Model optimization method based on transfer learning

Abstract

The embodiment of the invention discloses a model optimization method based on transfer learning. The method comprises the following steps: acquiring to-be-migrated learning data, wherein the to-be-migrated learning data comprises source domain data associated with pathological characteristics of kidney diseases; semantic alignment processing is carried out on the source domain data and target domain data, the target domain data is kidney disease pathology data, and a semantic alignment result is obtained; determining a feature distribution result corresponding to the target domain data, and determining a migration weight corresponding to the source domain data based on the feature distribution result and the semantic alignment result; and optimizing the target domain model based on the migration weight, the source domain data and the target domain data to obtain an optimized target domain model. According to the embodiment of the invention, the accuracy and robustness of the target domain model in a kidney disease pathological diagnosis task and the generalization ability of different data distributions can be remarkably improved.

Description

Technical Field

[0001] This application relates to the field of transfer learning technology, specifically to a model optimization method based on transfer learning. Background Technology

[0002] In the fields of medical imaging and artificial intelligence, large-scale pathology models have become an important tool for assisted diagnosis. Existing large-scale pathology models primarily rely on deep learning technology, training on massive amounts of pathological images and clinical data to achieve automated diagnosis and analysis of diseases. However, existing large-scale pathology models still face many challenges in practical applications, particularly in data annotation, model generalization ability, and integration with clinical applications.

[0003] In current technologies, large-scale pathology models have become important tools for assisted diagnosis in the fields of medical imaging and artificial intelligence. These models primarily rely on deep learning techniques, training on large amounts of pathological images and clinical data to achieve automated disease diagnosis and analysis. However, existing large-scale pathology models still face many challenges in practical applications, particularly in data annotation, model generalization ability, and integration with clinical applications. Therefore, the interpretability and practicality of these models need improvement. Summary of the Invention

[0004] To address the aforementioned technical problems, embodiments of this application provide a transfer learning-based model optimization method.

[0005] According to one aspect of the embodiments of this application, a model optimization method based on transfer learning is provided, comprising: acquiring data to be transferred, the data including source domain data associated with pathological features of kidney disease; performing semantic alignment processing on the source domain data and target domain data, the target domain data being pathological data of kidney disease, to obtain a semantic alignment result; determining the feature distribution result corresponding to the target domain data, and determining the transfer weight corresponding to the source domain data based on the feature distribution result and the semantic alignment result; and optimizing the target domain model based on the transfer weight, the source domain data, and the target domain data to obtain an optimized target domain model.

[0006] According to one aspect of the embodiments of this application, determining the feature distribution result corresponding to the target domain data and determining the migration weight corresponding to the source domain data based on the feature distribution result and the semantic alignment result includes: classifying the target domain data and determining the scarce category data corresponding to the target domain based on the classification result; and determining the migration weight corresponding to the source domain data based on the scarcity coefficient corresponding to the scarce category data and the semantic alignment result.

[0007] According to one aspect of the embodiments of this application, determining the transfer weight corresponding to the source domain data based on the scarcity coefficient corresponding to the scarce category data and the semantic alignment result includes: determining the feature mapping between the target domain data and the source domain data to the same semantic space based on the semantic alignment result, and obtaining a semantic mapping function; calculating the similarity between the target domain data and the source domain data based on the semantic mapping in the same semantic space; and determining the transfer weight corresponding to the source domain data based on the similarity.

[0008] According to one aspect of the embodiments of this application, determining the migration weight corresponding to the source domain data based on the scarcity coefficient corresponding to the scarce category data and the semantic alignment result includes: determining the scarce source domain data corresponding to the source domain data and the scarce category data; if the scarcity coefficient is greater than a preset coefficient threshold, then determining the migration weight corresponding to the scarce source domain data based on the scarcity coefficient.

[0009] According to one aspect of the embodiments of this application, the method further includes: if the source domain data is image data, constructing a feature extraction layer, the feature extraction including a convolutional network and a pooling layer corresponding to the convolutional neural network; determining the image data for pre-training based on the feature extraction layer to obtain the classification result of the source domain data; and determining the scarce source domain data corresponding to the scarce category data based on the classification result.

[0010] According to one aspect of the embodiments of this application, the method further includes: if the source domain data is image data, then normalizing the image data to unify the pixel values ​​of the image data to a preset range; within the preset range, standardizing the image data to obtain standardized source domain data; and determining source domain data associated with the pathological features of kidney disease based on the standardized source domain data.

[0011] According to one aspect of the embodiments of this application, the method further includes: performing image enhancement processing on the standardized source domain data to obtain enhanced source domain data, wherein the image enhancement processing includes one or more of rotation, flipping, and cropping; constructing a training dataset for a target domain model based on the enhanced source domain data, and training the target domain model based on the training dataset.

[0012] According to one aspect of the embodiments of this application, the step of performing semantic alignment processing on the source domain data and the target domain data, wherein the target domain data is pathological data of kidney disease, to obtain a semantic alignment result includes: extracting semantic features corresponding to the source domain data, and semantically mapping the semantic features to source domain features; extracting target domain features corresponding to the target domain data; and obtaining a semantic alignment result by applying a preset semantic mapping function to the source domain features and the target domain features.

[0013] In the technical solution provided in the embodiments of this application, semantic alignment technology is used to eliminate cross-domain expression differences in kidney disease pathological data to construct a shared semantic space. The value of source domain data migration is dynamically quantified and weights are assigned in combination with the target domain feature distribution. Ultimately, the accurate adaptation and efficient migration of source domain pathological knowledge to the target domain are achieved, thereby significantly improving the accuracy, robustness and generalization ability of the target domain model in kidney disease pathological diagnosis tasks.

