HSIL postoperative cutting edge state prediction method, system, equipment and medium

By constructing a deep learning network combined with multiple machine learning models, the HSIL postoperative margin status prediction method is solved, the prediction of high sensitivity and robustness is achieved, and the risk of postoperative lesion residual and recurrence of HSIL patients is reduced, and personalized treatment guidance is provided.

CN120452779APending Publication Date: 2025-08-08THE SECOND HOSPITAL OF SHANDONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202510529550.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-04-25
Publication Date
2025-08-08

AI Technical Summary

Technical Problem

There is a lack of accurate prediction models for the postoperative marginal status of LEEP in HSIL patients in the prior art, resulting in a high risk of lesion residual and recurrence after surgery, and the subjectivity of relying on colposcopy leads to frequent misdiagnosis and misdiagnosis.

Method used

A method for predicting the state of the margin after HSIL surgery is constructed. By obtaining the clinical data of the patient and colposcopic images, a deep learning network is used to combine multiple machine learning models, including feature extraction, feature fusion and prediction layers, to screen the final features and build a combined model for prediction.

Benefits of technology

It improves the prediction accuracy and robustness of postoperative margin status in HSIL patients, provides personalized treatment guidance, reduces the risk of lesion residual and recurrence, and enhances the objectivity and accuracy of colposcopy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120452779A_ABST
    Figure CN120452779A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of image processing, and discloses an HSIL postoperative cutting edge state prediction method, system and device and a medium, and the method comprises the steps: obtaining a data set which comprises clinical data of a plurality of patients, colposcope images and corresponding diagnosis labels; inputting the colposcope image of each patient into the risk prediction model for classification prediction to obtain an incisal edge prediction score; screening final features in the clinical data and the cutting edge prediction score based on a feature screening method; constructing a combined model, wherein the combined model is composed of a plurality of machine learning models; training the combined model based on the selected final features, performing verification evaluation, and selecting an optimal machine learning model as a cutting edge state prediction model based on an evaluation result; and executing an incisal edge state prediction task of the patient based on the incisal edge state prediction model. According to the method, the accuracy and effectiveness of preoperative evaluation of the HSIL patient by a doctor can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a method, system, device and medium for predicting the resection margin status after HSIL surgery. Background Art

[0002] Cervical cancer is the fourth most common malignancy in women. Efforts to eliminate cervical cancer include vaccination against high-risk human papillomavirus (HPV), cervical cancer screening and treatment of precancerous lesions, and early detection and prompt treatment of early-stage cervical cancer. HPV vaccination for women of appropriate age is a key approach to preventing cervical cancer. However, it will take considerable time to reduce the incidence of cervical cancer through large-scale HPV vaccination. Cervical intraepithelial lesions (CIN) are categorized as low-grade squamous intraepithelial lesions (LSIL) and high-grade squamous intraepithelial lesions (HSIL). An estimated 30% of HSILs may progress to cervical cancer. Currently, surgery is the mainstay of treatment for HSILs, including loop electrosurgical excision procedure (LEEP) and cold knife conization (CKC). Although surgical treatment is highly effective for HSILs, approximately 5% to 25% of patients experience residual disease, recurrence, or progression to higher-grade disease after surgery. Positive resection margins in HSILs are significantly associated with residual and recurrent disease. In addition, residual disease and recurrence caused by positive surgical margins in HSIL may affect the patient's fertility and lead to adverse pregnancy outcomes. Even if standardized surgical procedures are strictly followed, positive margins cannot be completely avoided. Therefore, developing an accurate model to predict the surgical margin status of HSIL patients after LEEP surgery can help gynecologists perform accurate and effective preoperative evaluations of HSIL patients. Colposcopy is an important tool for identifying cervical precancerous lesions, determining the extent and grade of lesions, and guiding biopsy. However, colposcopy relies heavily on the subjective experience of the colposcopist, and its overall performance is unsatisfactory. Due to the high complexity of cervical lesions and the scarcity of professional colposcopists, missed diagnoses and misdiagnoses often occur.

[0003] In recent years, artificial intelligence (AI) has been widely used in medical imaging research and practice. Artificial intelligence methods using machine learning or deep learning frameworks can extract relevant information from medical imaging datasets. This information indicates the patient's lesion grade and prognostic significance. However, in current technology, AI-based colposcopy research mainly focuses on lesion grade. There is no predictive model for the postoperative resection margin status of HSIL patients, and there is a lack of AI-based colposcopy research models for predicting the resection margin status after LEEP in HSIL patients. Summary of the Invention

[0004] The purpose of the present invention is to provide a method, system, device and medium for predicting the postoperative resection margin status of HSIL to solve the problems existing in the above-mentioned prior art.

