Self-adaptive liver focal lesion image analysis method based on trust perception condition confrontation domain, medium and program product

Abstract

The invention relates to the technical field of medical image engineering, in particular to a self-adaptive liver focal lesion image analysis method based on the trust perception condition confrontation domain, a medium and a program product, and the analysis method comprises the steps: extracting original image sample features and classification prediction probability, quantizing the mobility of data through GUM, and obtaining a target image; and then calculating probability weighted adversarial training loss, feature norm distribution difference, confidence-guided comparison alignment loss and mixed information-guided entropy loss, realizing feature norm distribution alignment of a source domain and a target domain, determining an optimal transmission strategy, and performing classification prediction so as to realize liver focal lesion image analysis. Compared with the prior art, the problem that cross-domain class distribution cannot be accurately aligned by an existing unsupervised field adaptive method can be relieved, and the migration performance of the model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical imaging engineering, and in particular, to a method, medium, and program product for analyzing liver focal lesion images based on trust-aware conditional adversarial domain adaptation. Background Art

[0002] With the continuous development of medical imaging technology, the image analysis of liver focal lesions plays an increasingly important role in clinical diagnosis. However, due to factors such as equipment differences, inconsistent imaging parameters, and patient individual differences in the image data collected by different medical institutions, there are obvious distribution differences between the source domain and the target domain, which poses a great challenge to the accurate diagnosis of liver focal lesions. To solve this problem, domain adaptation technology, as an effective transfer learning method, has been widely applied to the field of medical image analysis.

[0003] Currently, the conditional adversarial domain adaptation method has become the mainstream technical route for solving cross-domain medical image analysis problems. CN113344044A discloses a cross-species medical image classification method based on domain adaptation, which realizes the transfer learning of cross-species medical image data by constructing a constraint based on the distribution consistency between the source domain and the target domain and a domain invariance constraint based on conditional adversarial learning. CN119477815A proposes a method for analyzing liver focal lesion images based on unbalanced optimal transport, which extracts the features of the original image samples, embeds each image into a Gaussian distribution to encode the uncertainty, and realizes accurate pairwise matching through the calculation of unbalanced optimal transport loss and classification loss, effectively reducing the negative transfer phenomenon.

[0004] In terms of sample transferability evaluation, CN113159126A introduces an industrial Internet of Things device fault diagnosis method based on general domain adaptation. In the sample weight learning stage, according to the designed transferability metric, corresponding weights are assigned to the input monitoring data samples, and the weight represents the possibility that the sample belongs to the common label set of the source domain and the target domain. CN114528913A discloses a model transfer method based on trust and consistency, which performs model adaptive learning through a dual classification network and uses a trust and consistency mechanism for training optimization to solve the domain adaptation problem in the case of missing source domain data.

[0005] However, the existing technologies still have the following several key problems:

[0006] First, existing conditional adversarial domain adaptation strategies usually assign the same weight to all samples, ignoring the differences in the transfer difficulty among samples. In particular, for samples with high transfer difficulty and high prediction uncertainty, there is a lack of effective processing mechanisms, which seriously affects the domain alignment effect and the overall performance of the model. Although CN113159126A proposes the concept of sample weight learning, its weight assignment mechanism fails to fully consider the characteristics of the prediction probability distribution of samples, resulting in insufficient adaptability in complex medical imaging scenarios.

[0007] Second, existing methods have limitations in aligning the feature norm distributions. Aligning only the average feature norm cannot completely eliminate the domain differences. Especially in the high-dimensional feature space such as liver focal lesions, the complexity of the feature norm distribution may lead to a decrease in the effectiveness of features, affecting the adaptation performance of the model. Although CN119477815A introduces an uncertainty estimation mechanism, it fails to perform systematic alignment from the overall perspective of the feature norm distribution.

[0008] Third, cross-domain class distribution alignment remains a difficult problem. Many unsupervised domain adaptation methods can only approximately align the marginal distributions of the source domain and the target domain and cannot accurately align the cross-domain class distributions, resulting in a high prediction error rate when dealing with samples in the target domain that are near the class cluster boundary or far from the class center. The domain invariance constraint based on conditional adversarial learning proposed by CN113344044A still has room for improvement in class distribution alignment.

[0009] Finally, the method in the prior art of minimizing the conditional entropy to make the target features far from the decision boundary has limitations, which may lead to a deviation in the model optimization direction and even a decrease in the model performance in complex transfer tasks. Especially in highly complex medical imaging analysis tasks such as liver focal lesions, a single entropy minimization strategy is difficult to effectively guide the features to migrate towards the correct class center.

