Breast cancer dynamic risk assessment method and system based on adaptive model

Abstract

The invention relates to a breast cancer dynamic risk assessment method based on an adaptive model, and aims to improve the accuracy of early screening and risk prediction of breast cancer. The method comprises the following steps that firstly, clinical data of the unresected side mammary gland of an estrogen receptor positive breast cancer patient are obtained, and the data comprise the estrogen receptor state, the hormone treatment duration and mammary gland density classification; selecting an imaging detection mode according to the mammary gland density grading and the estrogen receptor state, and obtaining a calcification focus image of the target mammary gland; performing multi-parameter quantitative analysis on the image, and extracting parameters such as hydroxyapatite crystal proportion, calcification edge sharpness index and the like; and carrying out risk value calculation by using the trained adaptive risk assessment model, comparing the risk value with a preset risk grading threshold, and finally generating an assessment report based on a risk grade. According to the method, accurate dynamic risk assessment can be provided for the breast cancer patient, and personalized health management and intervention decision are supported.

Description

Technical Field

[0001] The present application relates to the medical field, and in particular to a method and system for dynamic risk assessment of breast cancer based on an adaptive model. Background Art

[0002] Breast cancer is one of the most common malignancies in women worldwide, with estrogen receptor-positive (ER+) breast cancer accounting for over 70% of all breast cancer cases. These patients are generally more sensitive to hormone therapy (such as tamoxifen and aromatase inhibitors) and have a relatively good prognosis after treatment. However, there is still a risk of new breast cancer in the other breast for many years after completing treatment. This risk requires long-term attention, especially in patients who have had one mastectomy.

[0003] Hormone therapy can cause calcification of breast cells, which may appear as calcification foci in imaging examinations. Under normal circumstances, doctors will judge whether it is a benign lesion based on the morphological characteristics of the calcification foci (such as uniformity of distribution, sharpness of edges, etc.). Benign calcification foci are usually evenly distributed with smooth edges, while malignant calcification foci are concentrated with sharp edges. However, the calcification morphology of ER-positive patients is special. Due to the influence of hormone therapy, the calcification foci of these patients are often denser and irregularly distributed, easily forming a contour similar to the lesion, thereby increasing the risk of misjudgment by doctors. In addition, the detection of patients with dense breasts (such as patients with breast density grade C or D) is more complicated because the density of breast tissue will block the calcification area, further affecting the accuracy of imaging judgment.

[0004] Existing breast cancer follow-up and recurrence risk monitoring strategies are usually based on routine imaging at fixed intervals, and fail to make dynamic and personalized adjustments based on the characteristics of the ER+ population. Currently, most breast cancer follow-up programs fail to dynamically model risks by combining data such as ER status, duration of hormone therapy, and changes in expression trends, which may underestimate or ignore some individuals with a high risk of potential recurrence. Existing methods make it difficult to focus on deep image feature analysis and risk quantification of the target area. Image feature analysis of early manifestations of breast cancer, such as calcifications, is usually separated from clinical information such as the patient's endocrine treatment background, lacking a unified risk modeling path. Due to privacy and security restrictions, it is difficult for different medical participants to directly share original patient data, resulting in insufficient training samples for the intelligent breast cancer risk model and difficulty in improving generalization capabilities.

[0005] Therefore, there is an urgent need for a breast cancer risk assessment method for ER+ patients that can integrate clinical and image multimodal information, dynamic modeling, and support multi-center safe collaborative training, so as to enhance the early identification ability of new lesions in the other breast and the clinical intervention value. Summary of the Invention

[0006] The purpose of this application is to overcome the above-mentioned problems of increased risk of misdiagnosis due to the complexity of calcification morphology, obscuration of calcification areas in patients with dense breasts, and lack of dynamic risk assessment and personalized intervention plans for ER-positive patients.

[0007] According to one aspect of the present application, a dynamic risk assessment method for breast cancer based on an adaptive model is provided, which is applied to patients with breast cancer who have undergone unilateral mastectomy and are estrogen receptor-positive. The dynamic risk assessment method comprises the following steps:

[0008] S10. Obtain clinical data of the unresected breast of a patient with estrogen receptor-positive breast cancer who has had one mastectomy, the clinical data including: estrogen receptor status test results, duration of hormone therapy, and breast density grading data based on the BI-RADS standard;

[0009] S20, selecting a corresponding imaging detection mode according to the breast density grading data and the estrogen receptor status detection result, and acquiring an image of the calcification focus of the target breast;

[0010] S30, performing multi-parameter quantitative analysis on the calcification focus image, wherein the parameters include: hydroxyapatite crystal ratio, calcification edge sharpness index, and calcification cluster spatial distribution entropy value;

[0011] S40, using the multi-parameter quantitative analysis results as input, calculating the risk value using the trained adaptive risk assessment model, and outputting the breast cancer risk value of the target breast;

[0012] S50: Compare the risk value with a preset risk classification threshold and generate an assessment report based on the risk level.

[0013] Preferably, the S10 further includes:

[0014] S11, performing a credibility correction on the estrogen receptor status test result, wherein the credibility correction is based on the duration of hormone treatment, the patient's medical history, and the past estrogen receptor expression trend;

[0015] Accordingly, the S40 is: taking the multi-parameter quantitative analysis result and the corrected estrogen receptor status as input, performing calculations using the trained adaptive risk assessment model, and outputting the patient's breast cancer risk value.

