Molecular subclass identification method for endometrial cancer pathological image

Abstract

The invention discloses an IECMS (International Electrotechnical Commission Mass Spectrometry) method for an endometrial cancer pathological image, and the method comprises the steps: constructing a pathological data set, carrying out the data preprocessing, extracting a region of interest of endometrial cancer according to a marker file, segmenting the region of interest into a plurality of instances, and carrying out the Vahadane color normalization technology for all instances, so as to ensure the color consistency of different instances; using a VMama feature extraction model to extract features of each instance; combining instance features in the region of interest ROI into packet-level features; converting the packet-level features into two-dimensional masked instance features through a dynamic mask strategy DMS module; learning correlation among instances through a context instance learning module CILM to obtain a feature fused with multiple kinds of context information, and converting the feature fused with multiple kinds of context information from a two-dimensional instance feature into a one-dimensional feature to obtain a packet-level mosaic feature; and classifying the packet-level mosaic features by using a multi-layer perceptron, and outputting a prediction result. According to the method, the context information can be better and more comprehensively understood.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image recognition, and in particular to a method for identifying molecular subclasses of endometrial cancer pathological images based on VMamba and context instance learning. Background Art

[0002] AB-MIL (Attention-based deep multiple instance learning) is a pathology image classification algorithm based on multiple instance learning. The algorithm uses a gated attention mechanism to aggregate instances in a bag. For each instance, the gated attention mechanism calculates the corresponding weights, which are normalized by the Softmax function to ensure that the sum of the weights is 1. The normalized weights reflect the degree of influence of each instance on the bag label. However, AB-MIL uses gated attention to assign weights to instances, only considering the morphological information of the instance itself, while ignoring the contextual information between instances. This hinders the capture of subtle and specific contextual information, which is important for comprehensive pathology analysis.

[0003] Zhang et al. (Clinical-grade endometrial cancer detection system via whole-slide images using deep learning) proposed a deep learning-based whole-slide image (WSIs) classification system for endometrial cancer (EC). The system is based on DeepLab v3 and ResNet 50, and introduces void space pyramid pooling with different expansion rates to enhance multi-scale detection capabilities. Zhang et al. used deep learning methods to classify EC, but only distinguished between cancer and non-cancer (binary classification) without further distinguishing EC molecular subclasses. However, EC molecular subclassification is difficult because the differences in cell morphology are subtle and the tumor microenvironment is complex.

[0004] High-quality EC WSIs are very scarce, which brings a great risk of overfitting during model training. The scarcity of data can cause the model to be over-specialized on the limited training data, resulting in poor generalization to new, unseen data. Summary of the invention

[0005] The purpose of the present invention is to provide a method for identifying molecular subclasses of endometrial carcinoma pathological images, which can accurately identify molecular subclasses by analyzing WSIs of EC, thereby achieving molecular subclass classification of EC.

[0006] The technical solution adopted by the present invention is:

[0007] A method for identifying molecular subtypes of endometrial cancer pathological images (IECMS) comprises the following steps:

[0008] Step S1: Construct a pathology dataset and perform data preprocessing. Extract the region of interest of endometrial cancer EC according to the label file and segment it into several instances. i,j , and apply Vahadane color normalization technology to all instances to ensure color consistency between different instances;

[0009] Step S2: Extract features from instances: Use VMamba feature extraction model to extract each instance s i,j The characteristics are expressed as E i,j ;

[0010] Step S3: Constructing package-level features: Constructing instance features E in the region of interest ROI i,j Combined into a package-level feature, denoted as X i ={E i,j};

[0011] Step S4: Context instance aggregation and classification: The package-level feature X i The dynamic mask strategy DMS module is used to convert the instance features into two-dimensional masked features. The context instance learning module CILM learns the correlation between instances to obtain features that integrate multiple context information. The features that integrate multiple context information are then converted from two-dimensional instance features to one-dimensional features to obtain package-level mosaic features. The package-level mosaic features are classified using a multi-layer perceptron and the prediction results are output.

[0012] Furthermore, step S1 specifically includes the following steps:

[0013] Step S1-1: Construct a data set D = {S1, Y1), (S2, Y2), ..., (S i ,Y i ),…,(S N ,Y N )}, N represents the number of WSIs in the dataset, S i represents the i-th WSI in the data set, Y i represents the corresponding label of the i-th WSI.

