Grape tire risk hierarchical prediction method based on multi-instance learning

Abstract

The invention provides a multi-instance learning-based grape tire risk hierarchical prediction method. The method comprises the steps of obtaining suspected grape tire pathological image data and clinical parameters; preprocessing the image and cutting the image into image blocks; extracting image block features by using a feature extraction network, and aggregating to generate packet-level features; high-discrimination features are obtained through a multi-instance feature aggregation module; coding the coding network into features, and generating multi-source fusion features with fusion; inputting into the classification network, outputting prediction probabilities of non-grape tyres, partial grape tyres and complete grape tyres, and calculating disease risk scores according to the prediction probabilities of the non-grape tyres, the partial grape tyres and the complete grape tyres; dividing into four risk levels of low, medium, medium and high and high according to the scores, and automatically triggering an auxiliary diagnosis prompt for cases falling into classified easy-to-confuse levels. Pixel-level labeling is not needed, clinical safety and interpretability are considered while classification accuracy is improved, and the method is suitable for grape tire intelligent auxiliary diagnosis of primary and superior medical institutions.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically a method for risk stratification prediction of hydatidiform mole based on multi-instance learning. Background Technology

[0002] Hydatidiform mole, a unique disease in gynecological pathology, falls under the category of gestational trophoblastic diseases. It is characterized by abnormal proliferation of chorionic trophoblastic cells and can be classified as complete or partial hydatidiform mole. Accurate diagnosis of hydatidiform mole and effective differentiation from similar non-neoplastic lesions are crucial for developing individualized postoperative monitoring plans and early identification of persistent or invasive gestational trophoblastic diseases. Currently, while STR genotyping is widely recognized as the "gold standard" for hydatidiform mole diagnosis, possessing excellent sensitivity and specificity, its complex procedures and high testing costs hinder its widespread adoption in routine clinical practice. Therefore, histopathological examination remains the primary method for hydatidiform mole screening. At present, pathologists primarily rely on manual interpretation of full-field digital pathology slides for differential diagnosis. However, due to the high morphological similarity between complete, partial, and non-hydatidiform moles, manual interpretation is not only highly subjective and inefficient but also susceptible to inter- and intra-observer variability. Although existing studies have attempted to introduce machine learning algorithms to assist in the classification of the three types of lesions, due to limitations such as small training data scale and simple model architecture, existing methods generally suffer from problems such as weak generalization ability, poor robustness and insufficient classification accuracy, making it difficult to meet actual clinical needs. Summary of the Invention

[0003] This invention provides a method for stratified prediction of hydatidiform mole risk based on multi-instance learning, characterized in that the method includes the following steps:

[0004] Step 1: Collect full-field digital pathological images and clinical parameter information of patients suspected of having hydatidiform mole, and review their data integrity and usability to form raw pathological image data to be classified. and clinical parameters ;

[0005] Step 2: Preprocess the pathological image and perform regular cropping on the effective tissue area obtained from the preprocessing to generate a set of image blocks of fixed size. ;

[0006] Step 3, collect image patch data Input to a feature extraction network pre-trained on large-scale pathological images In this process, the set of image patch feature vectors is obtained. By splicing The feature vectors of all individual image patches within the packet form a packet-level feature representation. ;

[0007] Step 4, Represent the packet-level features Input to a multi-instance feature aggregation module, output highly discriminative enhanced image features. ;

[0008] Step 5, set clinical parameters Input text feature encoding network In this process, clinical parameter characteristics are obtained. and enhanced image features Fusion yields multi-source fusion characteristics. ;

[0009] Step 6, integrate multi-source features Input to classification network The output sets the predicted probabilities of the pathological image belonging to three pathological states: non-hydatidiform mole, partial hydatidiform mole, and complete hydatidiform mole. ;

[0010] Step 7: Utilize the predicted probability set S of the three pathological states and the preset weighting coefficients. Calculate the clinically-guided disease risk score probability Based on the score, cases are automatically classified into four levels: low risk, medium risk, medium-high risk, and high risk. For cases whose risk scores fall into the medium risk range (where pathological status classification is easily confused), auxiliary diagnostic prompts are automatically triggered to achieve intelligent classification of hydatidiform mole.