[0016] Preferably, before S10, the step further includes:

[0017] S60. Training the adaptive risk assessment model based on historical case data, wherein the historical case data includes: images of calcification foci in the other breast of a patient with estrogen receptor-positive breast cancer whose one breast had been removed, multi-parameter quantitative analysis results, and corresponding clinical follow-up information after the patient's other breast was re-diagnosed with breast cancer.

[0018] Preferably, the S40 includes:

[0019] The adaptive risk assessment model is trained in a multi-center collaborative manner through a federated learning mechanism, wherein the type of the federated learning mechanism is horizontal federated learning.

[0020] Preferably, the multi-center collaborative training through the federated learning mechanism includes:

[0021] S41. Extracting an estrogen receptor-positive patient dataset from the local database of each participant, wherein the patient dataset has been subjected to differential privacy processing;

[0022] S42. Decrypt the patient data sets of each participant and aggregate them according to dynamic weights to generate a training data set. The dynamic weights include a basic weight and a correction factor. The basic weight is calculated based on the proportion of the number of cases of each participant to the total number of cases. If the historical diagnostic consistency of the participant is higher than 90%, the weight is increased by 20%; if the participant has more than 10% feedback correction records, the weight is decreased by 20%.

[0023] S43. Training the adaptive risk assessment model using the training data set.

[0024] Preferably, the patient dataset includes the patient's basic information, imaging data, clinical treatment records, calcification annotation and pathology verification data, dynamic follow-up data, and federated learning metadata.

[0025] Preferably, selecting the detection mode in step S20 includes:

[0026] When the breast density is classified as C or D, a digital breast tomosynthesis device is used to perform multi-angle scanning, a three-dimensional calcification distribution map is reconstructed using a sliding window method, and a frequency domain fusion algorithm is used to enhance the contrast of the calcification edge;

[0027] When the breast density is classified as Class A or Class B, a dual-energy X-ray molybdenum target is used to generate a calcification energy spectrum image, and the energy spectrum signals of hydroxyapatite and calcium salt are distinguished by energy spectrum separation technology.

[0028] Preferably, after S50, the step further includes:

[0029] S70: Outputting personalized inspection suggestions based on the risk level corresponding to the risk value, wherein:

[0030] When the risk level is low, it is recommended that the patient undergo routine breast imaging every 12 months;

[0031] When the risk level is moderate, it is recommended that the patient undergo enhanced breast imaging every 6 months and assess compliance with hormone therapy;

[0032] When the risk level is high, it is recommended that the patient undergo high-level breast imaging every 3 months, and make intervention decisions in combination with tumor marker monitoring and multidisciplinary consultation.

[0033] The present invention also provides a breast cancer risk assessment system dynamically adjusted according to an adaptive model, which implements the above-mentioned breast cancer dynamic risk assessment method based on the adaptive model, including:

[0034] Data acquisition module, used to collect clinical data and imaging data of target patients;

[0035] Image analysis module, used for parameter extraction and quantitative analysis of calcification images;

[0036] The model processing module includes a trained adaptive risk assessment model, which is used to receive parameter analysis results and output breast cancer risk values;

[0037] The assessment report generation module is used to generate a breast cancer risk assessment report based on the risk value.

[0038] Preferably, the model processing module includes: a federated learning training submodule, which is used to receive encrypted data from multiple participants and perform cross-center collaborative training.

[0039] This application has the following beneficial effects:

[0040] Provide personalized dynamic assessment pathways for specific high-risk groups (ER+ breast cancer patients after surgery) to improve the accuracy of new breast cancer risk monitoring;

[0041] Introducing multi-parameter analysis of calcification images and joint modeling of ER status to enhance the structure and scientific nature of risk prediction;

[0042] The model is trained based on large-scale case federated learning and has strong generalization capabilities and the ability to adapt to multi-center clinical data;

[0043] Output multi-level risk results and link them to subsequent inspection strategies to achieve a risk-driven precision medicine path;

[0044] To make up for the lack of sensitivity of traditional breast density grading and static imaging methods in people with high-density breasts. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the implementation methods of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the implementation methods or the description of the prior art. Obviously, the drawings described below are only some implementation methods of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1 This is a logic block diagram of a dynamic breast cancer risk assessment method based on an adaptive model according to one embodiment of the present application. DETAILED DESCRIPTION

[0047] To facilitate understanding of the present application, a more comprehensive description of the present application will be provided below with reference to the accompanying drawings. The accompanying drawings illustrate preferred embodiments of the present application. However, the present application may be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the disclosure of the present application.

[0048] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly attached to the other element or there may be an intermediate element. When an element is referred to as being "connected to" another element, it may be directly connected to the other element or there may be an intermediate element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein in the specification of this application are intended only to describe specific embodiments and are not intended to limit this application. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0050] Please refer to Figure 1 An embodiment of the present application provides a method for dynamic risk assessment of breast cancer based on an adaptive model, comprising the following steps:

[0051] S10. Obtain clinical data on the unresected breast of patients with estrogen receptor-positive breast cancer who have had one mastectomy. Clinical data include: estrogen receptor status test results, duration of hormone therapy, and breast density grading data based on the BI-RADS standard.

