An endometrial carcinoma molecular typing method based on cascaded multiple-instance learning
By employing ROI-driven instance construction and cascaded multi-instance learning methods, the problems of sparse instance localization and mixed multi-class discrimination in endometrial cancer subtyping were solved, achieving efficient and accurate molecular subtyping results that meet clinical needs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FUJIAN UNIV OF TECH
- Filing Date
- 2026-02-26
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies struggle to accurately locate sparse key cases in endometrial cancer classification. Parallel classification of multiple categories can lead to conflicting criteria and is difficult to align with clinical classification procedures, resulting in low classification efficiency and insufficient accuracy.
We employ ROI-driven instance construction, PatchVMamba feature encoding, and cascaded pair decision strategies. We define the region of interest using a binary mask, combine the PatchVMamba instance encoder and attention mechanism for feature extraction and weighted aggregation, and decompose the four-class classification task into a serial binary classification link, using a unified threshold for cascaded judgment.
It improves the accuracy of capturing subtle morphological features, enhances the interpretability and stability of typing results, reduces the fluctuation of typing accuracy under different scanning devices and staining batches, and conforms to clinical interpretation logic.
Smart Images
Figure CN122115974A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computational pathology and digital image processing, and in particular to a molecular subtyping method for endometrial cancer based on cascaded multi-instance learning. Background Technology
[0002] Currently, deep learning-based pathological image analysis methods are mainly divided into two categories: traditional machine learning algorithms based on handcrafted features and neural network models based on weakly supervised multi-instance learning. The following analysis examines the technical characteristics and shortcomings of representative solutions.
[0003] Existing multi-instance learning schemes using Transformers, represented by TransMIL-like methods, treat the entire WSI as a set of instances composed of image slices. They first extract instance-level deep features using a convolutional neural network, then model long-range dependencies between instances using a Transformer encoder, and finally aggregate slice-level features for classification using an attention mechanism. For example, TransMIL captures spatial relationships between instances through a self-attention structure, achieving high accuracy in breast cancer pathology image classification. However, effective information in WSIs is sparse and unevenly distributed, and the global attention of the Transformer is easily scattered by a large number of non-critical slices, making it difficult for the model to focus on subtle morphological cues related to molecular subtyping. When dealing with tasks with subtle morphological differences between subtypes, such as endometrial cancer, multi-class parallel classification can easily lead to mixed discrimination criteria, especially for subtypes like POLEmut that lack clear morphological markers, resulting in blurred classification boundaries. The quadratic complexity of the Transformer is inefficient for processing large-scale WSI data and is difficult to adapt to real-time clinical needs.
[0004] Furthermore, existing multi-instance learning schemes that integrate neighborhood structures or region re-embedding are represented by methods like CAMIL and RRTMIL. These schemes introduce neighborhood priors or region-level structural information during instance aggregation, improving feature discriminative power through contextual constraints. For example, CAMIL enhances aggregated representations by utilizing local spatial relationships between instances, while RRTMIL improves weakly supervised learning performance through region reorganization and re-encoding. However, in WSI, the proportion of background and non-tumor tissue is high, and region re-embedding still cannot completely avoid the dilution of global features by weakly informative instances, especially for diseases like endometrial cancer where lesion areas are sparse, making it easy for crucial evidence to be submerged. Most methods require pixel-level or region-level annotations to assist training, but obtaining fine-grained annotations in endometrial cancer molecular subtyping tasks is costly, resulting in insufficient model prediction stability and performance fluctuations under conditions with only slice-level annotations. At the same time, parallel multi-classification structures cause discriminative features of different subtypes to interfere with each other in the same aggregation space, making it difficult to reproduce the clinical "stepwise elimination" interpretation logic.