[0014] Step S1-2: Extract the region of interest ROI of endometrial cancer EC as a package according to the label file, represented by R i ; Package R iSplit into multiple small pieces i,j , represented by {s i,j |j=1,2,…,L}, L represents R i The number of image blocks in each small block s i,j is called an instance, s i,j Representative package R i The jth instance in ;

[0015] Step S1-3: Apply Vahadane color normalization technique to all instances to ensure color consistency between different instances.

[0016] Furthermore, the VMamba feature extraction model used in step S2 is divided into at least two stages, each stage includes stacking of more than two VSS blocks, and except for the first stage, the other stages first pass through a downsampling layer and then stack more than two VSS blocks.

[0017] Furthermore, the VSS block includes a first normalization layer, a first linear layer and a 2D depth-separable convolution layer connected in sequence. The input of the first normalization layer is connected to the original input. The output of the 2D depth-separable convolution layer is connected to the SS2D module via a SiLU activation function. The SS2D module scans the image block from four directions to efficiently aggregate global information. The output of the SS2D module is connected to the second normalization layer and the second linear layer in sequence. The output of the second linear layer is jump-connected to the original input; the output of the jump connection is connected to the third normalization layer and the FFN layer in sequence (the FFN layer is actually a linear transformation layer used to complete the dimensional transformation from input data to output data); the output of the FFN layer is added to the output of the jump connection as the output of the VSS block.

[0018] Specifically, the input data passes through a normalization layer and then is sent to a linear layer. The output of the linear layer passes through a 2D depth-separable convolution layer and then passes through the SiLU activation function. The activated data flows into the SS2D module, and its output passes through a normalization layer and then enters a linear layer. The output of the linear layer is jump-connected with the original input. At the same time, the result of the jump connection is sent to a normalization layer and a FFN, and their outputs are added to the results of the aforementioned jump connection. In this way, the VSS block can merge multiple different information processing paths to effectively extract and fuse rich features in visual data. Among them, the SS2D module specially designed for 2D data processing is the core. It can scan image blocks from four directions and efficiently aggregate global information, thereby enhancing the model's ability to recognize cell morphological differences and learn tumor microenvironment characteristics, effectively solving the limitations of traditional convolution methods.

[0019] Furthermore, step S4 specifically includes the following steps:

[0020] Step S4-1: Generate masked features and convert the packet-level features X i The instance features converted into two-dimensional masks by the dynamic mask strategy DMS module are represented as M i .

[0021] Step S4-2: The correlation between instances is learned through the context instance learning module CILM, and then the two-dimensional instance features are converted into one-dimensional features after a 1×1 standard convolution layer to obtain the package-level mosaic features.

[0022] Step S4-3: Predict the packet results, use a multi-layer perceptron to classify the packet-level mosaic features, and output the prediction results

[0023] Furthermore, the dynamic mask strategy (DMS) module generates more diverse relationships between instances by dynamically adjusting the mask ratio, which helps the model fully learn the relationships between instances, avoid overfitting, and improve the robustness and generalization ability of the model.

[0024] Specifically, the dynamic mask strategy (DMS) module is integrated into the IECMS method to enhance the model's understanding of the relationship between instances. DMS dynamically adjusts the mask ratio, and its core is to gradually increase the mask ratio instead of using a fixed mask ratio. This operation helps the model capture and encode the spatial relationship between instances, and combines convolution operations to fuse local and global information.

[0025] Furthermore, the specific steps of the context instance learning module CILM learning the correlation between instances in step S4-2 are as follows:

[0026] Step S4-2-1, define the key K (Key), query Q (Query) and value V (Value) as K = M i , Q=M i and V = M i W v , where M i is a parameter for the package-level feature X i The result obtained after the dynamic mask strategy DMS, W v is the weight of the convolution v;

[0027] Step S4-2-2, key K undergoes a specific k×k group of convolution operations to generate static context instance features C 1 ; Static contextualized instance feature C 1 The contextualized attention matrix H is obtained by joint learning with the query Q; the corresponding expression is as follows:

[0028] H=[C 1 ,Q]W ρ W σ , (4)

[0029] Specifically, the static contextualized instance feature C 1 It not only contains the contextual information of each adjacent instance, but also constructs the contextual attention matrix H, which further strengthens the connection between instances. 1 It is learned jointly with the query Q, rather than the traditional independent operation of query-key pairs, which ensures that CLIM can understand the relationship between instances more comprehensively.