[0014] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. In the drawings: Figure 1 This is an exemplary embodiment of the present application illustrating an implementation environment for performing model optimization based on transfer learning; Figure 2 This is a flowchart illustrating a model optimization method based on transfer learning, as shown in an exemplary embodiment of this application. Figure 3 This is a flowchart illustrating a model optimization method based on transfer learning, as shown in another exemplary embodiment of this application. Figure 4 This is a flowchart illustrating a model optimization method based on transfer learning, as shown in another exemplary embodiment of this application. Figure 5 This is a flowchart illustrating a model optimization method based on transfer learning, as shown in another exemplary embodiment of this application. Figure 6 This is a flowchart illustrating a model optimization method based on transfer learning, as shown in another exemplary embodiment of this application. Figure 7 This is a flowchart illustrating a model optimization method based on transfer learning, as shown in another exemplary embodiment of this application. Figure 8 This is a flowchart illustrating a model optimization method based on transfer learning, as shown in another exemplary embodiment of this application. Figure 9 This is a flowchart illustrating a model optimization method based on transfer learning, as shown in another exemplary embodiment of this application. Figure 10This is a block diagram illustrating a model optimization apparatus based on transfer learning, as shown in an exemplary embodiment of this application. Figure 11 A schematic diagram of the structure of a computer system suitable for implementing the electronic device of the present application is shown. Detailed Implementation

[0016] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0017] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0018] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily need to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0019] In this application, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0020] First, it's important to note that in the fields of medical imaging and artificial intelligence, large-scale pathology models have become crucial tools for assisted diagnosis. Existing large-scale pathology models primarily rely on deep learning technology, training on vast amounts of pathological images and clinical data to achieve automated disease diagnosis and analysis. However, existing large-scale pathology models still face numerous challenges in practical applications, particularly in data annotation, model generalization ability, and integration with clinical applications.

[0021] However, annotating pathological data requires specialized medical knowledge and a significant amount of time and effort. The annotation process is time-consuming and laborious, and the accuracy and consistency of the annotations are difficult to guarantee. For example, accurately annotating a single pathological slide may require a pathologist to spend several hours, and the annotation results may differ between different doctors. In particular, kidney diseases are diverse, and each disease has different pathological stages and manifestations. Existing large-scale pathological models often rely on specific datasets during training, making it difficult to adapt to new and unseen pathological data, resulting in a decrease in diagnostic accuracy in practical applications.

[0022] Transfer learning provides an efficient and low-cost solution for model optimization by transferring learned knowledge (such as feature representations, model parameters, or prior distributions) from a source domain (such as large-scale general datasets or relevant task data) to a target domain (such as small-scale or heterogeneous data in a specific scenario). Its core logic lies in breaking through the dependence of traditional model training on large-scale labeled data, and significantly improving the performance of the target model in scenarios with scarce data, distribution shifts, or similar tasks through strategies such as feature reuse, parameter fine-tuning, or relation transfer.

[0023] Figure 1 This is a schematic diagram illustrating an implementation environment for model optimization based on transfer learning, as shown in an exemplary embodiment of this application. Figure 1 As shown, during model training, transfer learning data can be obtained from the source domain model 110, which includes source domain data associated with pathological features of kidney disease. Then, the source domain data and the target domain data in the target domain model 120 are semantically aligned to obtain the semantic alignment result. The feature distribution result corresponding to the target domain data is determined, and the transfer weight corresponding to the source domain data is determined based on the feature distribution result and the semantic alignment result. Then, the target domain model is optimized according to the transfer weight, the source domain data, and the target domain data to obtain the optimized target domain model, thereby achieving the optimization of the target domain model.

[0024] In current technologies, large-scale pathology models have become important tools for assisted diagnosis in the fields of medical imaging and artificial intelligence. These models primarily rely on deep learning techniques, training on large amounts of pathological images and clinical data to achieve automated disease diagnosis and analysis. However, existing large-scale pathology models still face many challenges in practical applications, particularly in data annotation, model generalization ability, and integration with clinical applications. Therefore, the interpretability and practicality of these models need improvement.

[0025] To address these issues, embodiments of this application propose a model optimization method based on transfer learning, a model optimization device based on transfer learning, an electronic device, a computer-readable storage medium, and a computer program product, which will be described in detail below.

[0026] Please see Figure 2 , Figure 2 This is a flowchart illustrating a model optimization method based on transfer learning, as shown in an exemplary embodiment of this application. It should be understood that this method can also be applied to other exemplary implementation environments and specifically executed by devices in other implementation environments. This embodiment does not limit the implementation environment to which this method is applicable.

[0027] like Figure 2 As shown, in an exemplary embodiment, the model optimization method based on transfer learning includes at least steps S210 to S240, which are described in detail below: Step S210: Obtain the data to be transferred for learning, which includes source domain data associated with pathological features of kidney disease.

[0028] For example, knowledge can be transferred from mature models in other medical fields (such as cardiovascular diseases, neurological diseases, etc.). Source domain data that is highly correlated with the pathological features of kidney diseases can be selected, such as cardiovascular disease imaging data, bioinformatics data, etc. The feature extraction layer parameters of the source domain model can be transferred to the large pathological model, leveraging the rich features of the source domain data to improve the model's ability to recognize pathological images and reduce the dependence on the annotation of kidney pathological data.

[0029] Optionally, by mining cross-organ pathological information such as vascular calcification patterns in cardiovascular imaging, tissue stiffness features in liver fibrosis data, and protein deposition profiles in nervous system diseases, and combining gene and metabolite association networks with clinical course knowledge graphs, a multimodal source domain dataset highly aligned with core pathological features such as renal fibrosis and glomerulosclerosis can be constructed. Cross-domain feature transfer can be achieved by employing contrastive learning and dynamic weight adjustment strategies.

[0030] Step S220: Semantic alignment is performed between the source domain data and the target domain data, where the target domain data is pathological data of kidney disease, to obtain the semantic alignment result.

[0031] For example, a semantic mapping network is constructed, and features in the source domain data are aligned with features in the renal pathology data using semantic parsing techniques. For instance, edge features in cardiovascular disease images are matched with cell boundary features in renal pathology images, achieving accurate knowledge transfer. By minimizing the differences between the features of the source and target domains, the semantic mapping function is optimized to ensure that knowledge from the source domain can be accurately transferred to the target domain at the semantic level.