[0005] In summary, while existing methods have made progress in the analysis of pathological images of the breast and stomach, the morphological differences between subtypes of endometrial cancer are more subtle than those of breast cancer, and current methods lack the ability to accurately locate sparse key instances. Parallel classification mechanisms are prone to prediction instability when class similarity is high, and research on cascaded decision structures is still insufficient. When the WSI dataset is large, existing aggregation methods struggle to balance accuracy and efficiency and are not aligned with the subtyping workflow of clinical guidelines for endometrial cancer. Summary of the Invention
[0006] The purpose of this invention is to provide a molecular subtyping method for endometrial cancer based on cascaded multi-instance learning. Through ROI-driven instance construction, PatchVMamba feature encoding, and cascaded one-to-one decision-making strategies, the subtyping process is made more accurate, efficient, and clinically adaptable.
[0007] The technical solution adopted in this invention is:
[0008] A molecular subtyping method for endometrial cancer based on cascaded multi-instance learning includes the following steps:
[0009] Instance construction: Preprocess the whole slice image of endometrial cancer to generate a binary mask of the tumor region, extract image patches within the region of interest based on the binary mask, and construct an instance set by staining and normalizing the image patches;
[0010] Feature extraction: The PatchVMamba instance encoder is used to extract features from each image patch in the instance set to obtain the instance feature vector of each image patch, and then the instance feature bag of each full slice image is obtained.
[0011] Weighted aggregation: An attention mechanism is used to perform weighted aggregation of the instance feature bags to obtain slice-level feature representations;
[0012] Subtyping results: The slice-level features are input into three cascaded binary classifiers to obtain the probability values of the three subtypes of endometrial cancer (EC). Based on the probability values of the three subtypes, a unified threshold is used to perform cascaded subtype classification of endometrial cancer to output the final subtyping results.
[0013] Furthermore, the instance construction steps are implemented as follows:
[0014] Building a dataset S i Represents the WSI of the i-th full slice image. The denoting label represents the corresponding genotype label of the WSI of the i-th full slice image, and N represents the number of full slice images in the dataset;
[0015] Generate a binary mask with the tumor region as the region of interest (ROI) for each whole-slice image. To preserve tissue within the region of interest (ROI) for dicing; binary mask. The expression is: (1); where 1 represents the tumor region and 0 represents the non-tumor region; , These are the height and width of the full slice image, respectively;
[0016] The tumor region is divided into fixed-size regions based on the binary mask. Non-overlapping blocks; as a feasible implementation, non-overlapping blocks of 256×256 pixels are used.
[0017] After performing Vahadane staining normalization on each block of the full-slice image, the normalized blocks are summed to obtain the normalized block set for each full-slice image. As the corresponding collection of instances;
[0018] Specifically, the expression for the normalization process is: (3); This represents the Vahadane normalization operator. For a given slice, Cut into pieces for reference. For normalized chunks;
[0019] Normalized slice set The expression is: (4); To normalize the cutting, This represents the number of slices per full-slice image. (2).
[0020] Furthermore, the implementation of feature extraction includes:
[0021] Linear normalization is performed on the normalized segments of the full-slice image: ,in For LinearNorm operations;
[0022] The instance feature vectors of the corresponding segments are obtained by performing feature transformation on the N stacked visual state space modules in the PatchVMamba instance encoder. ;
[0023] The update formula for each visual state space module is:
[0024] (6);
[0025] (7);
[0026] in, For the first Output features of each visual state space module For the first Output features of each visual state space module; It is a two-dimensional selective scan operator, one of the core state-space computation units of PatchVMamba; This is a depthwise convolution. LN is the activation function; LN is the layer normalization function, used for normalization stabilization training and numerical scaling within residual blocks. It is a feedforward network. For the first Intermediate features of a visual state space module;
[0027] Instance feature vector The expression is: ; The feature output by the last visual state space module; N is the number of visual state space modules;
[0028] The instance feature vectors of each slice of the full slice image are obtained to form the instance feature bag of the corresponding full slice image. .
[0029] Specifically, any full-slice image can be encoded accordingly; that is, the instance feature bag of the i-th full-slice image is represented as... , ; Let be the instance feature bag of the i-th full-slice image; Let be the number of slices in the i-th full slice image; For the j-th slice of the i-th full slice image WSI Instance feature vectors.
