Medical image classification method and device based on high-order moment field alignment, and storage medium

Abstract

The invention relates to the technical field of medical image analysis, in particular to a medical image classification method and device based on high-order moment field alignment and a storage medium, and the method comprises the steps: extracting original medical image features of a labeled source domain and an unlabeled target domain; the extracted features are input into a high-order moment feature mixer module for cross-domain feature alignment; encoding the aligned high-order features into embedded representation and inputting the embedded representation into a self-adaptive classifier module for source domain classification prediction; constructing a cost matrix function through an improved partial optimal transmission strategy, and calculating domain alignment loss to sparise a coupling matrix; implementing intra-class compression and inter-class expansion, and dynamically aligning same-class centers of a source domain and a target domain; model parameters are optimized, and precise classification of the target domain medical images is achieved. Compared with the prior art, the method has the advantages that the problem of mismatching of feature alignment and the problem of negative migration in unsupervised field self-adaption can be effectively relieved, and the migration performance, robustness and adaptability of the model are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image analysis, and in particular, to a medical image classification method, device, and storage medium based on high-order moment domain alignment. Background Art

[0002] With the help of supervised learning, deep convolutional neural networks (CNNs) have made remarkable progress in various computer vision tasks. However, supervised learning methods rely on a large amount of manually labeled data, and obtaining such labeled data is both time-consuming and expensive. In addition, most existing supervised learning methods usually assume that the training set and the test set follow the same distribution. However, in reality, due to changes in image acquisition devices, shooting angles, or scenes, there is often a distribution shift between the training domain and the test domain. This distribution shift can significantly affect the generalization ability of the model and severely limit its deployment ability in practical applications.

[0003] To address this problem, unsupervised domain adaptation (UDA) has been proposed, whose goal is to improve the performance of the model in the target domain when the source domain (with labels) and the target domain (without labels) have different distributions. Currently, the mainstream UDA methods mainly reduce the domain difference through two types of strategies: one is to learn domain-invariant features by explicitly minimizing the distribution difference between the source domain and the target domain features; the other is to use an adversarial learning strategy, using a discriminator to make the features of the source domain and the target domain indistinguishable, thereby achieving implicit alignment.

[0004] With the popularization of digital pathology slide scanners, computer-aided diagnostic analysis (CAD) has become an effective way to reduce the misdiagnosis rate of pathologists and shorten the diagnosis time. The existing deep learning methods based on histopathological images mainly face the following limitations: On the one hand, existing unsupervised domain adaptation methods usually ignore sample category information and semantic consistency during feature alignment, which is prone to incorrect matching, especially when the inter-class distribution overlaps or the boundaries are blurred, resulting in hindered discriminative feature learning and affecting the classification performance. For example, CN118674984A proposes a classification method based on joint manifold alignment reconstruction. Although it focuses on sub-domain distribution alignment, it does not fully consider the high-order statistical characteristics of features, has limited modeling ability for complex feature distributions, and cannot effectively solve the incorrect matching problem in feature alignment.

[0005] On the other hand, many methods achieve the overall alignment of the feature distributions of the source domain and the target domain by extracting global domain offset features. However, this global alignment strategy is difficult to reflect the fine-grained semantic correspondence between the source domain and the target domain, and is prone to misalignment or feature confusion. CN114241239A proposes an unsupervised pre-adaptation image classification method based on manifold learning. Although it combines manifold learning to extract image features, it mainly focuses on improving feature transferability and discriminability through a manifold learning network and a high-order information matching network, and does not fully solve the problem of inaccurate semantic correspondence caused by global alignment in the feature alignment process, and has insufficient processing ability for the multi-modal distribution of features, and cannot accurately match the feature distributions of the source domain and the target domain.

[0006] In recent years, Optimal Transport (OT) has become an effective means widely used in domain adaptation due to its advantages in modeling the geometric structure between probability distributions. OT attempts to find an optimal coupling strategy to "transport" the distribution of the source domain to the target domain and minimize the transport cost. Nevertheless, the existing OT-based UDA methods still face the following challenges: (1) Wrong matching problem: Under the goal of globally minimizing the transport cost, optimal transport is prone to ignoring the class information and semantic consistency between samples, resulting in some class samples in the target domain being wrongly matched to classes with inconsistent semantics in the source domain. This wrong matching is particularly serious when the inter-class distribution overlaps or the boundaries are blurred, and ultimately will disrupt the learning of discriminative features and affect the downstream classification performance.