[0011] Specifically, in step 1, the acquisition of pathological image data to be classified and clinical parameter data Specifically, the process involves: collecting suspected hydatidiform mole samples that have undergone tissue sectioning and hematoxylin and eosin (H&E) staining by a pathologist; generating full-view digital pathology image source files using a high-resolution digital pathology scanner; subsequently, using visualization tools to verify the format compatibility and display integrity of the image files to ensure they can be loaded and viewed normally; images that pass the verification are used as valid input data, denoted as... Clinical parameter data Based on pathological images by pathologists The initial images obtained mainly include key morphological parameters such as the maximum diameter of the villi;

[0012] Specifically, in step 2, the image preprocessing of the pathological image involves: converting the pathological image to be processed to the HSV color space, performing binarization segmentation based on preset saturation and brightness thresholds, retaining pixels that meet the threshold conditions as effective tissue regions, and removing large areas of blank background and invalid hole regions through morphological operations; subsequently, the retained foreground image is input into the self-trained segmentation network. The network was trained on a large number of manually labeled samples of non-diagnostic areas such as dirt, bleeding, and decidua to accurately identify and remove these interfering components, retaining only the villous tissue regions with diagnostic value. Next, the obtained effective tissue regions were color-normalized using conventional staining normalization methods to reduce staining differences caused by different slide batches or scanners. Finally, the normalized tissue regions were slidably cropped according to a fixed window size based on the original spatial coordinates, generating several image patches that form a set. ,in M represents a single image patch, where M represents the first image patch. The total number of image blocks obtained by cropping a pathological image;

[0013] Specifically, in step 3, the data of each image patch instance set... Input to a feature extraction network pre-trained on large-scale pathological images Feature extraction is performed by using a feature extraction network pre-trained on large-scale pathological images as the feature extractor. ,for Each image block input Enter it into After forward propagation, the set of feature vectors of fixed dimension W from the output of its global pooling layer is extracted. ,in This represents the feature vector extracted from a single image patch; finally, all M feature vectors are concatenated sequentially to form a bag-level feature representation. This is used for subsequent multi-instance feature aggregation and classification tasks;

[0014] Specifically, in step 4, the packet-level feature representation... The input multi-instance feature aggregation module specifically involves: first, combining package-level features... as input sequence Input the first-level Transformer layer; to reduce computational complexity for large-scale instances, this layer employs a Nystrom approximation multi-head self-attention mechanism: low-dimensional landmark features are sampled from the query matrix Q and the key matrix K respectively. By constructing approximate attention weights through pseudo-inverse operations, a globally enhanced feature representation can be obtained. Subsequently, The M instance features are remapped into a two-dimensional grid structure based on their original spatial coordinates in the pathological image. On this structure, group convolutions with different kernel sizes are applied in parallel to capture tissue structures at different scales. The convolution results are then concatenated by channels and added to the original features to obtain enhanced features containing local spatial correlation information. This enables the integration of local spatial relationships and organizational morphology information. Subsequently, Input the second Transformer self-attention interaction layer, perform deep feature interaction, and output... After layer normalization, the final enhanced packet-level features are obtained. This is used for subsequent multi-source fusion operations;

[0015] Specifically, in step 5, the clinical parameters are... Input text feature encoding network The image features are then fused to obtain the fused features. Specifically, this means: transferring clinical parameter information After normalization, the input text feature encoding network The network mainly consists of pre-trained multilayer perceptrons, and its output dimension is D, representing clinical parameter features. and enhanced package-level image features splicing results in splicing features Then through the fusion projection layer The concatenated features are linearly transformed and then processed by an activation function to generate the final multi-source fusion features. ,in This is a bias term. This fusion feature It also integrates high-order semantic information from pathological images with key clinical parameters for subsequent classification and risk scoring.