[0030] Furthermore, the implementation of weighted aggregation includes:
[0031] Calculate the attention weights for each instance feature vector. The formula is:
[0032] (13);
[0033] in, and For learnable parameters, Let L be the feature vector of the k-th instance; L is the number of instance feature vectors, and L is equal to the number of blocks in the corresponding full slice image. ;
[0034] We perform weighted aggregation on the instance feature vectors in the instance feature bag to obtain the slice-level feature representation H:
[0035] (12);
[0036] Furthermore, the slice-level feature representation H is used by a classifier to obtain the class probability. :
[0037] (14);
[0038] in, and For classifier parameters, This indicates the number of categories corresponding to the binary classifier.
[0039] Furthermore, the binary classifiers for the three subtype classifications are NSMP subtype binary classifiers. MMRd subtype binary classifier and p53abn subtype binary classifier .
[0040] Furthermore, the cascading judgment rule is as follows:
[0041] When the probability value of the NSMP subtype of the whole slice image If the value is greater than the threshold T, then the endometrial cancer in the current whole slice image is of the NSMP subtype;
[0042] When the probability value of the NSMP subtype of the whole slice image The probability value of MMRd subtype not greater than the threshold T If the value is greater than the threshold T, then the endometrial cancer in the current whole slice image is of the MMRd subtype;
[0043] When the probability value of the MMRd subtype of the whole slice image The probability value of the p53abn subtype is not greater than the threshold T. If the value is greater than the threshold T, then the endometrial cancer in the current whole slice image is of the p53abn subtype;
[0044] In other cases, the endometrial cancer in the current whole-slice image is classified as POLEmut. That is, if none of the three conditions are met, the endometrial cancer in the current whole-slice image is determined to be the POLEmut subtype.
[0045] Furthermore, as a feasible implementation, the threshold T is preset to 0.5; the cascade strategy can reduce the decision complexity of the prediction task and improve the overall classification performance of the genotype.
[0046] This invention, employing the above technical solutions, offers the following significant advantages over existing technologies: 1) By employing an ROI-driven instance construction strategy, this invention effectively defines the tumor region and performs standardized staining normalization, reducing interference from background tissue and staining differences at the source, thereby increasing the information density of the instance set. Combined with the long-range spatial context modeling capabilities of the PatchVMamba encoder, the accuracy of capturing subtle morphological features is improved, significantly enhancing the classification and recognition performance for POLEmut and other methods lacking clear morphological cues. 2) The cascaded structure decomposes the four-class classification task into a serial binary classification link, which better aligns with the clinical "stepwise exclusion" interpretation logic. The cascaded structure ensures high consistency between the classification results and pathology expert diagnoses, surpassing traditional parallel classification methods. Simultaneously, the step-by-step decision-making mechanism provides doctors with a clear decision-making path, greatly enhancing the interpretability of the results. 3) The classification accuracy of this invention exhibits minimal fluctuations under different scanning devices and staining batches, significantly lower than existing methods, demonstrating broad clinical application prospects. Attached Figure Description
[0047] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments;
[0048] Figure 1 This is a schematic diagram illustrating the principle architecture of a molecular subtyping method for endometrial cancer based on cascaded multi-instance learning according to the present invention.
[0049] Figure 2 This is a schematic diagram of the structure of the PatchVMamba encoder of the present invention. Detailed Implementation
[0050] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.
[0051] This invention discloses a molecular subtyping method for endometrial cancer based on cascaded multi-instance learning. It achieves a reasoning process consistent with the "stepwise exclusion" path of clinical subtyping through the synergy of three elements: Region of Interest (ROI) driven instance construction, PatchVMamba feature encoding, and a cascaded One-vs-Rest (OvR) discriminant strategy for multi-instance decision-making. This invention reduces the reliance on fine-grained annotations such as pixel-level and region-level labels while relying solely on slide-level labels. Furthermore, it weakens evidence coupling and decision boundary entanglement caused by multi-class parallel competition through staged filtering and progressively shrinking candidate space, thereby improving classification performance.
