Liver focal lesion analysis method based on two progressive feature norm alignment strategies and storage medium

Abstract

The invention relates to a liver focal lesion analysis method based on two progressive feature norm alignment strategies and a storage medium. A histogram-guided norm alignment strategy and a transportation-guided norm alignment strategy are adopted. According to the strategy, in the histogram-guided norm alignment strategy, firstly, the histogram is used for estimating feature norm distribution of a source domain and a target domain, the difference between the feature norm distribution and the feature norm distribution is calculated, the two feature norms are aligned step by step through iteration, and it is ensured that the statistical characteristics of the features are fully considered. In a norm alignment strategy of transportation guidance, feature norms of a source domain and a target domain are modeled as independent distribution, the distance between the source domain and the target domain is accurately calculated by using an optimal transmission strategy, and the difference of the two distributions is gradually minimized by optimizing a coupling matrix so as to realize effective alignment. Compared with the prior art, the method has the advantages that the characteristic norm difference between the source domain and the target domain is effectively reduced, and the migration performance, robustness and adaptability of the model are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of medical imaging engineering, and particularly to a method for analyzing liver focal lesions and a storage medium based on two progressive feature norm alignment strategies. Background Art

[0002] In the fields of computer vision and machine learning, deep learning has made remarkable progress with the help of sufficient labeled data. Traditional supervised learning theories usually assume that the training set and the test set follow the same distribution. However, this ideal assumption does not hold in practical applications due to various reasons, such as sensors of different devices, different perspectives and positions when acquiring images. Due to the distribution shift problem, models trained based on existing labeled source domain data often experience a significant performance degradation when directly applied to other unlabeled domains, thus limiting their application and generalization in many practical scenarios.

[0003] Unsupervised Domain Adaptation (UDA), as a solution to this problem, aims to transfer shared knowledge across domains, seamlessly migrating knowledge from a domain rich in labels to the target domain, even if the latter lacks direct label information. UDA methods can be roughly divided into two categories: difference-based methods and adversarial learning-based methods. In recent years, adversarial domain adaptation methods have received considerable attention due to their ability to align complex distributions.

[0004] However, many previous adversarial domain adaptation methods inevitably damage the discriminative information contained in transferable features, limiting the potential of their adversarial learning. The reason is that, during the adversarial alignment process, the magnitude of the feature norm is often the main factor for the domain discriminator to distinguish between the source domain and the target domain. To train the domain discriminator, the feature extractor will pay more attention to generating features with similar norms rather than extracting semantically rich shared knowledge, affecting the ability of the feature extractor to learn more semantically rich transferable features. Some researchers have tried to solve the norm difference problem, and existing work usually adopts two strategies, but these two strategies also have their respective drawbacks:

[0005] (1) Using L2 normalization to map features to the spherical space: When the feature norm contains rich discriminative information, it is very difficult to find the optimal spherical radius R. When L2 normalization is too strict, it will inevitably damage the discriminative information contained in the feature norm;

[0006] (2) Matching the expected value of the feature norm: The strategy based on expectation is very sensitive to the restricted scalar R. When R is very small, the features learned by the model may become less informative. Directly matching the average feature norms of the two domains will also inevitably damage the discriminative information contained in the feature norm.

[0007] In summary, although existing adversarial domain adaptation methods have made some progress in feature distribution alignment, they are insufficient in maintaining discriminative information of features. The main problem lies in the difference in feature norms between the source domain and the target domain, which limits the performance and generalization ability of the model in the target domain. Summary of the Invention

[0008] The purpose of the present invention is to overcome the above-mentioned defects existing in the prior art and propose a liver focal lesion analysis method and storage medium based on two progressive feature norm alignment strategies, namely Histogram-guided Norms Alignment (HNA) and Transport-guided Norms Alignment (TNA). These two strategies model feature norms from the perspective of distribution, which not only helps to reduce norm differences but also fully utilizes the discriminative information contained in the norms. Gradually aligning the feature norm distributions of the two domains can effectively promote the model's learning of semantically rich shared features and significantly improve the model's transfer performance.