[0016] Specifically, in step 6, the multi-source fusion features Input to classification network The predicted probability set S is obtained by fusing features. The input is a classification network consisting of fully connected layers and a Softmax normalization function. The predicted probability distribution of the three pathological states corresponding to the target pathological image is obtained. ,in , , The predicted probabilities for non-hydatidiform mole, partial hydatidiform mole, and complete hydatidiform mole, respectively, are given, and the normalization conditions are satisfied. ;

[0017] Specifically, in step 7, the disease risk score is calculated based on the probability set S. The system automatically grades and provides results, specifically: First, it calculates a disease risk prediction score based on the corresponding probability distribution S. ,in Preset weighting coefficients were used to quantify the malignant potential of hydatidiform mole. Subsequently, based on previous experimental verification results, a risk score threshold range was set. Divided into four risk levels: low-risk area The model primarily identifies non-hydatidiform pregnancies with high accuracy; medium-risk areas. The predicted probabilities of non-hydatidiform pregnancies and partial hydatidiform pregnancies are similar, with blurred category boundaries, resulting in high uncertainty in model discrimination; medium- to high-risk areas The model primarily focuses on partial hydatidiform mole, and its accuracy is relatively high; high-risk areas The model shows high accuracy, primarily for complete hydatidiform moles. Notably, when the predicted result falls into the medium-risk zone, the system automatically identifies it as an "uncertain region" and does not directly output a definitive diagnosis. Instead, it triggers an auxiliary diagnostic suggestion mechanism: generating structured prompts and recommending that clinicians combine STR genotyping testing for final pathological confirmation. This avoids misdiagnosis or missed diagnosis due to high morphological similarity, significantly improving diagnostic safety and reliability. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0019] Figure 1 The flowchart is for a multi-instance learning-based risk stratification prediction method for hydatidiform mole.

[0020] Figure 2 This is a schematic diagram of the risk zone delineation results based on risk prediction scores; Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Currently, while STR genotyping is widely recognized as the gold standard for pathological classification of hydatidiform moles due to its high accuracy, its complex procedures, high costs, and demanding requirements for laboratory conditions and professional personnel hinder its widespread adoption in routine clinical practice. Therefore, in most primary care or resource-constrained medical institutions, hydatidiform mole classification still primarily relies on pathologists manually interpreting full-size digital pathology images. However, partial hydatidiform moles and non-hydatidiform moles are highly similar in histological morphology, making manual interpretation not only subjective and inefficient but also susceptible to inter- and intra-observer variability. Furthermore, existing classification methods based on traditional machine learning suffer from limitations such as small training data size and simple model architecture, generally exhibiting weak generalization ability, poor robustness, and insufficient classification accuracy, failing to meet actual clinical needs. To address the aforementioned problems, this invention innovatively proposes a hierarchical prediction method based on multi-instance learning without pixel-level annotation. By fusing full-view pathological images with key clinical parameters, an end-to-end risk stratification prediction framework is constructed. This framework automatically triggers auxiliary diagnostic prompts for cases whose risk scores fall into the intermediate-risk confusion zone, thereby improving classification accuracy while ensuring the safety and interpretability of clinical decisions. The specific steps are as follows:

[0023] S101: Collect full-view digital pathological images and clinical parameter information of patients suspected of having hydatidiform mole, and review their data integrity and availability to form raw pathological image data to be classified. and clinical parameters ;

[0024] Specifically, suspected hydatidiform mole samples were collected after pathologists had performed tissue sectioning and hematoxylin and eosin (H&E) staining. A high-resolution digital pathology scanner was used to generate full-view digital pathology image source files. Subsequently, visualization tools (such as OpenSlide) were used to verify the format compatibility and display integrity of the image files to ensure they could be loaded and viewed correctly. Images that passed the verification were used as valid input data and denoted as... Clinical parameter data Based on pathological images by pathologists The initial images obtained mainly consisted of a few morphological parameters, such as the maximum diameter of the villi.