[0052] S20, selecting a corresponding imaging detection mode according to the breast density grading data and the estrogen receptor status detection result, and acquiring an image of the calcification focus of the target breast;

[0053] S30, performing multi-parameter quantitative analysis on the calcification images, including the following parameters: hydroxyapatite crystal ratio, calcification edge sharpness index, and calcification cluster spatial distribution entropy;

[0054] S40, using the multi-parameter quantitative analysis results as input, calculating the risk value using the trained adaptive risk assessment model, and outputting the breast cancer risk value of the target breast;

[0055] S50: Compare the risk value with a preset risk classification threshold and generate an assessment report based on the risk level.

[0056] In this embodiment, it should be noted that after postoperative endocrine therapy, patients with estrogen receptor (ER)-positive breast cancer who have had one breast removed may still be at risk of developing new primary breast cancer or recurrent lesions in the other breast during long-term follow-up, even though the primary breast has been removed. This embodiment aims to provide a dynamic, accurate, and personalized risk assessment mechanism to enhance the clinical ability to identify and intervene in high-risk populations early. The details are as follows:

[0057] Step S10, data acquisition and preprocessing, involves obtaining clinical data from previous cases of ER-positive patients. This data includes: ER status test results, derived from immunohistochemistry testing, reflecting the current expression of breast tissue's dependence on estrogen; duration of hormone therapy, including the time patients took medications such as tamoxifen or aromatase inhibitors after surgery, to assist in determining trends in ER expression; and breast density grading, based on the BI-RADS classification standard (AD category), to assess the structural complexity and imaging difficulty of breast tissue. If ER test results show historical fluctuations or missing information, trend interpolation or confidence correction models are used to correct the test values.

[0058] Step S20, individualized imaging mode selection. When the breast density is classified as Class C or Class D, it indicates that the breast tissue is dense and the calcification foci are poorly visible. Digital breast tomography (DBT) is required to enhance image clarity through sliding window three-dimensional reconstruction and frequency domain fusion algorithm. When the breast density is Class A or Class B, dual-energy X-ray mammography is used to further generate energy spectrum images, and the energy spectrum separation algorithm is used to distinguish calcium salt deposits such as hydroxyapatite (HA) and calcium carbonate.

[0059] Step S30: Multi-parameter quantitative image analysis. This analysis analyzes the acquired calcification images and quantifies the following key indicators: the percentage of hydroxyapatite crystals, representing the primary component of malignant calcifications; the calcification edge sharpness index, which assesses the clarity and roughness of calcification edges and correlates with irregular growth of malignant tumors; and the spatial distribution entropy of calcification clusters, which measures the degree of dispersion of calcification points; the more concentrated the calcifications, the more likely they are to indicate lesion clustering. These parameters provide highly structured and quantifiable risk inputs for the assessment model.

[0060] The calculation method of "hydroxyapatite crystal ratio" in step S30 is:

[0061] The dual-energy attenuation coefficient ratio of the calcification lesion is obtained through spectral CT, and the dual-energy attenuation coefficient ratio is mapped to the preset hydroxyapatite concentration curve to obtain the hydroxyapatite crystal ratio;

[0062] The dual energy attenuation coefficient ratio is calculated by formula 1:

[0063]

[0064] Where DEI is the dual energy attenuation coefficient ratio;

[0065] μ low is the linear attenuation coefficient of low-energy X-rays;

[0066] μ high is the linear attenuation coefficient of high-energy X-rays;

[0067] The “calcification edge sharpness index” is calculated by formula 2:

[0068] Perform multi-scale gradient analysis on the edge of the calcification and use the Sobel operator to extract the edge gradient amplitude G(x, y);

[0069] The calcification edge sharpness index is calculated by formula 2:

[0070]

[0071] Where S index is the calcification edge sharpness index;

[0072] D(x, y) is the normalized distance from the edge point (x, y) to the centroid of the calcification;

[0073] σ is the distance attenuation coefficient.

[0074] Step S40: Model-based risk calculation. A pre-trained adaptive risk assessment model (e.g., a deep neural network model or an ensemble decision tree model) takes the above parameters and combines them with the corrected ER status to output a normalized risk value (0-1) representing the likelihood of malignant transformation in the unresected breast. This model, trained on extensive historical data, possesses a degree of self-updating (online fine-tuning) capability and can be continuously optimized based on newly added cases.

[0075] Step S50: Risk classification and report output. Compare the risk value with the set threshold (e.g., 0.3, 0.6) and classify it into:

[0076] Low risk (<0.3): Routine review every year is recommended;

[0077] Medium risk (0.3-0.6): 6-month follow-up and advanced imaging examinations are recommended;

[0078] High risk (>0.6): Further tissue biopsy or MRI enhancement examination is recommended.

[0079] The final assessment report not only includes the risk level, but also includes calcification change trend charts, ER expression trend charts, personalized review suggestions, etc., providing clinicians with intuitive decision-making support information.

[0080] The technical solution of this embodiment can achieve dynamic and personalized risk assessment of the remaining breast of patients with ER-positive breast cancer after surgery. Compared with traditional static assessment methods based on a single image or a single examination result, it has the following significant technical effects:

[0081] Improve the sensitivity and accuracy of risk identification: By introducing multi-dimensional clinical data (ER status, duration of hormone treatment, breast density level, etc.) and high-precision image parameters (such as HA ratio, sharpness index, spatial entropy value) for joint modeling, the model's ability to identify potential malignant evolution is improved, reducing the missed diagnosis rate and misjudgment rate.