[0030] Step S4-2-3, the value V (Value) and the information of the attention matrix H are fused through the self-attention mechanism to capture the dynamic feature interaction between instances to obtain the dynamic context instance feature C 2 , thereby enhancing the flexibility and generalization ability of the model; dynamic context instance feature C 2 The expression is as follows:

[0031] C 2 =VΘH, (5)

[0032] Step S4-2-4, passing the C through the instance channel selection module ICSM 1 and C 2 By adding them together, we can obtain feature F that integrates multiple context information. Feature F can fully reflect the complex relationship between instances and provide a richer and more accurate basis for subsequent classification decisions.

[0033] The instance channel selection module (ICSM) automatically identifies and emphasizes important instance features while suppressing irrelevant or redundant instance features, thereby generating a more focused and efficient overall feature representation.

[0034] The present invention adopts the above technical solutions and uses VMamba as a feature extraction module to effectively model global information and capture tumor microenvironment information. At the same time, a method of EC molecular subclassification is proposed to effectively distinguish four highly similar EC molecular subclasses. A data masking strategy is used to prevent network overfitting and encourage the model to learn more robust generalization features; thereby strengthening the relationship between learning instances and improving overall performance. Contextual learning is used to effectively capture the relationship between instances, so as to more comprehensively understand the contextual information in WSIs. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments;

[0036] Figure 1 A schematic diagram of the principle architecture of a method for identifying molecular subclasses of endometrial cancer pathological images according to the present invention;

[0037] Figure 2 It is a structural schematic diagram of the VSS block of the present invention;

[0038] Figure 3 It is a schematic diagram of the principle structure of the dynamic mask strategy of the present invention;

[0039] Figure 4 Schematic diagram of the structure of the channel selection module of the present invention. Implementation

[0041] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0042] like Figures 1 to 4 As shown in one, the present invention discloses a method for identifying molecular subclasses of endometrial cancer pathological images, which comprises the following steps:

[0043] Step S1: Construct a pathology dataset and perform data preprocessing. Extract the region of interest of endometrial cancer EC according to the label file and segment it into several instances. i,j , and apply Vahadane color normalization technology to all instances to ensure color consistency between different instances;

[0044] Step S2: Instance feature extraction: Use VMamba feature extraction model to extract each instance s i,j The characteristics are expressed as E i,j ;

[0045] Step S3: Constructing package-level features: Constructing instance features E in the region of interest ROI i,j Combined into a package-level feature, denoted as X i ={E i,j};

[0046] Step S4: Context instance aggregation and classification: The package-level feature X i The dynamic mask strategy DMS module is used to convert the instance features into two-dimensional masked features. The context instance learning module CILM learns the correlation between instances to obtain features that integrate multiple context information. The features that integrate multiple context information are then converted from two-dimensional instance features to one-dimensional features to obtain package-level mosaic features. The package-level mosaic features are classified using a multi-layer perceptron and the prediction results are output.

[0047] Furthermore, step S1 specifically includes the following steps:

[0048] Step S1-1: Construct a data set D = {S1, Y1), (S2, Y2), ..., (S i ,Y i ),…,(S N ,Y N )}, N represents the number of WSIs in the dataset, Si represents the i-th WSI in the data set, Y i represents the corresponding label of the i-th WSI.

[0049] Step S1-2: Extract the region of interest ROI of endometrial cancer EC as a package according to the label file, represented by R i ; Package R i Split into multiple small pieces i,j , represented by {s i,j |j=1,2,…,L}, L represents R i The number of image blocks in each small block s i,j is called an instance, s i,j Representative package R i The jth instance in ;

[0050] Step S1-3: Apply Vahadane color normalization technique to all instances to ensure color consistency between different instances.

[0051] Furthermore, the VMamba feature extraction model used in step S2 is divided into at least two stages, each stage includes stacking of more than two VSS blocks, and except for the first stage, the other stages first pass through a downsampling layer and then stack more than two VSS blocks.