[0005] To achieve the above objectives, the present invention provides a method for predicting the resection margin status after HSIL surgery, comprising:

[0006] Acquire a data set, the data set including clinical data, colposcopic images, and corresponding diagnostic labels of several patients;

[0007] Inputting each patient's colposcopic image into a risk prediction model for classification prediction to obtain a margin prediction score; wherein the risk prediction model includes a feature extraction module, a feature fusion module, and a prediction layer connected in sequence, and the feature extraction module includes a low-level feature representation extraction module and a global semantic information extraction module set in parallel;

[0008] screening the final features from the clinical data and the margin prediction score based on the feature screening method;

[0009] Building a combined model, the combined model consisting of several machine learning models;

[0010] The combined model is trained based on the final selected features, and a validation evaluation is performed, and based on the evaluation results, an optimal machine learning model is selected as a model for predicting the resection margin status;

[0011] A patient's resection margin status prediction task is performed based on the resection margin status prediction model.

[0012] Optionally, the risk prediction model training process specifically includes:

[0013] The colposcopic image is input into a risk prediction model, and training is performed with the goal of minimizing the loss between the initial prediction result after prediction classification and the diagnostic label corresponding to the colposcopic image to obtain a trained risk prediction model.

[0014] Optionally, inputting the colposcopy images of each patient into the risk prediction model for classification prediction specifically includes:

[0015] Dividing the colposcopic image into a first image and a second image of different resolutions, inputting the first image into a low-level feature representation extraction module, extracting low-level feature representation data of the first image through a CNN encoder, inputting the second image into a global semantic information extraction module, and extracting global semantic information of the second image through a ViT encoder;

[0016] The extracted low-level feature representation data and global semantic information are input into the feature fusion module for fusion to obtain fusion features, and the fusion features are input into the prediction layer for prediction to obtain a margin prediction score.

[0017] Optionally, the feature-based screening method screens the final features from the clinical data and the margin prediction score, specifically including:

[0018] The optimal features of clinical data and margin prediction scores were screened according to different feature screening methods, and the intersection of the best feature subsets screened by different feature screening methods was taken to obtain the final features.

[0019] Optionally, the combined model includes a logistic regression model, an extreme gradient boosting model, a random forest model and a support vector machine model.

[0020] A HSIL postoperative margin status prediction system, comprising:

[0021] A data acquisition module is used to obtain a data set, wherein the data set includes clinical data of several patients, colposcopic images and corresponding diagnostic labels;

[0022] A deep learning network construction module is used to input each patient's colposcopic image into a risk prediction model for classification prediction to obtain a margin prediction score; wherein the risk prediction model includes a feature extraction module, a feature fusion module, and a prediction layer connected in sequence, and the feature extraction module includes a low-level feature representation extraction module and a global semantic information extraction module arranged in parallel;

[0023] A combined model construction module is used to screen the final features from the clinical data and margin prediction scores according to the feature screening method; construct a combined model, which is composed of several machine learning models; train the combined model based on the selected final features, and perform verification and evaluation, and select the optimal machine learning model as the margin status prediction model based on the evaluation results; and perform the patient's margin status prediction task based on the margin status prediction model.

[0024] An electronic device includes a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the method for predicting the postoperative resection margin status of HSIL.

[0025] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for predicting the postoperative resection margin status of HSIL.

[0026] The technical effects of the present invention are:

[0027] This study constructed a comprehensive model by combining clinical and deep learning networks and compared four commonly used machine learning models to identify the optimal model for predicting surgical margin status in patients with HSIL after LEEP surgery. The clinical combined with deep learning network demonstrated higher sensitivity and greater robustness in predicting positive surgical margins. Finally, this example used SHAP to visualize the entire model prediction process from the overall to the individual level. This finding highlights the importance of building a comprehensive model to comprehensively assess surgical margin status after LEEP surgery.

