Breast cancer individual case management data extraction and analysis method based on natural language processing

Abstract

The invention relates to the technical field of medical data processing and case management, and discloses a breast cancer case management data extraction and analysis method based on natural language processing. According to the method, an electronic medical record text and a CT image are collected, the text is subjected to de-identification, term normalization and time standardization, and the image is subjected to intensity normalization and symmetry segmentation. And inputting the preprocessed text into the joint extraction model, and optimizing and outputting entities, relationships and events through a weighted loss function. Normal and abnormal classification and abnormal area positioning of the CT image are completed through an image pipeline. And then performing consistency alignment on the text and image results based on the side identification and the time anchor point to obtain a cross-modal structured result, and finally writing the cross-modal structured result into a database or a knowledge graph for calling in combination with a rule base. According to the method, the intelligent level and reliability of breast cancer individual case management are improved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing, and in particular to a method for extracting and analyzing breast cancer case management data based on natural language processing. Background Technology

[0002] With the advancement of modern medical technology, patient case management is gradually shifting from traditional manual recording and diagnosis to an intelligent and data-driven management model. In breast cancer case management, a large amount of information is involved, including the patient's electronic medical records, imaging data, and follow-up records. Traditional methods for extracting and analyzing this multimodal data largely rely on manual processing and basic rule engines, resulting in low processing efficiency and poor accuracy.

[0003] During data extraction, traditional information extraction techniques are unable to comprehensively and accurately identify and extract key information from breast cancer cases, particularly in the extraction of complex medical terminology, patient medical history, and treatment pathways. This leads to incomplete or erroneous data, affecting subsequent analysis and decision-making. Furthermore, existing multimodal data fusion methods are insufficient for imaging data. Systems typically lack effective text-image fusion techniques, failing to simultaneously process and correlate medical record text and imaging data. This reduces overall accuracy and prevents the full utilization of this data. Summary of the Invention

[0004] The purpose of this invention is to provide a method for extracting and analyzing breast cancer case management data based on natural language processing and medical image analysis. This method can effectively process and analyze multimodal data in breast cancer cases, including electronic medical record text and medical image data. Through precise text entity extraction, image anomaly detection, and data fusion techniques, this invention can improve the efficiency and accuracy of breast cancer case management.

[0005] To achieve the above objectives, this method includes:

[0006] S1: Collect multimodal case data, including electronic medical record text and medical CT images;

[0007] S2: Performs de-identification, cleaning, segmentation and sentence division, terminology and unit normalization, and time expression standardization on electronic medical record text; performs intensity normalization, geometric correction, and left and right breast region segmentation based on anatomical symmetry on CT images;

[0008] S3: The joint extraction model for preprocessed text input specifically includes:

[0009] S3.1 Low-level feature extraction, performing sequence encoding on character / word / subword multi-channel embedding to obtain long-short dependency representation;

[0010] S3.2 High-level feature extraction: A multi-channel self-attention mechanism is used to compute multiple sets of attention in parallel for the same sentence and fuse them with learnable channel weights to obtain sentence-level semantic representation.

[0011] S3.3 introduces a weighted learning loss function, where the weights of each class are used as parameter vectors and participate in training together with the cross-entropy loss function.

[0012] S3.4 Joint discrimination outputs entities, relationships between entities, event types, and trigger words related to case management;

[0013] S4: Input CT images into the imaging collaboration pipeline to obtain and label the normal / abnormal conclusions on the imaging side;

[0014] S5: Based on lateralization, anatomical partitioning, and time anchors, the text-side and image-side results are aligned for consistency, a consistency score is calculated, and the confidence of entities, relationships, and events in the text-side are fused and decoded accordingly to obtain cross-modal consistent case structured results;

[0015] S6: Based on the time standardization results, perform time-series alignment and sorting of events to construct patient-level case timelines;

[0016] S7: Calculate and determine case management indicators according to the rule base, including at least the time from diagnosis to treatment initiation, standard treatment achievement, follow-up compliance, occurrence of adverse events and determination of recurrence / metastasis;

[0017] S8: Write the merged entities, relationships, events, timelines, and indicator results into a structured format into a database or knowledge graph, and output interface data for the case management system to call.