[0032] Optionally, a cross-modal feature decoupling network is constructed to semantically encode source domain data such as vascular morphological features in cardiovascular imaging, stiffness distribution features in liver fibrosis elastography, and protein deposition patterns in nervous system diseases. At the same time, a self-supervised learning strategy is used to extract core semantic features of the target domain, such as glomerular structure and interstitial fibrosis, from kidney pathology WSI images and omics data. Then, a high-order semantic alignment between the feature spaces of the source and target domains is achieved using a contrastive learning framework based on maximizing mutual information. Finally, a dynamic weight adjustment mechanism is used to balance the contributions of cross-domain shared features and kidney pathology-specific features, outputting semantic alignment results that are both generalizable and task-specific.

[0033] Step S230: Determine the feature distribution results corresponding to the target domain data, and determine the transfer weights corresponding to the source domain data based on the feature distribution results and semantic alignment results.

[0034] For example, to address the imbalance in pathological data, source domain data is dynamically weighted. The weights of transferred knowledge are adjusted based on the distribution and demand of the target domain data. For instance, for rare kidney disease samples, the weights of relevant source domain knowledge are increased to improve the model's ability to identify rare diseases. By dynamically adjusting the weights, the model's learning ability on scarce samples is optimized, mitigating the overfitting problem on imbalanced data.

[0035] Optionally, the probability distribution density of the target domain data of renal pathology is quantified by kernel density estimation on key pathological features such as the degree of glomerular sclerosis and the area of ​​interstitial fibrosis. Combined with the mutual information contribution of source domain features and target domain semantic space in the semantic alignment results, a distribution matching algorithm based on Wasserstein distance is used to evaluate the similarity between the distribution of source domain data and target domain features. Then, the attention mechanism is used to dynamically allocate the transfer weights, so that source domain features (such as liver fibrosis stiffness features) that are strongly correlated with high-density feature regions of the target domain (such as severe fibrosis patterns) are given higher weights, while suppressing source domain noise features that are associated with low-density regions of the target domain (such as normal kidney tissue). Finally, a transfer weight matrix that takes into account both distribution adaptability and semantic consistency is output.

[0036] Step S240: Optimize the target domain model based on the migration weights, source domain data, and target domain data to obtain the optimized target domain model.

[0037] For example, by weighting and fusing dynamically assigned transfer weights with semantically aligned source domain features (such as cardiovascular calcification patterns and liver fibrosis stiffness features), and combining gradient information from target domain renal pathology data (WSI images and genomic data), a multi-task learning framework is used to simultaneously optimize the shared feature extraction layer and task-specific prediction head of the target domain model. At the same time, a meta-learning strategy is introduced to dynamically adjust the transfer weights to adapt to the time-varying characteristics of the target domain data distribution. Finally, an optimized model with significantly improved generalization performance is output on tasks such as glomerular sclerosis grading and interstitial fibrosis quantification.

[0038] Optionally, source domain data strongly correlated with renal pathology are first obtained from multi-source heterogeneous data, including vascular calcification patterns in cardiovascular imaging, stiffness distribution characteristics in liver fibrosis elastography, protein deposition profiles in neurological diseases, and cross-organ genomics (e.g., APOL1 gene mutations) and metabolomics data (e.g., trimethylamine oxide TMAO). Through contrastive learning and knowledge distillation techniques, the semantic features of the source domain data (e.g., fibrosis texture, inflammatory cell infiltration patterns) are multimodally semantically aligned with the target domain renal pathology data (WSI images, renal biopsy omics data). Specifically, graph neural networks (GNNs) are used to model cross-domain entity relationships (e.g., the pathological mechanism association between "cardiovascular calcification and renal artery stenosis"), and the semantic consistency of the feature space is constrained by the mutual information maximization criterion. Subsequently, kernel density estimation was used to quantify the probability distribution of target domain data on key pathological indicators such as glomerular sclerosis degree and interstitial fibrosis area. The similarity between the feature distributions of the source and target domains was measured using Wasserstein distance. A dynamic weight allocation module based on an attention mechanism was then designed. This module automatically adjusts the transfer weights based on the mutual information contribution between source domain features and high-density feature regions in the target domain (such as severe fibrosis patterns). For example, it assigns higher weights to liver fibrosis stiffness features to strengthen cross-organ fibrosis cognition, while suppressing noisy features associated with low-density regions in the target domain (such as normal kidney tissue). Finally, the weighted source domain features and target domain data are input into a multi-task learning framework. A shared feature extraction layer captures cross-domain common pathological patterns (such as the morphological commonality of vascular lesions), and task-specific prediction heads (such as U-Net segmentation of glomeruli and XGBoost quantification of fibrosis proportions) are used to optimize the model's performance on renal pathology diagnosis tasks. Simultaneously, a meta-learning strategy is introduced to dynamically update the transfer weights to adapt to the time-varying characteristics of the target domain data distribution (such as the pathological heterogeneity of different patient groups).

[0039] In some embodiments of this application, by semantically aligning the data to be transferred and learning, determining the source domain transfer weights based on the target domain feature distribution, and optimizing the target domain model accordingly, knowledge related to the pathology of kidney diseases in the source domain can be effectively integrated, thereby improving the diagnostic performance and generalization ability of the target domain model for kidney diseases.

[0040] Furthermore, based on the above embodiments, please refer to... Figure 3 In some exemplary embodiments provided in this application, the specific implementation process of determining the feature distribution result corresponding to the target domain data and determining the migration weight corresponding to the source domain data based on the feature distribution result and semantic alignment result may further include steps S310 and S320, which are described in detail below: Step S310: Classify the target domain data and determine the scarce category data corresponding to the target domain based on the classification results; Step S320: Determine the migration weights corresponding to the source domain data based on the scarcity coefficients of the scarce category data and the semantic alignment results.