[0025] S102: Perform image preprocessing on the pathological image, and perform regular cropping on the effective tissue area obtained from the preprocessing to generate a set of image blocks of fixed size. ;

[0026] Specifically, the pathological image to be processed is converted to the HSV color space, and binarized segmentation is performed based on preset saturation and brightness thresholds. Pixels that meet the threshold conditions are retained as valid tissue regions. Large areas of blank background and invalid hole regions are removed through morphological operations. Subsequently, the retained foreground image is input into a self-trained segmentation network. The network was trained on a large number of manually labeled samples of non-diagnostic areas such as dirt, bleeding, and decidua to accurately identify and remove these interfering components, retaining only the villous tissue regions with diagnostic value. Next, the obtained effective tissue regions were color-normalized using conventional staining normalization methods to reduce staining differences caused by different slide batches or scanners. Finally, the normalized tissue regions were slidably cropped according to a fixed window size based on the original spatial coordinates, generating several image patches that form a set. ,in M represents a single image patch, where M represents the first image patch. The total number of image blocks obtained by cropping a pathological image.

[0027] S103: Transfer image patch set data Input to a feature extraction network pre-trained on large-scale pathological images In this process, the set of image patch feature vectors is obtained. By splicing The feature vectors of all individual image patches within the packet form a packet-level feature representation. ;

[0028] Specifically, a feature extraction network pre-trained on large-scale pathological images is used as the feature extractor. ,for Each image block input Enter it into After forward propagation, the set of feature vectors of fixed dimension W from the output of its global pooling layer is extracted. ,in This represents the feature vector extracted from a single image patch; finally, all M feature vectors are concatenated sequentially to form a bag-level feature representation. This is used for subsequent multi-instance feature aggregation and classification tasks.

[0029] S104: Represent package-level features Input to a multi-instance feature aggregation module, output highly discriminative enhanced image features. ;

[0030] Specifically, firstly, the package-level features as input sequence Input the first-level Transformer layer; to reduce computational complexity for large-scale instances, this layer employs a Nystrom approximation multi-head self-attention mechanism: low-dimensional landmark features are sampled from the query matrix Q and the key matrix K respectively. Further, we can obtain ,in This is a pseudo-inverse operation. The dimension of the attention head is used to obtain the globally semantically enhanced feature representation. Subsequently, The M instance features are remapped into a two-dimensional grid structure based on their original spatial coordinates in the pathological image. On this structure, group convolutions with different kernel sizes are applied in parallel to capture tissue structures at different scales. The convolution results are then concatenated by channels and added to the original features to obtain enhanced features containing local spatial correlation information. This enables the integration of local spatial relationships and organizational morphology information. Subsequently, Input the second Transformer self-attention interaction layer, perform deep feature interaction, and output... After layer normalization, the final enhanced packet-level features are obtained. This is used for subsequent multi-source fusion operations.

[0031] S105: Clinical parameters Input text feature encoding network In this process, clinical parameter characteristics are obtained. and enhanced image features Fusion yields multi-source fusion characteristics. ;

[0032] Specifically, clinical parameter information After normalization, the input text feature encoding network The network mainly consists of pre-trained multilayer perceptrons, and its output dimension is D, representing clinical parameter features. and enhanced package-level image features splicing results in splicing features Then through the fusion projection layer The concatenated features are linearly transformed and then processed by an activation function to generate the final multi-source fusion features. ,in This is a bias term. This fusion feature It also integrates high-order semantic information from pathological images with key clinical parameters for subsequent classification and risk scoring.

[0033] S106: Incorporating multi-source fusion features Input to classification network The output sets the predicted probabilities of the pathological image belonging to three pathological states: non-hydatidiform mole, partial hydatidiform mole, and complete hydatidiform mole. ;

[0034] Specifically, the fusion features The input is a classification network consisting of fully connected layers and a Softmax normalization function. The predicted probability distribution of the three pathological states corresponding to the target pathological image is obtained. ,in , , The predicted probabilities for non-hydatidiform mole, partial hydatidiform mole, and complete hydatidiform mole, respectively, are given, and the normalization conditions are satisfied. ;

[0035] S107: Calculate the clinically-guided disease risk score probability using the predicted probability set S of three pathological states. Based on the score, cases are automatically classified into four levels: low risk, medium risk, medium-high risk, and high risk. For cases whose risk scores fall into the medium risk range (where pathological status classification is easily confused), auxiliary diagnostic prompts are automatically triggered to achieve intelligent classification of hydatidiform mole.

