Liver focal lesion image analysis method based on unbalanced optimal transmission

Abstract

The invention relates to a liver focal lesion image analysis method based on unbalanced optimal transmission. The analysis method comprises the following steps: extracting original image sample features, and embedding Gaussian distribution into each image to encode uncertainty; and then unbalanced optimal transmission loss calculation and classification loss calculation are carried out to realize accurate pairwise matching so as to reduce negative migration, and adaptive learning is carried out by using identification type perception information, so that an optimal transmission strategy is determined, and classification prediction is carried out. According to the analysis method, an unsupervised field adaptive network module, a feature uncertainty estimation mechanism module and a UOT strategy module are adopted. Compared with the prior art, the method has the advantages that the problem of pairwise wrong matching during feature distribution alignment in the existing optimal transmission strategy can be relieved, the uncertainty estimation of the noisy training sample is emphasized, and the migration performance, the robustness and the adaptability of the model can be remarkably improved by combining the feature uncertainty estimation and the UOT strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of medical imaging engineering, and in particular to a liver focal lesion imaging analysis method based on unbalanced optimal transmission. Background Art

[0002] Thanks to the rich resources of annotated data, deep convolutional neural networks have played an important role in the vast arena of computer vision. However, collecting and annotating these massive amounts of data is often accompanied by high costs and a long time investment. Although the advent of the big data era has enabled access to large training data sets across multiple domains, the distribution differences between domains due to factors such as perspective differences, sensor types, and geographic locations are still significant. This directly leads to the dilemma of a sharp drop in performance when the model trained on the fully annotated source domain is directly migrated to the unannotated target domain.

[0003] In order to overcome this obstacle and enhance the cross-domain generalization ability of the model, researchers have actively explored and innovated a variety of unsupervised domain adaptation (UDA) strategies, aiming to cleverly transfer knowledge from the label-rich domain to the target domain seamlessly, even if the latter lacks direct label information. In the field of UDA, a series of technologies designed to reduce the differences between domains have emerged, which use various statistical measures as bridges, including but not limited to maximum mean difference (MMD), correlation alignment (CORAL), Kullback-Leibler divergence (KL divergence) and adversarial training.

[0004] It is worth noting that Optimal Transport (OT), as a powerful probabilistic metric, has been widely used in the field of domain adaptation to quantify the differences between complex distributions. Its uniqueness lies in its ability to capture and encode geometric structures in the probability space, which makes it possible to accurately evaluate and adjust the differences in distributions between domains. However, the current domain adaptation methods based on OT still face the following three major defects:

[0005] (1) Limitations of global matching: Limited by the rigorous framework of OT theory, most existing works tend to perform comprehensive cross-domain sample matching, a process that often covers all sample pairs indiscriminately. This "one-size-fits-all" strategy may lead to incorrect or illogical sample pairing, thereby distorting the accurate assessment of distribution differences and may even induce negative transfer effects, that is, noise or invalid information in the source domain is improperly transferred to the target domain.

[0006] (2) Imbalanced distribution: Classic OT methods assume that the feature distributions of the two domains have the same "quality" (i.e., the number of samples is equal), which is particularly idealistic in real-world scenarios. In fact, the feature distributions between domains often show significant imbalance, so most existing methods are not effective in dealing with this imbalance, limiting their wide applicability in practical applications.

[0007] (3) Lack of uncertainty estimation: In UDA research, a key aspect that is often overlooked is uncertainty estimation, which is particularly important when dealing with datasets containing noise or outliers. These abnormal samples may be incorrectly encoded to extreme positions in the feature space, not only far away from the cluster center to which they should belong, but also close to the decision boundary, thereby interfering with the effective learning of the decision boundary and the reasonable construction of the feature space. If this uncertainty is not properly handled, it will directly affect the effect of domain alignment and the accuracy of the final classification.

[0008] 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, thereby realizing unsupervised diagnosis of focal liver lesions. Summary of the invention

[0009] The purpose of the present invention is to provide a method for analyzing liver focal lesions based on unbalanced optimal transmission in order to overcome the defects of the above-mentioned prior art. By combining feature uncertainty estimation and UOT strategy, the migration performance, robustness and adaptability of the model can be significantly improved, the pairwise mismatching problem of feature distribution alignment in the existing optimal transmission strategy can be alleviated, and more emphasis can be placed on uncertainty estimation of noisy training samples.