[0018] The multichannel self-attention mechanism specifically includes the following processing steps:

[0019] S3.2.1: Use an LSTM model to concatenate the learned long and short distance dependency information and output the hidden layer vector: Z_t = <z1,z2,z3,…,Z n >;

[0020] S3.2.2: Z_t serves as the input for multi-channel self-attention. The multi-channel self-attention mechanism is given the number of channels C, and in each channel C... i The self-attention weights are calculated once in each step, specifically in the following form:

[0021]

[0022] Here, α is a hyperparameter used to adjust the position embedding weights; softmax is a normalization function that converts the output of all channels into a probability distribution; Refers to channel C iThe set of learnable weight matrices in the LSTM includes the number of channels C as a key hyperparameter, which needs to be adjusted according to the actual business scenario; tanh is an activation function that performs a nonlinear transformation on the result of matrix multiplication; W_z is a learnable weight matrix used to perform a linear transformation on the hidden state output by the LSTM. It is the transpose of the LSTM output matrix, which, after matrix multiplication, yields a new vector representation;

[0023] S3.2.3: Add constraint terms:

[0024]

[0025] Where ||*|| represents the Frobenius norm of the matrix; the penalty term ||*|| is used as part of the loss function to train the model together. For each additional channel, the multi-channel self-attention mechanism will perform weighted fitting on the component i in the sentence.

[0026] S3.2.4: Calculate the weight matrix using a two-layer perceptron Will With Z t Multiply and then normalize to obtain the output H of the self-attention layer. * The H * It is a vector representation of sentence-level semantic features.

[0027] Joint discrimination employs sequence labeling and classification joint learning, named entity recognition uses conditional random fields for sequence decoding, and the probability distribution of relation and event types is output by the classification head and shares the low-level and high-level representations with the entity boundaries.

[0028] The loss function for weighted learning is the cross-entropy with class weights, where the class weights are calculated based on class frequency or the number of valid samples; the class weights ω C Depending on the original distribution of samples in each category, the number of samples N corresponding to category i is... i If the number of samples in category i is less than the mean of the total number of samples in all categories, it is assigned a weight greater than 1, thus amplifying the penalty for misclassification of samples in smaller categories; the number of samples N corresponding to category i. i If the sample size is greater than the average of all categories, assign a weight less than 1.

[0029] The image collaboration pipeline includes the following steps, executed sequentially:

[0030] S4.1: A medical image positive anomaly classification method, which outputs a normal / abnormal conclusion based on global intensity and texture representation. Specifically, this method includes a weak symmetry determination algorithm, with the following steps:

[0031] S4.1.1: Preprocess the medical image to obtain a standardized grayscale image, eliminate image noise and normalize the size to ensure the consistency of the input image;

[0032] S4.1.2: Based on the geometric symmetry of the left and right breast regions in medical images, the intersection degree of the gray-level histograms of the left and right parts of the image is calculated to obtain the symmetry features of the image;

[0033] S4.1.3: Input the symmetry features of the image into the weak symmetry determination algorithm to calculate a symmetry score S. sym This score is used to represent the degree of left-right symmetry of the image;

[0034] S4.1.4: If the symmetry score of the image is S sym Greater than the preset threshold T sym If S, then the image is considered to belong to the normal category; if S sym If the image is smaller than the threshold, it is considered to belong to the anomaly category and proceeds to the next step of anomaly localization and classification.

[0035] S4.2: Medical image anomaly localization and classification. Candidate regions are generated under anatomical priors and symmetry constraints, and feature aggregation is performed to locate suspicious regions. The specific steps of anomaly localization and classification are as follows:

[0036] S4.2.1: Anomaly Candidate Generation: Based on the symmetry differences of the image, the geometric structural differences between the left and right breast regions are calculated to generate preliminary anomaly candidate regions; for each candidate region, the region boundary is further refined through an edge detection algorithm to ensure the accuracy and effectiveness of the candidate regions;

[0037] S4.2.2: Symmetry Difference Calculation: For each candidate region, calculate its symmetry difference degree D. sym ;

[0038] S4.2.3: Precise localization of candidate regions: Further refine the localization of the abnormal regions output by the classification network, and use a convolutional neural network (CNN) for position optimization to obtain the precise coordinates and size of each abnormal region;

[0039] S4.2.4: Abnormal area identification. For areas identified as abnormal, draw borders on the image and set labels.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] This invention provides a method for extracting and analyzing breast cancer case management data based on natural language processing and medical image analysis. By employing a multi-channel self-attention mechanism and a weighted learning loss function, it can accurately extract key entities, relationships, and events from electronic medical records and perform comprehensive analysis in conjunction with medical image data. Through cross-modal consistency alignment technology, textual and image data are precisely fused, ensuring the consistency and integrity of multimodal data, thereby improving the reliability and accuracy of the analysis results. Furthermore, this invention utilizes advanced medical image positive anomaly classification and anomaly localization classification methods to accurately detect and locate abnormal regions under complex conditions, significantly improving the analysis accuracy of breast cancer image data.