[0036] Specifically, firstly, a disease risk prediction score is calculated based on the probability distribution S to which the disease belongs. ,in Preset weighting coefficients (which can be adjusted based on training results) are used to quantify the malignant potential of hydatidiform mole. Subsequently, based on previous experimental verification results, a risk score threshold range is set, and... Divided into four risk levels: low-risk area The model primarily identifies non-hydatidiform pregnancies with high accuracy; medium-risk areas. The predicted probabilities of non-hydatidiform pregnancies and partial hydatidiform pregnancies are similar, with blurred category boundaries, resulting in high uncertainty in model discrimination; medium- to high-risk areas Primarily targeting partial hydatidiform mole, the model achieves high accuracy; high-risk areas The model is highly accurate, primarily targeting complete hydatidiform moles. Notably, when the predicted result falls into the medium-risk zone, the system automatically identifies it as an "uncertain region" and does not directly output a definitive diagnosis. Instead, it triggers an auxiliary diagnostic suggestion mechanism: generating structured prompts and recommending that clinicians combine STR genotyping testing for final pathological confirmation. This avoids misdiagnosis or missed diagnosis due to high morphological similarity, significantly improving diagnostic safety and reliability.

[0037] like Figure 2 The diagram shown illustrates the risk zone delineation results based on risk prediction scores, specifically including:

[0038] The horizontal axis in the graph represents the pathology category (data label), and the vertical axis represents the disease risk prediction score obtained by the model. Different markers represent different pathological category labels: a dot (●) indicates a non-hydatidiform mole, a triangle (▲) indicates a partial hydatidiform mole, and an asterisk (★) indicates a complete hydatidiform mole. The threshold was determined earlier on the independent validation set by maximizing the balance between sensitivity and specificity. The model was used to determine the risk level, and four risk zones were defined: low-risk, medium-risk, medium-high-risk, and high-risk. The low-risk zone primarily contains samples of non-hydatidiform mole pregnancies, which are accurately identified by the model and can be directly classified as non-hydatidiform moles. The medium-risk zone shows a high degree of overlap between non-hydatidiform mole and partial hydatidiform mole samples, with similar morphological characteristics, leading to higher uncertainty in the model and a higher risk of misdiagnosis. The medium-high-risk zone is dominated by partial hydatidiform moles, indicating the presence of partial hydatidiform mole lesions. The high-risk zone is concentrated with samples of complete hydatidiform moles. Notably, the medium-risk zone is shown in the figure. Significant overlap was observed between samples of non-hydatidiform mole and partial hydatidiform mole, indicating that this region represents a typical "blurred" case set. In such cases, structured prompts are automatically generated, recommending that clinicians combine STR genotyping with the final pathological confirmation. This enhances the safety, reliability, and clinical applicability of this method in real-world clinical scenarios.

[0039] This invention addresses the clinical challenge of pathological diagnosis of hydatidiform mole, constructing an end-to-end intelligent classification framework for three clinically easily confused pathological types: complete hydatidiform mole, partial hydatidiform mole, and non-hydatidiform mole pregnancy. By fusing high-quality full-view digital pathological images with key clinical parameters (such as maximum chorionic villus diameter), combined with a feature encoder pre-trained on a large scale and a multi-instance learning mechanism based on a Transformer architecture, the ability to discriminate hydatidiform mole pathological images is effectively enhanced. Furthermore, a clinically guided risk score is constructed based on the predicted probability of different pathological categories, and an automated auxiliary diagnostic trigger mechanism is designed for the easily confused medium-risk range. This enables accurate identification of high-risk cases and safe warnings for cases with ambiguous boundaries, significantly improving classification accuracy and clinical applicability, and providing reliable technical support for the accurate diagnosis of hydatidiform mole.

