An artificial intelligence-based automatic identification method for gynecological screening abnormalities

Abstract

This invention discloses an automatic identification method for gynecological screening abnormalities based on artificial intelligence, belonging to the field of artificial intelligence technology. The method includes collecting multimodal raw data from gynecological screening and constructing a standardized raw screening dataset. The raw screening data is then preprocessed by category to obtain a preprocessed dataset. This invention constructs a complete process encompassing multimodal data acquisition and preprocessing, intelligent target segmentation, multi-layer feature fusion, automatic identification using an artificial intelligence model, and result verification report generation. Relying on a comprehensive architecture of standardized multimodal data processing, global-local fusion segmentation, and multi-feature weighted fusion intelligent identification, it eliminates excessive reliance on manual image reading, achieving fully automatic intelligent identification of gynecological screening abnormalities. This method adapts to the application needs of different screening devices and scenarios, improving the objectivity and stability of screening results.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically referring to an automatic identification method for gynecological screening abnormalities based on artificial intelligence. Background Technology

[0002] With the continuous development of medical technology, artificial intelligence technology is gradually being applied to the field of gynecological screening abnormality identification, aiming to solve the limitations of traditional manual screening methods.

[0003] However, existing automatic identification methods for gynecological screening abnormalities still have certain shortcomings. Current technologies rely on manual intervention, and the data collection stage lacks systematic integration and standardized labeling of multimodal data, making it difficult to provide comprehensive and reliable raw data support. Data preprocessing uses a single processing mode, which cannot adapt to the characteristics of different types of data, making it difficult to effectively eliminate noise and data bias, easily leading to identification errors in subsequent steps. Target region segmentation often uses a single algorithm, failing to balance global localization and local detail depiction, easily resulting in edge incompleteness and region offset, affecting the accuracy of feature extraction. Feature extraction methods are relatively simple, failing to achieve effective fusion of multi-layer semantic features, making it difficult to comprehensively and accurately represent lesion characteristics. Intelligent identification models lack effective combination of multi-scale adaptation and attention mechanisms, resulting in insufficient ability to identify hidden and atypical abnormal samples, easily leading to false positives and false negatives. Identification results lack multi-dimensional verification and targeted error correction mechanisms, making it difficult to guarantee the reliability of identification results. The final generated screening report lacks standardization and structure, failing to fully integrate various screening information and provide reference suggestions that meet clinical needs, making it difficult to meet the actual needs of physician diagnosis, patient retention, and clinical archiving. Therefore, this paper proposes an automatic identification method for gynecological screening abnormalities based on artificial intelligence. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic identification method for gynecological screening abnormalities based on artificial intelligence, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an automatic identification method for gynecological screening abnormalities based on artificial intelligence, comprising the following steps:

[0006] S1. Collect multimodal raw data of gynecological screening and construct a labeled raw screening dataset;

[0007] S2. Perform categorized preprocessing on the original screening data to eliminate various types of data noise, unify image format and size, correct abnormal medical record information, and screen valid data that meets quality standards to obtain the preprocessed dataset.

[0008] S3. A segmentation algorithm that integrates global and local optimization is used to segment the preprocessed image data into target regions, separating the lesion-related regions to be detected from the background regions, and obtaining a complete and clear target region image set.

[0009] S4. Extract multi-layer semantic features from the target region image, perform normalization on various features, and then perform weighted fusion to obtain high-dimensional fusion features that can accurately represent lesion features.

[0010] S5. Input the high-dimensional fusion features into the trained and optimized artificial intelligence recognition model, combine it with clinical medical record features to assist in the calculation, and output the preliminary recognition results with confidence.

[0011] S6. Verify the credibility of the preliminary identification results in multiple dimensions, correct errors for false positive and false negative samples, and combine manual review and optimization to obtain accurate identification results;

[0012] S7. Based on the accurate identification results, generate a standardized and structured screening report.

[0013] Preferably, in step S1, the collected cervical liquid-based cytology images are obtained using a sedimentation liquid-based preparation method, and then digitally scanned to generate whole-slice images. The collected gynecological ultrasound images cover key screening areas such as the ovaries, cervix, uterus, and adnexa. During the acquisition process, parameters such as the angle and scanning position of the ultrasound probe are recorded simultaneously. The acquisition process strictly follows aseptic operation procedures to avoid contamination of the cervical liquid-based sample with the ultrasound probe and cross-infection. At the same time, the acquisition time, acquisition equipment model, acquisition personnel, and testing institution information are recorded in detail. The constructed original screening dataset is divided into a training set, a validation set, and a test set, and the division process adopts a stratified sampling method. All datasets are labeled with corresponding medical diagnostic labels, and the label types are strictly divided according to the TBS diagnostic standards.

[0014] Preferably, in step S2, a differentiated preprocessing strategy is adopted for different types of raw data. Gaussian filtering is used to remove Gaussian noise from cervical liquid-based cytology images, median filtering is used to remove salt-and-pepper noise from gynecological ultrasound images, and all images are uniformly formatted by converting them into a specified format and using bilinear interpolation to uniformly adjust the image size.