[0041] For example, to address the model bias problem caused by the uneven distribution of categories in renal pathology data, a hierarchical classification strategy is first adopted to refine the target domain data: for core pathological types such as glomerular lesions (e.g., focal segmental glomerulosclerosis FSGS, membranous nephropathy MN) and interstitial injuries (e.g., acute tubular necrosis ATN, chronic interstitial nephritis CIN), a multi-level classifier is constructed (the first-level classification distinguishes between proliferative and atrophic lesions, and the second-level classification refines specific subtypes), and the classification robustness is improved through ensemble learning (e.g., the Stacking model integrates pathologist annotations and automated initial screening results). Based on the classification results, the scarcity coefficient is quantified by a weighted combination of information entropy and class frequency. For scarce categories with a sample size of less than 5% (such as C3 glomerulonephritis), a higher scarcity weight is assigned (e.g., coefficient = 1.5 × (1 - class frequency)). At the same time, combined with the correlation strength between source domain features and the semantic space of scarce categories in the semantic alignment results (e.g., the mutual information value between liver fibrosis features and rare subtypes of renal interstitial fibrosis), a dynamic transfer weight adjustment mechanism is designed: This mechanism uses a two-branch attention network. One branch focuses on the high-order semantic features of scarce categories (e.g., the immune complex deposition pattern of rare glomerulonephritis), and the other branch integrates cross-organ common features of source domain data (e.g., the morphological similarity between cardiovascular calcification and renal vascular sclerosis). Finally, a transfer weight matrix is ​​generated by weighted summation. Among them, source domain features strongly correlated with scarce categories (e.g., the contribution of liver fibrosis stiffness features to the prediction of CIN) are given higher weights (weight value increased by 30%), while generalized features associated with common categories maintain the basic weight.

[0042] Optionally, in some feasible embodiments of this application, the source domain data is dynamically weighted to address the imbalance in pathological data. Specifically, the weights of the source domain data can be defined as follows: The weights of the target domain data are And set the weight adjustment function. So that by using the weight adjustment function A, The weight adjustment function can dynamically adjust the weights based on the distribution and needs of the target domain data, prioritizing the transfer of knowledge related to scarce samples in the target domain. Furthermore, the weights of the target domain can be adjusted through the following optimization objectives: in, It is the first The loss function for each sample. It involves adjusting parameters to control the magnitude of weight changes. For example, for rare kidney disease samples, increasing the weight of relevant source domain knowledge can improve the model's ability to identify rare diseases.

[0043] In some embodiments of this application, by classifying the target domain data to determine scarce category data, and combining its scarcity coefficient with semantic alignment results to determine the source domain data migration weight, the source domain knowledge most critical to improving the performance of the target domain model can be accurately screened, effectively solving the problem of target domain data scarcity and enhancing the model's generalization ability.

[0044] Furthermore, based on the above embodiments, please refer to... Figure 4 In one exemplary embodiment provided in this application, the specific implementation process of determining the migration weight corresponding to the source domain data based on the scarcity coefficient corresponding to the scarce category data and the semantic alignment result may further include steps S410 to S430, which are described in detail below: Step S410: Based on the semantic alignment result, determine the feature mapping between the target domain data and the source domain data to the same semantic space, and obtain the semantic mapping function; Step S420: Under the same semantic space, calculate the similarity between the target domain data and the source domain data based on the semantic mapping. Step S430: Determine the migration weights corresponding to the source domain data based on similarity.

[0045] For example, in the development of medical image-assisted diagnostic systems, there is a problem of insufficient training data for kidney disease diagnostic models, while medical image diagnostic models in fields such as cardiovascular diseases and neurological diseases are relatively mature and have accumulated a large amount of data. To improve the performance of the kidney disease diagnostic model, it was decided to transfer knowledge from other medical fields. First, image data and bioinformatics data from fields such as cardiovascular diseases and neurological diseases are collected as source domain data, while kidney disease-related image data is acquired as target domain data. Next, semantic alignment is performed. The system deeply analyzes the semantic features of these different domain data. For example, the morphological and texture changes of blood vessels in cardiovascular disease images are compared with the abnormal manifestations of renal blood vessels and renal parenchyma in kidney disease images to find elements with similar semantics. Through this semantic alignment, the system can determine the feature mapping relationship between the target domain data (kidney disease data) and the source domain data (cardiovascular disease and neurological disease data), transforming the data features from different domains into the same semantic space. During this process, the deep learning model continuously optimizes the parameters, making the feature representations of the same semantics as close as possible in different domains, ultimately obtaining a semantic mapping function that can accurately achieve this transformation. Once all data is mapped to the same semantic space, the system calculates the similarity between the target domain data (kidney disease data) and the source domain data (cardiovascular disease and nervous system disease data) based on a semantic mapping function. For example, the similarity is measured by comparing the feature vector distance between vascular stenosis features in kidney disease images and vascular stenosis features in cardiovascular disease images in the semantic space; the closer the distance, the higher the similarity. Finally, the transfer weights corresponding to the source domain data are determined based on the calculated similarity. Source domain data with higher similarity contains knowledge and patterns that have greater reference value for the target domain (kidney disease diagnosis), and therefore will be assigned higher transfer weights. For source domain data that is highly correlated with the pathological features of kidney disease, such as stenosis and calcification features in cardiovascular disease images that are similar to kidney vascular lesions, larger transfer weights will be given. When the feature extraction layer parameters of the source domain model are subsequently transferred to the large-scale kidney disease pathology model, the feature extraction parameters corresponding to these high-weight source domain data will be more fully referenced and utilized, thereby helping the large-scale kidney disease pathology model to better learn effective feature representations and improve the model's diagnostic accuracy for kidney diseases.

[0046] In some embodiments of this application, a semantic mapping function is constructed using the semantic alignment results to map the data features of the target domain and the source domain to the same semantic space, and the similarity is calculated accordingly to determine the transfer weight. This can effectively mine the semantic association between the source domain and the target domain, improve the pertinence and effectiveness of the transferred knowledge, and thus enhance the performance of the target domain model.