[0010] In summary, there is an urgent need for a liver focal lesion imaging analysis method that can comprehensively consider issues such as sample transferability differences, feature norm distribution complexity, cross-domain class distribution alignment, and the limitations of conditional entropy minimization, so as to improve the accuracy and robustness of cross-domain medical imaging analysis. Summary of the Invention

[0011] To solve problems such as differences in sample transferability, complexity of feature norm distribution, cross-domain class distribution alignment challenges, and limitations of conditional entropy minimization, and to achieve the technical effects of quantifying the transferability of training samples, improving the model transfer performance, promoting class-level distribution alignment, and making more accurate predictions for target samples, a liver focal lesion image analysis method, medium, and program product based on Trust-aware Conditional Adversarial Domain Adaptation (TCADA) are proposed.

[0012] Overall, the present invention can effectively promote conditional adversarial adaptation, reduce the cross-domain feature norm difference, promote the precise alignment of class distributions, and effectively alleviate label drift and create a clearer decision boundary through probability-weighted adversarial training loss, FNDA strategy, CCA strategy, and MIE strategy.

[0013] The object of the present invention can be achieved by the following technical solutions:

[0014] In the first aspect of the present invention, a liver focal lesion image analysis method based on trust-aware conditional adversarial domain adaptation is provided, including the following steps:

[0015] Extract the features of the original liver image samples and the classification prediction probabilities, further quantify the transferability of the probability data through a Gaussian-uniform mixture model, process the outliers therein, and further calculate the probability-weighted adversarial training loss, feature norm distribution difference, confidence-guided contrast alignment loss, and mixed information-guided entropy loss to achieve the alignment of the feature norm distributions between the source domain and the target domain, determine the optimal transfer strategy, and perform liver image classification prediction to complete the liver focal lesion image analysis.

[0016] Furthermore, the specific process of achieving the alignment of the feature norm distributions between the source domain and the target domain includes the following steps:

[0017] S1: Construct an unsupervised domain adaptation framework based on a conditional domain adversarial network, capture the multimodal structure of the feature distribution through joint variable modeling, and establish an initial alignment channel between the source domain and the target domain;

[0018] S2: Based on the joint variable modeling adversarial network framework, implement probability-weighted adversarial training, including:

[0019] Adopt a transferability quantification module to dynamically evaluate the sample transfer difficulty and generate a sample transferability coefficient.

[0020] Construct a probability-weighted adversarial loss function, inject the transferability coefficient as a sample weight into the discriminator training process to form an adversarial learning mechanism driven by easily transferable samples;

[0021] S3: Synchronously perform cross-domain feature norm distribution alignment during adversarial training, including:

[0022] The statistical difference between the feature norm distribution of the source domain and the target domain is calculated through the optimal transfer algorithm with entropy regularization.

[0023] Adaptively adjusting the feature mapping parameters of the adversarial network based on the statistical difference to achieve dynamic distribution matching of cross-domain feature norms;

[0024] S4: Based on the feature distribution after S3 alignment, perform confidence-guided comparative alignment, including:

[0025] Generate confidence weights based on the pseudo-label confidence of the target domain samples,

[0026] Based on the confidence weight, the target domain features are gradually pulled to the neighborhood space of the corresponding source class center through the moving average method through the contrast loss function;

[0027] S5: Implementing hybrid information-guided entropy constraints during feature space optimization, including:

[0028] Based on the target domain feature distribution after contrast alignment adjustment, the mixed entropy value of the distance entropy from the target feature to the nearest source class center and the category prediction entropy is calculated;

[0029] The entropy minimization criterion is used to drive the target features to migrate to the center of the high-confidence source class, and establish a feature distribution far away from the decision boundary.

[0030] Furthermore, in S1, the joint variable is specifically: the original image is taken as x input, the feature extractor The extracted deep features f and the classification probability g predicted by the image classifier C are combined into a joint variable h, expressed as:

[0031] h=(f,g)

[0032] Among them, the depth feature Classification prediction probability

[0033] Furthermore, in S2, the specific process of the GUM (quantified data transferability) is as follows: modeling each training sample x i The classifier prediction entropy H(g i ) distribution p(H(g i )|x i ), expressed as:

[0034]

[0035] in:

[0036]

[0037] is a uniform distribution on [0, δ], π is the prior probability, and Σ is the variance of the Gaussian distribution .

[0038] Furthermore, the specific process of the probability-weighted adversarial training loss is as follows: Based on the posterior probability, introduce a weight strategy η(x i ) = 1 + r φ (x i ) to re-weight each training sample, so that the model gives priority to the data that is easy to transfer during training. η(x i ) is the weight strategy for each training sample x i , and r φ (x i ) is the posterior probability of sample x i ;

[0039] Probability-weighted adversarial training loss is defined as follows:

[0040]

[0041] where T(·) represents a multilinear mapping, represents the j-th target domain sample, represents the i-th source domain sample, represents the feature representation of the i-th source domain sample, D(·) represents the discriminator, represents the feature representation of the j-th target domain sample.