[0015] Preferably, in step S2, grayscale normalization is performed on the image to map the image pixel values ​​to a specified range, eliminating pixel value deviations caused by different acquisition devices and different lighting conditions. Normalization is implemented as follows:

[0016] ,

[0017] In the formula, This represents the pixel value at coordinates xy in the original image. This represents the minimum pixel value in the original image. This represents the maximum pixel value in the original image. This represents the pixel value at coordinates xy after normalization. Quality screening is performed on the preprocessed image by setting image sharpness threshold, contrast threshold, and noise content threshold. Anomaly handling is performed on the clinical medical record information, and the mean imputation method is used to fill in missing numerical data. Outliers in the medical records are removed according to principles. At the same time, the medical record information is standardized and coded to finally obtain the preprocessed dataset.

[0018] Preferably, in step S3, the preprocessed normalized image is... The input is a segmentation model, which captures global semantic information of the image through a self-attention mechanism to quickly determine the approximate range of the target region and outputs a global target region mask. ;

[0019] The global feature map and mask output by the segmentation model will be used. The U-Net model is input together, and the encoder-decoder structure and skip connections of U-Net are used to accurately capture the local details of the target region, correct the bias in global segmentation, and output a preliminary segmentation mask. The preprocessed image Mapped to an undirected graph V is the set of image pixels, E is the set of edges formed by adjacent pixels, and the weight of the edges is... Determined by the Euclidean distance of pixel grayscale values, superpixel blocks containing the target region are selected based on an internal difference threshold, forming the initial segmentation mask. Corrections are made; morphological processing is performed on the segmented target region to repair edge defects, resulting in a complete and clear target region image set and the final segmentation mask. .

[0020] Preferably, in step S4, the target region image and the final segmentation mask are used. As input, multi-layer semantic features of the target region are extracted, and texture features of the target region are extracted. The gray-level co-occurrence matrix is ​​used to extract texture feature parameters, including four core parameters: contrast, correlation, energy, and entropy. Among these, the entropy value... Used to characterize the complexity of the texture of the target region, the entropy value of the lesion region differs significantly from that of the normal region, implemented as follows:

[0021] ,

[0022] In the formula, This represents the gray levels of the gray-level co-occurrence matrix. This represents the element value in the i-th row and j-th column of the gray-level co-occurrence matrix, used to extract morphological features of the target region, including the area of ​​the target region. perimeter Circularity Parameters such as aspect ratio and contour complexity, including roundness. Used to characterize the shape regularity of the target area, the roundness of diseased cells usually differs significantly from that of normal cells.

[0023] Preferably, in step S4, an improved visual Transformer model is used to extract the depth features of the target region. The target region image is divided into image blocks of a specified size. Each image block is converted into a feature vector through an embedding layer and input into a Transformer encoder to extract multiple layers of depth features. Each layer of features captures the dependencies between image blocks through a self-attention mechanism, highlighting the feature differences of the lesion region.

[0024] The extracted texture features, morphological features, and depth features are normalized, mapping all features to the same dimension. A weighted fusion algorithm is then used to fuse the three features into a high-dimensional fusion feature. The implementation is as follows:

[0025] ,

[0026] In the formula, This represents the normalized texture feature vector. This represents the normalized morphological feature vector. This represents the normalized deep feature vector. , , These are the weighting coefficients for the three features.

[0027] Preferably, in step S5, high-dimensional fusion features are... The input model employs the multi-scale MIL method to process the features corresponding to the two magnification factors separately, generating feature embedding vectors at two scales.

[0028] Feature embeddings from two scales are concatenated to form a feature pyramid, which is then input into the MIL aggregator. An attention mechanism is used to weight the features, highlighting information relevant to abnormal lesions and suppressing irrelevant features. During model training, the cross-entropy loss function is used to calculate the error between the model's predicted values ​​and the true labels. Implementation as follows:

[0029] ,

[0030] In the formula, Indicates the number of training samples. This represents the true label of the i-th sample. This represents the model's predicted probability for the i-th sample, and the model parameters are updated using the Adam optimization algorithm.

[0031] During model training, an early stopping strategy is employed to prevent overfitting. Model training is stopped when the loss function on the validation set fails to decrease for several consecutive training epochs, and the optimal model parameters are saved. After model training is complete, high-dimensional fusion features from the test set are used. The trained model is input along with the standardized clinical case feature vector. The model performs multi-feature fusion operations and outputs the anomaly confidence and anomaly type prediction results for each sample, thus obtaining preliminary identification results.