[0082] Adapt to individual differences and match imaging methods: Intelligently select DBT or mammography spectral imaging based on breast density, overcoming the limitations of poor imaging quality in high-density breasts and ensuring the credibility and consistency of image analysis results. It is especially suitable for Chinese women with high breast density.

[0083] Introducing an ER expression credibility correction mechanism: Considering the impact of treatment and testing fluctuations on ER status, a credibility correction model or trend interpolation is introduced to improve the authenticity of ER status as a risk feature input, making the risk assessment results more consistent with the patient's physiological reality.

[0084] Dynamic learning and updating capabilities: The evaluation model has the ability to fine-tune online, and can optimize parameters and adjust weights as new data are continuously acquired in clinical practice, so that its risk judgment ability is continuously iteratively enhanced, in line with clinical trends driven by real-world data.

[0085] Output structured personalized assessment report: Risk assessment results not only output quantitative risk values and grading results, but also include image analysis results, ER expression trend charts, and review suggestions, etc., which help clinicians quickly understand the patient's current status and formulate precise management strategies.

[0086] In an optional embodiment, S10 also includes: S11, performing credibility correction on the estrogen receptor status detection result, and the credibility correction is based on the duration of hormone treatment, the patient's medical history records and previous estrogen receptor expression trends; accordingly, S40 is: using the multi-parameter quantitative analysis results and the corrected estrogen receptor status as input, using the trained adaptive risk assessment model to calculate, and outputting the patient's breast cancer risk value.

[0087] In this example, it should be noted that the reliability correction of estrogen receptor (ER) status test results is intended to address the issue of inaccurate ER status determination in some patients due to differences in testing time points, sample source limitations, or the influence of therapeutic interventions. Because ER expression in ER-positive breast cancer patients may fluctuate or exhibit apparent expression deviations during long-term hormone therapy, a single test result may not fully reflect the true receptor status of the target breast.

[0088] Credibility correction step S11 combines the following three types of information for multi-dimensional evaluation:

[0089] Duration of hormone therapy: The longer the duration of endocrine therapy (such as tamoxifen or aromatase inhibitors), the more pronounced its inhibitory or regulatory effect on ER expression, which can be used to infer the current trend of downregulation or activation of ER expression;

[0090] Patient medical history: such as whether there has been postoperative recurrence, tolerance to hormone therapy, family history of breast cancer, and genetic testing results (such as BRCA status). This background information can help adjust the confidence level of ER expression determination;

[0091] Past ER expression trends: By retrospectively modeling historical test results, if there are multiple test records, its expression fluctuation curve can be fitted to further correct the representativeness of the current test value.

[0092] The credibility-corrected ER status data not only improves the stability and authenticity of the input variables but also enhances the overall model's sensitivity and accuracy to risk characteristics. Accordingly, in step S40, the assessment model uses these more valuable corrected ER statuses as one of its core feature inputs, along with the structural parameters derived from image analysis, in the comprehensive calculation of the risk value, ensuring that the output is more consistent with the actual biological risk situation.

[0093] The implementation of the technical solution of this embodiment can effectively improve the reliability and explanatory power of the model in the biomarker input dimension while maintaining the basic structure of the original risk assessment model. It has the following technical advantages and clinical value: Enhance the stability of ER status as a risk factor: Through the credibility correction mechanism, it alleviates the misleading input caused by single-test sample deviation (such as inconsistent sampling location, failure to eliminate treatment effects, etc.), improves the feature contribution quality of ER status in the model, and prevents the "noise" of key variables.

[0094] Reflecting the dynamic characteristics of an individual's disease course: This embodiment introduces the duration of a patient's hormone treatment and historical ER expression trends, ensuring that the input of ER status is not just a result at a specific point in time, but rather reflects a continuous and dynamically changing biological trend, thereby improving the timeliness and pertinence of risk assessment.

[0095] Taking into account individual differences and treatment response characteristics: Some patients have unique circumstances, such as resistance to hormone therapy, early relapse, or a family genetic background. Conventional ER assessments may not cover these characteristics. The correction mechanism can address this deficiency by incorporating unstructured information, such as medical history, to make the assessment more personalized.

[0096] Simultaneous improvement in model sensitivity and prediction accuracy: The credibility-corrected ER status input not only enhances the risk model's responsiveness to low-level ER expression fluctuations, but also increases the credibility of the overall prediction output, significantly reducing ambiguity in risk value boundary determination (e.g., between low and medium risk).

[0097] Furthermore, this method enhances the assessment model's adaptability to dynamic clinical data, providing more proactive risk assessment support for long-term postoperative management of breast cancer. This approach is particularly important for cases in the clinical "gray zone," such as those with mild image abnormalities but uncertain ER expression. This example provides an important auxiliary judgment mechanism that effectively reduces the risk of misdiagnosis and missed diagnosis.

[0098] Furthermore, before S10, it also includes:

[0099] S60. The adaptive risk assessment model is trained based on historical case data, including: patients with estrogen receptor-positive breast cancer who had one breast removed, images of calcification foci in the other breast of patients who were re-diagnosed with breast cancer, multi-parameter quantitative analysis results, and corresponding clinical follow-up information.