[0042] Based on this, a case timeline is constructed using time-standardized event sequence alignment and sorting technology to ensure that all key diagnostic and treatment events are accurately recorded in chronological order, providing reliable support for subsequent case management and clinical decision-making. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the accompanying drawings used in the following description of the embodiments or examples will be briefly introduced. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the technical solutions shown in these drawings without creative effort.

[0044] Figure 1 A flowchart illustrating a method for extracting and analyzing breast cancer case management data based on natural language processing;

[0045] Figure 2 A schematic diagram of the structure of LSTM and BiLSTM for a method of extracting and analyzing breast cancer case management data based on natural language processing;

[0046] Figure 3 A comparison of the time complexity of the MSC and SVM-MPO methods for extracting and analyzing breast cancer case management data based on natural language processing;

[0047] Figure 4 Gray values ​​of normal-state CT images for a data extraction and analysis method for breast cancer case management based on natural language processing;

[0048] Figure 5 This refers to the grayscale values ​​of abnormal CT images used in a method for extracting and analyzing breast cancer case management data based on natural language processing. Detailed Implementation

[0049] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0050] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0051] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0052] Example 1

[0053] See Figures 1-2 This embodiment uses the electronic medical record text of breast cancer patients as input. The experimental corpus uses the DDI relation dataset shown in Table 1 for supervised training and validation. The DDI dataset for breast cancer-related drug use simulation includes drug interactions, containing information such as drug mechanisms, effects, recommendations, and no significant effects, making it suitable for studying drug interactions when used in combination. Table 1 is shown below:

[0054] Table 1 DDI Dataset

[0055]

[0056]

[0057] First, multimodal case data was collected. In this embodiment, text data came from the electronic medical records of breast cancer patients, including natural language records such as medical history descriptions, examination results, treatment plans, and adverse reactions; numerical data came from the WBCD dataset shown in Table 2, as follows:

[0058] Table 2 WBCD dataset:

[0059]

[0060] Each sample contains nine morphological features, including radius, texture, perimeter, area, smoothness, compactness, concavity, shape consistency, and density. Each sample is labeled as benign or malignant.

[0061] The collected data underwent preprocessing. First, patient identification information was removed from the electronic medical record text to achieve de-identification. Then, segmentation and sentence division were performed, breaking down long texts into semantically complete sentences. Medical terminology was normalized using a unified thesaurus; in this embodiment, "breast cancer" and "malignant breast tumor" were standardized to the standard term "breast cancer." Numerical units and time representations were standardized to ensure consistency across different records. Numerical features were Z-score standardized, transforming them into a zero-mean, unit-variance distribution to eliminate the adverse effects of dimensional differences on model training. After these processing steps, both textual and numerical data were ready for direct input into the model.

[0062] The preprocessed data is input into the joint extraction model. In the low-level feature extraction, the text-side input is embedded through three channels: character, word, and sub-word. The word vector dimension is set to 200, and the character and sub-word vectors are each 100. At the same time, positional embedding is introduced, using target entity pairs as anchors, calculating the relative distance and mapping it to a 50-dimensional sparse vector. Finally, the word-level input vector with a dimension of 450 is concatenated to form the input sequence. The nine features of the numerical input are also embedded and mapped to a 128-dimensional space, so that each feature of the samples in Table 2 has a vectorized representation.

[0063] In high-level feature extraction, the input sequence is fed into a BiLSTM network to obtain the hidden layer output matrix Z_t. <z1,z2,z3,…,Z n >, with a dimension of 256. In the electronic medical record sample, z i These correspond to the contextual semantics of disease, drug, and adverse reaction, respectively; subsequently, based on... Calculate the attention weights for C=4 channels respectively. Each channel independently learns its attention points, corresponding to disease, drug, adverse reaction, and time information in the text samples, and geometric dimensions, texture indices, morphological parameters, and overall compactness in the samples in Table 2. The attention points of each channel are multiplied and fused with the LSTM output according to a formula to obtain sentence-level or sample-level high-level representations, which are then used as classification input.