[0052] Furthermore, the VSS block includes a first normalization layer, a first linear layer and a 2D depth-separable convolution layer connected in sequence. The input of the first normalization layer is connected to the original input. The output of the 2D depth-separable convolution layer is connected to the SS2D module via a SiLU activation function. The SS2D module scans the image block from four directions to efficiently aggregate global information. The output of the SS2D module is connected to the second normalization layer and the second linear layer in sequence. The output of the second linear layer is jump-connected to the original input; the output of the jump connection is connected to the third normalization layer and the FFN layer in sequence (the FFN layer is actually a linear transformation layer used to complete the dimensional transformation from input data to output data); the output of the FFN layer is added to the output of the jump connection as the output of the VSS block.

[0053] Specifically, the input data passes through a normalization layer and then is sent to a linear layer. The output of the linear layer passes through a 2D depth-separable convolution layer and then passes through the SiLU activation function. The activated data flows into the SS2D module, and its output passes through a normalization layer and then enters a linear layer. The output of the linear layer is jump-connected with the original input. At the same time, the result of the jump connection is sent to a normalization layer and a FFN, and their outputs are added to the results of the aforementioned jump connection. In this way, the VSS block can merge multiple different information processing paths to effectively extract and fuse rich features in visual data. Among them, the SS2D module specially designed for 2D data processing is the core. It can scan image blocks from four directions and efficiently aggregate global information, thereby enhancing the model's ability to recognize cell morphological differences and learn tumor microenvironment characteristics, effectively solving the limitations of traditional convolution methods.

[0054] Furthermore, step S4 specifically includes the following steps:

[0055] Step S4-1: Generate masked features and convert the packet-level features X i The instance features converted into two-dimensional masks by the dynamic mask strategy DMS module are represented as M i .

[0056] Step S4-2: The correlation between instances is learned through the context instance learning module CILM, and then the two-dimensional instance features are converted into one-dimensional features after a 1×1 standard convolution layer to obtain the package-level mosaic features.

[0057] Step S4-3: Predict the packet results, use a multi-layer perceptron to classify the packet-level mosaic features, and output the prediction results

[0058] Furthermore, the dynamic mask strategy (DMS) module generates more diverse relationships between instances by dynamically adjusting the mask ratio, which helps the model fully learn the relationships between instances, avoid overfitting, and improve the robustness and generalization ability of the model.

[0059] Specifically, the dynamic mask strategy (DMS) module is integrated into the IECMS method to enhance the model's understanding of the relationship between instances. DMS dynamically adjusts the mask ratio, and its core is to gradually increase the mask ratio instead of using a fixed mask ratio. This operation helps the model capture and encode the spatial relationship between instances, and combines convolution operations to fuse local and global information.

[0060] Furthermore, the specific steps of the context instance learning module CILM learning the correlation between instances in step S4-2 are as follows:

[0061] Step S4-2-1, define the key K (Key), query Q (Query) and value V (Value) as K = M i , Q=M i and V = M i W v , where M i is a parameter for the package-level feature X i The result obtained after the dynamic mask strategy DMS, W v is the weight of the convolution v;

[0062] Step S4-2-2, key K undergoes a specific k×k group of convolution operations to generate contextualized instance features C 1 Contextualized instance feature C 1 The contextualized attention matrix H is obtained by joint learning with the query Q; the corresponding expression is as follows:

[0063] H=[C 1 ,Q]W ρ W σ , (4)

[0064] Specifically, the contextualized instance feature C 1 It not only contains the contextual information of each adjacent instance, but also constructs the contextual attention matrix H, which further strengthens the connection between instances. 1 It is learned jointly with the query Q, rather than the traditional independent operation of query-key pairs, which ensures that CLIM can understand the relationship between instances more comprehensively.

[0065] Step S4-2-3, the value V (Value) and the information of the attention matrix H are fused through the self-attention mechanism to capture the dynamic feature interaction between instances to obtain the dynamic context instance feature C 2 , thereby enhancing the flexibility and generalization ability of the model; dynamic context instance feature C 2 The expression is as follows:

[0066] C 2 =VΘH, (5)

[0067] Step S4-2-4, passing the C through the instance channel selection module ICSM 1 and C 2 By adding them together, we can obtain feature F that integrates multiple context information. Feature F can fully reflect the complex relationship between instances and provide a richer and more accurate basis for subsequent classification decisions.

[0068] The instance channel selection module (ICSM) automatically identifies and emphasizes important instance features while suppressing irrelevant or redundant instance features, thereby generating a more focused and efficient overall feature representation.