[0032] Preferably, in step S6, the preliminary identification results are subjected to multi-dimensional credibility verification, and the confusion matrix between the preliminary identification results and the real labels in the test set is calculated to obtain the precision, recall, and accuracy evaluation metrics. This is used to measure the model's ability to identify anomalous samples, and is implemented as follows:

[0033] ,

[0034] In the formula, This indicates the number of true positive samples. This indicates the number of false negative samples. For false positive samples, it combines the extracted high-dimensional fusion features. Calculate the cosine similarity of features between false positive samples and normal samples. False positive samples with similarity reaching a set value are removed;

[0035] For false negative samples, the trained AI recognition model is re-inputted, the model's confidence threshold is adjusted, and the recognition is performed again. If it is still recognized as normal, then the clinical medical record information and the target area image are combined for manual review by a physician with gynecological clinical diagnostic experience to confirm whether it is an abnormal sample. The consistency test is performed on the corrected recognition results to obtain the accurate result.

[0036] Preferably, in step S7, a standardized and structured gynecological screening report is generated based on the accurate identification results. The report strictly follows clinical reporting standards and is generated in a specified format.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] 1. This invention constructs a complete process of multimodal data acquisition and preprocessing, intelligent target segmentation, multi-layer feature fusion, automatic identification by artificial intelligence model, and generation of result verification reports. Relying on the overall architecture of multimodal data standardization processing, global and local fusion segmentation, and multi-feature weighted fusion intelligent identification, it gets rid of excessive reliance on manual image reading, realizes fully automatic intelligent identification of gynecological screening abnormalities, adapts to the application needs of different screening equipment and screening scenarios, and improves the objectivity and stability of screening results.

[0039] 2. This invention employs differentiated preprocessing methods for various types of data, including gynecological screening images and clinical medical records. It selects appropriate filtering methods for different types of image noise, while standardizing image format and size, performing grayscale normalization and data quality screening, and standardizing and correcting missing and abnormal information in medical records. This effectively removes irrelevant noise interference from the original data, unifies data differences caused by different acquisition devices, eliminates unqualified and invalid data, standardizes the structure of clinical medical record information, and ensures the reliability of the data foundation for the entire intelligent recognition process.

[0040] 3. This invention employs a global and local dual-architecture segmentation method that integrates the segmentation model and the U-Net model, combined with superpixel segmentation and morphological post-processing to achieve accurate segmentation of the target region. At the same time, it takes into account both global semantic localization and accurate depiction of local details, accurately separating the lesion target region and background region in gynecological screening images, fully preserving the target contour and detailed features of key screening sites, reducing the interference of irrelevant background information on feature extraction, and improving the purity and relevance of subsequent feature representation.

[0041] 4. This invention extracts deep multi-layer semantic features of texture morphology simultaneously and performs standardized weighted fusion. Then, it introduces an artificial intelligence model with multi-scale, multi-instance learning and attention mechanism to carry out anomaly recognition. It completes the preliminary discrimination by relying on feature association operation and model parameter iterative optimization. It comprehensively depicts the texture morphology and deep semantic characteristics of the lesion area from multiple dimensions. Through feature weighted fusion, it highlights the weight of lesion-related features and suppresses redundant feature interference. At the same time, it uses a multi-scale model architecture to adapt to the differences in lesion features under different magnification perspectives, which enhances the model's ability to identify hidden and atypical gynecological abnormality samples. It does not require manual setting of discrimination rules and learns the inherent lesion feature rules of the data autonomously, which greatly reduces the degree of manual intervention and improves the sensitivity and universality of anomaly recognition. Attached Figure Description

[0042] Figure 1 The present invention describes the operation flow of an artificial intelligence-based automatic identification method for gynecological screening abnormalities. Figure 1 ;

[0043] Figure 2 The present invention describes the operation flow of an artificial intelligence-based automatic identification method for gynecological screening abnormalities. Figure 2 ;

[0044] Figure 3 The present invention describes the operation flow of an artificial intelligence-based automatic identification method for gynecological screening abnormalities. Figure 3 ;

[0045] Figure 4 The present invention describes the operation flow of an artificial intelligence-based automatic identification method for gynecological screening abnormalities. Figure 4 . Detailed Implementation

[0046] 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.

[0047] Example

[0048] Please see Figures 1-4 As shown, the present invention provides a technical solution comprising the following steps:

[0049] S1. Collect multimodal raw data for gynecological screening, including whole slide images of cervical liquid-based cytology, ultrasound images of key gynecological sites, and corresponding complete clinical medical record information, and construct a standardized raw screening dataset.

[0050] S2. Perform categorized preprocessing on the original screening data to eliminate various types of data noise, unify image format and size, correct abnormal medical record information, and screen valid data that meets quality standards to obtain the preprocessed dataset.

[0051] S3. A segmentation algorithm that integrates global and local optimization is used to segment the preprocessed image data into target regions, separating the lesion-related regions to be detected from the background regions, and obtaining a complete and clear target region image set.

[0052] S4. Extract multi-layer semantic features from the target region image, perform normalization on various features, and then perform weighted fusion to obtain high-dimensional fusion features that can accurately represent lesion features.