[0010] The purpose of the present invention can be achieved by the following technical solutions:

[0011] The present invention provides a liver focal lesion image analysis method based on unbalanced optimal transmission, comprising: extracting original image sample features, and then embedding each image into a Gaussian distribution to encode uncertainty, thereby improving the representation space, enhancing the separability between categories and the compactness within categories;

[0012] Then, the unbalanced optimal transmission loss and classification loss calculations are performed to achieve accurate pairwise matching to reduce negative transfer. The discriminant class perception information is used for adaptive learning to determine the best transmission strategy and perform classification predictions to achieve imaging analysis of focal liver lesions.

[0013] Furthermore, the process of optimizing the unsupervised liver focal lesion image analysis model specifically includes:

[0014] S1: Unsupervised domain adaptation network construction: Two feature extractors are set to extract high-level features from the input liver images using a deep convolutional neural network;

[0015] Construct two adaptive classification networks and dynamically adjust parameters according to the extracted features to adapt to different data characteristics;

[0016] S2: Feature uncertainty estimation: The features are encoded using Gaussian distribution, where the parameters of the distribution are determined by feature statistics and are used to represent the uncertainty of the features. The parameters of the distribution are the mean and variance.

[0017] Define and calculate feature uncertainty loss to force the model to maintain uncertainty in feature representation;

[0018] S3: Application of unbalanced optimal transmission strategy, using Wasserstein distance to construct a cost matrix to measure the differences between different feature distributions; according to the results of the cost matrix, adaptively adjust the learning rate or hyperparameters to optimize the learning process of the model;

[0019] The model parameters are adjusted through an iterative process to minimize the UOT loss and achieve the best alignment of feature distributions.

[0020] Furthermore, the feature extractor G f The original image is input as x, and two depth representations are output, namely the mean μ and variance σ, which are used to describe the Gaussian distribution p(z|x), that is, p(z|x) = G f (x);

[0021] Then extract a sample s from the Gaussian distribution to represent the feature vector of the original input, that is, s~p(z|x);

[0022] The classification network G c Map the extracted samples s to the classification prediction G c (s).

[0023] Furthermore, the image analysis method adopts an unsupervised domain adaptive network module for unbalanced optimal transmission, a feature uncertainty estimation mechanism module, and a UOT strategy module;

[0024] The unsupervised domain adaptive network module of the unbalanced optimal transmission includes two feature extractors and two adaptive classification networks;

[0025] The feature uncertainty estimation mechanism module is used to calculate distribution representation, re-setting parameter skills, source supervision classification loss and feature uncertainty loss;

[0026] The UOT strategy module is used for adaptive large-scale learning based on the cost matrix of Wasserstein distance and implements the optimization steps;

[0027] The unsupervised domain adaptive network module of imbalanced optimal transmission is used to extract original image samples, and the uncertainty estimation mechanism module is used to embed the image into a Gaussian distribution model to encode uncertainty, thereby improving the representation space and enhancing the separability between categories and the compactness within categories;

[0028] The UOT strategy module is used to improve the migration performance and robustness of the model to reduce negative migration.

[0029] Furthermore, the distribution representation is specifically as follows: the uncertainty of the data is encoded as a multivariate Gaussian distribution p(z i |x i ), so that the representation of a single image is from The random embedding s sampled from i , rather than a fixed point estimate;

[0030] The multivariate Gaussian distribution p(z i |x i ) is shown in the following formula:

[0031] p(z i |x i )=N(z i ;μ i ,(σ i ) 2 I)

[0032] Among them, μ i and (σ i ) 2 All from the i-th input image θ1 and θ2 represent network parameters; I represents an identity matrix;

[0033] The resetting parameters are specifically as follows: extract a random noise ε from the standard normal distribution, that is, ε~N(0,I), and then obtain a sample s i As an equivalent sampling, it is expressed as:

[0034] s i =μ i +εσ i ,ε~N(0,I)

[0035] Among them, μ i is the mean, σ i is the variance.