[0042] Furthermore, the specific entropy-regularized optimal transport algorithm is as follows: Introduce a strict convex regularization term E(γ) to accelerate the solution of the optimal transport problem, expressed as:

[0043]

[0044] where is the optimal transport matrix between the source feature norm distribution ψ s and the target feature norm distribution ψ T , Π(ψ s , ψ T ) is the set of feasible transport matrices between ψ S and ψ T , expressed as: is an N-dimensional column vector with all elements being 1, and each term in represents the pairwise distance between the source sample and the target sample feature norms in the loss matrix : λ otγ is a given regularization parameter used to control the smoothness of the transportation plan, and E(γ) is the entropy regularization term, expressed as:

[0045]

[0046] where γ ij represents the transportation probability from source sample i to target sample j,

[0047] The statistical difference degree of the feature norm distribution is defined as:

[0048]

[0049] where is the pairwise distance between source sample i and target sample j.

[0050] Furthermore, the confidence-guided contrast alignment means assigning different priorities according to the confidence level of the target sample pseudo-label and is defined as:

[0051]

[0052] where τ is a temperature coefficient, is the weighting factor, and the distance metric dis(·,·) between features uses a multi-kernel strategy to fully capture the distance between representations using different kernels, defined as follows:

[0053]

[0054] where the feature kernel related to the feature mapping is a linear combination of m positive definite kernels and is the feature mapping, and the kernel set is expressed as:

[0055]

[0056] where the constraint on the parameters {α u} is used to ensure that the resulting linear combination is characteristic;

[0057] The moving average method gradually updates the class center through iteration, i.e.:

[0058]

[0059] where represents the source class center, is the label of the i-th sample in the source domain, and λ s is the learning rate used to update the class center, is the class center of the k-th category in the t-th iteration, where k ∈ {1, 2, 3, …, K}. If the condition is satisfied, then δ(condition) = 1; otherwise, δ(condition) = 0.

[0060] The confidence weight reflects the feature and the class center of the k-th category in the source domain similarity. The confidence weight calculation formula is:

[0061]

[0062] where sim(·,·) represents the cosine similarity.

[0063] Furthermore, the mixed information-guided entropy loss (MIE loss) combines the information of the shared feature layer and the output layer, better guiding the target samples to the most likely class centers and promoting the target output to be close to the one-hot vector. The MIE loss is defined as:

[0064]

[0065] where N is the total number of target samples, represents the target sample belonging to the prediction probability of category k, is the confidence weight.

[0066] The second aspect of the present invention provides a storage medium containing computer-executable instructions. When the computer-executable instructions of the storage medium are executed by a computer processor, they are used to execute the liver focal lesion image analysis method based on trust-aware conditional adversarial domain adaptation as described above.

[0067] The third aspect of the present invention provides a computer program product, including a computer program. When the computer program is executed by a processor, it implements the liver focal lesion image analysis method based on trust-aware conditional adversarial domain adaptation as described above.

[0068] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0069] 1. The present invention can quantify the transferability of training samples. TCADA first studies and models the transferability of samples from the perspective of probability distribution, enabling the model to pay more attention to easily transferable samples during training and encouraging the model to learn more cross-domain shared features.

[0070] 2. The present invention can improve the transfer performance of the model. TCADA proposes the FNDA strategy, which simplifies the domain adaptation process by gradually aligning the feature norms across domains. This strategy can also be used as a preprocessing step and integrated into existing UDA methods, thereby significantly improving the transfer performance.

[0071] 3. The present invention can promote the alignment of class-level distributions. TCADA proposes the CCA strategy, which assigns priorities based on the posterior probability of GUM according to the pseudo-label confidence of different target samples, prompting target samples with higher-confidence pseudo-labels to play a greater role in the class distribution alignment process and gradually optimizing the cross-domain class center alignment of target features.

[0072] 4. The present invention can make more accurate predictions for target samples. TCADA proposes the MIE regularization term, which combines the information of the shared feature layer and the output layer, guides the target data to approach the most likely class center, and prompts the target output to be close to the one-hot vector. The results of multiple experiments show that its performance is better than the conditional entropy term widely used in UDA. Description of the Drawings

[0073] Figure 1 is the implementation flowchart of the liver focal lesion image analysis method based on trust-aware conditional adversarial domain adaptation in Embodiment 1;

[0074] Figure 2 is the schematic diagram of the TCADA network architecture in Embodiment 1;

[0075] Figure 3 is the t-SNE embedding visualization result graph of the comparison between TCADA and other methods in Embodiment 1. Detailed Embodiments

[0076] The following further elaborates on the specific embodiments of the present invention through examples. These examples are implemented on the premise of the solution described in the present invention, and detailed implementation manners and specific operation processes are given, but the protection scope of the present invention is not limited to the following examples.

[0077] The present invention is further described below in conjunction with the drawings and specific embodiments. Structural / module names, control modes, algorithms, process flows, or composition ratios and other features not clearly described in this technical solution are regarded as common technical features disclosed in the prior art.