[0053] S5. Input the high-dimensional fusion features into the trained and optimized artificial intelligence recognition model, combine it with clinical medical record features to assist in the calculation, and output the preliminary recognition results with confidence.

[0054] S6. Verify the credibility of the preliminary identification results in multiple dimensions, correct errors for false positive and false negative samples, and combine manual review and optimization to obtain accurate identification results;

[0055] S7. Based on the accurate identification results, generate a standardized and structured screening report.

[0056] In this embodiment, in S1, the cervical liquid-based cytology images are obtained by sedimentation liquid-based preparation method, and full-slice images are generated by digital scanning. The acquired gynecological ultrasound images cover key screening sites such as ovaries, cervix, uterine body and adnexa. Parameters such as ultrasound probe angle and scanning position are recorded simultaneously during the acquisition process.

[0057] Clinical medical record information includes core contents such as basic patient information, gynecological history, menstrual cycle information, HPV test results, and previous screening records; the collection process strictly follows aseptic operation procedures to avoid contamination of cervical liquid-based samples and cross-infection of ultrasound probes, and at the same time, the collection time, collection equipment model, collection personnel and testing institution information are recorded in detail; the constructed original screening dataset is divided into training set, validation set and test set, and the division process adopts stratified sampling to ensure that the distribution of samples of different lesion types and different age groups is uniform within each dataset;

[0058] All datasets are labeled with corresponding medical diagnostic tags. The tag types are strictly classified according to the TBS diagnostic criteria, including normal low-grade squamous intraepithelial lesions, high-grade squamous intraepithelial lesions, squamous cell carcinoma, and glandular cell abnormalities. The labeling work is completed by licensed physicians with gynecological clinical diagnostic experience, and each labeled sample is reviewed and confirmed by at least two senior gynecologists.

[0059] In this embodiment, in S2, a differentiated preprocessing strategy is adopted for different types of raw data. Gaussian filtering is used to remove Gaussian noise from cervical liquid-based cytology images, and median filtering is used to remove salt-and-pepper noise from gynecological ultrasound images. The size of the filter kernel is adaptively adjusted according to the image noise intensity. During the filtering process, the lesion details in the image are preserved, and feature loss caused by over-filtering is avoided.

[0060] All images are processed to unify their format, converting them to a specified format and using bilinear interpolation to adjust their size uniformly. This ensures consistency of image data input to subsequent algorithms and eliminates size differences caused by different acquisition devices.

[0061] In this embodiment, in step S2, grayscale normalization processing is performed on the image to map the image pixel values ​​to a specified range, eliminating pixel value deviations caused by different acquisition devices and different lighting conditions. The normalization is implemented as follows:

[0062] ,

[0063] In the formula, This represents the pixel value at coordinates xy in the original image. This represents the minimum pixel value in the original image. This represents the maximum pixel value in the original image. This represents the pixel value at coordinates xy after normalization.

[0064] Specifically, the preprocessed images undergo quality screening by setting image sharpness thresholds, contrast thresholds, and noise content thresholds. Computer vision algorithms are used to automatically detect and remove blurry images with insufficient contrast or excessive noise. Anomaly handling is performed on clinical medical record information by using mean imputation to fill in missing numerical data and removing outliers from medical records according to principles. At the same time, the medical record information is standardized and coded to obtain the preprocessed dataset.

[0065] In this embodiment, in step S3, the preprocessed normalized image is... The input is a segmentation model, which captures global semantic information of the image through a self-attention mechanism to quickly determine the approximate range of the target region and outputs a global target region mask. ;

[0066] The global feature map and mask output by the segmentation model will be used. The target region is input into a U-Net model, and the encoder-decoder structure and skip connections of U-Net are used to accurately capture local details of the target region, correct biases in global segmentation, achieve accurate segmentation of the target region and background region, and output a preliminary segmentation mask. During the segmentation process, a graph-based superpixel segmentation algorithm is used to assist in optimization.

[0067] Preprocessed image Mapped to an undirected graph V is the set of image pixels, E is the set of edges formed by adjacent pixels, and the weight of the edges is... Determined by the Euclidean distance of pixel grayscale values, it is implemented as follows:

[0068] ,

[0069] In the formula, Each pixel coordinates Each pixel coordinates For pixels The normalized gray value, For pixels The normalized grayscale value; by generating a minimum spanning tree, pixels are aggregated into superpixel blocks, and the internal differences of each superpixel block are calculated. ,in Let c be the minimum spanning tree. Let e ​​be the weight of edge e;

[0070] Superpixel blocks containing the target region are selected based on the internal difference threshold, and the initial segmentation mask is used. Corrections are made; morphological processing is performed on the segmented target region. First, dilation is used to fill small holes inside the target region, and then erosion is used to remove small noise areas at the edges of the target region, repairing edge defects and obtaining a complete and clear target region image set and the final segmentation mask. The target area precisely covers key screening sites such as cervical cells, ovaries, and uterus.