[0052] like Figure 1 As shown, this invention discloses a specific implementation of a molecular subtyping method for endometrial cancer based on cascaded multi-instance learning, including the following steps:
[0053] Instance construction: Preprocess the whole slice image of endometrial cancer to generate a binary mask of the tumor region, extract image patches within the region of interest based on the binary mask, and construct an instance set by staining and normalizing the image patches;
[0054] Feature extraction: The PatchVMamba instance encoder is used to extract features from each image patch in the instance set to obtain the instance feature vector of each image patch, and then the instance feature bag of each full slice image is obtained.
[0055] Weighted aggregation: An attention mechanism is used to perform weighted aggregation of the instance feature bags to obtain slice-level feature representations;
[0056] Subtyping results: The slice-level features are input into three cascaded binary classifiers to obtain the probability values of the three subtypes of endometrial cancer (EC). Based on the probability values of the three subtypes, a unified threshold is used to perform cascaded subtype classification of endometrial cancer to output the final subtyping results.
[0057] Furthermore, the instance construction steps are implemented as follows:
[0058] Building a dataset S i Represents the WSI of the i-th full slice image. The denoting label represents the corresponding genotype label of the WSI of the i-th full slice image, and N represents the number of full slice images in the dataset;
[0059] Generate a binary mask with the tumor region as the region of interest (ROI) for each whole-slice image. To preserve tissue within the region of interest (ROI) for dicing; binary mask. The expression is: (1); where 1 represents the tumor region and 0 represents the non-tumor region; in this embodiment , These are the height and width of the full slice image, respectively;
[0060] The tumor region was divided into fixed-size regions. Non-overlapping blocks; as a feasible implementation, non-overlapping blocks of 256×256 pixels are used.
[0061] After performing Vahadane staining normalization on each block of the full-slice image, the normalized blocks are summed to obtain the normalized block set for each full-slice image. As the corresponding collection of instances;
[0062] Specifically, the expression for the normalization process is: (3); This represents the Vahadane normalization operator. For a given slice, Cut into pieces for reference. For normalized chunks;
[0063] Normalized slice set The expression is:
[0064] (4);
[0065] in, For the j-th normalized block, This represents the number of slices per full-slice image. (2); In this embodiment , These represent the height and width of the region of interest in pixel coordinates, respectively. and A fixed height and width for a single slice.
[0066] The core function of the aforementioned binary mask M is to filter out the "tumor region" from the entire WSI, thus allowing subsequent slicing to be performed only within the regions where the mask is 1, preventing a large amount of background or non-target tissue from entering the instance set. Specifically, a binary mask for the tumor region is first generated, where "1 represents the tumor region and 0 represents the non-tumor region." Then, the tumor region is divided into non-overlapping slices of fixed size based on the binary mask, and the slices are stained and normalized to construct the instance set. First, the spatial range of sliceable slices is defined using the binary mask M (sampling only in the region where M=1), and then a patch set is generated with a fixed size (256×256) of non-overlapping slices to form the instance set.
[0067] Furthermore, the implementation of feature extraction includes:
[0068] like Figure 2 As shown, PatchVMamba is used as the instance encoder. Using the Visual State Space (VSS) module and Selective Scan in 2D (SS2D) as core components, long-range spatial context is modeled with near-linear complexity to enhance the discriminative power of instance representations. The specific steps are as follows:
[0069] Linear normalization is performed on the normalized segments of the full-slice image: ,in For LinearNorm operations;
[0070] The instance feature vectors of the corresponding segments are obtained by performing feature transformation on the N stacked visual state space modules in the PatchVMamba instance encoder. ;
[0071] Each visual state space module is updated using residual updates, with the following formula:
[0072] (6);
[0073] (7);
[0074] in, For the first Output features of each visual state space module For the first Output features of each visual state space module; It is a two-dimensional selective scan operator, one of the core state-space computation units of PatchVMamba; This is a depthwise convolution. LN is the activation function; LN is the layer normalization function, used for normalization stabilization training and numerical scaling within residual blocks. It is a feedforward network. For the first Intermediate features of a visual state space module;
[0075] After stacking N VSS modules, the instance feature vector is obtained. The expression is: ; The features output by the last visual state space module;
[0076] The instance feature vectors of each slice of the full slice image are obtained to form the instance feature bag of the corresponding full slice image. L represents the number of feature vectors, that is, the number of patches selected from a WSI and sent to the encoder. The encoder maps each patch to a feature vector, thus obtaining L feature vectors, L = . This represents the Lth feature vector (that is, the Lth instance embedding).