[0007] (2) Coarse-grained alignment problem: Traditional OT methods usually perform sample alignment from the perspective of the overall distribution and lack the ability to model local structure or class-level information. This global alignment strategy is difficult to accurately reflect the fine-grained semantic correspondence between the source domain and the target domain, especially when there is a multi-modal structure within a class or the inter-class distributions are similar, it is prone to misalignment or feature confusion.

[0008] (3) Negative transfer problem: Due to the natural distribution differences between the source domain and the target domain, some samples in the target domain may be close to the classification decision boundary or in low-density regions, and these samples are prone to be misclassified during training. When the model is trained only based on the labels provided by the source domain and fails to effectively adapt to the structural features of the target domain samples, it may lead to the generalization of wrong decisions, thus triggering negative transfer, that is, the performance of the model on the target domain actually decreases.

[0009] In summary, to promote the development of OT-based domain adaptation methods to a higher level, it is necessary to overcome the above challenges and explore more flexible, robust and adaptable strategies to achieve more accurate and efficient inter-domain knowledge transfer, so as to realize unsupervised medical image classification. Summary of the Invention

[0010] The object of the present invention is to overcome the defects of the above-mentioned existing technologies, and provide a medical image classification method, device, and storage medium based on high-order moment domain alignment. The present invention can effectively alleviate the problems of feature alignment mismatch and negative transfer in unsupervised domain adaptation, and significantly improve the transfer performance, robustness, and adaptability of the model by combining high-order matrix feature alignment, partial optimal transport strategy, and class center-based clustering regularization method.

[0011] The object of the present invention can be achieved by the following technical solutions: The first aspect of the present invention provides a medical image classification method based on high-order moment domain alignment, which realizes medical image classification through an unsupervised domain adaptation network model. The unsupervised domain adaptation network model includes a feature extractor module, a high-order moment feature mixer module, an adaptive classifier module, and a discriminative class center clustering module. The medical image classification method based on high-order moment domain alignment includes the following steps: S1. Respectively extract the original medical image features of the labeled source domain and the unlabeled target domain through the feature extractor module; S2. Input the features extracted in S1 into the high-order moment feature mixer module, and perform cross-domain feature alignment by combining statistical features of different orders; S3. Encode the aligned high-order features in S2 into an embedding representation through the feature extractor module, and input them into the adaptive classifier module for source domain classification prediction, and synchronously calculate the standard classification loss; S4. Based on the embedding representation generated in S3, construct a cost matrix function through an improved partial optimal transport strategy, and calculate the domain alignment loss to sparsify the coupling matrix; S5. Utilize the embedding representation generated in S3 and the transmission result in S4, implement in-class compression and between-class expansion through the discriminative class center clustering module, dynamically align the source domain and the target domain of the same class center, and calculate the clustering loss; S6. Synthesize the standard classification loss in S3, the domain alignment loss in S4, and the clustering loss in S5, optimize the model parameters through backpropagation, and classify the medical images in the target domain based on the optimized model.

[0012] Further, in S1, the labeled source domain is , the unlabeled target domain is , and the two domains share the same label space ; The samples in the source domain and the target domain follow different probability distributions and , that is, ; Wherein is the i-th medical image sample in the source domain, is the corresponding clinical diagnosis label, is the total number of samples in the source domain, is the total number of samples in the target domain, is the j-th medical image sample in the target domain, is the source domain label space, is the target domain label space, is the joint distribution of the source domain medical image data, is the marginal distribution of the target domain medical image data.

[0013] Furthermore, in S1, the feature extractor module adopts a two-stream feature extractor, where: the first feature extractor is used to extract deep features, and the second feature extractor encodes the fused high-order features into image embeddings .