[0071] In this embodiment, in step S4, the target region image and the final segmentation mask are used. As input, multi-layer semantic features of the target region are extracted, and texture features of the target region are extracted. The gray-level co-occurrence matrix is ​​used to extract texture feature parameters, including four core parameters: contrast, correlation, energy, and entropy. Among these, the entropy value... Used to characterize the complexity of the texture of the target region, the entropy value of the lesion region differs significantly from that of the normal region, implemented as follows:

[0072] ,

[0073] In the formula, This represents the gray levels of the gray-level co-occurrence matrix. This represents the element value in the i-th row and j-th column of the gray-level co-occurrence matrix, i.e., the probability that pixels with gray values ​​i and j appear adjacently within the target region. The calculation only considers the segmentation mask. Target area pixels within the range, excluding background interference;

[0074] Extract morphological features of the target region, including the area of ​​the target region. perimeter Circularity Parameters such as aspect ratio and contour complexity, including roundness. Used to characterize the shape regularity of the target region, the roundness of diseased cells usually differs significantly from that of normal cells, as achieved by:

[0075] ,

[0076] In the formula, Indicates the area of ​​the target region. This indicates the perimeter of the target area.

[0077] In this embodiment, in step S4, an improved visual Transformer model is used to extract the depth features of the target region. The target region image is divided into image blocks of a specified size. Each image block is converted into a feature vector through an embedding layer and input into the Transformer encoder to extract multi-layer depth features. Each layer of features captures the dependency relationship between image blocks through a self-attention mechanism, highlighting the feature differences of the lesion region.

[0078] The extracted texture features, morphological features, and depth features are normalized, mapping all features to the same dimension. A weighted fusion algorithm is then used to fuse the three features into a high-dimensional fusion feature. The implementation is as follows:

[0079] ,

[0080] In the formula, This represents the normalized texture feature vector. This represents the normalized morphological feature vector. This represents the normalized deep feature vector. , , These are the weight coefficients for the three features, and The weight coefficients were determined by grid search on the validation set, with the optimization objective being to minimize the model recognition error. The fused high-dimensional features were then used for subsequent anomaly identification.

[0081] In this embodiment, in step S5, the artificial intelligence recognition model adopts a deep learning model that integrates multi-scale MIL and attention mechanisms. The model training process uses high-dimensional fusion features. Using clinical medical record features as the core input, and combining them with training to assist in improving recognition accuracy;

[0082] High-dimensional fusion features The input model employs the multi-scale MIL method to process the features corresponding to the two magnification factors separately, generating feature embedding vectors at two scales.

[0083] Feature embeddings from two scales are concatenated to form a feature pyramid, which is then input into the MIL aggregator. An attention mechanism is used to weight the features, highlighting information relevant to abnormal lesions and suppressing irrelevant features. During model training, the cross-entropy loss function is used to calculate the error between the model's predicted values ​​and the true labels. Implementation as follows:

[0084] ,

[0085] In the formula, Indicates the number of training samples. This represents the true label of the i-th sample. This represents the model's predicted probability for the i-th sample, i.e., the confidence level of the anomaly in the initial identification result output by the model. The Adam optimization algorithm is used to update the model parameters, and the parameter update is implemented as follows:

[0086] ,

[0087] In the formula, This represents the updated model parameters. Indicates the current model parameters. Indicates the learning rate. This represents the gradient of the cross-entropy loss function under the current parameters;

[0088] During model training, an early stopping strategy is employed to prevent overfitting. Model training is stopped when the loss function on the validation set fails to decrease for several consecutive training epochs, and the optimal model parameters are saved. After model training is complete, high-dimensional fusion features from the test set are used. The trained model is input along with the standardized clinical medical record feature vector. The model performs multi-feature fusion operations and outputs the abnormality confidence and abnormality type prediction results for each sample to obtain preliminary identification results. The abnormality confidence is set with a threshold. When the confidence reaches the threshold, it is judged as a preliminary abnormality; otherwise, it is judged as normal. At the same time, the coordinates of the abnormal area of ​​each preliminary abnormal sample are recorded.

[0089] In this embodiment, in step S6, the initial identification results are verified for credibility across multiple dimensions. The confusion matrix between the initial identification results and the real labels in the test set is calculated to obtain three core evaluation metrics: precision, recall, and accuracy. To measure the model's ability to identify anomalous samples and avoid missed diagnoses, it is implemented as follows:

[0090] ,

[0091] In the formula, This indicates the number of true positive samples. This indicates the number of false negative samples, the sample type determination using reused labels, and the preliminary identification results;

[0092] For false positive samples, combine the extracted high-dimensional fusion features Calculate the cosine similarity of features between false positive samples and normal samples. False positive samples with similarity reaching a set value are removed. Cosine similarity is achieved as follows:

[0093] ,

[0094] In the formula, This represents the high-dimensional fusion feature vector of a false positive sample. This represents a high-dimensional fusion feature vector of a normal sample. , These represent the magnitudes of the corresponding eigenvectors;

[0095] For false negative samples, the trained AI recognition model is re-inputted, the model's confidence threshold is adjusted, and the recognition is performed again. If it is still recognized as normal, it is manually reviewed by a physician with gynecological clinical diagnostic experience, combining clinical medical record information and target area images, to confirm whether it is an abnormal sample. The consistency of the corrected recognition results is tested using the Kappa coefficient. When the Kappa coefficient reaches the set value, it is determined to be an accurate recognition result; otherwise, the error correction is repeated until the consistency requirements are met, and an accurate result is obtained.