[0064] In this embodiment, the loss function used is a common loss function employed by the softmax classifier, specifically cross-entropy:

[0065]

[0066] In this embodiment, if the input features of sample number 3 are: radius 13.5, texture 19.8, perimeter 85.2, area 540.0, smoothness 0.102, compactness 18.7, concavity 0.027, and shape consistency 0.055, its true label t i The corresponding malignancy category is encoded as [0,1], and the model's softmax output predicts the result. The cross-entropy component of the loss value is calculated to be 0.094; this sample was correctly classified, resulting in a low loss value. Furthermore, during training, L2 regularization is used to limit the parameter size, maintaining the model's stability and generalization performance. If the model mistakenly classifies the sample as benign, then... The cross-entropy loss of 1.897 significantly improved the loss value, driving the model to adjust in the correct direction during parameter updates. Joint discrimination was performed. On the text side, a conditional random field was used for sequence decoding, outputting the boundaries and categories of entities such as diseases, drugs, and adverse reactions; the relation classification head output the relationship categories between entity pairs. In the electronic medical record sample "Patient received paclitaxel treatment for breast cancer recurrence and subsequently developed peripheral neuropathy," the model output the entities "breast cancer recurrence [disease]," "paclitaxel [treatment]," and "peripheral neuropathy [adverse reaction]," and determined the relation "paclitaxel → peripheral neuropathy" as Effective with a confidence level of 0.87. In the numerical feature input of sample number 3 in Table 2, after fusing high-level representations, the model output "malignant" with a confidence level of 0.91, consistent with the pathological annotation.

[0067] See Figure 3 In terms of overall experimental results, this embodiment achieved a classification accuracy of 89.3% and an area under the curve (AUC) of 0.91 on the WBCD dataset. The recall rate for malicious samples improved from 81.2% in the unweighted model to 88.6%. On electronic medical record text samples, the accuracy of entity relation extraction was also better than that of the single-channel model, and it was able to stably identify overlapping entities and multiple relations in complex sentences.

[0068] Example 2

[0069] This embodiment incorporates bilateral breast CT images of breast cancer patients as input, employs a multi-stage image classification method based on symmetry theory, and progressively achieves normal / abnormal discrimination, abnormal region localization, and feature description, which are then fused with the text-based results in the subsequent consistency alignment stage.

[0070] See Figure 4 , Figure 5 In the first stage, based on the overall morphological features of the image, a compactness C is defined to describe the regularity of the tumor contour edges. The formula is as follows: In this embodiment, the tumor area A = 314 and the perimeter p = 80. Substituting these values ​​into the formula yields C = 0.39, indicating that the tumor edge exhibits significant irregularity. Roughness R0 is defined as: In this embodiment, R0 = 0.072 was calculated, indicating that there is a large fluctuation in the edge; the image was classified as "abnormal", and the conclusion is consistent with the pathological annotation.

[0071] Further, abnormal region localization is performed within the abnormal samples. Anomaly candidate regions are generated within the ROI, and the symmetry difference of these candidate regions is calculated. A convolutional neural network is then used for position optimization to obtain the precise coordinates of the abnormal region. In this embodiment, the candidate region is located in the left breast area. The CNN localization output bounding box coordinates are (x1, y1) = (128, 96) and (x2, y2) = (176, 144). This region is highly consistent with the manually labeled results, with an IoU value of 0.87. The model ultimately labels this region as "abnormal-malignant" and draws a border and label on the image.

[0072] This embodiment was tested on 200 CT images, achieving a positive anomaly classification accuracy of 91.2% and an average IoU of 0.82 for anomaly region localization. This demonstrates that the method can reliably distinguish between normal and abnormal images and achieve precise localization of suspicious regions. The results are stored in a structured format, including coordinates, labels, and confidence scores, for subsequent cross-modal consistency alignment with text-based extraction results.

[0073] Example 3

[0074] In the cross-modal analysis phase, this embodiment fuses the results from the electronic medical record text side and the CT image side, and sequentially completes consistency alignment, timeline construction, index calculation and structured output.

[0075] The system aligns text and image results for consistency based on lateralization, anatomical partitioning, and temporal anchors. For the text side, extracted entities include "breast cancer recurrence [disease]", "paclitaxel [treatment]", and "peripheral neuropathy [adverse reaction]". On the image side, an abnormal area was located in the left breast region and classified as "abnormal-malignant" with a confidence level of 0.89. The system matches the text "left breast recurrence" with the abnormal conclusion on the left side of the image and combines this with the temporal information "received treatment after recurrence," calculating a consistency score of 0.84. Based on this score, the confidence levels of entities, relationships, and events on the text side are updated and integrated, ultimately generating cross-modal consistent structured case results to ensure that the text and image conclusions corroborate each other.