[0014] Furthermore, in S2, the high-order moment feature mixer module realizes cross-domain feature alignment through the following operations: Uses multiple 1×1 randomly initialized convolutional kernels as projectors to generate statistical features of different orders step by step; Integrates features of each order through a cascaded high-order module to avoid degradation of high-order features.

[0015] Furthermore, in S2, the high-order moment feature mixer module includes multiple 1×1 convolutional kernels, and the multiple 1×1 convolutional kernels are used as random projectors to estimate the K-order moment features: , where represents the Hadamard product, represents a 1×1 convolutional kernel with randomly initialized weights, represents a K-order moment feature, is the depth feature map extracted by the first feature extractor of the convolutional neural network, representing the height, width, and channel dimensions respectively.

[0016] Furthermore, in S2, the information of different orders of moments is integrated through a cascaded high-order module ; The cascaded high-order module is used to estimate high-order moment features and prevent high-order features from degrading into low-order features during a large number of Hadamard product operations, specifically expressed as: , In the cascaded high-order module The output end integrates information of different moments for characterizing complex features, specifically expressed as: .

[0017] Furthermore, in S3, the input of the adaptive classifier module is the image embedding encoded by high-order features, and the adaptive classifier module outputs the final classification prediction result , and the standard classification loss on the source domain is defined as: , where is the standard cross-entropy loss, is the batch size, represents the adaptive classifier operation, represents the high-order moment feature mixer operation, represents the first feature extractor operation, represents the second feature extractor operation.

[0018] Furthermore, in S4, the cost matrix function constructed by the improved partial optimal transport strategy is expressed as: , where is a trade-off parameter; the domain alignment loss is: , where represents the optimized cross-domain transport weight.

[0019] Furthermore, in S5, the discriminative class center clustering module guides the alignment of the same-class centers in the source domain and the target domain through intra-class compression and inter-class expansion; the function of the clustering loss is expressed as: , where is the distance constraint boundary for controlling the distance between clusters in the source domain, is the distance constraint boundary for making the class centers of the same class closer, , represents the -th class center of the deep features in the source domain, , is the number of classes, is 's pseudo-label, is the estimated class center in the target domain; In S6, by integrating the standard classification loss in S3, the domain alignment loss in S4, and the clustering loss in S5, and optimizing the model parameters through backpropagation, the total loss function adopted is expressed as: , where the hyperparameters and are balance parameters.

[0020] In the second aspect of the present invention, an electronic device is provided, including a memory and a processor. The processor is configured to execute the program in the memory to implement the medical image classification method based on high-order moment domain alignment as described above.

[0021] In the third aspect of the present invention, a storage medium containing computer-executable instructions is provided. When the computer-executable instructions in the storage medium are executed by a computer processor, they are used to execute the medical image classification method based on high-order moment domain alignment as described above.

[0022] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention can significantly alleviate the problem of incorrect matching of cross-domain features. By introducing an improved Partial Optimal Transport (POT) strategy, restricting the transport quality and introducing a sparse coupling mechanism, it effectively avoids the misalignment of low-quality or heterogeneous samples, thereby achieving more accurate and reliable cross-domain matching, which has a positive effect on improving the accuracy and transfer performance of feature alignment.

[0023] 2. The present invention can improve the discriminative ability of features and the clarity of class structure. The designed discriminative class center clustering mechanism is based on the true labels of the source domain and the pseudo-labels of the target domain, guiding the aggregation of similar samples and the separation of different samples, and promoting the consistency of the class centers of the source domain and the target domain, making the deep embedding space have stronger intra-class compactness and inter-class separability, providing a guarantee for the subsequent classifier to learn a clearer decision boundary.

[0024] 3. The present invention can enhance the model's ability to model complex distributions and expressiveness. By proposing a High-Order Moment Mixer, fusing the statistical information of different orders, capturing the non-linear changes and high-order relationships of features, and transcending the limitations of traditional first-order or second-order feature representations, thereby improving the model's adaptation and generalization ability to complex or non-Gaussian distribution data.