[0036] Furthermore, the source supervised classification loss refers to defining the standard classification loss on the source domain as:

[0037]

[0038] in, represents the standard cross entropy loss, n s is the number of samples.

[0039] Furthermore, the feature uncertainty loss is used to force the model to maintain the uncertainty level of the training samples, preventing the random representation of the model from being degraded to a deterministic point representation during the training process;

[0040] The feature uncertainty loss is specifically: use the entropy of the estimated Gaussian distribution to measure each well-behaved sample x i At a given variance The uncertainty level at , is calculated as:

[0041]

[0042] Among them, diag(·) represents the diagonal vector of the input matrix, D is the total feature dimension, and k represents each dimension. The margin m is introduced to represent the boundary of the overall uncertainty:

[0043]

[0044] Where n is the batch size, and represent the entropy of the estimated Gaussian distribution of the i-th sample in the source batch and the j-th sample in the target batch, respectively.

[0045] Furthermore, the cost matrix based on Wasserstein distance is:

[0046]

[0047] Among them, L ce (·,·) denotes the cross entropy loss; parameters α1 and α2 are scalars measuring the contribution of the distance term.

[0048] Furthermore, the adaptive large-scale learning is specifically: using the cross inner product based on the category center to generate the quality a associated with each feature vector i ,b j (i,j∈[1,n]);

[0049] The category center It is approximately estimated as:

[0050]

[0051] The mass of each source sample is calculated using the following formula:

[0052]

[0053] Among them, k′∈{1,2,3,…C′}, C′ represents the number of categories in the target batch; if but otherwise is an equivalent sampling representation.

[0054] Furthermore, the optimization step refers to optimizing the representation space and the label space simultaneously by reducing the distance between samples of the same category in the source domain and the target domain;

[0055] The domain difference is calculated using the following formula:

[0056]

[0057] Where η is a given regularization coefficient, which represents the optimization scheme γ * smoothness; C is the number of categories in the source batch; a, b are given non-negative vectors; τ is a given regularization parameter; KL represents KL divergence; γ *T For γ * The transposed matrix of .

[0058] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0059] 1. The present invention can model feature uncertainty. Uncertainty-guided Joint Unbalanced Optimal Transport (UJUOT) is the first method to explore and model feature uncertainty in the UDA (unsupervised domain adaptation) image classification task. By encoding uncertainty in feature embeddings, intra-class compactness and inter-class separability between source and target domain features are promoted, which is very beneficial for improving domain alignment and learning of decision boundaries.

[0060] 2. The present invention can reduce false matches. UJUOT has designed a UOT (unbalanced optimal transmission) strategy based on an adaptive quality (or weight) learning scheme, which effectively alleviates the problem of false feature matching. By adaptively learning the quality (or importance) of each feature, UJUOT can achieve more accurate pairwise matching, thereby improving the accuracy of domain adaptation.

[0061] 3. The present invention can improve the migration performance of the model. Combining feature uncertainty estimation and UOT strategy, UJUOT significantly improves the migration performance of the model. In multiple standard benchmark tests, UJUOT's migration performance is better than existing methods, demonstrating its superiority in unsupervised domain adaptation tasks.

[0062] 4. The present invention can improve the robustness and adaptability of the model. UJUOT introduces the FUE (Feature Uncertainty Estimation) mechanism, which represents each image as a distribution in the latent space, enhancing the compactness and separation of the learned deep representation. This not only makes domain alignment easier, but also greatly improves the robustness and adaptability of the network in the presence of abnormal or noisy samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is a structural block diagram of the unsupervised liver focal lesion image analysis method based on unbalanced optimal transmission in Example 1;

[0064] Figure 2 is a flowchart for implementing the unsupervised liver focal lesion image analysis method based on unbalanced optimal transmission in Example 1;

[0065] Figure 3 is a schematic diagram of the UJUOT network architecture in Example 1;

[0066] Figure 4 This is a t-SNE embedding visualization result diagram comparing UJUOT in Example 1 with other methods. DETAILED DESCRIPTION

[0067] The specific implementation methods of the present invention are described in detail below through examples. These examples are implemented on the premise of the scheme described in the present invention, and provide detailed implementation methods and specific operating processes, but the protection scope of the present invention is not limited to the following examples.