[0069] The specific principle of the present invention is described in detail below:

[0070] like Figure 1 As shown, the method of the present invention consists of three parts: data processing, feature extraction, and context instance aggregation. Specifically, the present invention takes into account the subtle differences between different molecular subclasses and the complexity of the tumor microenvironment. The classic ResNet50 architecture is not sufficient to fully extract and distinguish these features. Therefore, the introduction of VMamba as a feature extraction module can significantly improve the performance of the model in the EC molecular subclass identification task, especially when dealing with samples with high similarity and complex tumor microenvironment. The present invention targets the problems of few high-quality data of WSIs and imbalanced molecular subclass data. Existing networks often have overfitting problems during the training process, resulting in poor generalization performance. Therefore, we designed a plug-and-play dynamic masking strategy (DMS) to simulate more instance relationships and enhance the diversity and complexity of the data. The present invention takes into account that the classification of EC molecular subclasses cannot only focus on the morphological features of the instances, but also the relationships between the instances. In view of the fact that the existing methods only pay attention to the importance of morphological features and ignore the impact of the relationships between instances on the classification performance, a contextual instance learning module (CILM) is designed.

[0071] In order to efficiently extract cell morphology and tumor microenvironment features, we introduced a feature extraction method based on a visual state space model. This method is suitable for analyzing complex and subtly different EC pathological images. As an advanced visual backbone network, the core advantage of VMamba is that it can use the visual state space for efficient visual representation learning. Specifically, VMamba is able to capture the global contextual information in the image by building a comprehensive visual state space model. This global context-awareness capability helps to extract richer features from complex pathological images, which is crucial for identifying and distinguishing subtle molecular subtypes. The package-level features are defined as follows:

[0072] E i,j =VM(s i,j ) (1)

[0073] X i =VM(R i )={E i,j} (2)

[0074] Among them, R i represents the area of ​​the EC of the ith WSI; each R i is divided into several image blocks, denoted as {s i,j|j=1,2,…,L}, where these small blocks are regarded as instances and r i Considered as a package, L stands for R i The number of image blocks in s i,j Representative package R i The jth instance in ; VM represents the feature extraction operation based on VMamba extractor; E i,j Representative package R i The instance-level features of the jth instance in {E i,j} indicates package R i All instance-level feature sets in X i Representative package R i Package-level features.

[0075] The VMamba network can be divided into multiple stages, and except for the first stage, each stage consists of a downsampling layer followed by a series of VSS blocks stacked. Figure 2 As shown in the figure, the structure of VSS, first, the input data passes through a normalization layer and then is sent to a linear layer. The output of the linear layer passes through a 2D depth-separable convolution layer and then passes through the SiLU activation function. The activated data flows into the SS2D module, and its output passes through a normalization layer and then enters a linear layer. The output of the linear layer is jump-connected with the original input. At the same time, the result of the jump connection is sent to a normalization layer and a FFN, and their outputs are added to the results of the aforementioned jump connection. In this way, the VSS block can merge multiple different information processing paths and effectively extract and fuse rich features in visual data. Among them, the SS2D module specially designed for 2D data processing is the core. It can scan image blocks from four directions and efficiently aggregate global information, thereby enhancing the model's ability to recognize cell morphological differences and learn tumor microenvironment characteristics, effectively solving the limitations of traditional convolution methods.

[0076] Based on the dynamic mask strategy module (DMS): In medical image analysis, high-quality pathological image resources are often scarce, which makes deep learning models prone to overfitting, especially when dealing with complex relationships between instances, the generalization performance of the model is limited. This method proposes a dynamic mask strategy (DMS) such as Figure 3 As shown in the figure, as a pluggable module, it is integrated into the IECMS method to enhance the model's understanding of the relationship between instances. DMS dynamically adjusts the mask ratio. The core is that the mask ratio gradually increases instead of using a fixed mask ratio. This operation helps the model capture and encode the spatial relationship between instances, and combines convolution operations to fuse local and global information. The gray in the figure represents the masked instance. The packet-level feature M after masking i Expressed as:

[0077] M i =DMS (X i ) (3)

[0078] Among them, M i Indicates the masked packet-level features, D MS Represents an implementation of a dynamic masking strategy.