[0100] S40 includes: performing multi-center collaborative training on the adaptive risk assessment model through a federated learning mechanism, wherein the type of the federated learning mechanism is horizontal federated learning.

[0101] Multi-center collaborative training through federated learning mechanisms includes:

[0102] S41. Extract the estrogen receptor-positive patient dataset from the local database of each participant, where the patient dataset has been processed with differential privacy; the patient dataset includes the patient's basic information, imaging data, clinical treatment records, calcification annotation and pathology verification data, dynamic follow-up data, and federated learning metadata.

[0103] S42. Decrypt the data sets of each participant and aggregate them according to dynamic weights to generate a training data set. The dynamic weights include a basic weight and a correction factor. The basic weight is calculated based on the proportion of the number of cases of each participant to the total number of cases. If the historical diagnostic consistency of the participant is higher than 90%, the weight is increased by 20%; if the participant has more than 10% feedback correction records, the weight is decreased by 20%;

[0104] S43. Train the adaptive risk assessment model using the training data set.

[0105] In this example, it's important to note that the introduction of a federated learning mechanism for multi-center collaborative training of the adaptive risk assessment model aims to address the data silos, sample limitations, and privacy protection challenges inherent in traditional AI model training. This is particularly true for breast cancer risk prediction, where the limited number of cases per center, heterogeneous data structures, and the inconvenience of sharing patient privacy data directly impact the model's generalization and clinical adaptability.

[0106] To this end, the embodiment improves the training quality and application value of the risk assessment model through the following technical designs:

[0107] Typical cases were selected as the training basis: the historical case data selected were ER-positive patients who had one breast removed and were re-diagnosed with breast cancer on the other side. These patients had a clear bilateral breast pathological process. The contralateral calcification images and follow-up information provided constituted high-quality training samples, which was conducive to the model learning the imaging characteristics and quantitative parameter patterns of potential high-risk calcifications.

[0108] Ensure data privacy and compliance security: Before providing patient data, each center performs differential privacy processing on the data to avoid patient identity leakage; the federated learning mechanism adopts a horizontal federated learning architecture, only updating model parameters locally, and the original data is not transmitted during the parameter aggregation process, effectively preventing the leakage of sensitive information.

[0109] Introducing multi-dimensional collaborative data: Each patient dataset not only contains imaging and pathology annotation data, but also integrates clinical treatment records, dynamic follow-up information, and metadata required for federated learning, making model training closer to actual clinical scenarios and enhancing its adaptability to multi-source heterogeneous data.

[0110] Construct a dynamic weight aggregation mechanism: adopt a "basic weight + correction factor" weight strategy, guided by improving the quality of training data. The basic weight reflects the contribution of cases; the correction factor regulates the data quality of participants from the two perspectives of "historical diagnostic consistency" and "feedback correction record"; the weight increase / decrease strategy guides high-quality data to play a greater role in model training, thereby improving the robustness and accuracy of the final model.

[0111] Strengthen the generalization and adaptability of the model: Through the joint optimization of multi-center training data in step S43, the adaptive risk assessment model can better adapt to the real application environment in different regions, different equipment, and different diagnostic processes, and has good generalizability.

[0112] Linked with subsequent ER credibility correction and quantitative analysis to enhance risk prediction accuracy: Since the model training has integrated real recurrence data and a complete follow-up path, its output risk value can better form a logical closed loop with subsequent steps (such as the corrected ER status and image parameter analysis results), thereby improving the ability to identify the malignant evolution trend of early microscopic calcifications.

[0113] The implementation of the technical solution of this embodiment significantly improves the generalization ability, data security and medical credibility of the breast cancer risk assessment model: it achieves "high-quality, high-privacy" collaboration in model training: by introducing a horizontal federated learning mechanism, all medical participants can collaborate in model training without sharing original data, effectively protecting patient privacy and data compliance (in compliance with GDPR, HIPAA and other regulatory requirements), while ensuring that the model can access larger-scale, cross-population data distribution.

[0114] Fully utilize real recurrence case training data: This embodiment is based on real cases of "unilateral mastectomy + recurrence on the other side", and uses its calcification focus images and follow-up results together for model training, so that the model has the clinical insight of "identifying future disease risks from potentially unaffected breasts".

[0115] Establish an interpretable risk model input system: The training dataset contains not only calcification images and structural parameters, but also estrogen receptor expression status, treatment response records and follow-up information. This ensures that risk assessment is not only based on image features, but also reflects the patient's molecular characteristics and disease evolution dynamics, enhancing the medical interpretability of the model.

[0116] The dynamic weight mechanism improves the fairness and robustness of model training: by setting basic weights, it ensures that participants with a larger number of cases contribute more to the model; at the same time, it introduces diagnostic consistency and feedback correction records as correction factors to quantitatively control the diagnostic quality and accuracy, avoiding participants with "large data volume but poor quality" from misleading the model; and realizes a quality-oriented training collaboration mechanism to effectively improve the model convergence speed and prediction accuracy.

[0117] Cross-center training improves generalization capabilities: Through collaborative training of multi-center data, the model overfitting problem caused by single-center sample bias is effectively reduced, so that the trained adaptive risk assessment model can maintain stable predictive performance in more different regions, different races and treatment strategies.