[0078] Embodiment 1

[0079] This embodiment provides a method for liver focal lesion image analysis based on trust-aware conditional adversarial domain adaptation, including: extracting the features of the original image samples and the classification prediction probabilities, quantifying the transferability of the data through GUM (Gaussian-uniform mixture model, GUM), and effectively processing outliers;

[0080] Then calculate the probability-weighted adversarial training loss, the difference in feature norm distribution, the confidence-guided contrast alignment loss, and the mixed information-guided entropy loss to align the source domain and target domain feature norm distributions, determine the optimal transfer strategy, and perform classification prediction, so as to realize liver focal lesion image analysis.

[0081] Specifically, the process of aligning the source domain and target domain feature norm distributions specifically includes:

[0082] S1: Unsupervised domain adaptation network construction: Taking the conditional domain adversarial network as the backbone, constructing joint variables to capture the complex multimodal structure of the feature distribution and more effectively align the source domain and target domain;

[0083] S2: Probability-weighted adversarial training loss: Using GUM to quantify the transferability of the data and distinguish easily transferable samples from difficult-to-transfer samples;

[0084] Define the probability-weighted adversarial training loss so that the model pays more attention to easily transferable samples during training;

[0085] S3: Cross-domain feature norm distribution alignment (Feature norm distribution alignment, FNDA): Gradually align the feature norms of the source domain and target domain through entropy-regularized optimal transport;

[0086] Define and calculate the statistical difference between the feature norm distributions, and adaptively adjust the hyperparameters according to the calculation results to promote the alignment of the source domain and target domain feature norm distributions of the model;

[0087] S4: Confidence-guided Contrastive Alignment (CCA): Assign different weights according to the confidence of the pseudo-labels, and use confidence-guided contrastive learning to gradually optimize the cross-domain class center alignment of the target features.

[0088] S5: Mixed information-guided entropy (MIE): Encourage the target features to stay away from the decision boundary and guide them to the most likely source class center.

[0089] In specific implementation, the image analysis method adopts a trust-aware conditional adversarial domain adaptation module, a probability-weighted adversarial training loss module, a cross-domain feature norm distribution alignment module, a CCA module, and a mixed information-guided entropy (MIE) module;

[0090] The trust-aware conditional adversarial domain adaptation module includes a joint variable and a source supervised classification loss;

[0091] The probability-weighted adversarial training loss module includes a Gaussian-uniform mixture model, a posterior probability, an expectation maximization algorithm, and a probability-weighted adversarial training loss;

[0092] The cross-domain feature norm distribution alignment module includes entropy-regularized optimal transport and the statistical difference between feature norm distributions;

[0093] The confidence-guided contrast alignment module includes confidence-guided contrast learning and a moving average method;

[0094] The mixed information-guided entropy (MIE) module includes a confidence weight and a mixed information-guided entropy loss.

[0095] In specific implementation, the joint variable is specifically: taking the original image as the input x, and combining the depth feature f extracted by the feature extractor and the classification probability g predicted by the image classifier C into a joint variable h, which is expressed as:

[0096] h = (f, g)

[0097] where the depth feature classification prediction probability

[0098] In specific implementation, the source supervised classification loss is specifically:

[0099]

[0100] where represents the standard cross-entropy loss, N represents the batch size, represents the i-th source domain sample, is the corresponding label.

[0101] In specific implementation, the GUM is specifically: modeling the distribution of the classifier prediction entropy H(g i ) of each training sample x i , which is expressed as:

[0102]

[0103] where:

[0104]

[0105] is a uniform distribution on [0, δ], π is the prior probability, and ∑ is the variance of the Gaussian distribution .

[0106] In specific implementation, the posterior probability is specifically: Use a random variable e i ∈ {0, 1}, where 1 and 0 respectively indicate that the sample is easy or difficult to transfer; the probability r i that the sample x φ (x i ) = P φ (e i = 1 | x i ) can be expressed as:

[0107]

[0108] where the parameter set φ = {π, ∑, δ}.

[0109] In specific implementation, the Expectation-Maximization (EM) algorithm is an iterative optimization algorithm that can be used to estimate the parameters of the model:

[0110]

[0111] In specific implementation, the probability-weighted adversarial training loss is specifically: Based on the posterior probability, introduce a weight strategy η(x i ) = 1 + r φ (x i ) to re-weight each training sample, so that the model gives priority to data that is easy to transfer during training; the probability-weighted adversarial training loss is defined as follows:

[0112]

[0113] where T(·) represents a multilinear mapping, represents the j-th target domain sample, represents the i-th source domain sample, represents the feature representation of the i-th source domain sample, D(·) represents the discriminator, represents the feature representation of the j-th target domain sample.