[0025] 4. The present invention can achieve leading performance on multiple mainstream unsupervised domain adaptive datasets. Through comprehensive empirical research on datasets such as Office-31, Office-Home, and Digits, the effectiveness and stability of the present invention in multiple tasks are verified. At the same time, systematic ablation experiments further prove the independent contributions and collaborative values of each module in the overall framework, demonstrating the broad potential and promotion value of this method in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a structural block diagram of the unsupervised medical image classification method based on high-order moment domain alignment in Embodiment 1; Figure 2 It is an implementation flowchart of the unsupervised medical image classification method based on high-order moment domain alignment in Embodiment 1; Figure 3 It is a schematic diagram of the HOMDA network architecture in Embodiment 2; Figure 4 It is a t-SNE embedding visualization result graph of the comparison between HOMDA and other methods in Embodiment 2. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0027] Overall, the present invention provides a medical image classification method based on high-order moment domain alignment. The classification method is as follows: extract the image features of the source domain and the target domain, then use the high-order moment feature mixer for feature alignment, and enhance the discrimination ability of the features and enrich the feature representation by combining statistical features of different orders; then encode the high-order features into embedding representations and input them into a classifier for classification prediction; then limit the transmission quality and sparsify the coupling matrix through an improved partial optimal transport strategy to reduce incorrect matches; then perform intra-class compression and inter-class expansion on the embedding features through discriminative class center-guided clustering regularization, guide the alignment of the same-class centers in the source domain and the target domain, and enhance the discriminability and migration stability of the features, so as to obtain accurate medical image classification results. The analysis method uses a feature extractor module, a high-order matrix feature mixer module, an adaptive classifier module, and a discriminative class center-guided clustering module.

[0028] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Features such as component models, material names, connection structures, control methods, algorithms, etc. that are not clearly described in this technical solution are regarded as common technical features disclosed in the prior art.

[0029] Embodiment 1 This embodiment provides a medical image classification method based on high-order moment domain alignment, as Figure 1 and Figure 2As shown in the figure, it includes: extracting the image features of the source domain and the target domain, then using a high-order moment feature mixer for feature alignment, and enhancing the discriminative ability of features and enriching feature representation by combining statistical features of different orders; then encoding the high-order features into an embedded representation and inputting them into a classifier for classification prediction; then restricting the transmission quality and sparsifying the coupling matrix through an improved partial optimal transport strategy, significantly reducing incorrect matches; and then performing intra-class compression and inter-class expansion on the embedded features through discriminative class center-guided clustering regularization, guiding the alignment of the same-class centers in the source domain and the target domain, enhancing the discriminative ability and migration stability of features, so as to obtain accurate classification prediction results, thereby realizing medical image classification. The specific process includes: S1: Construction of the unsupervised domain adaptation network HOMDA: Set two feature extractors, and use a two-stream convolutional neural network to extract features from the input images and encode them into an embedded representation; Construct an adaptive classifier, perform classification prediction according to the input features of the source domain, and calculate the standard classification loss; S2: High-order moment feature mixing: Use the random Maclaurin decomposition method with multiple 1×1 convolutional kernels as random projectors to calculate high-order statistical features, and use the cascaded high-order module H to integrate the information of different moments; S3: Application of the partial optimal transport strategy: Use the method of introducing virtual points and regularization optimization to construct a cost matrix, and calculate the final optimized target matrix; calculate the final partial optimal domain alignment loss to optimize the learning process of the model; S4: Class center-guided clustering regularization: Use the momentum update strategy to gradually update the class centers, strengthen intra-class compactness and inter-class separation in the feature space, and reduce the negative transfer caused by pseudo-label errors; calculate the module loss function to optimize the learning process of the model; Calculate the overall loss function, adjust the model parameters through backpropagation, and achieve the best alignment of feature distributions.

[0030] According to the criteria of unsupervised domain adaptation, define the labeled source domain as and the unlabeled target domain as . The two domains share the same label space . The samples in the source domain and the target domain follow different probability distributions and , that is .