[0118] Constructing a model framework suitable for early screening scenarios: The training samples in this embodiment come from patients with contralateral recurrence after surgery. The images of their calcification foci can be regarded as potential lesion images before recurrence. The model trained on this basis has a stronger "risk prediction" ability, which is in line with the actual clinical needs of early screening and dynamic monitoring of breast cancer.

[0119] By introducing a federated learning mechanism to achieve high-quality cross-center training, combined with real clinical recurrence data and credible corrected ER status input, a more credible, generalizable and practical adaptive risk assessment model for breast cancer is constructed.

[0120] In an optional embodiment, selecting the detection mode in step S20 includes:

[0121] When breast density is classified as C or D, a digital breast tomosynthesis device is used for multi-angle scanning, and a three-dimensional calcification distribution map is reconstructed using the sliding window method. The frequency domain fusion algorithm is used to enhance the contrast of the calcification edge.

[0122] When the breast density is classified as A or B, a dual-energy X-ray molybdenum target is used to generate a calcification energy spectrum image, and the energy spectrum separation technology is used to distinguish the energy spectrum signals of hydroxyapatite and calcium salts.

[0123] In this embodiment, it should be noted that breast density is an important factor affecting the accuracy of breast calcification detection and imaging clarity. Tissue structures corresponding to different density levels have large differences in X-ray absorption. Therefore, the optimal detection mode should be intelligently selected according to the breast density level to improve the recognition rate of calcification foci and the accuracy of quantitative analysis.

[0124] Specifically, when the density is classified as C or D (high-density breast), this type of breast tissue is dense, and traditional two-dimensional mammography is prone to tissue overlap, obscuring tiny calcifications and reducing the detection rate. To this end, this embodiment uses Digital Breast Tomosynthesis (DBT) to perform multi-angle scanning to obtain images from different perspectives. Using a sliding window method, the multi-angle images are locally reconstructed to generate a three-dimensional calcification distribution map, which can effectively reduce the interference caused by tissue overlap.

[0125] To enhance the edge features of calcifications, a frequency domain fusion algorithm was further introduced. This algorithm fuses features of images from different perspectives in the frequency domain to enhance the high-frequency signal response, thereby improving the contrast between the calcification edge and background tissue, making small or low-contrast calcifications more clearly visible in the reconstructed image.

[0126] When the density is classified as Class A or Class B (low-density breast), the breast fat content is high, X-ray penetration is strong, and the effect of tissue overlap is minimal. Dual-energy X-ray imaging can be used. This technology uses two different energies of X-rays to image the tissue and generate a calcification spectrum image.

[0127] Spectral decomposition technology can distinguish the energy responses of different components in an image, thereby identifying characteristic material signals. Specifically, this technology can distinguish two common chemical components in breast calcifications: hydroxyapatite (associated with malignant calcifications) and calcium salt deposits (more common in benign lesions), providing a more physical basis for subsequent risk classification and lesion nature assessment.

[0128] In summary, this embodiment, by combining intelligent selection of detection modes based on breast density, not only improves the imaging resolution of calcification foci in high-density breasts, but also enhances the ability to identify calcification components in low-density breasts. This establishes a breast calcification detection pathway with precise classification and clear characteristics, providing more reliable input data for subsequent image analysis, risk modeling, and personalized diagnosis and treatment.

[0129] Implementation of the technical solution of this embodiment enables personalized detection mode selection for patients with different breast density levels, thereby improving the accuracy and image quality of breast calcification detection. Specifically, it can improve the detection rate of calcifications in high-density breasts: by using digital breast tomosynthesis equipment to perform multi-angle imaging of Class C or Class D high-density breasts, combined with a sliding window reconstruction algorithm and frequency domain fusion processing, it can effectively reduce artifact interference caused by tissue overlap, enhance the edge clarity of calcifications and the three-dimensional presentation of their spatial distribution, and facilitate the detection and location of early microcalcifications.

[0130] It can also enhance the ability to identify the calcification properties of low-density breasts: For Class A or Class B low-density breasts, dual-energy X-ray mammography combined with energy spectrum separation technology can achieve preliminary differentiation of the composition of calcification materials at the image level, especially in distinguishing hydroxyapatite (indicating potential malignancy) from other calcium salts (possibly benign calcifications), providing a basic physical basis for judging the nature of the lesions.

[0131] It can also optimize the data input quality of subsequent risk assessment models: due to the precise adaptation of the detection mode, the acquired image data and quantitative analysis parameters are more representative and stable, thereby providing a better data basis for subsequent image feature extraction, ER status credibility correction and risk prediction modeling, and improving the accuracy and clinical application value of the entire breast cancer risk prediction process.

[0132] In a specific implementation, after S50, the method further includes:

[0133] S70: Output personalized inspection suggestions based on the risk level corresponding to the risk value, where:

[0134] When the risk level is low, patients are recommended to undergo routine breast imaging every 12 months;

[0135] When the risk level is moderate, patients are recommended to undergo enhanced breast imaging every 6 months and to assess compliance with hormone therapy;

[0136] When the risk level is high, patients are recommended to undergo high-level breast imaging every 3 months, and intervention decisions should be made in combination with tumor marker monitoring and multidisciplinary consultation.