[0096] In this embodiment, in step S7, a standardized and structured gynecological screening report is generated based on the accurate identification results. The report strictly follows clinical reporting standards and covers four core parts: patient basic information, screening data, identification results, and clinical reference suggestions.

[0097] Patient basic information includes patient name, age, ID number, contact information, screening date, etc., and the collected information can be reused.

[0098] The screening data information includes the collection time, collection device model, screening location, original data type, key parameters of the preprocessing and segmentation process, and reuses relevant information from S1-S3.

[0099] The identification results include abnormality type, abnormal location, abnormal area, identification confidence level, identification time, etc., and also include a visually labeled image of the abnormal area, clearly showing the location and extent of the abnormality to assist physicians in clinical diagnosis. Clinical reference suggestions are generated specifically according to the abnormality type and confidence level. For samples with low-grade squamous intraepithelial lesions and confidence levels within the set range, it is recommended to have regular follow-up examinations and monitor lesion changes regularly.

[0100] For high-grade squamous intraepithelial lesions, squamous cell carcinoma with abnormal glandular cells, and low-grade lesion samples with confidence levels reaching the set value, it is recommended to conduct further targeted examinations such as pathological biopsy and HPV typing, and to develop a treatment plan based on clinical symptoms.

[0101] For normal samples, it is recommended to conduct routine gynecological screenings regularly to prevent diseases. The report is generated in a specified format, supports the addition of physician electronic signatures, and also supports export, printing, and encrypted storage functions, making it convenient for physicians to view, retain, and archive for clinical records.

[0102] Working Principle: This system collects various types of raw gynecological screening data, including whole-slice images of cervical liquid-based cytology, ultrasound images of key gynecological sites, and corresponding complete clinical medical records. It integrates various basic data related to gynecological screening to construct a standardized raw screening dataset. The collection process follows standardized procedures to ensure the integrity and standardization of the collected data. The dataset is rationally divided, and medical diagnostic labels are annotated and reviewed. For the collected multimodal raw data, a categorized preprocessing strategy is adopted, with differentiated processing methods for different data types. Image data undergoes noise removal processing, using appropriate filtering methods to eliminate various types of noise. Image format and size are standardized, and grayscale normalization is performed to eliminate data deviations caused by different collection conditions. Unqualified images are then removed through quality screening. Anomalies in clinical medical record information are handled. Missing data is filled, outliers are removed, and standardized coding is performed. Finally, valid data meeting quality standards are selected to form a preprocessed dataset. Using the preprocessed image data as input, a segmentation algorithm integrating global and local optimization is employed to segment the target region. The normalized image is input into the segmentation model, which uses a self-attention mechanism to capture global semantic information, determine the approximate range of the target region, and output a global mask. The global feature map and mask are then input into the U-Net model, leveraging its structural advantages to capture local detail features, correct segmentation biases, and output a preliminary mask. Combined with a graph-based superpixel segmentation algorithm for further optimization, pixel aggregation and internal difference calculation are used to filter relevant superpixel blocks in the target region, correcting the preliminary mask. Finally, morphological processing is used to repair edge defects in the target region, separating lesion-related regions from the background region, resulting in a complete and clear image set of the target region.