[0031] In specific implementation, for the medical image classification method based on high-order moment domain alignment in this embodiment, the image classification method adopts a feature extractor module, a high-order matrix feature mixer module, an adaptive classifier module, and a discriminative class center-guided clustering module; the feature extractor module includes two feature extractors, one for extracting deep features and the other for encoding the mixed high-order features into an embedding representation; the high-order matrix feature mixer module is used for feature alignment, calculating high-order statistical features and combining statistical features of different orders; the adaptive classifier module is used for feature classification prediction of the source domain; the discriminative class center-guided clustering module is used for compact clustering and separation between classes, and gradually updates the class center by adopting a momentum update strategy; the feature extractor module is used for extracting original image samples, the high-order matrix feature mixer module is used for enhancing the discrimination ability of features and enriching the feature representation; the discriminative class center-guided clustering module is used for reducing the negative transfer caused by pseudo-label errors.

[0032] The following are the explanations of each module involved in the present invention: The feature extractor module is an important part of the unsupervised domain adaptation network model, and its main function is to extract discriminative features from medical images. It can process labeled source domain data and unlabeled target domain data, and extract the deep features and high-order features of the original medical images.

[0033] The role of the high-order moment feature mixer module is to enhance the discrimination ability of features and enrich the feature representation. It combines statistical features of different orders, uses multiple 1×1 convolutional kernels as random projectors to estimate high-order moment features, and integrates the information of different order moments. In this way, it can capture the non-linear changes and high-order relationships of features, thereby improving the adaptation and generalization ability of the model to complex or non-Gaussian distribution data. The adaptive classifier module is used for classification prediction according to the features extracted by the feature extractor module. It receives the feature encodings processed by the high-order moment feature mixer module, and these feature encodings contain the alignment information of the source domain and the target domain. The adaptive classifier module can learn according to the label information of the source domain and transfer its knowledge to the target domain, so as to classify the medical images of the target domain.

[0034] The discriminative class center clustering module aims to enhance the discrimination and transfer stability of features through intra-class compression and inter-class expansion. It gradually updates the class center by adopting a momentum update strategy, strengthening the intra-class compactness and inter-class separation in the feature space. This module can reduce the negative transfer caused by pseudo-label errors, ensuring that the same-class samples are closer in the feature space and different-class samples are farther apart.

[0035] The unsupervised domain adaptation network based on high-order moment domain alignment is as Figure 3As shown, a two-stream convolutional neural network is adopted. The first feature extractor of the feature extractor module maps the original input image to deep features ; the second feature extractor encodes the fused high-order features into image embeddings .

[0036] According to the criteria of unsupervised domain adaptation, the labeled source domain is defined as , and the unlabeled target domain is . The two domains share the same label space . The samples in the source domain and the target domain follow different probability distributions and , that is . Among them is the i-th medical image sample in the source domain, is the corresponding clinical diagnosis label, is the total number of samples in the source domain, is the total number of samples in the target domain, is the j-th medical image sample in the target domain, is the source domain label space, is the target domain label space, is the joint distribution of the source domain medical image data, is the marginal distribution of the target domain medical image data.

[0037] The high-order moment feature mixer module adopts a random Maclaurin decomposition scheme to solve the problem of excessive time consumption of the Kronecker product of high-dimensional features, including multiple 1×1 convolutional kernels as random projectors to estimate the K-order moment features: Among them represents the Hadamard (element-wise) product, represents a 1×1 convolutional kernel with randomly initialized weights, represents a K-order moment feature, is the depth feature map extracted by the first feature extractor of the convolutional neural network, representing the height, width, and channel dimensions respectively. [[ID=,67]]

[0038] After that, the estimated high-order features are input into the cascade module to prevent the high-order features from degenerating into low-order features during a large number of Hadamard product operations: , in the cascade module The output end integrates information of different moments and is used to characterize complex features: , The input of the adaptive classifier module is the image embedding encoded by high-order features, and the output is the final classification prediction result . The standard classification loss on the source domain is defined as: , where is the standard cross-entropy loss, is the batch size, represents the operation of the adaptive classifier, represents the operation of the high-order moment feature mixer, represents the operation of the first feature extractor, represents the operation of the second feature extractor.