[0137] In this embodiment, it should be noted that the output of personalized examination recommendations is based on the refined management goal of the patient's risk level, which aims to achieve early warning and effective intervention of the risk of breast cancer recurrence or new onset through a dynamic grading strategy. Specifically, the risk value is derived by an adaptive risk assessment model that comprehensively considers multidimensional information such as image structure parameters, corrected ER status, clinical history, etc. The output risk level division refers to previous epidemiological big data statistics and clinical intervention standards to ensure the scientificity and practicality of the grading basis. For patients with low risk levels, conventional breast imaging (such as digital mammography) can meet the monitoring needs, which can effectively detect potential lesions and avoid waste of resources and psychological burden on patients. For medium-risk patients, such patients may have decreased tolerance to hormone therapy or mild abnormalities in imaging indicators. Therefore, it is necessary to increase the frequency of review and use more sensitive enhanced imaging methods (such as MRI or CE-MAM). At the same time, combined with hormone therapy compliance assessment, it provides a basis for whether to adjust the treatment plan. For individuals with higher risk levels, a comprehensive strategy of "close monitoring + active intervention" needs to be established. Not only does the imaging frequency and level need to be increased, but dynamic monitoring of tumor markers (such as CA15-3, CEA, etc.) should also be combined. At the same time, comprehensive consultation by a multidisciplinary team (MDT) can achieve personalized adjustment of treatment and follow-up pathways and reduce the risk of recurrence and missed diagnosis.

[0138] The technical solution implemented in this embodiment provides a dynamic risk assessment method for breast cancer based on an adaptive model, which combines multi-source data analysis with intelligent model prediction to improve the accuracy of early detection of breast cancer. This method uses a deep learning model to perform risk assessment by integrating the patient's clinical data and imaging data, and introduces a credibility correction mechanism to consider factors such as the duration of hormone therapy, the patient's medical history, and ER expression trends to ensure the reliability of the assessment results. In addition, the system implements multi-center data collaborative training through a federated learning mechanism, which improves the generalization ability of the model while ensuring patient privacy. Based on the risk level assessment results, the system can output personalized inspection recommendations, provide different inspection frequencies and intervention plans for low-, medium-, and high-risk patients, and help clinicians develop precise personalized intervention strategies to improve breast cancer monitoring results.

[0139] In a specific embodiment, the present invention further provides a breast cancer dynamic risk assessment system based on an adaptive model, which implements the above-mentioned breast cancer dynamic risk assessment method based on an adaptive model, including:

[0140] Data acquisition module, used to collect clinical data and imaging data of target patients;

[0141] Image analysis module, used for parameter extraction and quantitative analysis of calcification images;

[0142] The model processing module includes a trained adaptive risk assessment model, which is used to receive parameter analysis results and output breast cancer risk values;

[0143] The assessment report generation module is used to generate a breast cancer risk assessment report based on the risk value.

[0144] The model processing module includes: a federated learning training submodule, which is used to receive encrypted data from multiple participants and conduct cross-center collaborative training.

[0145] In this embodiment, it's important to note that the system design fully considers the needs of patient data privacy protection and multi-center data sharing. It utilizes a federated learning mechanism, enabling different healthcare participants to share model training experience without directly exchanging patient data, thereby improving the accuracy and generalization of risk assessment models. Specifically, federated learning allows each participant to perform data training locally, sharing only model parameters rather than raw data, thus preventing the leakage of sensitive patient information.

[0146] Furthermore, the data acquisition module supports the collection of multiple data types, ensuring that key information such as a patient's clinical data, imaging data, and pathology records are fully captured and provided for subsequent analysis and evaluation. The image analysis module can extract the characteristics of calcifications from breast images and perform quantitative analysis, providing accurate input data for the model processing module.

[0147] When conducting assessments, the model processing module not only relies on traditional analysis based on historical data, but also dynamically learns individual differences and treatment responses of patients to ensure real-time updates and adaptive adjustments to breast cancer risk assessments, especially customized processing for factors such as hormone therapy compliance and breast density for different patients.

[0148] Finally, the assessment report generation module provides clinicians with a detailed assessment report based on the calculated breast cancer risk value. The report includes the patient's risk level, regular examination recommendations, possible intervention measures, etc., providing strong decision-making support for the patient's subsequent treatment and monitoring.

[0149] The technical solution of this embodiment continuously optimizes assessment results through multi-dimensional data input (including clinical data, imaging data, and historical case data), combined with the dynamic learning mechanism of the adaptive model. The system can adjust the assessment model in real time to better adapt to the individual characteristics of different patients, thereby improving the accuracy of breast cancer risk assessment.

[0150] The system supports collaborative training among multiple healthcare stakeholders through a federated learning mechanism. By training data locally and sharing encrypted model parameters, it effectively prevents the leakage of sensitive patient data. It also enables collaborative optimization across different medical centers, improving the model's generalization and accuracy.

[0151] By quantitatively analyzing patient imaging data and accurately processing calcification images, the system can promptly identify potential breast cancer risk points. Combined with the patient's medical history, treatment records, and other information, it enables early screening and accurate diagnosis, providing timely intervention and treatment recommendations.

[0152] Based on the assessment results, the system can generate personalized examination recommendations, suggesting different examination frequencies and intervention measures based on the patient's risk level. Low-risk patients are recommended to undergo regular follow-up examinations, while medium-risk patients are recommended to undergo more frequent examinations and assess hormone therapy compliance. High-risk patients are recommended to undergo advanced breast imaging and multidisciplinary consultations, helping doctors develop personalized health management plans for their patients.