[0114] In specific implementation, the entropy-regularized optimal transport is specifically: Introduce a strictly convex regularization term E(γ) to accelerate the solution of the optimal transport problem, expressed as:

[0115]

[0116] where is the source feature norm distribution ψS The optimal transport matrix between the target feature norm distribution ψ T is denoted as ∏(ψ s , ψ T ) and is the set of feasible transport matrices between ψ S and ψ T , which is expressed as: is an N-dimensional column vector with all elements equal to 1. Each

[0117] in represents the pairwise distance between the source sample and the target sample feature norms in the loss matrix : λ ot is a given regularization parameter used to control the smoothness of the transport plan γ, and E(γ) is the entropy regularization term, which is expressed as:

[0118]

[0119] where γ ij represents the transport probability from source sample i to target sample j.

[0120] Specifically, the statistical difference between the feature norm distributions is defined as:

[0121]

[0122] where is the pairwise distance between source sample i and target sample j.

[0123] Specifically, the confidence-guided contrastive learning means that different priorities are assigned according to the confidence level of the target sample pseudo-label , which is defined as:

[0124]

[0125] where τ is a temperature coefficient, is a weighting factor, and the distance metric dis(·,·) between features uses a multi-kernel strategy, which can fully capture the distance between representations using different kernels and is more conducive to class-level distribution alignment, and is defined as follows:

[0126]

[0127] where the feature kernel related to the feature mapping is a linear combination of m positive definite kernels , is the feature mapping, and the kernel set is expressed as:

[0128]

[0129] Among them, the constraint on the parameter {α u} is used to ensure that the resulting linear combination is characteristic.

[0130] In specific implementation, the moving average method gradually updates the class center through iteration, that is:

[0131]

[0132] Where represents the source class center, is the label of the i-th sample in the source domain, and λ s is the learning rate for updating the class center. is the class center of the k-th category in the t-th iteration and k ∈ {1, 2, 3,..., K}. If the condition is satisfied, then δ(condition) = 1; otherwise, δ(condition) = 0.

[0133] In specific implementation, the confidence weight reflects the similarity between the feature and the class center of the k-th class in the source domain , and the calculation formula is:

[0134]

[0135] Where sim(·, ·) represents the cosine similarity.

[0136] In specific implementation, the MIE loss combines the information of the shared feature layer and the output layer, better guiding the target sample to the most likely class center and promoting the target output to be close to the one-hot vector. The MIE loss is defined as:

[0137]

[0138] Where N is the total number of target samples, represents the predicted probability that the target sample belongs to the category k, and is 's confidence weight.

[0139] Example 2

[0140] This embodiment provides a method for liver focal lesion image analysis based on trust-aware conditional adversarial domain adaptation. The image analysis method employs an unsupervised domain adaptation network module, a probability-weighted adversarial training loss module, a cross-domain FNDA strategy module, a CCA module, and a MIE module. The unsupervised domain adaptation network module is used to extract features from the original image. The probability-weighted adversarial training loss module is used to quantify the transferability of data. The cross-domain FNDA module is used to gradually align the feature norms of the source domain and the target domain. The CCA module is used to align the target features with the most likely source class centers. The MIE module is used to guide the target samples to the most likely class centers and to prompt the target output to be close to the one-hot vector.

[0141] The trust-aware conditional adversarial domain adaptation module includes a feature extractor, a domain discriminator, an image classifier, a joint variable, and a source supervised classification loss. As Figure 2 shown, the CDAN architecture is adopted.

[0142] The joint variable is specifically: taking the original image as x input, and combining the deep features f extracted by the feature extractor and the classification probability g predicted by the image classifier C into a joint variable h, which is expressed as:

[0143] h = (f, g)

[0144] where the deep features the classification prediction probability

[0145] The source supervised classification loss refers to defining the standard cross-entropy loss of the labeled data on the source domain as:

[0146]

[0147] where represents the standard cross-entropy loss, N represents the batch size, represents the i-th source domain sample, is the corresponding label.

[0148] The probability-weighted adversarial training loss module includes GUM, posterior probability, expectation maximization algorithm, and probability-weighted adversarial training loss.

[0149] GUM is specifically: modeling the distribution of the classifier prediction entropy H(g i ) of each training sample x i , which can effectively handle outliers and has high robustness when describing different types of samples, and is expressed as:

[0150]

[0151] where:

[0152]

[0153] is a uniform distribution on [0, δ], π is the prior probability, and ∑ is the variance of the Gaussian distribution of

[0154] The posterior probability is specifically: To simulate the transferability of each training sample, a random variable e i ∈ {0, 1}, where 1 and 0 indicate that the sample is easy or difficult to transfer respectively; for the sample x i the probability r φ (x i ) = P φ (e i = 1|x i ) can be expressed as:

[0155]

[0156] where the parameter set φ = {π, ∑, δ}.