[0079] Contextual Instance Learning Module (CILM): The CLIM module aims to deeply explore the correlation between instances in a package and significantly improve the classification accuracy and generalization performance of the model by enhancing the expression ability of instance features. Figure 4 As shown. Define the key K (Key), query Q (Query) and value V (Value) as K = M i , Q=M i and V = M i W v , where M i is a parameter for the package-level feature X i The result obtained after the dynamic mask strategy DMS is v is the weight of the convolution v; the key K undergoes a specific k×k group of convolution operations to produce a static context instance feature C 1 Specifically, the static context instance feature C 1 It not only contains the contextual information of each adjacent instance, but also constructs the contextual attention matrix H, which further strengthens the connection between instances. 1 It is learned jointly with the query Q, rather than the traditional independent operation of query-key pairs, which ensures that CLIM can understand the relationship between instances more comprehensively.

[0080] Contextualized instance feature C 1 The contextualized attention matrix H is obtained by joint learning with the query Q; the corresponding expression is as follows:

[0081] H=[C 1 ,Q]W ρ W σ , (4)

[0082] Then, the information of the value V (Value) and the attention matrix H is fused through the self-attention mechanism to capture the dynamic feature interaction between instances and obtain the dynamic context instance feature C 2 , thereby enhancing the flexibility and generalization ability of the model; dynamic context instance feature C 2 The expression is as follows:

[0083] C 2 =VΘH, (5)

[0084] In order to improve the discriminative power of features, the present invention introduces an instance channel selection module (ICSM) to automatically identify and emphasize important instance features while suppressing irrelevant or redundant instance features, thereby generating a more focused and efficient overall feature representation. 1 and C 2 By adding them together, we can obtain the feature F that integrates multiple context information. Feature F can fully reflect the complex relationship between instances and provide a richer and more accurate basis for subsequent classification decisions. Finally, the recognition result is obtained through a 1×1 standard convolutional layer and a multi-layer perceptron.

[0085] F=I CSM (C 1 )+I CSM (C 2 ), (6)

[0086]

[0087] Among them, I CSM Represents the implementation of the instance channel selection module; MLP represents a multi-layer perceptron, and CONV represents a 1×1 standard convolutional layer.

[0088] The present invention has the following advantages over the prior art: 1) The feature extraction networks of the AB-MIL and other methods in the prior art are all based on convolution methods, which cannot establish global information well. The similarity between the molecular subtypes of EC is very high, which makes the classification of the molecular subtypes of EC more difficult. The difference in cell morphology of EC is very subtle, and the tumor microenvironment is extremely complex. The present invention integrates VMamba into a method specifically used for the classification of molecular subtypes of EC for the first time, and proposes a new IECMS method, which can effectively distinguish four molecular subtypes with high similarity, not only can capture the difference in cell morphology, but also can handle complex tumor microenvironment characteristics. 2) Histopathological images are very scarce, which increases the risk of model overfitting during training. The present invention proposes a dynamic mask strategy DMS, which gradually increases the proportion of masked data during training, so that the model can discover more information about instances and their relationships. This strategy helps prevent overfitting during training and enables the model to learn more robust and generalized features. 3) Traditional MIL methods usually assume that instances are independent and identically distributed, focus on their morphological features, and ignore the contextual information of instances. The present invention proposes a new contextual instance learning module (CILM), which can effectively capture the relationships between instances, thereby achieving a better and more comprehensive understanding of the contextual information in WSIs.

[0089] Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. In the absence of conflict, the embodiments in the present application and the features in the embodiments can be combined with each other. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians of the art without making creative work are within the scope of protection of the present application.