Claims

1. A method for stratified prediction of hydatidiform mole risk based on multi-instance learning, characterized in that, The method includes the following steps: Step 1: Collect full-field digital pathological images and clinical parameter information of patients suspected of having hydatidiform mole, and review their data integrity and usability to form raw pathological image data to be classified. and clinical parameters ; Step 2: Preprocess the pathological image and perform regular cropping on the effective tissue area obtained from the preprocessing to generate a set of image blocks of fixed size. ; Step 3, collect image patch data Input to a feature extraction network pre-trained on large-scale pathological images In this process, the set of image patch feature vectors is obtained. By splicing The feature vectors of all individual image patches within the packet form a packet-level feature representation. ; Step 4, Represent the packet-level features Input to a multi-instance feature aggregation module, output highly discriminative enhanced image features. ; Step 5, set clinical parameters Input text feature encoding network In this process, clinical parameter characteristics are obtained. and enhanced image features Fusion yields multi-source fusion characteristics. ; Step 6, fusion features from multiple sources Input to classification network The output sets the predicted probabilities of the pathological image belonging to three pathological states: non-hydatidiform mole, partial hydatidiform mole, and complete hydatidiform mole. ; Step 7: Utilize the predicted probability set S of the three pathological states and the preset weighting coefficients. Calculate the clinically-guided disease risk score probability Based on the score, cases are automatically classified into four levels: low risk, medium risk, medium-high risk, and high risk. For cases whose risk scores fall into the medium risk range (where pathological status classification is easily confused), auxiliary diagnostic prompts are automatically triggered to achieve intelligent classification of hydatidiform mole.

2. The method for stratified prediction of hydatidiform mole risk based on multi-instance learning as described in claim 1, characterized in that, In step 1, the pathological image data to be classified is obtained. and clinical parameter data Specifically, this involves collecting suspected hydatidiform mole samples that have been sectioned and stained with hematoxylin and eosin (H&E) by a pathologist, and generating full-view digital pathology image source files using a high-resolution digital pathology scanner. Subsequently, a visual tool was used to verify the format compatibility and display integrity of the image files to ensure that they could be loaded and viewed normally. The image that passes the verification is used as valid input data, denoted as... Clinical parameter data Based on pathological images by pathologists The initial images obtained mainly include key morphological parameters such as the maximum diameter of the villi.

3. The method for stratified prediction of hydatidiform mole risk based on multi-instance learning as described in claim 1, characterized in that, In step 2, the image preprocessing of the pathological image specifically involves: converting the pathological image to be processed to the HSV color space, performing binarization segmentation based on preset saturation and brightness thresholds, retaining pixels that meet the threshold conditions as effective tissue areas, and removing large areas of blank background and invalid hole areas through morphological operations. The retained foreground image is then fed into the self-trained segmentation network. The network was trained on a large number of manually labeled samples of non-diagnostic areas such as dirt, bleeding, and decidua to accurately identify and remove these interfering components, retaining only the villous tissue regions with diagnostic value. Next, the obtained effective tissue regions were color-normalized using conventional staining normalization methods to reduce staining differences caused by different slide batches or scanners. Finally, the normalized tissue regions were slidably cropped according to a fixed window size based on the original spatial coordinates, generating several image patches that form a set. ,in M represents a single image patch, where M represents the first image patch. The total number of image blocks obtained by cropping a pathological image.

4. The method for stratified prediction of hydatidiform mole risk based on multi-instance learning as described in claim 1, characterized in that, In step 3, the data of each image patch instance set is... Input to a feature extraction network pre-trained on large-scale pathological images Feature extraction is performed by using a feature extraction network pre-trained on large-scale pathological images as the feature extractor. ,for Each image block input Enter it into After forward propagation, the set of feature vectors of fixed dimension W from the output of its global pooling layer is extracted. ,in This represents the feature vector extracted from a single image patch; finally, all M feature vectors are concatenated sequentially to form a bag-level feature representation. This is used for subsequent multi-instance feature aggregation and classification tasks.