[0077] Specifically, any full-slice image can be encoded accordingly; that is, the instance feature bag of the i-th full-slice image is represented as... , ; Let be the instance feature bag of the i-th full-slice image; Let L be the number of slices in the i-th full-slice image, i.e., the number of patches selected from a WSI and fed into the encoder. The encoder maps each patch to a feature vector, thus obtaining L feature vectors, where L = ... ; For the j-th slice of the i-th full slice image WSI Instance feature vectors. PatchVMamba These are learnable parameters.
[0078] Furthermore, in the bag-level weighted aggregation stage, attention-based MIL aggregation is used to aggregate the instance set X into a bag-level representation H. The specific implementation includes:
[0079] Calculate the attention weights for each instance feature vector. The formula is:
[0080] (13);
[0081] in, and These are learnable parameters, used to map instance features to a scalar score; Let L be the feature vector of the k-th instance; L is the number of instance feature vectors, and L is equal to the number of blocks in the corresponding full slice image. ;
[0082] We perform weighted aggregation on the instance feature vectors in the instance feature bag to obtain the slice-level feature representation H:
[0083] (12);
[0084] Furthermore, the binary classifiers for the three subtype classifications are NSMP subtype binary classifiers. MMRd subtype binary classifier and p53abn subtype binary classifier .
[0085] Furthermore, the slice-level feature representation H is used by a classifier to obtain the class probability. :
[0086] (14);
[0087] in, and For classifier parameters, This indicates the number of categories corresponding to the binary classifier.
[0088] Furthermore, to improve the stability and controllability of classification, this invention decomposes the original four-class classification task into a serial OvR binary classification link, and adopts a unified threshold T and node-level calibration during the inference stage. The output probabilities of the three binary classifiers are:
[0089] (15);
[0090] (16);
[0091] (17);
[0092] Furthermore, the cascading judgment rule is as follows:
[0093] When the probability value of the NSMP subtype of the whole slice image If the value is greater than the threshold T, then the endometrial cancer in the current whole slice image is of the NSMP subtype;
[0094] When the probability value of the NSMP subtype of the whole slice image The probability value of MMRd subtype not greater than the threshold T If the value is greater than the threshold T, then the endometrial cancer in the current whole slice image is of the MMRd subtype;
[0095] When the probability value of the MMRd subtype of the whole slice image The probability value of the p53abn subtype is not greater than the threshold T. If the value is greater than the threshold T, then the endometrial cancer in the current whole slice image is of the p53abn subtype;
[0096] In other cases, the endometrial cancer in the current whole-slice image is classified as POLEmut. That is, if none of the three conditions are met, the endometrial cancer in the current whole-slice image is determined to be the POLEmut subtype.
[0097] The corresponding expression is:
[0098] (18);
[0099] Subtype represents the final subtyping result.
[0100] Furthermore, as a feasible implementation, the threshold T is preset to 0.5; the cascade strategy can reduce the decision complexity of the prediction task and improve the overall classification performance of the genotype.
[0101] This invention uses a binary mask to define the effective tissue range, generating fixed-size non-overlapping slices only within the mask area (where the mask is set to 1). The slices are then color-normalized to form a standardized set for subsequent modeling, reducing the interference of background and staining differences on the results. Each slice is input into a PatchVMamba instance encoder, which models the spatial context using linear normalization, VSS, and two-dimensional selective scanning to obtain a high-dimensional instance feature vector. This vector is then used to construct the instance feature set representation of the entire slice. This invention employs attention-based multi-instance aggregation to weightedly converge instance features, obtaining a slice-level representation and outputting class probabilities. Based on this, the four-class classification is reconstructed into a sequential pair of binary classification inference links, using a unified threshold for step-by-step determination to obtain the final classification result. This reduces interference from multi-class parallel competition and improves decision controllability.