[0157] The Expectation-Maximization (EM) algorithm is an iterative optimization algorithm that can be used to estimate the parameters of the model:

[0158] If m i ~ B(1, 0.5), and B is the Bernoulli distribution, then the variable has the following mixture probability distribution:

[0159]

[0160] The EM algorithm can be used to estimate the parameter set φ of GUM and its iterative formula:

[0161]

[0162] where, is the responsibility assignment for each sample, π (l) is the mixture weight of the class, is the estimated latent variable, δ (l) is the variance, ∑ (l+1) is the newly estimated covariance matrix, and ρ1 and ρ2 represent the first-order and second-order central data moments respectively:

[0163]

[0164] The probability-weighted adversarial training loss is specifically: Based on the posterior probability, a weight strategy η(x i ) = 1 + r φ (x i)Re - weight each training sample so that the model gives priority to data that is easy to transfer during training; Probability - weighted adversarial training loss is defined as follows:

[0165]

[0166] where \(T(\cdot)\) represents a multilinear mapping, represents the \(j\) - th target - domain sample, represents the \(i\) - th source - domain sample, represents the feature representation of the \(i\) - th source - domain sample, \(D(\cdot)\) represents the discriminator, represents the feature representation of the \(j\) - th target - domain sample.

[0167] The cross - domain feature norm distribution alignment module includes entropy - regularized optimal transport and the statistical difference between feature norm distributions.

[0168] The entropy - regularized optimal transport is specifically: introduce a strictly convex regularization term \(E(\gamma)\) to accelerate the solution of the optimal transport problem, expressed as:

[0169]

[0170] where is the optimal transport matrix between the source feature norm distribution \(\psi\) S and the target feature norm distribution \(\psi\) T , \(\Pi(\psi\) s , \(\psi\) T ) is the set of feasible transport matrices between \(\psi\) S and \(\psi\) T , expressed as: is an \(N\) - dimensional column vector with all elements equal to 1. Each

[0171] in represents the pairwise distance between the source sample and the target sample feature norms in the loss matrix : \(\lambda\) ot is a given regularization parameter used to control the smoothness of the transport plan \(\gamma\), and \(E(\gamma)\) is the entropy - regularization term, expressed as:

[0172]

[0173] The statistical difference between feature norm distributions is defined as:

[0174]

[0175] The CCA module includes confidence - guided contrastive learning and the moving average method.

[0176] Confidence-guided contrastive learning means that different priorities are assigned according to the confidence level of the pseudo-labels of the target samples and the classifier prediction probability can also be used as a measure of the confidence of the pseudo-label indicating whether the sample is correctly labeled. Therefore, confidence-guided contrastive learning is defined as:

[0177]

[0178] where τ is a temperature coefficient, is a weighting factor, and the distance metric dis(·,·) between features uses a multi-kernel strategy, which can utilize different kernels to fully capture the distance between representations and is more conducive to class-level distribution alignment, defined as follows:

[0179]

[0180] where the feature kernel related to the feature mapping is a linear combination of m positive definite kernels is the feature mapping, and the kernel set is expressed as:

[0181]

[0182] where the constraint on the parameters {α u} is used to ensure that the resulting linear combination is characteristic.

[0183] The moving average method gradually updates the class center iteratively, i.e.:

[0184]

[0185] where represents the source class center, is the label of the i-th sample in the source domain, and λ s is the learning rate for updating the class center. is the class center of the k-th class in the t-th iteration and k ∈ {1, 2, 3,..., K}. If the condition is satisfied, then δ(condition) = 1; otherwise, δ(condition) = 0.

[0186] The MIE module includes confidence weights and MIE losses.

[0187] The confidence weight reflects the similarity between the feature and the class center of the k-th class in the source domain, and the calculation formula is:

[0188]

[0189] Among them, sim(·, ·) represents the cosine similarity.

[0190] The MIE loss combines the information of the shared feature layer and the output layer, better guiding the target samples to the most likely class centers and promoting the target output to be close to the one-hot vector. The MIE loss is defined as:

[0191]

[0192] Among them is the confidence weight of

[0193] Example 3

[0194] Corresponding to Example 2, this example provides a computer program product for liver focal lesion image analysis based on trust-aware conditional adversarial domain adaptation. The computer program product includes an unsupervised domain adaptation network module, a probability-weighted adversarial training loss module, a cross-domain FNDA strategy module, a CCA module, and an MIE module; the unsupervised domain adaptation network module is used to extract features from the original images, the probability-weighted adversarial training loss module is used to quantify the transferability of the data, the cross-domain FNDA module is used to gradually align the feature norms of the source domain and the target domain, the CCA module is used to align the target features with the most likely source class centers, and the MIE module is used to guide the target samples to the most likely class centers and promote the target output to be close to the one-hot vector.

[0195] The trust-aware conditional adversarial domain adaptation module includes a feature extractor, a domain discriminator, an image classifier, a joint variable, and a source supervised classification loss; as Figure 2 shown, the CDAN architecture is adopted.