[0039] The improved partial optimal transport strategy is used to align the representations of the source domain and the target domain, further alleviating the problem of incorrect matching. Define to represent the transport probability mass. The improved partial optimal transport strategy reduces the mismatching situation by restricting the probability mass transported from the source domain to the target domain; the mass , making cross-domain samples from the same class more likely to match. The improved optimal transport strategy attempts to find an optimal coupling matrix that satisfies the following improved constraint conditions : , where represents or , which can further reduce the small mass mismatching between source samples and target samples. Under this constraint condition, cross-domain samples from the same class can be well matched, while some incorrect mapping relationships between cross-domain samples from different classes are effectively cut off. The improved POT optimization problem can now be formulated as: , where represents the cost required for the probability mass to be transported from to .

[0040] By adding virtual points and , expanding the cost matrix , and taking group-lasso regularization for simplification , a sparse transport mapping can be generated. Where , is a fixed positive scalar, representation of the th row and the first columns of the corresponding sub-vector. Assume that the masses of the added virtual points are and . Thus, it can be obtained that the improved partial optimal transport problem only needs to solve: , where , are all extended vectors, is the regularization parameter. The cost function of the partial optimal transport strategy proposed by the present invention is a weighted mixture of distances in the feature space and the label space ， Therefore, the cost function obtained by the improved partial optimal transport is expressed as: , where is a trade-off parameter.

[0041] The domain difference between the source domain and the target domain, that is, the domain alignment loss is defined as: , The discriminative class center clustering loss function is a loss function designed in the discriminative class center clustering module, which can make the samples from the same class closer in the embedding space and the samples from different classes farther apart: , where is the distance constraint boundary, which is used to control the distance between clusters in the source domain. is the distance constraint boundary, which is used to make the class centers of the same class closer ( ). represents the th class center of the deep features in the source domain, , is the number of classes. is the pseudo label of, is the estimated class center in the target domain.

[0042] Ideally, the class centers should be accurately estimated based on the entire training set, but in practice it is both time-consuming and unrealistic. Therefore, in each iteration, the following moving average method is adopted to update : , , where is the learning frequency for updating the class center, is the In the center of the category in the next iteration, note that when the condition is met, otherwise

[0043] For the target domain, samples with higher confidence are selected from the target samples whose pseudo-label probabilities are higher than the threshold for class center estimation. The class centers of the target domain are initialized with these reliable samples, and then updated in each iteration in the following way: , , The total loss function of the HOMDA model can be expressed as: , where the hyperparameters and are balance parameters.

[0044] Example 2 This example provides an electronic device equipped with a high-performance memory and a high-speed processor. The memory is used to store the programs and data sets required to implement the medical image classification method based on high-order moment domain alignment. The processor is responsible for executing the programs in the memory to implement a series of complex operations, from the extraction of image features in the source domain and target domain, to high-order moment feature mixing, classification prediction, domain alignment loss calculation, and clustering optimization.

[0045] The electronic device receives medical image data through a high-speed network interface and, with its powerful computing power, efficiently completes the unsupervised domain adaptation process. Specifically, the processor first calls the feature extraction module to extract the original image features from the labeled source domain and the unlabeled target domain. Then, these features are aligned through the high-order moment feature mixer module to enhance the discriminative ability of the features. Next, the processor uses the adaptive classifier module to classify and predict the source domain images and calculates the classification loss based on the prediction results. At the same time, through an improved partial optimal transport strategy, the processor constructs a cost matrix function, calculates the domain alignment loss to sparsify the coupling matrix, and reduces misalignments. In addition, the processor also uses the discriminative class center clustering module to perform intra-class compression and inter-class expansion, dynamically aligning the same-class centers of the source domain and the target domain. Finally, by integrating all the loss functions, the model parameters are optimized through backpropagation, thus achieving accurate classification of the target domain medical images. This electronic device demonstrates excellent classification performance and high computing power in practical applications and is suitable for large-scale medical image analysis scenarios.