[0102] This invention limits the source of the cut pieces to a tumor mask on the input side and performs uniform staining normalization and specification constraints on the cut pieces. This reduces the interference of non-critical information and batch differences before feature learning, increases the information density of the instance set, and makes the aggregation stage more focused on the tissue morphology clues related to EC classification, thus mitigating the problem of weak information diluting strong information.
[0103] This invention structurally combines feature encoding, attention convergence, probability output, and cascaded discrimination. While maintaining the interpretability and aggregation advantages of MIL (Multi-Level Interpretation), it introduces a step-by-step judgment rule with unified thresholds, making the decision-making process more controllable, the output easier to calibrate and verify, and improving the overall inference stability and reliability under conditions of high similarity categories in EC (Extremely Similar Classes). Simultaneously, this invention adjusts the four-class classification from a one-time parallel discrimination to a serial binary classification link, allowing each step to complete the judgment within a narrower candidate set. This reduces the confusion of discrimination criteria and attention drift caused by simultaneous competition among multiple categories, and is more conducive to forming more stable discrimination logic and output results in scenarios with subtle differences in EC morphology.
[0104] This invention, employing the above technical solutions, offers the following significant advantages over existing technologies: 1) By employing an ROI-driven instance construction strategy, this invention effectively defines the tumor region and performs standardized staining normalization, reducing interference from background tissue and staining differences at the source, thereby increasing the information density of the instance set. Combined with the long-range spatial context modeling capabilities of the PatchVMamba encoder, the accuracy of capturing subtle morphological features is improved, significantly enhancing the classification and recognition performance for POLEmut and other methods lacking clear morphological cues. 2) The cascaded structure decomposes the four-class classification task into a serial binary classification link, which better aligns with the clinical "stepwise exclusion" interpretation logic. The cascaded structure ensures high consistency between the classification results and pathology expert diagnoses, surpassing traditional parallel classification methods. Simultaneously, the step-by-step decision-making mechanism provides doctors with a clear decision-making path, greatly enhancing the interpretability of the results. 3) The classification accuracy of this invention exhibits minimal fluctuations under different scanning devices and staining batches, significantly lower than existing methods, demonstrating broad clinical application prospects.
[0105] Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Without conflict, the embodiments and features in the embodiments of this application can be combined with each other. The components of the embodiments of this application described and illustrated herein can generally be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.
Claims
1. A molecular subtyping method for endometrial cancer based on cascaded multi-instance learning, characterized in that: Includes the following steps: Instance construction: Preprocess the whole slice image of endometrial cancer to generate a binary mask of the tumor region, extract image patches within the region of interest based on the binary mask, and construct an instance set by staining and normalizing the image patches; Feature extraction: The PatchVMamba instance encoder is used to extract features from each image patch in the instance set to obtain the instance feature vector of each image patch, and then the instance feature bag of each full slice image is obtained. Weighted aggregation: An attention mechanism is used to perform weighted aggregation of the instance feature bags to obtain slice-level feature representations; Subtyping results: The slice-level features are input into three cascaded binary classifiers to obtain the probability values of the three subtypes of endometrial cancer respectively; based on the probability values of the three subtypes, a unified threshold is used to perform cascaded subtype classification of endometrial cancer to output the final subtyping results.