[0196] Example 4

[0197] The computer-executable instructions included in the storage medium in this example are specifically configured to: construct a conditional domain adversarial architecture through the unsupervised domain adaptation network module, the architecture includes a feature extractor, a domain discriminator, and an image classifier, where multi-modal feature interaction is realized between the feature extractor and the domain discriminator through a joint variable generation module; the probability-weighted adversarial training loss module is operated to: call the GUM quantization sub-module to calculate the sample transferability coefficient during each forward propagation, and inject this coefficient into the cross-entropy loss function of the domain discriminator to form a dynamically weighted adversarial gradient update rule; the cross-domain FNDA strategy module incorporates an entropy-regularized optimal transport algorithm to dynamically adjust the channel attention weights of the feature extractor according to the difference degree of the feature norm distribution during the backpropagation stage.

[0198] The storage medium is further configured with data flow control instructions, such that the CCA module and the MIE module form a cascaded optimization link: First, the CCA module generates a class center alignment weight matrix based on the pseudo-label confidence, and updates the target domain prototype vector library by the moving average method; Subsequently, the MIE module receives the updated prototype vectors, calculates the mixed entropy value of the KL divergence entropy and the prediction entropy between the target features and the nearest source class center, and drives the topological reconstruction of the feature space through the backpropagation optimizer; Finally, the image classifier outputs the probability distribution map of liver lesions based on the optimized target domain feature vectors, and the storage medium writes the classification result into the DICOM file header and generates a structured diagnostic report.

[0199] Application Example 1

[0200] In a specific case of applying the system of this embodiment, the source domain and target domain data are input, and TCADA extracts the deep features and classification prediction probabilities of the source domain and target domain data through the CDAN architecture; GUM is introduced to quantify the transferability of the data and effectively handle outliers; the FNDA strategy is applied to eliminate the cross-domain feature norm differences and improve the conditional adversarial adaptation process between the source domain and the target domain; the CCA strategy is applied to assign different priorities to the target sample pseudo-labels, prompting the target features to align with the most likely source class centers; the MIE combines the information of the shared feature layer and the output layer, effectively alleviating label drift and creating a clearer decision boundary. In this process, TCADA continuously optimizes the network parameters by calculating the source supervised classification loss, the probability weighted adversarial training loss, the feature norm distribution difference, the CCA loss, and the MIE loss, and finally achieves a high-performance classification task on the unlabeled target domain. As Figure 3 shown, using t-SNE to visualize the features of the source domain and the target domain helps to intuitively display the alignment of the feature distributions of the TCADA model between the source domain and the target domain, and enhance the within-class compactness and between-class separability. Compared with the ResNet-50 model, the DANN model, the CDAN+E model, the PRONOUN model, and the RSDA model, the TCADA model can align the source domain and the target domain more accurately.

[0201] 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 efforts. 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 method for liver focal lesion image analysis based on trust perception conditional adversarial domain adaptation, characterized in that, It includes the following steps: Extract the features of the original liver image samples and the classification prediction probabilities, further quantify the transferability of the probability data through a Gaussian-uniform mixture model, process the outliers therein, and further calculate the probability-weighted adversarial training loss, the difference in feature norm distributions, the confidence-guided contrast alignment loss, and the mixture information-guided entropy loss, so as to align the feature norm distributions of the source domain and the target domain, determine the optimal transfer strategy, and perform liver image classification prediction to complete the analysis of liver focal lesion images.

2. The method for analyzing liver focal lesion images based on trust perception conditional adversarial domain adaptation according to claim 1, wherein, The specific process of calculating the probability-weighted adversarial training loss, the difference in feature norm distributions, the confidence-guided contrast alignment loss, and the mixture information-guided entropy loss to align the feature norm distributions of the source domain and the target domain includes the following steps: S1: Construct an unsupervised domain adaptation framework based on a conditional domain adversarial network, capture the multimodal structure of the feature distribution through joint variable modeling, and establish an initial alignment channel between the source domain and the target domain; S2: Based on the joint variable modeling adversarial network framework constructed in S1, implement probability-weighted adversarial training, including: Adopt a dynamic evaluation of the transferability of the quantified data to measure the sample transfer difficulty and generate a sample transferability coefficient; Construct a probability-weighted adversarial loss function, inject the transferability coefficient as a sample weight into the discriminator training process, and form an adversarial learning mechanism driven by easily transferable samples; S3: During the adversarial training process, synchronously perform cross-domain feature norm distribution alignment, including: Calculate the statistical difference degree of the feature norm distributions of the source domain and the target domain through an entropy-regularized optimal transport algorithm; Based on the statistical difference degree, adaptively adjust the feature mapping parameters of the adversarial network to achieve dynamic distribution matching of cross-domain feature norms; S4: Based on the feature distribution aligned in S3, perform confidence-guided contrast alignment, including: Generate confidence weights according to the pseudo-label confidence of the target domain samples; Based on the confidence weights, gradually pull the target domain features to the neighborhood space corresponding to the source class center through a contrast loss function by means of moving average; S5: Implement the constraint of the mixture information-guided entropy loss during the feature space optimization process, including: Based on the target domain feature distribution adjusted by contrast alignment, calculate the mixed entropy value of the distance entropy from the target feature to the nearest source class center and the category prediction entropy; Drive the target feature to migrate to the high-confidence source class center through the entropy minimization criterion, and establish a feature distribution far from the decision boundary.