[0103] This model extracts multi-layer semantic features of the target region using the target region image and the final segmentation mask as input. It employs a gray-level co-occurrence matrix to extract texture feature parameters, capturing the texture characteristics of the target region; extracts morphological feature parameters to characterize the shape features of the target region; and uses an improved visual Transformer model to extract depth features, capturing deep semantic information and lesion-related feature differences. The extracted features are standardized, mapping all features to the same dimension, and a weighted fusion algorithm is used to fuse the three features, resulting in a high-dimensional fusion feature that accurately represents lesion characteristics. This high-dimensional fusion feature is input into the trained and optimized AI recognition model, and combined with standardized clinical medical record features to assist in computation, improving recognition accuracy. The model uses a structure that integrates multi-scale, multi-instance learning and attention mechanisms. The high-dimensional fusion feature is processed at different magnification levels to generate multi-scale feature embedding vectors, which are then concatenated to form a feature pyramid. The attention mechanism highlights lesion-related features and suppresses irrelevant interference. During model training, the prediction error is calculated using a loss function, and the model parameters are updated using an optimization algorithm. An early stopping strategy is used to prevent overfitting. After training, the abnormality confidence and abnormality type prediction results for each sample are output, yielding preliminary recognition results. The preliminary recognition results are then analyzed. Multi-dimensional credibility verification is performed, and core evaluation indicators are obtained by calculating the confusion matrix to measure the reliability of the preliminary identification results. For false positive samples, the similarity between the extracted high-dimensional fusion features and the features of normal samples is calculated, and false positive samples that meet the set similarity requirements are removed to reduce misjudgments. For false negative samples, the model is re-inputted and the confidence threshold is adjusted for re-identification. If they are still judged as normal, they are manually reviewed in conjunction with clinical medical record information and target area images to confirm the sample type. The corrected identification results are subjected to consistency testing until the set consistency requirements are met to obtain accurate identification results. The accurate identification results are used as the core... Following clinical reporting guidelines, the system generates standardized and structured gynecological screening reports. These reports integrate multiple information sources, reusing collected information as basic patient information, reusing relevant information from S1 to S3 as screening data, and reusing relevant content from precise identification results as identification outcome information. They also include visually labeled images of abnormal areas. Based on the type and confidence level of the abnormality, targeted clinical reference suggestions are generated to adapt to the clinical needs of different samples. The reports are generated in a specified format and support electronic signatures, export, printing, and encrypted storage, meeting the needs of physicians for review, patients for retention, and clinical archiving, thus completing the entire automated gynecological screening abnormality identification process.

[0104] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.

[0105] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An automatic identification method for gynecological screening abnormalities based on artificial intelligence, characterized in that, Includes the following steps: S1. Collect multimodal raw data of gynecological screening and construct a labeled raw screening dataset; S2. Perform categorized preprocessing on the original screening data to obtain the preprocessed dataset; S3. A segmentation algorithm that integrates global and local optimization is used to segment the preprocessed image data into target regions, separating the lesion-related regions to be detected from the background regions, and obtaining a complete and clear target region image set. S4. Extract multi-layer semantic features from the target region image, perform normalization on various features, and then perform weighted fusion to obtain high-dimensional fusion features that can accurately represent lesion features. S5. Input the high-dimensional fusion features into the trained and optimized artificial intelligence recognition model, combine it with clinical medical record features to assist in the calculation, and output the preliminary recognition results with confidence. S6. Verify the credibility of the preliminary identification results in multiple dimensions, correct errors for false positive and false negative samples, and combine manual review and optimization to obtain accurate identification results; S7. Based on the accurate identification results, generate a standardized and structured screening report.

2. The method for automatic identification of abnormalities in gynecological screening based on artificial intelligence according to claim 1, characterized in that: In S1, the collected cervical liquid-based cytology images are obtained using the sedimentation liquid-based preparation method and generated into whole-slice images through digital scanning. The collected gynecological ultrasound images cover key screening sites such as the ovaries, cervix, uterus, and adnexa. During the acquisition process, the ultrasound probe angle and scanning position parameters are recorded simultaneously. The acquisition process strictly follows aseptic operation procedures, and the acquisition time, acquisition equipment model, acquisition personnel, and testing institution information are recorded in detail. The constructed original screening dataset is divided into training set, validation set, and test set. All datasets are labeled with corresponding medical diagnostic labels, and the label types are strictly classified according to the TBS diagnostic standards.

3. The method for automatic identification of abnormalities in gynecological screening based on artificial intelligence according to claim 1, characterized in that: In S2, a differentiated preprocessing strategy is adopted for different types of raw data. Gaussian filtering is used to remove Gaussian noise from cervical liquid-based cytology images, median filtering is used to remove salt-and-pepper noise from gynecological ultrasound images, and all images are uniformly formatted by converting them into a specified format and using bilinear interpolation to uniformly adjust the image size.

4. The method for automatic identification of abnormalities in gynecological screening based on artificial intelligence according to claim 1, characterized in that: In step S2, the image undergoes grayscale normalization processing, mapping the image pixel values ​​to a specified range to eliminate pixel value deviations caused by different acquisition devices and different lighting conditions. The normalization is implemented as follows: ， In the formula, This represents the pixel value at coordinates xy in the original image. This represents the minimum pixel value in the original image. This represents the maximum pixel value in the original image. This represents the pixel value at coordinates xy after normalization. Quality screening is performed on the preprocessed image by setting image sharpness threshold, contrast threshold, and noise content threshold. Anomaly handling is performed on the clinical medical record information, and the mean imputation method is used to fill in missing numerical data. Outliers in the medical records are removed according to principles. At the same time, the medical record information is standardized and coded to finally obtain the preprocessed dataset.