2. The molecular subtyping method for endometrial cancer based on cascaded multi-instance learning according to claim 1, characterized in that: The instance construction steps are implemented as follows: Obtain full-slice images and corresponding fractal labels to construct a dataset. S i This represents the i-th full slice image. The i-th full-slice image represents the corresponding fractal label, and N represents the number of full-slice images in the dataset. Generate a binary mask with the tumor region as the region of interest for each whole-slice image. Binary mask The expression is: (1); where 1 represents the tumor region and 0 represents the non-tumor region; , These are the height and width of the full slice image, respectively; The tumor region is divided into fixed-size regions based on a binary mask. Non-overlapping cuts, and A fixed height and width for each individual slice; After performing Vahadane staining normalization on each slice, the normalized slices are summed to obtain the normalized slice set for each full-slice image. As the corresponding set of instances.
3. The molecular subtyping method for endometrial cancer based on cascaded multi-instance learning according to claim 2, characterized in that: The expression for the block normalization process is: (3); This represents the Vahadane normalization operator. For a given slice, Cut into pieces for reference. For normalized chunks; Normalized slice set The expression is: (4); Let j be the number of chunks in each full slice image, and j be the index of the chunk in the current full slice image. (2), , These represent the height and width of the region of interest in pixel coordinates, respectively.
4. The molecular subtyping method for endometrial cancer based on cascaded multi-instance learning according to claim 3, characterized in that: The implementation of feature extraction includes: Linear normalization is performed on the normalized segments of the full-slice image: ,in For LinearNorm operations; The instance feature vectors of the corresponding segments are obtained by performing feature transformation on the N stacked visual state space modules in the PatchVMamba instance encoder. ; The instance feature vectors of each slice of the full slice image are obtained to form the instance feature bag of the corresponding full slice image. , Let L be the feature vector of the Lth instance, where L is the number of instance feature vectors, and L is equal to the number of blocks in the corresponding full slice image. .
5. The molecular subtyping method for endometrial cancer based on cascaded multi-instance learning according to claim 4, characterized in that: The update formula for each visual state space module is: （6）； （7）； in, For the first Output features of each visual state space module For the first Output features of each visual state space module; It is a two-dimensional selective scanning operator. For depthwise separable convolution, Here, LN is the activation function, and LN is the layer normalization function. It is a feedforward network; For the first Intermediate features of a visual state space module; Instance feature vector The expression is: ； in, The feature output by the last visual state space module is N, where N is the number of visual state space modules.
6. The molecular subtyping method for endometrial cancer based on cascaded multi-instance learning according to claim 1, characterized in that: The implementation of weighted aggregation includes: Calculate instance feature vectors Attention weights The formula is: （13）； in, and For learnable parameters, Let L be the feature vector of the k-th instance; L is the number of instance feature vectors, and L is equal to the number of blocks in the corresponding full slice image. ; We perform weighted aggregation on the instance feature vectors in the instance feature bag to obtain the slice-level feature representation H: （12）； The slice-level feature representation H is then processed by a binary classifier to obtain the class probabilities. : （14）； in, and For classifier parameters, This indicates the number of categories corresponding to the binary classifier.
7. The molecular subtyping method for endometrial cancer based on cascaded multi-instance learning according to claim 1, characterized in that: The binary classifiers for the three subtype classifications are the NSMP subtype binary classifiers. MMRd subtype binary classifier and p53abn subtype binary classifier .
8. A molecular subtyping method for endometrial cancer based on cascaded multi-instance learning according to claim 1 or 7, characterized in that: The cascading judgment rule is: When the probability value of the NSMP subtype of the whole slice image If the value is greater than the threshold T, then the endometrial cancer in the current whole slice image is of the NSMP subtype; When the probability value of the NSMP subtype of the whole slice image The probability value of MMRd subtype not greater than the threshold T If the value is greater than the threshold T, then the endometrial cancer in the current whole slice image is of the MMRd subtype; When the probability value of the MMRd subtype of the whole slice image The probability value of the p53abn subtype is not greater than the threshold T. If the value is greater than the threshold T, then the endometrial cancer in the current whole slice image is of the p53abn subtype; If none of the three conditions are met, the endometrial cancer in the current whole-slice image is determined to be of the POLEmut subtype.
9. The molecular subtyping method for endometrial cancer based on cascaded multi-instance learning according to claim 8, characterized in that: The threshold T is 0.5.