[0009] The purpose of the present invention can be achieved through the following technical solutions:

[0010] The first aspect of the present invention provides a liver focal lesion analysis method based on two progressive feature norm alignment strategies, mainly using the histogram-guided norm alignment strategy and the transport-guided norm alignment strategy, and then performing the analysis of liver focal lesions. The main steps include:

[0011] Obtain source domain features and target domain features extracted from liver medical image data;

[0012] In the histogram-guided norm alignment strategy, first use histograms to estimate the feature norm distributions of the source domain and the target domain respectively and calculate their differences, and gradually align these two feature norms through iteration to ensure full consideration of the statistical characteristics of the features;

[0013] In the transport-guided norm alignment strategy, model the feature norms of the source domain and the target domain as independent distributions, accurately calculate the distance between them using the optimal transport strategy, and gradually minimize the difference between the two distributions by optimizing the coupling matrix to achieve effective alignment.

[0014] Perform the analysis of liver focal lesions based on the aligned features.

[0015] Furthermore, the specific process of the histogram-guided norm alignment strategy includes:

[0016] S1-1: Estimation of the Feature Norm Distributions in the Source Domain and the Target Domain: Calculate the normalized feature norm sets of the source domain and the target domain;

[0017] Calculate the weights using a Gaussian kernel function;

[0018] Estimate the feature norms of the source domain and the target domain using a simple histogram with uniformly spaced bin intervals;

[0019] S1-2: Calculation of the Difference in Feature Norms between the Source Domain and the Target Domain: Calculate the difference between the feature norm distributions of the source domain and the target domain using the Kullback-Leibler divergence;

[0020] Minimize the difference between the two distributions and gradually align the feature norms of the source domain and the target domain.

[0021] Furthermore, in the histogram-guided norm alignment strategy, in the step of estimating the feature norm distributions of the source domain and the target domain, a simple histogram and are used to estimate the feature norm distributions of the source domain and the target domain and for calculating the difference in feature norms between the source domain and the target domain

[0022] Furthermore, in the histogram-guided norm alignment strategy, the step of estimating the feature norm distributions of the source domain and the target domain is specifically as follows: Calculate the normalized feature norm sets S n and T n of the source domain and the target domain respectively; Calculate the weights using a Gaussian kernel function to calculate the Q-dimensional simple histogram and for estimating the feature norms of the source domain and the target domain;

[0023] The normalized feature norm sets S n and T n are specifically as follows:

[0024]

[0025] where represents the maximum element in the source domain S and the target domain T, that is, so as to limit the normalized feature norm of each sample to [0,1], which is beneficial to the estimation of the probability distribution;

[0026] The nodes k1 = 0, k2,..., k and of the Q-dimensional simple histogram Q are uniformly distributed between [0,1] with a step size of

[0027] The weight Specifically:

[0028]

[0029] where k q represents the q-th node of the histogram, and η represents the bandwidth of the Gaussian kernel function;

[0030] The histogram The value h s is:

[0031]

[0032] where i is the number of all source domain samples in the batch.

[0033] Furthermore, in the histogram-guided norm alignment strategy, the feature norm difference between the source domain and the target domain is defined as:

[0034]

[0035] where, represents the Kullback-Leibler divergence.

[0036] The transport-guided norm alignment strategy, the specific process includes:

[0037] S2-1: Model the source domain and target domain feature norms as distributions and

[0038] S2-2: Use the optimal transport strategy to accurately calculate the difference between the source domain and target domain feature norm distributions: Based on the Kantorovich relaxation operation, combined with the quicksort algorithm, calculate the coupling matrix between the two distributions;

[0039] In the cost matrix, calculate the paired distance of the feature norms;

[0040] Accurately calculate the difference between the source domain and target domain feature norm distributions;

[0041] S2-3: Minimize the difference between the two distributions and gradually align the feature norms of the source domain and the target domain.