[0028] The present invention combines colposcopic image features and clinical information features to identify LEEP margin status and stratify the recurrence risk of HSIL patients. SHAP provides a bridge for personalized prediction and may help provide personalized treatment and clinical decision-making for HSIL patients preparing to undergo LEEP surgery. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0030] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0031] Figure 1 Schematic diagram of the deep learning network structure in an embodiment of the present invention;

[0032] Figure 2 is a GradCAM heat map in an embodiment of the present invention;

[0033] Figure 3 1 is a diagram showing comparison results of different models in an embodiment of the present invention;

[0034] Figure 4 is the DCA curve in the embodiment of the present invention;

[0035] Figure 5 This is an implementation flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0036] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as limiting the present invention, but rather as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0037] It should be understood that the terms described herein are intended only to describe particular embodiments and are not intended to limit the present invention. In addition, for numerical ranges herein, it should be understood that each intermediate value between the upper and lower limits of the range is also specifically disclosed. Each smaller range between any intermediate value within a stated value or stated range and any other stated value or intermediate value within the stated range is also encompassed by the present invention. The upper and lower limits of these smaller ranges may be independently included or excluded within the scope.

[0038] It will be apparent to those skilled in the art that various modifications and variations may be made to the specific embodiments of the present invention without departing from the scope or spirit of the invention. Other embodiments will be apparent to those skilled in the art from the present invention. The present description and examples are intended to be illustrative only.

[0039] The words “include,” “including,” “have,” “contain,” etc. used in this article are open-ended terms, meaning including but not limited to.

[0040] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0041] like Figure 1 - Figure 5 As shown, this embodiment provides a method for predicting the postoperative margin status of HSIL, including: obtaining a data set, the data set including clinical data, colposcopic images and corresponding diagnostic labels of several patients; inputting the colposcopic images of each patient into a risk prediction model for classification prediction to obtain a margin prediction score; wherein the risk prediction model includes a feature extraction module, a feature fusion module and a prediction layer connected in sequence, and the feature extraction module includes a low-level feature representation extraction module and a global semantic information extraction module set in parallel; based on a feature screening method, screening the final features from the clinical data and the margin prediction score; constructing a combined model, the combined model consisting of several machine learning models; training the combined model based on the selected final features, and performing verification and evaluation, and selecting the optimal machine learning model as the margin status prediction model based on the evaluation results; and executing the patient's margin status prediction task based on the margin status prediction model.

[0042] Positive resection margins were identified as a significant and independent predictor of residual disease after surgery. Notably, patients with positive resection margins faced an 11.36-fold increased relative risk of recurrence one year after treatment compared with those with negative resection margins. Furthermore, managing positive resection margins is a significant challenge for women who are still pursuing fertility plans, as any further surgical treatment carries the risk of preterm birth.

[0043] Preoperative assessment of positive resection margins is challenging and controversial for clinicians. Previous studies have shown that several clinical factors may be associated with positive resection margins. This model provides a robust and clinically valuable tool for predicting positive resection margins and can guide clinical decision-making and patient management after LEEP.

[0044] In this example, a deep learning network based on preoperative colposcopy images was constructed. A comprehensive model was constructed by combining clinical and deep learning networks, and four commonly used machine learning models were compared to determine the best model for predicting the resection margin status of patients with HSIL after LEEP. The clinical + deep learning network has higher sensitivity and stronger robustness in predicting positive resection margins. Finally, this example uses SHAP to visualize the entire model prediction process from the overall to the individual level. This finding emphasizes the importance of constructing a comprehensive model to comprehensively evaluate the resection margin status after LEEP. The results of this example show that the comprehensive model is superior to a single model.

[0045] In this example, 5167 patients with HSIL who underwent LEEP surgery at five centers were recruited. A total of 3129 patients from Qilu Hospital of Shandong University were selected as the training cohort, while external validation cohort 1 (EVC1) included 578 patients from the Second Hospital of Shandong University and 1237 patients from Zhangzhou Hospital of Fujian Medical University. External validation cohort 3 (EVC3) consisted of 97 patients from Yantai Zhifu Maternal and Child Health Hospital and 126 patients from the Second Affiliated Hospital of Shandong First Medical University.

[0046] Clinical Data Collection: In this example, detailed clinical information and colposcopic images were collected from patients with HSIL undergoing LEEP, including age, reproductive history, transformation zone (TZ) type, Thinprep cytology test (TCT) results, HPV status, and colposcopic images. Participants had HSIL confirmed by biopsy, underwent colposcopic LEEP, and had clear surgical margins. A positive surgical margin was diagnosed if postoperative pathology showed CIN at or near the specimen margin, particularly if the CIN lesion was within 1 mm of the margin.

[0047] Colposcopic image acquisition and preprocessing: Colposcopic diagnoses were classified as normal / benign, low-grade, or high-grade. Histopathological diagnoses were made by highly experienced histopathologists at local institutions. Results were classified as normal, LSIL, and HSIL according to the revised World Health Organization (WHO) criteria.