5. The method for stratified prediction of hydatidiform mole risk based on multi-instance learning as described in claim 1, characterized in that, In step 4, the packet-level feature representation is... The input multi-instance feature aggregation module specifically involves: first, combining package-level features... as input sequence Input the first-level Transformer layer; to reduce computational complexity for large-scale instances, this layer employs a Nystrom approximation multi-head self-attention mechanism: low-dimensional landmark features are sampled from the query matrix Q and the key matrix K respectively. By constructing approximate attention weights through pseudo-inverse operations, a globally enhanced feature representation can be obtained. Subsequently, The M instance features are remapped into a two-dimensional grid structure based on their original spatial coordinates in the pathological image. On this structure, group convolutions with different kernel sizes are applied in parallel to capture tissue structures at different scales. The convolution results are then concatenated by channels and added to the original features to obtain enhanced features containing local spatial correlation information. This enables the integration of local spatial relationships and organizational morphology information. Subsequently, Input the second Transformer self-attention interaction layer, perform deep feature interaction, and output... ; After layer normalization (LN), the final enhanced packet-level features are obtained. This is used for subsequent multi-source fusion operations.

6. The method for stratified prediction of hydatidiform mole risk based on multi-instance learning as described in claim 1, characterized in that, In step 5, the clinical parameters are... Input text feature encoding network The image features are then fused to obtain the fused features. Specifically, this means: transferring clinical parameter information After normalization, the input text feature encoding network The network mainly consists of pre-trained multilayer perceptrons, and its output dimension is D, representing clinical parameter features. and enhanced package-level image features splicing results in splicing features Then through the fusion projection layer The concatenated features are linearly transformed and then processed by an activation function to generate the final multi-source fusion features. ,in This is a bias term. This fusion feature It also integrates high-level semantic information from pathological images with key clinical parameters for subsequent classification and risk scoring.

7. The method for stratified prediction of hydatidiform mole risk based on multi-instance learning as described in claim 1, characterized in that, In step 6, the multi-source fusion features are... Input to classification network The predicted probability set S is obtained by fusing features. The input is a classification network consisting of fully connected layers and a Softmax normalization function. The predicted probability distribution of the three pathological states corresponding to the target pathological image is obtained. ,in , , The predicted probabilities for non-hydatidiform mole, partial hydatidiform mole, and complete hydatidiform mole, respectively, are given, and the normalization conditions are satisfied. .

8. The method for stratified prediction of hydatidiform mole risk based on multi-instance learning as described in claim 1, characterized in that, In step 7, the disease risk score is calculated based on the probability set S. The system automatically grades and provides results, specifically: First, it calculates a disease risk prediction score based on the corresponding probability distribution S. ,in Preset weighting coefficients were used to quantify the malignant potential of hydatidiform mole. Subsequently, based on previous experimental verification results, a risk score threshold range was set. Divided into four risk levels: low-risk area The model primarily identifies non-hydatidiform pregnancies with high accuracy; medium-risk areas. The predicted probabilities of non-hydatidiform pregnancies and partial hydatidiform pregnancies are similar, with blurred category boundaries, resulting in high uncertainty in model discrimination; medium- to high-risk areas The model primarily focuses on partial hydatidiform mole, demonstrating high accuracy; high-risk areas... The model is highly accurate, primarily targeting complete hydatidiform moles. Notably, when the predicted result falls into the medium-risk zone, the system automatically identifies it as an "uncertain region" and does not directly output a definitive diagnosis. Instead, it triggers an auxiliary diagnostic suggestion mechanism: generating structured prompts and recommending that clinicians combine STR genotyping testing for final pathological confirmation. This avoids misdiagnosis or missed diagnosis due to highly similar morphologies, significantly improving diagnostic safety and reliability.

Images