[0042] In the transport-guided norm alignment strategy, the optimal transport strategy performs the Kantorovich relaxation operation to calculate the coupling matrix π and between the source domain and target domain feature norm models * , which is used to accurately calculate the feature norm difference between the source domain and the target domain

[0043] Further, in the norm alignment strategy of the transport guidance, the calculation process based on the optimal transport strategy specifically includes: calculating the modeling of the feature norm distributions of the source domain and the target domain and the coupling matrix π * between them; calculating the pairwise distance C ij of the feature norms;

[0044] The coupling matrix π * is specifically:

[0045]

[0046] where represents and the set of joint probability distributions between them;

[0047] The cost function represents the cost of moving the probability mass from to and its discrete form can be expressed as follows:

[0048]

[0049] where is the optimal coupling matrix between the source domain and the target domain feature norm distributions and ;

[0050] is and a set of feasible coupling matrices between them, that is:

[0051]

[0052] where represents an N-dimensional column vector all of whose elements are 1;

[0053] Each term C in the cost matrix ij represents the pairwise distance of the feature norms, that is

[0054]

[0055] Further, in the norm alignment strategy of the transport guidance, the difference between the accurately calculated source domain and target domain feature norm distributions and is:

[0056]

[0057] Furthermore, the step of analyzing liver focal lesions based on the aligned features includes:

[0058] The aligned features are input into the pre-trained liver lesion classification model to determine the type of focal liver lesions;

[0059] Using image segmentation algorithms, the lesion area in the liver image is segmented based on alignment features, and its geometric features are analyzed to assist in determining the nature and development stage of the lesion;

[0060] Correlate alignment features with patient clinical information to comprehensively assess the impact of lesions on patient health and their development trends;

[0061] Generative adversarial network technology is used to generate virtual liver lesion samples based on alignment features and compare them with real samples to explore the potential characteristics and patterns of lesions.

[0062] A second aspect of the present invention provides a storage medium comprising computer executable instructions, which, when executed by a computer processor, is used to perform a liver focal lesion analysis method based on two progressive feature norm alignment strategies as described above.

[0063] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0064] 1. This paper introduces the histogram-guided norm alignment (HNA) and transport-guided norm alignment (TNA) strategies, emphasizing the importance of feature norm in adversarial domain adaptation and effectively reducing the norm difference between the source domain and the target domain. However, existing adversarial domain adaptation methods often focus on feature alignment and ignore the impact of feature norm.

[0065] 2. The present invention can effectively preserve discriminant information. Many traditional methods may damage the discriminant information in the transferable features when performing feature alignment. The present invention can effectively preserve the discriminant information in the features by modeling the feature norm from the perspective of distribution.

[0066] 3. The present invention can improve the migration ability of the model. The test results on multiple data sets show that after adopting the HNA and TNA strategies, the migration performance of the model on multiple standard data sets is significantly improved, demonstrating its superiority in unsupervised domain adaptation tasks.

[0067] 4. The present invention has excellent flexibility and robustness. HNA and TNA strategies can be flexibly integrated into different adversarial domain adaptation frameworks, such as DANN, CDAN+E, etc. This flexibility enables the present invention to be widely applied to various domain adaptation tasks, enhancing its practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 The structural block diagram of the liver focal lesion analysis method based on two progressive feature norm alignment strategies, namely the histogram-guided norm alignment strategy (HNA) and the transport-guided norm alignment strategy (TNA), in Example 1;

[0069] Figure 2 It is a schematic diagram of the global network architecture of the liver focal lesion analysis method based on two progressive feature norm alignment strategies in the DANN framework in Example 2;

[0070] Figure 3 It is a t-SNE embedding visualization result graph of the comparison between the application of the liver focal lesion analysis method based on two progressive feature norm alignment strategies in the CDAN+E framework and other methods in Example 2. Detailed implementation manners

[0071] Overall, the present invention provides a liver focal lesion analysis method based on two progressive feature norm alignment strategies, the histogram-guided norm alignment strategy and the transport-guided norm alignment strategy, as Figure 1 shown, including:

[0072] Obtain source domain features and target domain features extracted from liver medical image data;

[0073] In the histogram-guided norm alignment strategy, first use histograms to estimate the source domain feature norm distribution and the target domain feature norm distribution respectively, and then minimize the difference between the two distributions to gradually align the source domain and target domain feature norms;

[0074] The specific process of the histogram-guided norm alignment strategy includes:

[0075] S1-1: Estimation of source domain and target domain feature norm distributions: Calculate the normalized source domain and target domain feature norm sets;

[0076] Use a Gaussian kernel function to calculate weights;

[0077] Use a simple histogram with evenly spaced class intervals to estimate the source domain and target domain feature norms;

[0078] The source domain and target domain feature norm distribution estimation step uses a simple histogram and to estimate the source domain and target domain feature norm distributions and for calculating the feature norm difference between the source domain and the target domain

[0079] S1-2: Calculation of the feature norm difference between the source domain and the target domain: Use the Kullback-Leibler divergence to calculate the difference between the feature norm distributions of the source domain and the target domain;

[0080] Minimize the difference between the two distributions and gradually align the feature norms of the source domain and the target domain.