[0046] Example 3 This embodiment provides a storage medium containing computer-executable instructions, which is specifically a high-speed solid-state storage device installed with dedicated image classification software. All the program codes and relevant data required to implement the medical image classification method based on high-order moment domain alignment are stored in this storage medium. When the storage medium is read by a computer system, the computer processor executes the stored instructions, thereby realizing a series of complex operations from the extraction of source domain and target domain image features to high-order moment feature mixing, classification prediction, domain alignment loss calculation, and clustering optimization. Specifically, the processor first calls the feature extractor module to extract the original image features from the labeled source domain and the unlabeled target domain, and then aligns these features through the high-order moment feature mixer module to enhance the discriminative ability of the features. Next, the processor uses the adaptive classifier module to perform classification prediction on the source domain images and calculates the classification loss based on the prediction results. At the same time, through the improved partial optimal transport strategy, the processor constructs a cost matrix function and calculates the domain alignment loss to sparsify the coupling matrix and reduce misalignments. In addition, the processor also uses the discriminative class center clustering module to perform intra-class compression and inter-class expansion, and dynamically aligns the same-class centers of the source domain and the target domain. Finally, by integrating all the loss functions, the model parameters are optimized through backpropagation, thereby realizing the accurate classification of the target domain medical images. This storage medium shows excellent compatibility and high execution speed in practical applications, is suitable for various medical image analysis scenarios, and can effectively improve the accuracy and efficiency of diagnosis.

[0047] Application Example 1 In a specific case of the method in an application embodiment, the source domain and target domain data are input, and HOMDA extracts the features of the source domain and target domain data through the first feature extractor in the two-stream convolutional neural network with shared weights; then uses the high-order moment feature mixer for feature alignment to enhance the discriminative ability of the features and enrich the feature representation; applies the improved partial optimal transport strategy to limit the transport quality and sparsify the coupling matrix, significantly reducing misalignments; and then performs intra-class compression and inter-class expansion on the embedded features through discriminative class center-guided clustering regularization to guide the alignment of the same-class centers of the source domain and the target domain, enhancing the discriminative ability and transfer stability of the features. In this process, HOMDA continuously optimizes the network parameters by calculating the classification loss, POT loss, and discriminative class center clustering loss, and finally achieves a high-performance classification task on the unlabeled target domain.

[0048] As Figure 4 shown, t-SNE is used to map the features of the source domain and the target domain to a two-dimensional plane for easy observation and comparison of their distributions, which helps to intuitively understand how the HOMDA model processes the differences between domains and how the improved POT mechanism affects the feature representation. Compared with the UJUOT model, the DeepJDO model, and the JUMBOT model, the transfer accuracy has been significantly improved.

[0049] Compared with the prior art, the present invention can effectively alleviate the feature alignment mismatch problem and the negative transfer problem in unsupervised domain adaptation. On the Digits dataset, the average accuracy of the method provided by the present invention is 98.1%; on the Office-31 dataset, the average accuracy is 89.5%; on the Office–Home dataset, the average accuracy is 70.2%, significantly improving the migration performance, robustness and adaptability of the model.

[0050] The above description of the embodiments is to enable those of ordinary skill in the art to understand and use the invention. It is obvious that those skilled in the art can easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without creative labor. Therefore, the present invention is not limited to the above embodiments, and the improvements and modifications made by those skilled in the art without departing from the scope of the present invention should be within the protection scope of the present invention.

Claims

1. A medical image classification method based on high-order moment domain alignment realizes medical image classification through an unsupervised domain adaptation network model. The unsupervised domain adaptation network model includes a feature extractor module, a high-order moment feature mixer module, an adaptive classifier module, and a discriminative class center clustering module, and is characterized in that The medical image classification method based on high-order moment domain alignment includes the following steps: S1. Respectively extract the original medical image features of the labeled source domain and the unlabeled target domain through the feature extractor module; S2. Input the features extracted in S1 into the high-order moment feature mixer module, and perform cross-domain feature alignment by combining statistical features of different orders; S3. Encode the aligned high-order features in S2 into embedding representations through the feature extractor module, and input them into the adaptive classifier module for source domain classification prediction, and synchronously calculate the standard classification loss; S4. Based on the embedding representations generated in S3, construct a cost matrix function through an improved partial optimal transport strategy, and calculate the domain alignment loss to sparsify the coupling matrix; S5. Utilize the embedding representations generated in S3 and the transmission results in S4, implement intra-class compression and inter-class expansion through the discriminative class center clustering module, dynamically align the source domain and the target domain of the same class center, and calculate the clustering loss; S6. Integrate the standard classification loss in S3, the domain alignment loss in S4, and the clustering loss in S5, optimize the model parameters through backpropagation, and classify the medical images of the target domain based on the optimized model.