[0047] Furthermore, based on the above embodiments, please refer to... Figure 5 In one exemplary embodiment provided in this application, the specific implementation process of determining the migration weight corresponding to the source domain data based on the scarcity coefficient corresponding to the scarce category data and the semantic alignment result may further include steps S510 and S520, which are described in detail below: Step S510: Determine the scarce source domain data corresponding to the source domain data and the scarce category data; Step S520: If the scarcity coefficient is greater than the preset coefficient threshold, then determine the migration weight corresponding to the scarce source domain data based on the scarcity coefficient.

[0048] For example, in the development of medical image-assisted diagnostic systems, there is a problem of insufficient training data for kidney disease diagnostic models, especially since data on certain specific pathological types of kidney disease is extremely scarce. Meanwhile, medical image diagnostic models in fields such as cardiovascular disease and neurological diseases are relatively mature and have accumulated a large amount of data. To improve the performance of the kidney disease diagnostic model, it was decided to transfer knowledge from other medical fields. First, image data and bioinformatics data from fields such as cardiovascular disease and neurological disease were collected as source domain data, while kidney disease-related image data was acquired as target domain data. When analyzing this data, it was necessary to identify scarce source domain data that correspond to scarce categories of kidney disease data within the source domain data. Various features of the source domain data were analyzed in depth; for example, the features exhibited by different lesion types in cardiovascular disease image data were compared in detail with various pathological features in kidney disease image data. For those features that correspond to scarce categories in kidney disease image data, similar data were searched in the source domain data; these similar data were then identified as scarce source domain data. For example, a rare kidney tumor in kidney disease may exhibit specific morphological and textural features in imaging. If similar features exist in cardiovascular disease imaging data, these would be labeled as scarce source domain data corresponding to that scarce category. Next, the scarcity coefficient of these scarce source domain data is calculated. This coefficient comprehensively considers various factors, such as the proportion of this type of data in the source domain data and the degree of matching with scarce category data in the target domain. Then, the calculated scarcity coefficient is compared with a preset coefficient threshold. If the scarcity coefficient is greater than the preset threshold, it indicates that this type of scarce source domain data has high reference value for the target domain (kidney disease diagnosis). At this point, the migration weight corresponding to the scarce source domain data is further determined based on the scarcity coefficient. The higher the scarcity coefficient, the scarcer the data and the stronger its correlation with scarce category data in the target domain, and therefore, it will be assigned a higher migration weight. For example, for a type of lesion data that is extremely rare in kidney disease but has similar features in cardiovascular disease imaging data, due to its high scarcity coefficient, when the feature extraction layer parameters of the source domain model are subsequently transferred to the kidney disease pathology model, the feature extraction parameters corresponding to these high-weight scarce source domain data will be more fully referenced and utilized. This will help the kidney disease pathology model learn more effective feature representations for these scarce pathological types, thereby improving the model's diagnostic accuracy for various pathological conditions of kidney disease, especially scarce pathological types.

[0049] In some embodiments of this application, by accurately locating the scarce source domain data corresponding to the scarce category and reasonably determining the transfer weight based on the scarcity coefficient, the role of key scarce data in knowledge transfer can be effectively highlighted, thereby improving the learning effect and overall performance of the target model on the scarce category.

[0050] Furthermore, based on the above embodiments, please refer to... Figure 6 In one exemplary embodiment provided in this application, the specific implementation process of the above-mentioned model optimization method based on transfer learning may further include steps S610 to S630, which are described in detail below: Step S610: If the source domain data is image data, then construct a feature extraction layer. Feature extraction includes a convolutional network and a pooling layer corresponding to the convolutional neural network. Step S620: Based on the feature extraction layer, pre-train the image data to obtain the classification results of the source domain data; Step S630: Determine the scarce source domain data corresponding to the scarce category data based on the classification results.

[0051] For example, in the development of medical image-assisted diagnostic systems, there is a lack of training data for kidney disease diagnostic models, especially a scarcity of data for certain specific pathological types. In contrast, medical image diagnostic models for cardiovascular diseases, neurological diseases, and other fields are relatively mature and have accumulated a large amount of data. To improve the performance of the kidney disease diagnostic model, it was decided to transfer knowledge from these fields.

[0052] First, image data from cardiovascular diseases, neurological diseases, and related fields are collected as source domain data, while image data related to kidney diseases is acquired as target domain data. Since the source domain data is image data, a feature extraction layer needs to be constructed. This feature extraction layer includes a convolutional neural network (CNN) and corresponding pooling layers. The CNN can automatically learn local features in the image, extracting feature information at different levels, such as edges and textures, through the sliding operation of multiple convolutional kernels. The pooling layer is used to downsample the feature map output by the convolutional layer, reducing the amount of data while retaining the main features, and enhancing the model's translation invariance. After constructing the feature extraction layer, the image data is pre-trained based on this layer. During pre-training, the source domain image data is input into the feature extraction layer, and through convolution and pooling operations, the feature representation of the image is extracted. Then, these features are input into subsequent classifiers, such as fully connected layers and a softmax function, to classify the images and obtain the classification results of the source domain data. This classification result reflects the different categories to which the source domain image data belongs, such as different lesion types in cardiovascular disease image data.

[0053] Next, based on the classification results, the scarce source domain data corresponding to the scarce category data is determined. The distribution of each category in the source domain data is analyzed. Categories with a small number of samples in the source domain data but belonging to rare pathological types in kidney disease imaging data are marked as scarce categories. Then, data belonging to these scarce categories are selected from the classification results of the source domain data; this data constitutes the scarce source domain data. For example, in cardiovascular disease imaging data, the number of samples for a certain rare type of heart lesion is very small, while in kidney disease imaging data, a corresponding type of kidney lesion also belongs to a scarce category. Therefore, the data belonging to this rare heart lesion type in the classification results of cardiovascular disease imaging data is identified as the scarce source domain data corresponding to the scarce category data in kidney disease. These identified scarce source domain data will play a crucial role in the subsequent model transfer process. The feature extraction layer parameters of the source domain model will be transferred to the large-scale pathological model of kidney disease. Since these scarce source domain data and scarce category data of the target domain are highly correlated, by transferring their corresponding feature extraction parameters, the large-scale pathological model of kidney disease can better learn effective feature representations for these scarce pathological types, thereby improving the model's diagnostic accuracy for various pathological conditions of kidney disease, especially scarce pathological types.