[0081] In the transportation-guided norm alignment strategy, first model the source domain feature norm and the target domain feature norm as two distributions respectively, then use the optimal transportation strategy to accurately calculate the difference between them, and then minimize the difference between the two distributions to gradually align the feature norms of the source domain and the target domain.

[0082] The transportation-guided norm alignment strategy, the specific process includes:

[0083] S2-1: Model the source domain and target domain feature norms as distributions respectively and

[0084] S2-2: Use the optimal transportation strategy to accurately calculate the difference between the source domain and target domain feature norm distributions: Based on the Kantorovich relaxation operation, combined with the quicksort algorithm, calculate the coupling matrix between the two distributions;

[0085] The optimal transportation strategy performs the Kantorovich relaxation operation to calculate the coupling matrix π and between the modeled source domain and target domain feature norms * , which is used to accurately calculate the feature norm difference between the source domain and the target domain

[0086] In the cost matrix, calculate the paired distance of the feature norms;

[0087] Accurately calculate the difference between the source domain and target domain feature norm distributions;

[0088] S2-3: Minimize the difference between the two distributions and gradually align the feature norms of the source domain and the target domain.

[0089] Based on the aligned features, analyze the focal liver lesions.

[0090] The following uses examples to illustrate in detail the specific implementation manners of the present invention. These examples are implemented on the premise of the solution described in the present invention, and give the detailed implementation manners and specific operation processes, but the protection scope of the present invention is not limited to the following examples.

[0091] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. In the present technical solution, features such as the names of structures / modules, control modes, algorithms, process flows, or composition ratios that are not clearly described are regarded as common technical features disclosed in the prior art.

[0092] Embodiment 1

[0093] This embodiment provides a method for analyzing liver focal lesions based on two progressive feature norm alignment strategies, namely, histogram-guided norm alignment strategy (HNA) and transport-guided norm alignment strategy (TNA).

[0094] The method of this embodiment focuses on processing source domain and target domain features extracted from liver medical images (such as CT, MRI images, etc., which may contain relevant features of liver focal lesions) to achieve accurate analysis of liver focal lesions. The specific strategies are as follows:

[0095] In the histogram-guided norm alignment strategy, first, the feature norm distributions of the source domain and the target domain are estimated using histograms respectively, and their differences are calculated. The two feature norms are gradually aligned through iteration to ensure full consideration of the statistical characteristics of the features. In the transport-guided norm alignment strategy, the feature norms of the source domain and the target domain are modeled as independent distributions, the distance between the two is accurately calculated using the optimal transport strategy, and the difference between the two distributions is gradually minimized by optimizing the coupling matrix to achieve effective alignment.

[0096] This embodiment uses DANN (Domain-Adversarial Neural Network) as the basic architecture to apply HNA and TNA, which consists of the following three main components: feature extractor domain discriminator classifier When the input image is x, the feature extractor extracts the feature representation as The classifier predicts the probability as where d represents the feature dimension and K represents the number of classifications. The training objective of the DANN framework can be formally expressed as:

[0097]

[0098] where is the supervised classification loss of the labeled source domain, is the standard cross-entropy loss, is the adversarial training loss. N is the batch size, and λ d is the trade-off parameter.

[0099] In the histogram-guided norm alignment strategy, the steps for estimating the feature norm distributions of the source domain and the target domain are specifically as follows: Calculate the normalized feature norm sets S n and T n of the source domain and the target domain respectively; Calculate the weights through the Gaussian kernel function; Calculate the Q-dimensional simple histograms and for estimating the feature norms of the source domain and the target domain.