[0068] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments. Any features such as structure / module name, control mode, algorithm, process or composition ratio not clearly described in this technical solution are regarded as common technical features disclosed in the prior art.

[0069] Example 1

[0070] This embodiment provides a method for analyzing liver focal lesions images based on unbalanced optimal transmission. Figure 1 and Figure 2 As shown, it includes: extracting features of original image samples, and then embedding each image into a Gaussian distribution to encode uncertainty, thereby improving the representation space, enhancing the separability between categories and the compactness within categories;

[0071] Then, the unbalanced optimal transmission loss calculation and classification loss calculation are performed to achieve accurate pairwise matching to reduce negative transfer. The identification class perception information is used for adaptive learning to determine the best transmission strategy and perform classification prediction to achieve liver focal lesion imaging analysis. The specific steps include:

[0072] S1: Unsupervised domain adaptation network construction: Two feature extractors are set to extract high-level features from the input liver images using a deep convolutional neural network;

[0073] Construct two adaptive classification networks and dynamically adjust parameters according to the extracted features to adapt to different data characteristics;

[0074] The feature extractor G f The original image is input as x, and two depth representations are output, namely the mean μ and variance σ, which are used to describe the Gaussian distribution p(z|x), that is, p(z|x) = G f (x); then extract a sample s from the Gaussian distribution to represent the feature vector of the original input, that is, s~p(z|x); the classification network G c Map the extracted samples s to the classification prediction G c (s).

[0075] S2: Feature uncertainty estimation: The features are encoded using Gaussian distribution, where the parameters of the distribution are determined by feature statistics and are used to represent the uncertainty of the features. The parameters of the distribution are the mean and variance.

[0076] Define and calculate feature uncertainty loss to force the model to maintain uncertainty in feature representation;

[0077] S3: Application of unbalanced optimal transmission strategy, using Wasserstein distance to construct a cost matrix to measure the differences between different feature distributions; according to the results of the cost matrix, adaptively adjust the learning rate or hyperparameters to optimize the learning process of the model;

[0078] The model parameters are adjusted through an iterative process to minimize the UOT loss and achieve the best alignment of feature distributions.

[0079] Example 2

[0080] The present embodiment provides a liver focal lesion image analysis method based on unbalanced optimal transmission, the image analysis method adopts an unbalanced optimal transmission unsupervised domain adaptive network module, a feature uncertainty estimation mechanism module and a UOT strategy (unbalanced optimal sensing strategy) module; the unbalanced optimal transmission unsupervised domain adaptive network module is used to extract original image samples, and the uncertainty estimation mechanism module is used to embed the image into a Gaussian distribution model to encode uncertainty, thereby improving the representation space and enhancing the separability between categories and the compactness within categories; the UOT strategy module is used to improve the migration performance and robustness of the model to reduce negative migration.

[0081] The unsupervised domain adaptation network module of imbalanced optimal transfer includes two feature extractors and two adaptive classification networks; Figure 3As shown in Figure 2, a two-stream twin CNN architecture is adopted, and the source domain and the target domain share weights.

[0082] The feature uncertainty estimation (FUE) mechanism module includes distribution representation, parameter resetting technique, source supervision classification loss and feature uncertainty loss.

[0083] The distribution representation is as follows: the uncertainty of the data is encoded as a multivariate Gaussian distribution p(z i |x i ), so that the representation of a single image is from The random embedding s sampled from i , rather than a fixed point estimate; the multivariate Gaussian distribution p(z i |x i ) is shown in the following formula:

[0084] p(z i |x i )=N(z i ;μ i ,(σ i ) 2 I)

[0085] Among them, μ i and (σ i ) 2 All from the i-th input image θ1 and θ2 represent network parameters; I represents an identity matrix.

[0086] The specific parameters of the resetting are: extract a random noise ε from the standard normal distribution, that is, ε~N(0,I), and then get a sample s i As an equivalent sampling, it is expressed as:

[0087] s i =μ i +εσ i ,ε~N(0,I)

[0088] Among them, μ i is the mean, σ i is the variance.

[0089] The source supervised classification loss refers to the standard classification loss on the source domain defined as:

[0090]

[0091] in, represents the standard cross entropy loss.