[0076] The events were aligned chronologically based on the time standardization results. The textual time expression "paclitaxel treatment started two weeks after diagnosis" was standardized to an absolute timestamp during preprocessing. The system associated this time point with the event node to construct a patient-level case timeline, with the events in the following order: "March 1, 2022: Diagnosis of relapse → March 14, 2022: Paclitaxel treatment started → May 10, 2022: Peripheral neuropathy appeared." This timeline clearly reflects the sequential relationship between disease progression, treatment implementation, and adverse events.

[0077] The system calculates and determines case management indicators according to a preset rule base. Specific results are as follows: the time delay from diagnosis to treatment initiation is 13 days, below the 21-day threshold, indicating timely treatment; the paclitaxel regimen meets standardized treatment criteria, indicating compliance; follow-up records are complete, and adherence is good; an abnormal area was detected on imaging, and combined with the textual "relapse" event, a relapse is determined; the adverse event "peripheral neuropathy" has occurred, recorded as positive. The final generated case management indicators include: [Time delay = 13 days, Compliance with standardized criteria = Yes, Follow-up adherence = High, Adverse event = Peripheral neuropathy, Relapse / metastasis = Relapse].

[0078] The fusion results are then written into a structured format into the database and knowledge graph. The structured results include: text-based entities, relationships, and events; coordinates and classification labels of abnormal areas in images; timeline nodes and sequence; and case management indicators. These results are output through a standardized interface for use by the hospital's case management system, enabling automated storage and analysis of patient data throughout the entire process. For example, the entries recorded in the database might be:

[0079] · Specific conditions: [recurrent breast cancer], [paclitaxel], [peripheral neuropathy]

[0080] · Relationship: [Paclitaxel → Peripheral neuropathy, Effect]

[0081] · Events: [Paclitaxel treatment started on March 14, 2022], [Peripheral neuropathy appeared on May 10, 2022]

[0082] · Abnormal imaging findings: [Left breast region coordinates (128, 96) - (176, 144), malignant, confidence level 0.89]

[0083] · Timeline: [Diagnosis → Treatment → Adverse Reactions]

[0084] · Management indicators: [Duration = 13 days, Standardized treatment achieved = Yes, Follow-up compliance = High, Adverse events = Peripheral neuropathy, Relapse / metastasis = Relapse]

[0085] This embodiment runs on data from multiple patient cases and can automatically generate consistent structured results and timelines across modalities, as well as output standardized indicators, providing reliable support for the refined management of breast cancer cases.

Claims

1. A method for extracting and analyzing breast cancer case management data based on natural language processing, characterized in that, include: S1: Collect multimodal case data, including electronic medical record text and medical CT images; S2: Performs de-identification, cleaning, segmentation and sentence division, terminology and unit normalization, and time expression standardization on electronic medical record text; performs intensity normalization, geometric correction, and left and right breast region segmentation based on anatomical symmetry on CT images; S3: The joint extraction model for preprocessed text input specifically includes: S3.1 Low-level feature extraction, performing sequence encoding on character / word / subword multi-channel embedding to obtain long-short dependency representation; S3.2 High-level feature extraction: A multi-channel self-attention mechanism is used to compute multiple sets of attention in parallel for the same sentence and fuse them with learnable channel weights to obtain sentence-level semantic representation. S3.3 introduces a weighted learning loss function, where the weights of each class are used as parameter vectors and participate in training together with the cross-entropy loss function. S3.4 Joint discrimination outputs entities, relationships between entities, event types, and trigger words related to case management; S4: Input CT images into the imaging collaboration pipeline to obtain and label the normal / abnormal conclusions on the imaging side; S5: Based on lateralization, anatomical partitioning, and time anchors, the text-side and image-side results are aligned for consistency, a consistency score is calculated, and the confidence of entities, relationships, and events in the text-side are fused and decoded accordingly to obtain cross-modal consistent case structured results; S6: Based on the time standardization results, perform time-series alignment and sorting of events to construct patient-level case timelines; S7: Calculate and determine case management indicators according to the rule base, including at least the time from diagnosis to treatment initiation, standard treatment achievement, follow-up compliance, occurrence of adverse events and determination of recurrence / metastasis; S8: Write the merged entities, relationships, events, timelines, and indicator results into a structured format into a database or knowledge graph, and output interface data for the case management system to call.