[0048] Clinical Factor Analysis and Modeling: Univariate logistic regression analysis was used to select clinical factors, and variables with p-values < 0.05 were retained. These variables were then subjected to multivariable logistic regression analysis to identify the most predictive factors. A multivariable logistic regression model was constructed based on the final selected variables to develop a risk prediction model.

[0049] Construction of deep learning network: Figure 1 The network structure diagram is shown. The architecture proposed in this example is based on two encoders: a convolutional neural network (CNN) and a visual transformer (ViT). This encoder captures rich local and global information, ensuring that the network can more accurately learn salient features. The CNN encoder captures low-level feature representations from high-resolution input images, while the ViT encoder is used to learn global semantic information from lower-resolution input images.

[0050] In the CNN encoder, this embodiment adopts a network structure called InceptionNeXt. InceptionNeXt is an innovative deep learning model that combines the advantages of the Inception architecture and ConvNeXt design. Its network structure is characterized by decomposing the large kernel convolution into multiple parallel branches in an Inception-style manner. Specifically, InceptionNeXt uses four parallel branches in the channel dimension, including a small square kernel, two orthogonal kernels (usually rectangular kernels in different directions), and an identity mapping. This design not only reduces the computational burden, but also achieves effective control of parameters and MACs (matrix multiplication and addition). InceptionNeXt significantly improves computational efficiency while maintaining high accuracy, achieving an excellent balance between speed and performance.

[0051] In the ViT encoder, this embodiment adopts a network structure called Swin-Transformer. Swin-Transformer is a visual Transformer model based on the self-attention mechanism. It has a hierarchical network architecture and efficiently extracts image features through a windowing method. The input image is first divided into non-overlapping image blocks, and then these image blocks are mapped to a high-dimensional vector space through linear embedding. These image blocks are then processed by multiple Swin-Transformer blocks, each of which contains window-based multi-head self-attention (W-MSA), shifted window multi-head self-attention (SW-MSA), layer normalization (LN), and multi-layer perceptron (MLP) modules. In addition, the model also contains a patch merging layer that reduces the resolution and increases the channel dimension by merging the features of adjacent image blocks. Swin-Transformer has many advantages, such as efficient computational performance, powerful feature extraction capabilities, and multi-scale feature fusion. First, the windowed self-attention mechanism significantly reduces computational overhead, enabling the model to be applied to high-resolution images. Second, through its hierarchical design and patch merging layer, Swin-Transformer is able to extract multi-scale features and transfer information between windows, achieving a larger receptive field.

[0052] More specifically, the input image (H=480, W=320, C=3) is divided into two data with different resolutions, respectively and And pass it to two encoders. In the CNN encoder, the feature map is gradually downsampled to learn more complex features. A similar approach is also taken in the ViT encoder. Finally, this embodiment obtains two outputs, which are respectively and Their dimensions are (15×10×512) and (14×14×256) respectively.

[0053] In order to fuse these two features, this embodiment proposes a new module, namely the cross fusion module (CFM), to fuse the output features of the two encoders and complete the fusion of local and global semantic representations. In the CFM module, this embodiment first adjusts the two types of features to the same spatial size through a sampling operation. Then, this embodiment uses a layer normalization operation to generate a query (Q), a key (K), and a value (V). By fusing (element-by-element addition) the tensors in each tuple, i.e. (Q R , Q S ), (K R , K S ), (V R , V S), this embodiment obtains a result tensor rich in local details. To learn global semantic information, this embodiment applies a self-attention (SA) mechanism that efficiently computes point-by-point relationships between these result tensors. A linear projection layer is applied to the SA output, reshaped to its original size, and input into a convolutional layer. Throughout the fusion process, two skip connections are used to enhance information flow and facilitate training.

[0054] Combination Model Construction: Integrating multimodal data often improves performance. Therefore, this example integrates deepLEEP scores and clinical factors to construct a combined model. To determine the optimal model, this example uses four machine learning algorithms to construct the combined model: logistic regression (LR), extreme gradient boosting (XGBoost), random forest (RF), and support vector machine (SVM).

[0055] The steps for building the combined model include:

[0056] 1. First, clinical factors (age, reproductive history, transformation zone (TZ) type, Thinprep cytology test (TCT) results, and HPV status) and deepLEEP prediction values were integrated. Multiple feature screening methods, including LASSO, random forest, and XGBoost, were used to screen for the optimal features. The intersection of the best feature subsets screened by different feature screening methods was taken, and finally the three optimal features of deepLEEP, TCT, and HPV were screened.