[0092] The feature uncertainty loss is used to force the model to maintain the uncertainty level of the training samples, preventing the model from degrading random representations to deterministic point representations during training. The feature uncertainty loss is specifically: use the entropy of the estimated Gaussian distribution to measure each well-behaved sample x i At a given variance The uncertainty level at , is calculated as:

[0093]

[0094] Among them, diag(·) represents the diagonal vector of the input matrix, D is the total feature dimension, and k represents each dimension. The margin m is introduced to represent the boundary of the overall uncertainty:

[0095]

[0096] Where n is the batch size, and represent the entropy of the estimated Gaussian distribution of the i-th sample in the source batch and the j-th sample in the target batch, respectively.

[0097] The Unbalanced Optimal Transducer (UOT) strategy module includes a cost matrix based on Wasserstein distance, adaptive large-scale learning, and optimization steps.

[0098] The cost matrix based on Wasserstein distance is as follows: In the traditional optimal transmission (OT) strategy, the source data characteristics and the target data characteristics are defined as Dirac mass respectively: Among them, the quality a i ,b j ≥0, indicating a vector Represents a i ,b j The position of each entry γ of the transmission scheme γ ij Defines a i and b j The quality of transmission between. The classic OT problem can be expressed as:

[0099]

[0100]

[0101] The cost function matrix C is used to calculate the distance between samples in two different domains. The square of the Euclidean distance is generally used to determine the distance between samples. In the UOT strategy, a "soft" divergence metric is used to replace the "hard" mass conservation, thus transforming the classic OT problem into:

[0102]

[0103] where τ>0 is a given regularization parameter and KL is the Kullback–Leibler divergence, defined as:

[0104]

[0105] Due to the introduction of feature uncertainty, the distance between two samples is no longer the distance between two points, but the distance between two Gaussian distributions. Therefore, the square of Wasserstein distance is used instead of the square of Euclidean distance as the cost function, that is:

[0106]

[0107] in, B represents the Bures metric between positive definite matrices, and the calculation formula is as follows:

[0108]

[0109] Where tr(·) represents the trace of the matrix, (∑) 1 / 2 is the square root of the matrix. In order to simplify the calculation in this model, only the covariance matrix is ​​modeled as a diagonal matrix, that is, ∑ = σ 2 I. In this case, the Bures metric is equal to the Hellinger distance:

[0110]

[0111] In general, the cost function matrix formula is:

[0112]

[0113] Among them, L ce (·,·) represents the cross entropy loss; parameters α1 and α2 are scalars that measure the contribution of the distance term. The label of the target sample Unknown, but can be predicted by the flexible maximum activation function To estimate.

[0114] Adaptive large-scale learning refers to the intuitive assumption that the quality is determined by the feature similarity between the source data and the target data. On this basis, an adaptive large-scale learning scheme is proposed, which uses the cross inner product based on the category center to generate the quality a associated with each feature vector. i ,b j (i,j∈[1,n]).

[0115] In order to make full use of the domain structure information to assist the OT process, Defined as the kth category center of the feature embedding, where k∈{1,2,3,…C}, C represents the number of categories in the source batch. Due to the randomness of samples in a mini-batch, the category center of the source batch It can be approximated by averaging the deep features of samples in the source batch:

[0116]

[0117] Among them, if but otherwise Represents the number of samples of the same type in the source batch. The quality of the target feature vector is determined by calculating the correlation with the source category center:

[0118]

[0119] Using pseudo labels of target samples to generate the quality associated with each source feature vector. Similarly, the category center can be approximately estimated as:

[0120]

[0121] Where k′∈{1,2,3,…C′}, C′ represents the number of categories in the target batch. but otherwise

[0122] The mass of each source sample is calculated using the following formula:

[0123]

[0124] The optimization step is to achieve UOT in a differentiable way, considering the entropy regularized UOT problem as follows:

[0125]

[0126] Where η is a given regularization coefficient, which indicates the smoothness of the transmission plan γ. In this network, η is empirically fixed at 0.001. H(γ) is the entropy regularization, which is calculated as follows:

[0127]

[0128] In order to obtain the optimal solution of γ * , consider the Fenchel-Legendre dual form of the entropy regularized UOT:

[0129]

[0130] Among them, the function F * (·) and G * (·) is the Legendre conjugate pair of KL, and the calculation formula is as follows:

[0131]

[0132] Since a and b are given non-negative vectors, calculating the optimal solution to the above maximization objective is equivalent to calculating the optimal solution to the following dual entropy regularized UOT problem:

[0133]

[0134] The Sinkhorn algorithm is then applied to approximate the optimal solution to the original UOT problem. The exact transmission plan γ is encoded by the optimal dual vectors u and v as follows:

[0135]

[0136] Through the above analysis, the UOT strategy optimizes the representation space and label space at the same time by reducing the distance between samples of the same category between the source domain and the target domain;

[0137] The domain difference is calculated using the following formula:

[0138]

[0139] Among them, η is the given regularization coefficient, which represents the optimization scheme γ * smoothness; C is the number of categories in the source batch; a, b are given non-negative vectors; τ is a given regularization parameter; KL represents KL divergence; γ *T For γ * The transposed matrix of .

[0140] In a specific case of applying the system of this embodiment, the source domain and target domain data are input, and UJUOT extracts the features of the source domain and target domain data through a two-stream CNN architecture with shared weights. At the same time, FUE is introduced to quantify the uncertainty of the features and enhance the robustness of the model. The UOT strategy is applied to optimize the feature matching between the source domain and the target domain, and the problem of erroneous pair transmission is reduced through adaptive learning of feature quality. In this process, UJUOT continuously optimizes the network parameters by calculating the classification loss, FUE loss, and UOT loss, and finally achieves high-performance classification tasks on the unlabeled target domain. Figure 4As shown in the figure, t-SNE is used to map the features of the source domain and the target domain to a two-dimensional plane to facilitate observation and comparison of their distributions, which helps to intuitively understand how the UJUOT model handles differences between domains and how the feature uncertainty estimation (FUE) mechanism affects feature representation. Compared with the DeepJDO model and the JUMBOT model, the migration accuracy of the UJUOT model has been significantly improved.

[0141] The above description of the embodiments is to facilitate the understanding and use of the invention by those skilled in the art. 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 work. Therefore, the present invention is not limited to the above embodiments, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A method for analyzing liver focal lesions images based on unbalanced optimal transmission, characterized in that: include: Extract features from original image samples and then embed each image into a Gaussian distribution to encode uncertainty, thereby improving the representation space and enhancing inter-category separability and intra-category compactness. Then, the unbalanced optimal transmission loss and classification loss calculations are performed to achieve accurate pairwise matching to reduce negative transfer. The discriminant class perception information is used for adaptive learning to determine the best transmission strategy and perform classification predictions to achieve imaging analysis of focal liver lesions.

2. The method for analyzing liver focal lesions images based on unbalanced optimal transmission according to claim 1, characterized in that: The specific process includes: S1: Unsupervised domain adaptation network construction: Two feature extractors are set to extract high-level features from the input liver images using a deep convolutional neural network; Construct two adaptive classification networks and dynamically adjust parameters according to the extracted features to adapt to different data characteristics; S2: Feature uncertainty estimation: The features are encoded using Gaussian distribution, where the parameters of the distribution are determined by feature statistics and are used to represent the uncertainty of the features. The parameters of the distribution are the mean and variance. Define and calculate feature uncertainty loss to force the model to maintain uncertainty in feature representation; S3: Application of unbalanced optimal transmission strategy, using Wasserstein distance to construct a cost matrix to measure the differences between different feature distributions; according to the results of the cost matrix, adaptively adjust the learning rate or hyperparameters to optimize the learning process of the model; The model parameters are adjusted through an iterative process to minimize the UOT loss and achieve the best alignment of feature distributions.

3. The method for analyzing liver focal lesions images based on unbalanced optimal transmission according to claim 2, characterized in that: The feature extractor G f The original image is input as x, and two depth representations are output, namely the mean μ and variance σ, which are used to describe the Gaussian distribution p(z|x), that is, p(z|x) = G f (x); Then extract a sample s from the Gaussian distribution to represent the feature vector of the original input, that is, s~p(z|x); The classification network G c Map the extracted samples s to the classification prediction G c (s).