[0100] Since the model optimization is carried out on small batches, the source domain feature norm set and the target domain feature norm set can be calculated in each iteration, so that two normalized feature norm sets can be obtained.

[0101] The normalized feature norm sets S n and T n are specifically:

[0102]

[0103] where represents the maximum element in the source domain S and the target domain T, that is Thus, the normalized feature norm of each sample is restricted to [0,1], which is beneficial to the estimation of the probability distribution;

[0104] The nodes k1 = 0, k2,..., k and of the Q-dimensional simple histogram are evenly distributed between [0,1] with a step size of Q

[0105] The weight is specifically:

[0106]

[0107] where k q represents the q-th node of the histogram, and η represents the bandwidth of the Gaussian kernel function;

[0108] The value h of the histogram s is:

[0109]

[0110] where i is the number of all source domain samples in the batch.

[0111] In the histogram-guided norm alignment strategy, the feature norm difference between the source domain and the target domain is defined as:

[0112] ​

[0113] Among them, represents the Kullback-Leibler divergence.

[0114] In the transportation-guided norm alignment strategy, the calculation process based on the optimal transportation strategy specifically includes: calculating the source domain and target domain feature norm distribution modeling and the coupling matrix π * between; calculating the pairwise distance C of the feature norms ij ;

[0115] The coupling matrix π * is specifically:

[0116]

[0117] where represents and the set of joint probability distributions between;

[0118] The cost function represents the cost of moving the probability value from to and its discrete form can be expressed as follows:

[0119]

[0120] where is the optimal coupling matrix between the source domain and target domain feature norm distributions and ;

[0121] is and a set of feasible coupling matrices between, that is:

[0122]

[0123] where represents an N-dimensional column vector with all elements equal to 1;

[0124] The cost matrix each item C in ij represents the pairwise distance of the feature norms, that is

[0125]

[0126] In the transportation-guided norm alignment strategy, the accurately calculated source domain and target domain feature norm distributions and The differences between them are as follows:

[0127]

[0128] When applying HNA in DANN, the overall training objective of the framework can be expressed as: When applying TNA in DANN, the overall training objective of the framework can be expressed as: Where the hyperparameters λ d , β h , β t are used to balance the contributions of their respective corresponding terms.

[0129] Based on the aligned features obtained after processing by the histogram-guided norm alignment strategy and the transport-guided norm alignment strategy, a series of precise analyses for liver focal lesions can be carried out.

[0130] First, input these aligned features into a pre-trained liver lesion classification model. This model is trained with a large amount of medical image data containing liver focal lesion information and has high accuracy and reliability. The model will make a preliminary type judgment on the liver focal lesion based on the input features, such as distinguishing different types of lesions like liver cysts, hepatic hemangiomas, liver cancers, etc.

[0131] Next, use an image segmentation algorithm to accurately segment the lesion area in the liver image based on the aligned features. By analyzing geometric features such as the size, shape, and boundary clarity of the segmented lesion area, it further assists in judging the nature and development stage of the lesion. For example, a lesion with a regular circular shape and clear boundary may be benign, while a lesion with an irregular shape and blurred boundary has a relatively higher possibility of being malignant.

[0132] Meanwhile, in combination with the clinical information database, perform a correlation analysis between the aligned features and the patient's clinical information such as age, gender, medical history, blood test indicators, etc. Through multi-dimensional data fusion, comprehensively evaluate the degree of impact of liver focal lesions on the patient's health and predict the development trend of the lesion, such as whether there is a risk of metastasis and whether it will grow rapidly.

[0133] In addition, the generative adversarial network (GAN) technology in deep learning can be used to generate virtual liver lesion samples based on the aligned features. By comparing and analyzing with real samples, further explore the potential features and variation rules of liver focal lesions, providing a more comprehensive and in-depth reference basis for clinical diagnosis and treatment plan formulation.