[0054] In some embodiments of this application, image source domain data is pre-trained by constructing a feature extraction layer containing convolutional networks and pooling layers to obtain classification results, thereby identifying scarce source domain data. This can effectively mine image data features, accurately locate key scarce data that is valuable to the target task, and improve knowledge transfer efficiency and model performance.

[0055] Furthermore, based on the above embodiments, please refer to... Figure 7 In one exemplary embodiment provided in this application, the specific implementation process of the above-mentioned model optimization method based on transfer learning may further include steps S710 to S730, which are described in detail below: Step S710: If the source domain data is image data, then the image data is normalized to unify the pixel values ​​of the image data to a preset range. Step S720: Within a preset range, the image data is standardized to obtain standardized source domain data; Step S730: Determine source domain data associated with pathological features of kidney disease based on standardized source domain data.

[0056] For example, firstly, image data from fields such as cardiovascular diseases and neurological diseases are collected as source domain data, while image data related to kidney diseases is acquired as target domain data. Since the source domain data is image data, it needs to be normalized. Normalization unifies the pixel values ​​of the image data to a preset range, such as scaling the pixel values ​​to between 0 and 1. This is done because image data from different sources may have different pixel value ranges; normalization can eliminate this difference, bringing the data to a uniform scale, facilitating subsequent processing and analysis. For instance, some images have pixel value ranges from 0 to 255; by normalizing and dividing by 255, the pixel values ​​can be mapped to the range of 0 to 1.

[0057] After normalization, the image data is standardized within a preset range. Standardization involves centering and scaling the data to ensure it has a mean of 0 and a standard deviation of 1. This reduces the differences between data points, improving the training performance and stability of the model. For example, for normalized image data, the mean and standard deviation of each pixel are calculated. Then, the mean is subtracted from the pixel value, and the result is divided by the standard deviation to obtain the standardized source domain data.

[0058] Next, based on the standardized source domain data, source domain data associated with the pathological features of kidney disease are identified. This involves in-depth analysis of the pathological features of kidney disease imaging data, such as the morphology, size, and location of kidney tumors, and abnormal changes in kidney blood vessels. Then, within the standardized source domain data, data with similar pathological features are searched. For example, certain cardiac lesions in cardiovascular disease imaging data may resemble kidney tumors in morphology and texture, or certain vascular abnormalities caused by lesions in neurological disease imaging data may match the vascular lesion characteristics in kidney disease. These source domain data associated with the pathological features of kidney disease are then selected.

[0059] In the subsequent model transfer process, these identified source domain data, which are associated with the pathological features of kidney diseases, will play a crucial role. The feature extraction layer parameters of the source domain model will be transferred to the large-scale kidney disease pathology model. Because these source domain data and target domain data are correlated in terms of pathological features, transferring their corresponding feature extraction parameters can help the large-scale kidney disease pathology model better learn effective feature representations, thereby improving the model's diagnostic accuracy for various pathological conditions of kidney diseases.

[0060] In some embodiments of this application, the source domain data, which is an image, is first normalized to a uniform range of pixel values, and then standardized to obtain standardized data. This can eliminate data differences and improve data quality, thereby more accurately identifying effective source domain data associated with the pathological features of kidney disease, and helping to improve model knowledge transfer and performance.

[0061] Furthermore, based on the above embodiments, please refer to... Figure 8 In one exemplary embodiment provided in this application, the specific implementation process of the above-mentioned model optimization method based on transfer learning may further include steps S810 and S820, which are described in detail below: Step S810: Perform image enhancement processing on the standardized source domain data to obtain enhanced source domain data. Image enhancement processing includes one or more of rotation, flipping, and cropping. Step S820: Construct a training dataset for the target domain model based on the enhanced source domain data, and train the target domain model based on the training dataset.

[0062] For example, image data from fields such as cardiovascular diseases and neurological diseases are collected as source domain data, while image data related to kidney diseases are acquired as target domain data. Since the source domain data is image data, it needs to be normalized to unify the pixel values ​​to a preset range, such as 0 to 1, to eliminate the influence of differences in pixel value ranges from different source image data. Next, the image data is standardized within the preset range to ensure the data has a mean of 0 and a standard deviation of 1, reducing data variance and improving subsequent processing results. After standardization, image enhancement processing is performed on the standardized source domain data. Image enhancement processing includes one or more of rotation, flipping, and cropping. Rotation can simulate images taken from different angles, such as rotating the image by 90, 180, or 270 degrees; flipping operations include horizontal and vertical flipping, which can increase data diversity; cropping operations extract a portion of the image to simulate different imaging ranges. Through these enhancement operations, more image data with variations but retaining key pathological features can be generated, resulting in enhanced source domain data.

[0063] Next, a training dataset for the target domain model is constructed based on the enhanced source domain data. Since the enhanced source domain data is associated with the pathological features of kidney diseases, this data can provide rich feature information for the target domain model. When constructing the training dataset, the enhanced source domain data is appropriately labeled and organized according to the needs of kidney disease diagnosis tasks, ensuring it meets the input requirements of the target domain model. Then, the target domain model can be trained based on the constructed training dataset. During training, the target domain model learns the feature representations in the enhanced source domain data; these features have significant reference value for the diagnosis of kidney diseases. Through continuous iterative optimization, the target domain model gradually masters the ability to extract effective features from image data and make accurate diagnoses, thereby improving the model's diagnostic accuracy for various pathological conditions of kidney diseases. Simultaneously, because knowledge is transferred from mature models in other medical fields, the feature extraction layer parameters of the source domain model are transferred to the large pathological model, and training with the enhanced source domain data can further accelerate the model's convergence speed and improve its performance.