2. The method for extracting and analyzing breast cancer case management data based on natural language processing according to claim 1, characterized in that, The multi-channel self-attention mechanism specifically includes the following processing steps: S3.2.1: Use an LSTM model to concatenate the learned long and short distance dependency information and output the hidden layer vector: Z t = <z1,z2,z3,…,Z n >; S3.2.2: Z_t serves as the input for multi-channel self-attention. The multi-channel self-attention mechanism is given the number of channels C, and in each channel C... i The self-attention weights are calculated once in each step, specifically in the following form: Here, α is a hyperparameter used to adjust the position embedding weights; softmax is a normalization function that converts the output of all channels into a probability distribution; Refers to channel C i A set of learnable weight matrices is provided, with the number of channels C being a key hyperparameter that needs to be adjusted according to the actual business scenario; tanh is an activation function that performs a nonlinear transformation on the result of matrix multiplication; W z It is a learnable weight matrix used to perform linear transformations on the hidden state output by the LSTM; It is the transpose of the LSTM output matrix, which, after matrix multiplication, yields a new vector representation; S3.2.3: Add constraint terms: Where ||*|| represents the Frobenius norm of the matrix; the penalty term ||*|| is used as part of the loss function to train the model together. For each additional channel, the multi-channel self-attention mechanism will perform weighted fitting on the component i in the sentence. S3.2.4: Calculate the weight matrix using a two-layer perceptron Will With Z t Multiply and then normalize to obtain the output H of the self-attention layer. * The H * It is a vector representation of sentence-level semantic features.

3. The method for extracting and analyzing breast cancer case management data based on natural language processing according to claim 1, characterized in that, The joint discrimination adopts sequence labeling and classification joint learning, named entity recognition uses conditional random fields for sequence decoding, and the relationship and event type are output by the classification head probability distribution and share the bottom and high-level representations with the entity boundary.

4. The method for extracting and analyzing breast cancer case management data based on natural language processing according to claim 1, characterized in that, The loss function for weighted learning is the cross-entropy with class weights, where the class weights are calculated based on class frequency or the number of valid samples; the class weights ω C Depending on the original distribution of samples in each category, the number of samples N corresponding to category i is... i If the number of samples in category i is less than the mean of the total number of samples in all categories, it is assigned a weight greater than 1, thus amplifying the penalty for misclassification of samples in smaller categories; the number of samples N corresponding to category i. i If the sample size is greater than the average of all categories, assign a weight less than 1.

5. The method for extracting and analyzing breast cancer case management data based on natural language processing according to claim 1, characterized in that, The image collaboration pipeline includes the following steps executed sequentially: S4.1: A positive anomaly classification method for medical images, which outputs normal / abnormal conclusions based on global intensity and texture representation; S4.2: Medical image anomaly localization and classification, generating candidate regions and performing feature aggregation under anatomical priors and symmetry constraints to locate suspicious regions.

6. The method for extracting and analyzing breast cancer case management data based on natural language processing according to claim 5, characterized in that, The medical image positive anomaly classification method specifically includes a weak symmetry determination algorithm, the specific steps of which are as follows: S4.1.1: Preprocess the medical image to obtain a standardized grayscale image, eliminate image noise and normalize the size to ensure the consistency of the input image; S4.1.2: Based on the geometric symmetry of the left and right breast regions in medical images, the intersection degree of the gray-level histograms of the left and right parts of the image is calculated to obtain the symmetry features of the image; S4.1.3: Input the symmetry features of the image into the weak symmetry determination algorithm to calculate a symmetry score S. sym This score is used to represent the degree of left-right symmetry of the image; S4.1.4: If the symmetry score of the image is S sym Greater than the preset threshold T sym If S, then the image is considered to belong to the normal category; if S sym If the image is smaller than the threshold, it is considered to belong to the anomaly category and proceeds to the next step of anomaly localization and classification.

7. The method for extracting and analyzing breast cancer case management data based on natural language processing according to claim 5, characterized in that, The specific steps for anomaly localization and classification are as follows: S4.2.1: Anomaly Candidate Generation: Based on the symmetry differences of the image, the geometric structural differences between the left and right breast regions are calculated to generate preliminary anomaly candidate regions; for each candidate region, the region boundary is further refined through an edge detection algorithm to ensure the accuracy and effectiveness of the candidate regions; S4.2.2: Symmetry Difference Calculation: For each candidate region, calculate its symmetry difference degree D. sym ; S4.2.3: Precise localization of candidate regions: Further refine the localization of the abnormal regions output by the classification network, and use a convolutional neural network for position optimization to obtain the precise coordinates and size of each abnormal region; S4.2.4: Abnormal area identification. For areas identified as abnormal, draw borders on the image and set labels.

Images