[0134] Example 2

[0135] The present embodiment provides a storage medium containing computer executable instructions, and the computer executable instructions stored in the storage medium, when executed by a computer processor, can be used to execute the above-mentioned liver focal lesion analysis method based on two progressive feature norm alignment strategies. Specifically, when the computer processor reads and executes the instructions in the storage medium, the operation will be carried out according to the following process. First, for the norm alignment strategy part guided by the histogram, the processor will use the histogram based on the instructions to estimate the feature norm distribution of the source domain and the target domain (both data related to liver focal lesions) respectively, and calculate the difference between them, and then gradually align the two feature norms in an iterative manner, so as to fully consider the statistical characteristics related to the characteristics of liver focal lesions. Next, in terms of the transport-guided norm alignment strategy, the processor will model the feature norms of the source domain and the target domain as independent distributions, use the optimal transmission strategy to accurately calculate the distance between the two, and gradually minimize the difference between the two distributions by optimizing the coupling matrix to achieve effective alignment of the feature norms. Finally, based on the aligned features, the processor will perform a series of analysis operations on focal liver lesions, such as inputting the features into a pre-trained liver lesion classification model to determine the type of lesion, using an image segmentation algorithm to segment the lesion area and analyze its geometric features, correlating clinical information to evaluate the impact and trend of the lesion, and using generative adversarial network technology to explore the potential patterns of the lesions, etc., ultimately providing medical personnel with comprehensive analysis results to assist in making diagnostic conclusions and treatment recommendations.

[0136] Application Example 1

[0137] In a specific case of applying this application example system, the HNA and TNA strategies are applied to the DANN and CDAN+E frameworks, such as Figure 2 As shown. In the histogram-guided norm alignment strategy, the histogram is first used to estimate the source domain feature norm distribution and the target domain feature norm distribution respectively, and then the difference between the two distributions is calculated to gradually align the feature norms of the source domain and the target domain; in the transport-guided norm alignment strategy, the source domain feature norm and the target domain feature norm are first modeled as two distributions respectively, and then the optimal transport strategy is used to accurately calculate the difference between the two, and then the difference between the two distributions is minimized to gradually align the feature norms of the source domain and the target domain. In this process, with the help of the HNA and TNA strategies, the model can more effectively reduce the norm difference, allowing the feature extractor to focus more on learning domain-invariant features. As shown Figure 3As shown, t-SNE is used to map the features of the source domain and the target domain to a two-dimensional plane to facilitate the observation and comparison of their distributions, which helps to intuitively understand the degree of class alignment when the CDAN+E model applies the expectation-based strategy (Exp), the sphere-based strategy (Sphere), and the HNA and TNA strategies. For CDAN+E+Exp and CDAN+E+Sphere, the class distributions are not well aligned, and there are a large number of discrete points on the decision boundary, resulting in poor generalization results of the model. In contrast, the decision boundaries generated by the HNA and TNA strategies proposed in the present invention are clearer and the class alignment is more accurate, indicating that the present invention is very strong in enhancing the ability of the adversarial learning model to extract discriminative domain-invariant features.

[0138] The above description of the embodiments is to enable those of ordinary skill in the art to understand and use the invention. Obviously, those who are familiar with the technology in this field 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 according to the disclosure of the present invention should be within the protection scope of the present invention.

Claims

1. A method for analyzing liver focal lesions based on two progressive feature norm alignment strategies, adopting a histogram-guided norm alignment strategy and a transport-guided norm alignment strategy, and then performing the analysis of liver focal lesions, characterized in that, The method for analyzing focal liver lesions specifically includes the following steps: Obtain the source domain features and target domain features extracted from liver medical image data; In the histogram-guided norm alignment strategy, first use histograms to estimate the feature norm distributions of the source domain and the target domain respectively and calculate their differences, and gradually align these two feature norms through iteration to ensure full consideration of the statistical characteristics of the features; In the transport-guided norm alignment strategy, model the feature norms of the source domain and the target domain as independent distributions, accurately calculate the distance between the two using the optimal transport strategy, and gradually minimize the difference between the two distributions by optimizing the coupling matrix to achieve effective alignment; Analyze focal liver lesions based on the aligned features.

