5 sites with more sample numbers among the 20 sites are selected, which are LEUVEN, NYU, UCLA, UM and USM.

[0080] In order to verify the effectiveness of the method of the application, the application is compared with traditional machine learning methods and commonly used domain adaptation methods, wherein the traditional methods include SVM and KNN, and the domain adaptation methods include TCA, JDA and JDOT; wherein TCA is used to align the edge distribution of source domain and target domain data by using maximum mean difference; JDA simultaneously aligns the conditional distribution on the basis of TCA; JDOT uses the optimal transport method to align the conditional distribution and edge distribution of source domain and target domain data; the experimental settings select one site as target domain data and the rest as source domain data, and the non-multisource domain method is to directly combine multiple source domain samples together as a source domain; the classification accuracy of various methods is shown in table 1.

[0081] Table 1 Comparison test results of the method of the application and traditional classification methods and domain adaptation methods

[0082]

[0083] As can be seen from Table 1, the classification accuracy of the domain self-adaption method is obviously higher than that of the traditional method, the domain self-adaption method has obvious classification accuracy improvement in processing cross-domain problems; meanwhile, through comparison of the three domain self-adaption methods, it can be known that the distribution difference between different site data can be reduced through aligning the edge distribution and the conditional distribution; compared with the single-source domain self-adaption method, the multi-source domain self-adaption method used in the application has very obvious classification accuracy improvement on all target domains, because the application aligns the feature distribution between the single source domain and the multiple target domain data, eliminates the heterogeneity of different site data, and constructs a target function with the minimum transmission cost between the source domain and each target domain data as the target, optimizes the classifier, and is helpful to improve the classification accuracy.

[0084] The unmentioned part of the application is applicable to the prior art.

Claims

1. A multi-site unsupervised domain adaptive medical image classification method, characterized in that, The method includes the following steps: The first step is to acquire resting-state functional magnetic resonance imaging (fMRI) images of the same brain tissue from multiple sites, with each site containing multiple samples; preprocess the samples to construct a functional connectivity matrix; and process the functional connectivity matrix to obtain the feature vectors of the samples. The second step is to construct the source domain feature matrix and the target domain feature matrix, and then select the source domain features and the target domain features. Step 2.1: Group all preprocessed samples by station, randomly select samples from one station as target domain data, and use all samples from the remaining stations as source domain data. Each station corresponds to a set of source domain data, and each set of source domain data contains multiple source domain samples. Stack the feature vectors of all source domain samples in the same set of source domain data according to the sample dimension to form a source domain feature matrix. Similarly, stack the feature vectors of all target domain samples according to the sample dimension to form a target domain feature matrix. Step 2.2: Select features from the source and target domains, extract key feature information, and obtain the source domain feature matrix and target domain feature matrix after feature selection. The third step is to construct a classifier; an objective function is constructed with the goal of aligning the feature distributions of multiple source domains with a single target domain. The expression for the objective function is as follows: In the formula, W D The Wasserstein distance is represented and is related to the transmission cost matrix and transmission plan matrix; x represents the target domain feature distribution estimation. t y represents the set of features in the target domain. t The set of pseudo-labels representing the target domain is generated by classifier f; α j This represents the weight of the j-th source domain; This represents the estimated feature distribution of the j-th source domain. Let c represent the feature set and label set of the j-th source domain, respectively; θ represents the classifier parameters; j γ j Let β represent the transmission cost matrix and transmission plan matrix between the target domain and the j-th source domain, respectively. j Let J represent the j-th component of the distance normalization vector β, and J represent the number of source domains. The classifier is initialized, and pseudo-labels for the target domain are generated using the initialized classifier. The transmission cost matrix and transmission plan matrix between the target domain and each source domain are calculated. The initial values ​​of the weights for each source domain are set to 1 / J, and the classifier is initialized based on these weights. The target domain feature matrix after feature selection is input into the initialized classifier to generate a set of pseudo-labels for the target domain. The distance matrix between the target domain label space and the j-th source domain label space is calculated based on the pseudo-label set of the target domain and the label set of the j-th source domain. Then, the Euclidean distance matrix between the target domain feature set and the j-th source domain feature set is calculated. The two distance matrices are added together to obtain the transmission cost matrix between the target domain and the j-th source domain, expressed as: in, Represents the target domain feature set x t With the feature set of the j-th source domain The Euclidean distance matrix between them Let λ represent the distance matrix between the target domain label space and the j-th source domain label space. t Indicates hyperparameters; The transmission cost matrix is ​​normalized, and the maximum value is selected from each normalized transmission cost matrix. These maximum values ​​are then used to form a distance normalization vector β∈R. J×1 ,β j Let be the component of β, representing the maximum value in the j-th normalized transmission cost matrix; Mapping each sample of the source domain data to each sample of the target domain data corresponds to a transfer vector; therefore, the source domain contains n. s n source domain samples were obtained in total. s There are J transmission vectors, which form a transmission plan matrix; a transmission plan matrix is ​​generated between each source domain and the destination domain, for a total of J transmission plan matrices. The transmission plan matrix between the destination domain and the j-th source domain is denoted as . n t The total number of samples in the target domain is represented by γ. The initial transfer plan is to uniformly map each sample in the source domain data to each sample in the target domain data. Therefore, the transfer plan matrix γ is... j Each element in the set takes the value 1 / n. s ×n t ; The fourth step involves training the classifier using the source domain feature matrix after feature selection, and then optimizing the classifier based on the objective function to obtain the optimized classifier. The target domain data is then input into the optimized classifier to output the classification result.

2. The multi-site unsupervised domain adaptive medical image classification method according to claim 1, characterized in that, For classifier optimization, the objective function includes two parameters to be optimized: the source domain weight vector α and the classifier parameters θ. First, the two parameters are randomly initialized, and the step size of the source domain weight vector is set to μ. α The step size of the classifier parameter θ is μ. θ ; Next, with the source domain weight vector α fixed, the classifier parameters θ are updated through gradient descent. The optimization process is expressed as formula (7): in, express Partial derivatives with respect to θ; Then, with the classifier parameters θ fixed, the source domain weight vector α is updated through gradient descent. The optimization process is expressed as formula (8): in, This means calculating the corresponding weight components in each of the J source domains. express Partial derivatives with respect to α; Repeat the above update process until the objective function converges, and you will get the optimized classifier.

3. The multi-site unsupervised domain adaptive medical image classification method according to claim 1 or 2, characterized in that, The classifier consists of four fully connected layers. The first three fully connected layers are followed by a ReLU activation function, and the last fully connected layer is followed by a Softmax activation function. The classifier's expression is: F1 = Relu(Linear(X)) (1) F2 = Relu(Linear(F1)) (2) F3 = Relu(Linear(F2)) (3) output=Softmax(Linear(F3)) (4) In this context, Linear(.) represents a fully connected layer, Relu(.) represents the Relu activation function, softmax(.) represents the softmax activation function, and F1 to F3 represent the outputs of the first three fully connected layers, respectively. This represents the output matrix of the classifier.

Images