[0057] 2. Use four machine learning methods, including logistic regression (LR), extreme gradient boosting (XGBoost), random forest (RF) and support vector machine (SVM), to build different machine learning models based on the above three optimal features.

[0058] 3. Use the validation set to find the optimal model as the final model;

[0059] 4. Use the test set to evaluate the performance of the optimal model.

[0060] Model Performance Evaluation and Statistical Analysis: Each model was evaluated using multiple metrics, including receiver operating characteristic (ROC) curves, area under the curve (AUC), accuracy, F1 score, sensitivity, and specificity. Differences in area under the curve (AUC) were statistically compared using the DeLong test. The net reclassification index (NRI) was used to quantify the incremental improvement of the combined model relative to the other models. Decision curve analysis (DCA) was used to assess the net clinical benefit of the model.

[0061] Results: Baseline characteristics of the study cohort: Table 1 summarizes the baseline clinical characteristics of all 5167 patients. In the training and validation cohorts, 494 (15.8%), 82 (14.2%), 186 (15.0%), and 34 (15.2%) patients, respectively, had positive margins after LEEP. In the training cohort, age ≥48, TCT ≥ASCUS, TCT ≥HSIL, HPV 16 / 18, gravidity ≥3, parity ≥2, TZ III type, and colposcopically assessed HSIL were statistically significantly different between the positive and negative margin groups (P < 0.05), as shown in Table 1.

[0062] Table 1 Baseline characteristics of all patients in the training and validation cohorts

[0063]

[0064]

[0065] This example proposes a deep learning framework that integrates the InceptionNeXt CNN encoder and the Swin TransformerViT encoder to exploit their complementary local and global feature extraction capabilities. This example uses the Gradient Weighted Class Activation Map (GradCAM) to demonstrate the contribution of each encoder. Figure 2 As shown, the result visualization highlights the ability of the CNN encoder in accurately localizing salient local features, while the ViT encoder demonstrates its effectiveness in considering a wider semantic context. The integration of these two encoders results in more detailed and comprehensive feature extraction, thereby enhancing the network's predictive power. Table 2 lists the AUC, accuracy, sensitivity, specificity, and F1 score of each model. Compared with the clinical model, the deepLEEP model showed superior performance, with AUCs of 0.910, 0.853, 0.803, and 0.828 in the training set, validation set, external test set 1, and external test set 2, respectively, while the AUCs of the clinical model were 0.687, 0.691, 0.641, and 0.655, respectively (DeLong test, all P < 0.05).

[0066] Table 2 Comparison results

[0067]

[0068] To explore the potential of multimodal data in predicting positive surgical margin status. This example integrates deepLEEP with HPV16 / 18, TCT≥HSIL and colposcopically assessed HSIL through data fusion, and adopts four machine learning methods. This example systematically evaluates the AUC of each combined model compared with deepLEEP. As shown in Table 2, the AUC values of all combined models are better than deepLEEP, emphasizing their enhanced predictive power. In order to rigorously evaluate the statistical significance of these performance differences, this example performed a DeLong test. The analysis showed that, except for the RF-C model, all combined models showed statistically significant improvements compared with deepLEEP (p<0.05). In the training cohort, the XGBoost model showed the best predictive performance. The detailed comparison results are shown in Figure 2. Figure 3 C. Therefore, this example selects the XGBoost model as the final combination model, and its AUC values are 0.932, 0.887, 0.834, and 0.878, respectively. To further illustrate the potential value of the combination model, this example uses NRI to compare the incremental difference between the XGBoost model and other models, and uses DCA to show the net benefit. Figure 4 As shown, DCA further shows that in all cohorts, the XGBoost model provides a larger net benefit than the single model within a reasonable threshold probability range.

[0069] In this example, a non-invasive comprehensive model based on machine learning was developed and validated. This model combines colposcopic image features and clinical information features to identify LEEP margin status and stratify the recurrence risk of HSIL patients. SHAP provides a bridge for personalized prediction and may help provide personalized treatment and clinical decision-making for HSIL patients preparing for LEEP surgery.