2. The medical image classification method based on high-order moment domain alignment according to claim 1, wherein In S1, the labeled source domain is , and the unlabeled target domain is . The two domains share the same label space ; Samples in the source domain and the target domain follow different probability distributions and , that is ; wherein is the i-th medical image sample in the source domain, is the corresponding clinical diagnosis label, is the total number of source domain samples, is the total number of target domain samples, is the j-th medical image sample in the target domain, is the source domain label space, is the target domain label space, is the joint distribution of the source domain medical image data, is the marginal distribution of the target domain medical image data; In S1, the feature extractor module adopts a two-stream feature extractor, where: the first feature extractor is used to extract depth features, and the second feature extractor encodes the fused high-order features into an image embedding .

3. A medical image classification method based on high-order moment domain alignment according to claim 1, characterized in that, In S2, the high-order moment feature mixer module realizes cross-domain feature alignment through the following operations: Use multiple 1×1 randomly initialized convolutional kernels as projectors to generate statistical features of different orders step by step; Integrate features of each order through a cascaded high-order module to avoid degradation of high-order features.

4. A medical image classification method based on high-order moment domain alignment according to claim 3, characterized in that In S2, the high-order moment feature mixer module includes multiple 1×1 convolutional kernels, and the multiple 1×1 convolutional kernels are used as random projectors to estimate the K-order moment features: ， Among them represents the Hadamard product represents a 1×1 convolutional kernel with randomly initialized weights represents a K-th moment feature is the depth feature map extracted by the first feature extractor of the convolutional neural network and represent the height, width, and channel dimensions respectively 5. A medical image classification method based on high-order moment domain alignment according to claim 4, characterized in that In S2, information of different moments is integrated by cascading high-order modules ; The cascaded high-order module is used to estimate high-order moment features and prevent high-order features from degrading into low-order features during a large number of Hadamard product operations, and is specifically described as follows: ， Integrates information of different moments at the output end of the cascaded high-order module to characterize complex features, which is specifically described as follows: 。 6. The medical image classification method based on high-order moment domain alignment according to claim 1, wherein, In S3, the input of the adaptive classifier module is the image embedding encoded by high-order features, and the adaptive classifier module outputs the final classification prediction result , and the standard classification loss on the source domain is defined as: ， wherein is the standard cross-entropy loss, is the batch size, represents the adaptive classifier operation, represents the high-order moment feature mixer operation, represents the first feature extractor operation, represents the second feature extractor operation.

7. A medical image classification method based on high-order moment domain alignment according to claim 1, characterized in that In S4, the cost matrix function constructed by the improved partial optimal transport strategy is expressed as: ， wherein is a trade-off parameter; The domain alignment loss is: ， Among them represents the optimized cross-domain transmission weight.

8. A medical image classification method based on high-order moment domain alignment according to claim 1, characterized in that In S5, the discriminative class center clustering module guides the alignment of the source domain and the target domain of the same class center through intra-class compression and inter-class expansion; The functional representation of the clustering loss is: ， Among them is the distance constraint boundary, which is used to control the distance between clusters in the source domain, is the distance constraint boundary, which is used to make the class centers of the same class closer, , represents the -th class center of the deep features in the source domain, , is the number of classes, is 's pseudo label, is the estimated class center in the target domain; In S6, when integrating the standard classification loss in S3, the domain alignment loss in S4, and the clustering loss in S5, the total loss function adopted in optimizing the model parameters through backpropagation is expressed as: ， Among them, the hyperparameters and are balance parameters.

9. An electronic device, comprising a memory and a processor, characterized in that, The processor is used to execute the program in the memory, thereby implementing the medical image classification method based on high-order moment domain alignment as described in any one of claims 1 to 8.

10. A storage medium containing computer-executable instructions, characterized in that, When the computer-executable instruction storage medium is executed by a computer processor, it is used to execute the medical image classification method based on high-order moment domain alignment as described in any one of claims 1 to 8.

Images