[0064] In some embodiments of this application, image enhancement processing such as rotation, flipping, and cropping of standardized source domain data can expand data diversity. The target domain model training dataset constructed based on this can effectively improve the model's generalization ability and help train a more accurate and reliable target domain model for kidney disease diagnosis.

[0065] Furthermore, based on the above embodiments, please refer to... Figure 9 In one exemplary embodiment provided in this application, the above-described semantic alignment process of source domain data and target domain data, where the target domain data is pathological data of kidney disease, to obtain the semantic alignment result may further include steps S910 to S930, which are described in detail below: Step S910: Extract the semantic features corresponding to the source domain data, and semantically map the data semantic features to the source domain features; Step S920: Extract the target domain features corresponding to the target domain data; Step S930: Obtain semantic alignment results by using a preset semantic mapping function to combine the source domain features and the target domain features.

[0066] For example, for source domain data, we need to extract its corresponding semantic features. This can be achieved through feature extraction layers in deep learning models, such as convolutional neural networks (CNNs). CNNs can automatically learn local features in images, extracting semantic features at different levels, such as edges, textures, and shapes, through the sliding operation of multiple convolutional kernels. These semantic features reflect key information in the source domain data, such as the morphology of blood vessels and the location of lesions in cardiovascular disease images. After extracting the semantic features of the source domain data, these features are semantically mapped to obtain source domain features. The purpose of semantic mapping is to transform the source domain features into a unified semantic space for subsequent comparison and matching with target domain features.

[0067] Next, for the target domain data, namely kidney disease imaging data, the corresponding target domain features are extracted. Similarly, the feature extraction layer of a deep learning model is used to extract features related to kidney pathology, such as the morphology of kidney tumors and abnormal changes in kidney blood vessels, based on the characteristics of kidney disease images. Finally, the source domain features and target domain features are processed through a pre-defined semantic mapping function to obtain semantic alignment results. The pre-defined semantic mapping function is designed based on the semantic understanding of the source and target domain data; it can map source and target domain features to the same semantic space, so that features with similar semantics have similar representations in this space.

[0068] Optionally, in some embodiments of this application, a semantic mapping network can be constructed to align features in the source domain data with features in the renal pathology data using semantic parsing techniques. Specifically, the source domain data features can be defined as follows: The characteristic of the number of households in the target domain is represented as Then construct the semantic mapping function. To be used through semantic mapping functions Make: The semantic mapping function can be implemented using a multilayer perceptron (MLP), with the goal of minimizing the difference between features in the source and target domains. The mapping network is then trained using the following optimization objective: in, and These are the first and second domains of the source and target domains, respectively. The feature representation of each sample is then used. Furthermore, by training a semantic mapping network, edge features in the source domain are matched with cell boundary features in the target domain, achieving accurate knowledge transfer.

[0069] In some embodiments of this application, by extracting data features from the source domain and the target domain and using a preset semantic mapping function to achieve semantic alignment, the semantic relationship between the two can be accurately mined, laying a solid foundation for the subsequent effective transfer of source domain knowledge to improve the performance of the target domain model.

[0070] Figure 10 This is a block diagram illustrating a model optimization apparatus based on transfer learning, as shown in an exemplary embodiment of this application. This apparatus can also be applied to other exemplary implementation environments and specifically configured in other devices; this embodiment does not limit the implementation environment to which the apparatus is applicable.

[0071] like Figure 10 As shown, this exemplary transfer learning-based model optimization device includes: an acquisition module 1010 for acquiring transfer learning data, which includes source domain data associated with pathological features of kidney disease; a semantic alignment module 1020 for performing semantic alignment processing on the source domain data and target domain data, where the target domain data is pathological data of kidney disease, to obtain a semantic alignment result; a determination module 1030 for determining the feature distribution result corresponding to the target domain data, and determining the transfer weight corresponding to the source domain data based on the feature distribution result and the semantic alignment result; and an optimization module 1040 for optimizing the target domain model based on the transfer weight, the source domain data, and the target domain data to obtain an optimized target domain model.

[0072] According to one aspect of the embodiments of this application, the determination module 1030 is further configured to classify the target domain data and determine the scarce category data corresponding to the target domain based on the classification results; and determine the migration weight corresponding to the source domain data based on the scarcity coefficient corresponding to the scarce category data and the semantic alignment results.

[0073] According to one aspect of the embodiments of this application, the determining module 1030 is further configured to: determine the feature mapping between the target domain data and the source domain data to the same semantic space based on the semantic alignment result, and obtain a semantic mapping function; calculate the similarity between the target domain data and the source domain data based on the semantic mapping in the same semantic space; and determine the migration weight corresponding to the source domain data based on the similarity.

[0074] According to one aspect of the embodiments of this application, the determination module 1030 is further configured to determine the scarce source domain data corresponding to the source domain data and the scarce category data; if the scarcity coefficient is greater than a preset coefficient threshold, then determine the migration weight corresponding to the scarce source domain data based on the scarcity coefficient.

[0075] According to one aspect of the embodiments of this application, the determination module 1030 is further configured to: if the source domain data is image data, construct a feature extraction layer, wherein the feature extraction includes a convolutional network and a pooling layer corresponding to the convolutional neural network; pre-train the image data based on the feature extraction layer to obtain the classification result of the source domain data; and determine the scarce source domain data corresponding to the scarce category data based on the classification result.

[0076] According to one aspect of the embodiments of this application, the optimization module 1040 is further configured to: if the source domain data is image data, perform normalization processing on the image data to unify the pixel values ​​of the image data to a preset range; within the preset range, perform standardization processing on the image data to obtain standardized source domain data; and determine source domain data associated with the pathological characteristics of kidney disease based on the standardized source domain data.

[0077] According to one aspect of the embodiments of this application, the optimization module 1040 is further configured to perform image enhancement processing on the standardized source domain data to obtain enhanced source domain data, wherein the image enhancement processing includes one or more of rotation, flipping, and cropping; and construct a training dataset for a target domain model based on the enhanced source domain data, so as to train the target domain model based on the training dataset.