2. The method for analyzing liver focal lesions based on two progressive feature norm alignment strategies according to claim 1, wherein In the histogram-guided norm alignment strategy, the specific process includes: S1-1: Estimation of source domain and target domain feature norm distributions: Calculate the normalized source domain and target domain feature norm sets; Use the Gaussian kernel function to calculate weights; Use a simple histogram with evenly spaced bin widths to estimate the feature norms of the source domain and the target domain; S1-2: Calculation of source domain and target domain feature norm differences: Use the Kullback-Leibler divergence to calculate the difference between the source domain and target domain feature norm distributions; Minimize the difference between the two distributions and gradually align the source domain and target domain feature norms.

3. The method for analyzing liver focal lesions based on two progressive feature norm alignment strategies according to claim 2, characterized in that In the step of estimating the characteristic norm distributions of the source domain and the target domain, a simple histogram is used and to estimate the characteristic norm distributions of the source domain and the target domain and for calculating the characteristic norm difference between the source domain and the target domain 4. The method for analyzing liver focal lesions based on two progressive feature norm alignment strategies according to claim 3, wherein The steps of the source domain and target domain feature norm distribution estimation specifically include: Calculate the normalized feature norm sets S n and T n ; Calculate weights through Gaussian kernel function Compute the Q-dimensional simple histogram and be used to estimate the feature norms of the source domain and the target domain; The normalized feature norm set S n and T n Specifically: where represents the maximum element in the source domain S and the target domain T, is the feature norm of the source domain sample , is the feature norm of the target domain sample , that is The Q-dimensional simple histogram and the nodes k1 = 0, k2, ……, k Q = 1 are uniformly distributed between [0, 1] with a step size of The said weight Specifically: where k q represents the q-th node of the histogram, and η represents the bandwidth of the Gaussian kernel function; The histogram with a value h s is as follows: where i is the number of all source domain samples in the batch.

5. The method for analyzing liver focal lesions based on two progressive feature norm alignment strategies according to claim 3, characterized in that Feature norm difference between the source domain and the target domain It is defined as: Among them, represents the Kullback-Leibler divergence, represents the distribution probability of the input instance x in the source domain, represents the distribution probability of the input instance x in the target domain.

6. The method for analyzing liver focal lesions based on two progressive feature norm alignment strategies according to claim 1, characterized in that In the transport-guided norm alignment strategy, the specific process includes: S2-1: Model the source domain and target domain feature norms as the source domain feature norm distribution and the target domain feature norm distribution S2-2: Accurately calculate the difference between the source domain feature norm distribution and the target domain feature norm distribution specifically as follows: Based on the Kantorovich relaxation operation, combined with the quicksort algorithm, calculate the coupling matrix between the two distributions; In the cost matrix, calculate the paired distances of the feature norms; Accurately calculate the difference between the source domain and target domain feature norm distributions; S2-3: Minimize the difference between the two distributions and gradually align the source domain and target domain feature norms.

7. The method for analyzing liver focal lesions based on two progressive feature norm alignment strategies according to claim 6, wherein the optimal transport strategy is based on the Kantorovich relaxation operation to calculate the feature norm modeling between the source domain and the target domain and the coupling matrix π * therebetween, which is used to accurately calculate the feature norm difference between the source domain and the target domain 8. The method for analyzing focal liver lesions based on two progressive feature norm alignment strategies according to claim 6, wherein The specific calculation process based on the optimal transport strategy specifically includes: Modeling the norm distribution of source domain and target domain features and the coupling matrix π * ; Calculate the pairwise distances C of the feature norms ij ; The coupling matrix π * Specifically: wherein represents and a set of joint probability distributions between; Cost function represents the cost of moving the probability mass from to and its discrete form can be expressed as follows: wherein is the optimal coupling matrix between the source domain and the target domain feature norm distributions and ; is and a set of feasible coupling matrices between, namely: wherein represents an N-dimensional column vector with all elements being 1; Cost matrix Each term C ij represents the pairwise distance of the feature norms, that is 9. The method for analyzing focal liver lesions based on two progressive feature norm alignment strategies according to claim 6, characterized in that Feature norm distributions in the source and target domains with precise calculations and the difference between is as follows:

10. A storage medium containing computer-executable instructions, characterized in that, When the storage medium of the computer-executable instructions is executed by a computer processor, it is used to execute a method for analyzing focal liver lesions based on two progressive feature norm alignment strategies as described in any one of claims 1 to 9.

Images