Claims

1. A method for identifying molecular subclasses of endometrial cancer pathological images, characterized by: It includes the following steps: Step S1: Construct a pathology dataset and perform data preprocessing. Extract the region of interest of endometrial cancer EC according to the label file and segment it into several instances. i,j , and apply Vahadane color normalization technology to all instances to ensure color consistency between different instances; Step S2: Extract features from instances: Use VMamba feature extraction model to extract each instance s i,j The characteristics are expressed as E i,j ; Step S3: Constructing package-level features: Constructing instance features E in the region of interest ROI i,j Combined into a package-level feature, denoted as X i ={E i,j }; Step S4: Context instance aggregation and classification: The package-level feature X i The instance features are converted into two-dimensional masks through the dynamic mask strategy DMS module; The context instance learning module CILM learns the correlation between instances to obtain features that integrate multiple context information, and then converts the features that integrate multiple context information from two-dimensional instance features to one-dimensional features to obtain package-level mosaic features; Use a multi-layer perceptron to classify packet-level mosaic features and output prediction results 2. The method for identifying molecular subclasses of endometrial cancer pathological images according to claim 1, characterized in that: Step S1 specifically includes the following steps: Step S1-1: Construct a data set D = {S1, Y1), (S2, Y2), ..., (S i ,Y i ),…,(S N ,Y N )}, N represents the number of WSIs in the dataset, S i represents the i-th WSI in the data set, Y i represents the corresponding label of the i-th WSI; Step S1-2: Extract the region of interest (ROI) of endometrial cancer EC as a package according to the label file, represented by R i ; Package R i Split into multiple small pieces i,j , represented by {s i,j |j=1,2,…,L}, L represents R i The number of image blocks in each small block s i,j is called an instance, s i,j Representative package R i The jth instance in ; Step S1-3: Apply Vahadane color normalization technique to all instances to ensure color consistency between different instances.

3. The method for identifying molecular subclasses of endometrial cancer pathological images according to claim 1, characterized in that: The VMamba feature extraction model used in step S2 is divided into at least two stages, each stage includes stacking of more than two VSS blocks, and except for the first stage, the other stages first pass through a downsampling layer and then stack more than two VSS blocks.

4. The method for identifying molecular subclasses of endometrial cancer pathological images according to claim 3, characterized in that: The VSS block includes a first normalization layer, a first linear layer and a 2D depth-separable convolution layer connected in sequence. The input of the first normalization layer is connected to the original input. The output of the 2D depth-separable convolution layer is connected to the SS2D module via a SiLU activation function. The SS2D module scans the image blocks from four directions to efficiently aggregate global information. The output of the SS2D module is connected to the second normalization layer and the second linear layer in sequence. The output of the second linear layer is jump-connected to the original input; the output of the jump connection is connected to the third normalization layer and the FFN layer in sequence. The FFN layer is used to complete the dimensional transformation from input data to output data; the output of the FFN layer is added to the output of the jump connection as the output of the VSS block.

5. The method for identifying molecular subclasses of endometrial cancer pathological images according to claim 1, characterized in that: The dynamic mask strategy DMS module generates more diverse relationships between instances by dynamically adjusting the mask ratio, which helps the model fully learn the relationships between instances, avoid overfitting, and improve the robustness and generalization ability of the model.

6. The method for identifying molecular subclasses of endometrial cancer pathological images according to claim 1, characterized in that: Step S4 The specific steps include: Step S4-1: Generate masked features and convert the packet-level features X i The instance features converted into two-dimensional masks by the dynamic mask strategy DMS module are represented as M i ; Step S4-2: The correlation between instances is learned through the context instance learning module CILM to obtain a feature F that integrates multiple context information. The feature F is then converted into a one-dimensional feature after passing through a 1×1 standard convolution layer to obtain a packet-level mosaic feature. Step S4-3: Predict the packet results, use a multi-layer perceptron to classify the packet-level mosaic features, and output the prediction results 7. The method for identifying molecular subclasses of endometrial cancer pathological images according to claim 6, characterized in that: The specific steps of the context instance learning module CILM learning the correlation between instances in step S4-2 are as follows: Step S4-2-1, define the key K, query Q and value V as K=M respectively i , Q=M i and V = M i W v , where M i is a parameter for the package-level feature X i The result obtained after the dynamic mask strategy DMS, W v is the weight of the convolution v; Step S4-2-2, key K undergoes a specific k×k group of convolution operations to generate static context instance features C 1 ; Static context instance feature C 1 The contextual attention matrix H is obtained by joint learning with the query Q. The corresponding expression is as follows: H=[C 1 ,Q]W ρ W σ , (4) Step S4-2-3, the information of the value V and the attention matrix V is fused through the self-attention mechanism to capture the dynamic feature interaction between instances to obtain the dynamic context instance feature C 2 , dynamic context instance feature C 2 The expression is as follows: C 2 =VΘH, (5) Step S4-2-4, the static context instance feature C passed through the instance channel selection module ICSM 1 and dynamic context instance feature C 2 Add them together to obtain the feature F that integrates multiple context information.

Images