[0078] According to one aspect of the embodiments of this application, the semantic alignment module 1020 is further configured to: extract semantic features corresponding to source domain data, and semantically map the semantic features to source domain features; extract target domain features corresponding to target domain data; and obtain semantic alignment results by using a preset semantic mapping function to combine the source domain features and target domain features.

[0079] It should be noted that the model optimization device based on transfer learning provided in the above embodiments and the model optimization method based on transfer learning provided in the above embodiments belong to the same concept. The specific ways in which each module and unit performs operations have been described in detail in the method embodiments, and will not be repeated here. In practical applications, the model optimization device based on transfer learning provided in the above embodiments can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. This is not a limitation here.

[0080] Embodiments of this application also provide an electronic device, including: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the electronic device enables the transfer learning-based model optimization method provided in the above embodiments.

[0081] Figure 11A schematic diagram of a computer system suitable for implementing the embodiments of this application is shown. It should be noted that... Figure 11 The computer system 1100 of the electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0082] like Figure 11 As shown, the computer system 1100 includes a Central Processing Unit (CPU) 1101, which can perform various appropriate actions and processes, such as executing the methods described in the above embodiments, based on programs stored in Read-Only Memory (ROM) 1102 or programs loaded from storage portion 1108 into Random Access Memory (RAM) 1103. The RAM 1103 also stores various programs and data required for system operation. The CPU 1101, ROM 1102, and RAM 1103 are interconnected via a bus 1104. An Input / Output (I / O) interface 1105 is also connected to the bus 1104.

[0083] The following components are connected to I / O interface 1105: an input section 1106 including a keyboard, mouse, etc.; an output section 1107 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 1108 including a hard disk, etc.; and a communication section 1109 including a network interface card such as a LAN (Local Area Network) card, modem, etc. The communication section 1109 performs communication processing via a network such as the Internet. A drive 1110 is also connected to I / O interface 1105 as needed. Removable media 1111, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 1110 as needed so that computer programs read from them can be installed into storage section 1108 as needed.

[0084] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including a computer program for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 1109, and / or installed from removable medium 1111. When the computer program is executed by central processing unit (CPU) 1101, it performs various functions defined in the system of this application.

[0085] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable computer program. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The computer program contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0086] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. Each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0087] The units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0088] Another aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned transfer learning-based model optimization method. This computer-readable storage medium may be included in the electronic device described in the above embodiments, or it may exist independently and not incorporated into the electronic device.

[0089] Another aspect of this application provides a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the transfer learning-based model optimization method provided in the various embodiments described above.

[0090] The above description is merely a preferred exemplary embodiment of this application and is not intended to limit the implementation of this application. Those skilled in the art can easily make corresponding modifications or alterations based on the main concept and spirit of this application. Therefore, the scope of protection of this application should be determined by the scope of protection claimed in the claims.

Claims

1. A model optimization method based on transfer learning, characterized in that, The method comprises: obtaining to-be-transferred learning data, the to-be-transferred learning data comprising source domain data associated with kidney disease pathological features; performing semantic alignment processing on the source domain data and target domain data, the target domain data being kidney disease pathological data, to obtain a semantic alignment result; determining a feature distribution result corresponding to the target domain data, and determining a transfer weight corresponding to the source domain data based on the feature distribution result and the semantic alignment result; optimizing a target domain model based on the transfer weight, the source domain data and the target domain data, to obtain an optimized target domain model.

2. The method of claim 1, wherein, The determination of the feature distribution result corresponding to the target domain data and the determination of the transfer weight corresponding to the source domain data based on the feature distribution result and the semantic alignment result comprise: performing classification processing on the target domain data, and determining rare category data corresponding to the target domain based on a classification processing result; determining the transfer weight corresponding to the source domain data based on a rarity coefficient corresponding to the rare category data and the semantic alignment result.

3. The method of claim 2, wherein, The determination of the transfer weight corresponding to the source domain data based on the rarity coefficient corresponding to the rare category data and the semantic alignment result comprises: determining, based on the semantic alignment result, that feature mapping between the target domain data and the source domain data is to the same semantic space, to obtain a semantic mapping function; calculating, in the same semantic space, a similarity between the target domain data and the source domain data based on the semantic mapping function; determining the transfer weight corresponding to the source domain data based on the similarity.

4. The method of claim 2, wherein, The determination of the transfer weight corresponding to the source domain data based on the rarity coefficient corresponding to the rare category data and the semantic alignment result comprises: determining rare source domain data corresponding to the source domain data and the rare category data; if the rarity coefficient is greater than a preset coefficient threshold, determining the transfer weight corresponding to the rare source domain data based on the rarity coefficient.

5. The method of claim 4, wherein, The method further comprises: if the source domain data is image data, constructing a feature extraction layer, the feature extraction comprising a convolutional network and a pooling layer corresponding to the convolutional neural network; determining pre-training of the image data based on the feature extraction layer, to obtain a classification result of the source domain data; determining rare source domain data corresponding to the rare category data based on the classification result.

6. The method of claim 1, wherein, The method further comprises: if the source domain data is image data, performing normalization processing on the image data, to unify pixel values of the image data to a preset range; performing standardization processing on the image data in the preset range, to obtain standardized source domain data; determining source domain data associated with kidney disease pathological features based on the standardized source domain data.

7. The method of claim 6, wherein, The method further comprises: performing image enhancement processing on the standardized source domain data, to obtain enhanced source domain data, the image enhancement processing comprising one or more of rotation, flipping and cropping; constructing a training data set of the target domain model based on the enhanced source domain data, to train the target domain model based on the training data set.

8. The method of claim 1, wherein, The semantic alignment processing of the source domain data and target domain data is performed, the target domain data is kidney disease pathological data, and a semantic alignment result is obtained, including: Extracting semantic features corresponding to the source domain data, and performing semantic mapping of the semantic features to source domain features; Extracting target domain features corresponding to the target domain data; The source domain features and the target domain features are obtained through a preset semantic mapping function to obtain a semantic alignment result.

Images