[0070] This embodiment may also provide a system for predicting the resection margin status after HSIL surgery, including:

[0071] A data acquisition module is used to obtain a data set, wherein the data set includes clinical data of several patients, colposcopic images and corresponding diagnostic labels;

[0072] A deep learning network construction module is used to input each patient's colposcopic image into a risk prediction model for classification prediction to obtain a margin prediction score; wherein the risk prediction model includes a feature extraction module, a feature fusion module, and a prediction layer connected in sequence, and the feature extraction module includes a low-level feature representation extraction module and a global semantic information extraction module arranged in parallel;

[0073] A combined model construction module is used to screen the final features from the clinical data and margin prediction scores according to the feature screening method; construct a combined model, which is composed of several machine learning models; train the combined model based on the selected final features, and perform verification and evaluation, and select the optimal machine learning model as the margin status prediction model based on the evaluation results; and perform the patient's margin status prediction task based on the margin status prediction model.

[0074] This embodiment may further provide an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the method for predicting the postoperative resection margin status of HSIL.

[0075] This embodiment further provides a computer-readable storage medium storing a computer program, which implements the method for predicting the postoperative resection margin status of HSIL when executed by a processor.

[0076] The above description is merely a preferred embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for predicting the surgical margin status after HSIL surgery, characterized in that: include: Acquire a data set, the data set including clinical data, colposcopic images, and corresponding diagnostic labels of several patients; Inputting each patient's colposcopic image into a risk prediction model for classification prediction to obtain a margin prediction score; wherein the risk prediction model includes a feature extraction module, a feature fusion module, and a prediction layer connected in sequence, and the feature extraction module includes a low-level feature representation extraction module and a global semantic information extraction module set in parallel; screening the final features from the clinical data and the margin prediction score based on the feature screening method; Building a combined model, the combined model consisting of several machine learning models; The combined model is trained based on the final selected features, and a validation evaluation is performed, and based on the evaluation results, an optimal machine learning model is selected as a model for predicting the resection margin status; A patient's resection margin status prediction task is performed based on the resection margin status prediction model.

2. The method for predicting the surgical margin status after HSIL surgery according to claim 1, characterized in that: The training process of the risk prediction model specifically includes: The colposcopic image is input into a risk prediction model, and training is performed with the goal of minimizing the loss between the initial prediction result after prediction classification and the diagnostic label corresponding to the colposcopic image to obtain a trained risk prediction model.

3. The method for predicting the surgical margin status after HSIL surgery according to claim 1, characterized in that: The colposcopy images of each patient are input into the risk prediction model for classification prediction, specifically including: Dividing the colposcopic image into a first image and a second image of different resolutions, inputting the first image into a low-level feature representation extraction module, extracting low-level feature representation data of the first image through a CNN encoder, inputting the second image into a global semantic information extraction module, and extracting global semantic information of the second image through a ViT encoder; The extracted low-level feature representation data and global semantic information are input into the feature fusion module for fusion to obtain fusion features, and the fusion features are input into the prediction layer for prediction to obtain a margin prediction score.

4. The method for predicting the resection margin status after HSIL surgery according to claim 1, wherein: The feature-based screening method screens the final features from the clinical data and the margin prediction score, specifically including: The optimal features of clinical data and margin prediction scores were screened according to different feature screening methods, and the intersection of the best feature subsets screened by different feature screening methods was taken to obtain the final features.

5. The method for predicting the surgical margin status after HSIL surgery according to claim 1, characterized in that: The combined model consists of a logistic regression model, an extreme gradient boosting model, a random forest model and a support vector machine model.

6. A system for predicting the surgical margin status after HSIL surgery, characterized in that: include: A data acquisition module is used to obtain a data set, wherein the data set includes clinical data of several patients, colposcopic images and corresponding diagnostic labels; A deep learning network construction module is used to input each patient's colposcopic image into a risk prediction model for classification prediction to obtain a margin prediction score; wherein the risk prediction model includes a feature extraction module, a feature fusion module, and a prediction layer connected in sequence, and the feature extraction module includes a low-level feature representation extraction module and a global semantic information extraction module arranged in parallel; A combined model construction module is used to screen the final features from the clinical data and margin prediction scores according to the feature screening method; construct a combined model, which is composed of several machine learning models; train the combined model based on the selected final features, and perform verification and evaluation, and select the optimal machine learning model as the margin status prediction model based on the evaluation results; and perform the patient's margin status prediction task based on the margin status prediction model.

7. An electronic device, characterized in that: The electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute a method for predicting the postoperative resection margin status of HSIL according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that The computer program is stored therein, and when the computer program is executed by a processor, the method for predicting the resection margin status after HSIL surgery according to any one of claims 1 to 5 is implemented.