3. The method for analyzing liver focal lesion images based on trust perception conditional adversarial domain adaptation according to claim 2, characterized in that, In S1, the joint variable is specifically: taking the original image as the input x, combining the depth feature f extracted by the feature extractor and the classification probability g predicted by the image classifier C into a joint variable h, and the joint variable h is expressed as: h = (f,g) Among them, the depth feature Classification prediction probability 4. A method for analyzing liver focal lesion images based on trust perception conditional adversarial domain adaptation according to claim 2, characterized in that In S2, the specific process of the transferability of the quantization data includes: modeling the classifier prediction entropy H(g i ) of each training sample x i ) and the distribution p(H(g i )|x i ), expressed as: where: is a uniform distribution on [0, δ], π is the prior probability, and Σ is the variance of the Gaussian distribution ​ 5. The method for analyzing liver focal lesion images based on trust perception conditional adversarial domain adaptation according to claim 3, characterized in that, The specific calculation process of the probability-weighted adversarial training loss is as follows: Based on the posterior probability, introduce a weight strategy η(x i ) = 1 + r φ (x i ) to re-weight each training sample, so that the model gives priority to the data that is easy to transfer during the training process. Among them, η(x i ) is the weight strategy for each training sample x i , and r φ (x i ) is the posterior probability of the sample x i ; Probability-Weighted Adversarial Training Loss is as follows: where \(T(\cdot)\) represents a multilinear mapping, represents the \(j\)-th target domain sample, represents the \(i\)-th source domain sample, represents the feature representation of the \(i\)-th source domain sample, and \(D(\cdot)\) represents the discriminator, represents the feature representation of the \(j\)-th target domain sample.

6. The method for analyzing liver focal lesion images based on trust perception conditional adversarial domain adaptation according to claim 2, wherein In S3, the entropy-regularized optimal transport algorithm is specifically: introduce a strict convex regularization term E(γ) to accelerate the solution of the optimal transport problem, expressed as: Among them is the optimal transport matrix between the source feature norm distribution ψ s and the target feature norm distribution ψ T , Π(ψ s , ψ T ) is the set of feasible transport matrices between ψ S and ψ T , denoted as: is an N-dimensional column vector with all elements equal to 1, and each entry in represents the pairwise distance between the source sample and the target sample feature norms in the loss matrix : λ ot is a given regularization parameter used to control the smoothness of the transport plan γ, and E(γ) is the entropy regularization term, denoted as: where γ ij represents the transmission probability from the source sample i to the target sample j, The statistical difference degree of the feature norm distribution is defined as: wherein is the pairwise distance between the source sample i and the target sample j.

7. A method for analyzing liver focal lesion images based on trust perception conditional adversarial domain adaptation according to claim 2, characterized in that In S4, the confidence-guided contrastive alignment loss assigns different priorities according to the confidence level of the pseudo-labels of the target samples and is defined as: where τ is a temperature coefficient, is a weighting factor, and the distance metric dis(·,·) between features uses a multi-kernel strategy to fully capture the distance between representations using different kernels, which is defined as follows: The feature kernel related to the feature mapping is a linear combination of m positive definite kernels and is the feature mapping, and the kernel set is expressed as: Among them, the constraint on the parameter {α u} is used to ensure that the resulting linear combination is characteristic; The moving average method updates the class center iteratively, that is: Among them represents the source class center, is the label of the i-th sample in the source domain, λ s is the learning rate for updating the class center, is the class center of the k-th category in the t-th iteration and k ∈ {1, 2, 3, …, K}. If the condition is satisfied, then δ(condition) = 1; otherwise, δ(condition) = 0. The confidence weight reflects the feature and the class center of the k-th class in the source domain The similarity of, and the confidence weight calculation formula is: where sim(·,·) represents the cosine similarity.

8. A method for analyzing liver focal lesion images based on trust perception conditional adversarial domain adaptation according to claim 2, characterized in that The mixture information-guided entropy loss is defined as: where N is the total number of target samples, denotes the target sample prediction probability belonging to class k, is the confidence weight.

9. 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 the method for analyzing liver focal lesion images based on trust-aware conditional adversarial domain adaptation according to any one of claims 1 to 8.

10. A computer program product comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method for analyzing liver focal lesion images based on trust perception conditional adversarial domain adaptation as described in any one of claims 1 to 8.

Images