4. The method for analyzing liver focal lesions images based on unbalanced optimal transmission according to claim 1, characterized in that: The image analysis method adopts an unsupervised domain adaptive network module of unbalanced optimal transmission, a feature uncertainty estimation mechanism module and a UOT strategy module; The unsupervised domain adaptive network module of the unbalanced optimal transmission includes two feature extractors and two adaptive classification networks; The feature uncertainty estimation mechanism module is used to calculate distribution representation, re-setting parameter skills, source supervision classification loss and feature uncertainty loss; The UOT strategy module is used for adaptive large-scale learning based on the cost matrix of Wasserstein distance and implements the optimization steps; The unsupervised domain adaptive network module of imbalanced optimal transmission is used to extract original image samples, and the uncertainty estimation mechanism module is used to embed the image into a Gaussian distribution model to encode uncertainty, thereby improving the representation space and enhancing the separability between categories and the compactness within categories; The UOT strategy module is used to improve the migration performance and robustness of the model to reduce negative migration.

5. The method for analyzing liver focal lesions images based on unbalanced optimal transmission according to claim 4, characterized in that: The distribution representation is specifically as follows: the uncertainty of the data is encoded as a multivariate Gaussian distribution p(z i |x i ), so that the representation of a single image is from The random embedding s sampled from i , rather than a fixed point estimate; The multivariate Gaussian distribution p(z i |x i ) is shown in the following formula: p(z i |x i )=N(z i ;μ i ,(σ i ) 2 I) Among them, μ i and (σ i ) 2 All from the i-th input image θ1 and θ2 represent network parameters; I represents an identity matrix; The resetting parameters are specifically as follows: extract a random noise ε from the standard normal distribution, that is, ε~N(0,I), and then obtain a sample s i As an equivalent sampling, it is expressed as: S i =μ i +εσ i ,ε~N(0,I)。 6. The method for analyzing liver focal lesions images based on unbalanced optimal transmission according to claim 4, characterized in that: The source supervised classification loss refers to the standard classification loss on the source domain defined as: in, represents the standard cross entropy loss, n s is the number of samples.

7. The method for analyzing liver focal lesions images based on unbalanced optimal transmission according to claim 4, characterized in that: The feature uncertainty loss is used to force the model to maintain the uncertainty level of the training samples, preventing the random representation of the model from being degraded to a deterministic point representation during the training process; The feature uncertainty loss is specifically: use the entropy of the estimated Gaussian distribution to measure each well-behaved sample x i At a given variance The uncertainty level at , is calculated as: Among them, diag(·) represents the diagonal vector of the input matrix, D is the total feature dimension, and k represents each dimension. The margin m is introduced to represent the boundary of the overall uncertainty: Where n is the batch size, and represent the entropy of the estimated Gaussian distribution of the i-th sample in the source batch and the j-th sample in the target batch, respectively.

8. The method for analyzing liver focal lesions images based on unbalanced optimal transmission according to claim 4, characterized in that: The cost matrix based on Wasserstein distance is: Among them, L ce (·,·) denotes the cross entropy loss; parameters α1 and α2 are scalars measuring the contribution of the distance term.

9. The method for analyzing liver focal lesions images based on unbalanced optimal transmission according to claim 4, characterized in that: The adaptive large-scale learning is specifically as follows: using the cross inner product based on the category center to generate the quality a associated with each feature vector i ,b j (i,j∈[1,n]); The category center It is approximately estimated as: The mass of each source sample is calculated using the following formula: Among them, k′∈{1,2,3,…C′}, C′ represents the number of categories in the target batch; if but otherwise The sample represented by equivalent sampling.

10. The method for analyzing liver focal lesions images based on unbalanced optimal transmission according to claim 4, characterized in that: The optimization step is to optimize the representation space and the label space simultaneously by reducing the distance between samples of the same category between the source domain and the target domain; The domain difference is calculated using the following formula: Among them, η is the given regularization coefficient, which represents the optimization scheme γ * smoothness; C is the number of categories in the source batch; a, b are given non-negative vectors; τ is a given regularization parameter; KL represents KL divergence; γ *T For γ * The transposed matrix of .

Images