5. The method for automatic identification of abnormalities in gynecological screening based on artificial intelligence according to claim 1, characterized in that: In step S3, the preprocessed normalized image Input the whole segmentation model, capture global semantic information of the image through a self-attention mechanism, quickly determine the approximate range of the target region, and output the global target region mask; The global feature map and mask output from the segmentation model are input together into the U-Net model. Utilizing U-Net's encoder-decoder structure and skip connections, the local detail features of the target region are accurately captured, correcting biases in global segmentation, and outputting a preliminary segmentation mask. The preprocessed image is then processed. Mapped to an undirected graph V is the set of image pixels, E is the set of edges formed by adjacent pixels, and the weight of the edges is... Determined by the Euclidean distance of pixel grayscale values, superpixel blocks containing the target region are selected based on the internal difference threshold, and the initial segmentation mask is corrected. Morphological processing is performed on the segmented target region to repair edge defects in the target region, resulting in a complete and clear target region image set and the final segmentation mask.

6. The method for automatic identification of abnormalities in gynecological screening based on artificial intelligence according to claim 1, characterized in that: In step S4, the target region image and the final segmentation mask are used as input to extract multi-layer semantic features and texture features of the target region. Texture feature parameters are extracted using a gray-level co-occurrence matrix, including four core parameters: contrast correlation, energy, and entropy. Among these, the entropy value... Used to characterize the complexity of the texture of the target region, the entropy value of the lesion region differs significantly from that of the normal region, implemented as follows: ， In the formula, This represents the gray levels of the gray-level co-occurrence matrix. This represents the element value in the i-th row and j-th column of the gray-level co-occurrence matrix, used to extract morphological features of the target region, including the area of ​​the target region. perimeter Circularity Parameters such as aspect ratio and contour complexity, including roundness. Used to characterize the shape regularity of the target area, the roundness of diseased cells usually differs significantly from that of normal cells.

7. The method for automatic identification of abnormalities in gynecological screening based on artificial intelligence according to claim 6, characterized in that: In S4, an improved visual Transformer model is used to extract the depth features of the target region. The target region image is divided into image blocks of a specified size. Each image block is converted into a feature vector through an embedding layer and input into the Transformer encoder to extract multi-layer depth features. Each layer of features captures the dependency relationship between image blocks through a self-attention mechanism, highlighting the feature differences of the lesion region. The extracted texture features, morphological features, and depth features are normalized, mapping all features to the same dimension. A weighted fusion algorithm is then used to fuse the three features into a high-dimensional fusion feature. The implementation is as follows: ， In the formula, This represents the normalized texture feature vector. This represents the normalized morphological feature vector. This represents the normalized deep feature vector. , , These are the weighting coefficients for the three features.

8. The method for automatic identification of abnormalities in gynecological screening based on artificial intelligence according to claim 1, characterized in that: In S5, high-dimensional fusion features are... The input model employs the multi-scale MIL method to process the features corresponding to the two magnification factors separately, generating feature embedding vectors at two scales. Feature embeddings from two scales are concatenated to form a feature pyramid, which is then input into the MIL aggregator. An attention mechanism is used to weight the features, highlighting information relevant to abnormal lesions and suppressing irrelevant features. During model training, the cross-entropy loss function is used to calculate the error between the model's predicted values ​​and the true labels. Implementation as follows: ， In the formula, Indicates the number of training samples. This represents the true label of the i-th sample. This represents the model's predicted probability for the i-th sample, and the model parameters are updated using the Adam optimization algorithm. During model training, an early stopping strategy is employed to prevent overfitting. Model training is stopped when the loss function on the validation set fails to decrease for several consecutive training epochs, and the optimal model parameters are saved. After model training is complete, high-dimensional fusion features from the test set are used. The trained model is input along with the standardized clinical case feature vector. The model performs multi-feature fusion operations and outputs the anomaly confidence and anomaly type prediction results for each sample, thus obtaining preliminary identification results.

9. The method for automatic identification of abnormalities in gynecological screening based on artificial intelligence according to claim 1, characterized in that: In step S6, the preliminary identification results are verified for credibility across multiple dimensions. The confusion matrix between the preliminary identification results and the real labels in the test set is calculated to obtain the precision, recall, and accuracy evaluation metrics. This is used to measure the model's ability to identify anomalous samples, and is implemented as follows: ， In the formula, This indicates the number of true positive samples. This indicates the number of false negative samples. For false positive samples, it combines the extracted high-dimensional fusion features. Calculate the cosine similarity of features between false positive samples and normal samples. False positive samples with similarity reaching a set value are removed; For false negative samples, the trained AI recognition model is re-inputted, the model's confidence threshold is adjusted, and the recognition is performed again. If it is still recognized as normal, then the clinical medical record information and the target area image are combined for manual review by a physician with gynecological clinical diagnostic experience to confirm whether it is an abnormal sample. The consistency test is performed on the corrected recognition results to obtain the accurate result.

10. The method for automatic identification of abnormalities in gynecological screening based on artificial intelligence according to claim 1, characterized in that: In S7, a standardized and structured gynecological screening report is generated based on the accurate identification results. The report strictly follows clinical reporting standards and is generated in a specified format.

Images