[0153] Federated learning's encrypted data transmission and training methods effectively prevent the leakage and misuse of patients' personal data, ensuring patient privacy. Data is stored and processed only locally, and only encrypted model update data is shared, ensuring data security.

[0154] The assessment report not only provides the breast cancer risk level but also provides specific examination and treatment recommendations. This allows doctors to make informed clinical decisions based on the report, improve early detection rates for breast cancer, and tailor treatment plans to the patient's specific circumstances.

[0155] The above-described embodiments merely represent several embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the spirit of the present application, and such modifications and improvements are all within the scope of protection of the present application.

Claims

1. A dynamic risk assessment method for breast cancer based on an adaptive model, applied to patients with breast cancer who have undergone unilateral mastectomy, characterized by: The patient is an estrogen receptor-positive patient, and the dynamic risk assessment method comprises the following steps: S10, obtaining clinical data of the patient's unresected breast, the clinical data including: estrogen receptor status test results, duration of hormone therapy, and breast density grading data based on the BI-RADS standard; S20, selecting a corresponding imaging detection mode according to the breast density grading data and the estrogen receptor status detection result, and acquiring an image of the calcification focus of the target breast; S30, performing multi-parameter quantitative analysis on the calcification focus image, wherein the parameters include: hydroxyapatite crystal ratio, calcification edge sharpness index, and calcification cluster spatial distribution entropy value; S40, using the multi-parameter quantitative analysis results as input, calculating the risk value using the trained adaptive risk assessment model, and outputting the breast cancer risk value of the target breast; S50: Compare the risk value with a preset risk classification threshold and generate an assessment report based on the risk level.

2. The method for dynamic breast cancer risk assessment based on an adaptive model according to claim 1, characterized in that: The S10 further includes: S11, performing a credibility correction on the estrogen receptor status test result, wherein the credibility correction is based on the duration of hormone treatment, the patient's medical history, and the past estrogen receptor expression trend; Accordingly, the S40 is: taking the multi-parameter quantitative analysis result and the corrected estrogen receptor status as input, performing calculations using the trained adaptive risk assessment model, and outputting the patient's breast cancer risk value.

3. The method for dynamic risk assessment of breast cancer based on an adaptive model according to claim 1, wherein: The S10 and above also include: S60. Training the adaptive risk assessment model based on historical case data, wherein the historical case data includes: images of calcification foci in the other breast of a patient with estrogen receptor-positive breast cancer whose one breast had been removed, multi-parameter quantitative analysis results, and corresponding clinical follow-up information after the patient's other breast was re-diagnosed with breast cancer.

4. The method for dynamic breast cancer risk assessment based on an adaptive model according to claim 3, wherein: The S40 includes: The adaptive risk assessment model is trained in a multi-center collaborative manner through a federated learning mechanism, wherein the type of the federated learning mechanism is horizontal federated learning.

5. The method for dynamic risk assessment of breast cancer based on an adaptive model according to claim 4, wherein: The multi-center collaborative training through the federated learning mechanism includes: S41. Extracting an estrogen receptor-positive patient dataset from the local database of each participant, wherein the patient dataset has been subjected to differential privacy processing; S42. Decrypt the patient data sets of each participant and aggregate them according to dynamic weights to generate a training data set. The dynamic weights include a basic weight and a correction factor. The basic weight is calculated based on the proportion of the number of cases of each participant to the total number of cases. If the historical diagnostic consistency of the participant is higher than 90%, the weight is increased by 20%; if the participant has more than 10% feedback correction records, the weight is decreased by 20%. S43. Training the adaptive risk assessment model using the training data set.

6. The method for dynamic breast cancer risk assessment based on an adaptive model according to claim 5, characterized in that: The patient dataset includes the patient's basic information, imaging data, clinical treatment records, calcification annotation and pathology verification data, dynamic follow-up data, and federated learning metadata.

7. The method for dynamic breast cancer risk assessment based on an adaptive model according to claim 1, wherein: The selection of the detection mode in S20 includes: When the breast density is classified as C or D, a digital breast tomosynthesis device is used to perform multi-angle scanning, a three-dimensional calcification distribution map is reconstructed using a sliding window method, and a frequency domain fusion algorithm is used to enhance the contrast of the calcification edge; When the breast density is classified as Class A or Class B, a dual-energy X-ray molybdenum target is used to generate a calcification energy spectrum image, and the energy spectrum signals of hydroxyapatite and calcium salt are distinguished by energy spectrum separation technology.

8. The method for dynamic breast cancer risk assessment based on an adaptive model according to claim 1, wherein: The S50 further includes: S70: Outputting personalized inspection suggestions based on the risk level corresponding to the risk value, wherein: When the risk level is low, it is recommended that the patient undergo routine breast imaging every 12 months; When the risk level is moderate, it is recommended that the patient undergo enhanced breast imaging every 6 months and assess compliance with hormone therapy; When the risk level is high, it is recommended that the patient undergo high-level breast imaging every 3 months, and make intervention decisions in combination with tumor marker monitoring and multidisciplinary consultation.

9. A breast cancer dynamic risk assessment system based on an adaptive model, which implements the breast cancer dynamic risk assessment method based on an adaptive model according to claims 1-8, characterized in that: include: Data acquisition module, used to collect clinical data and imaging data of target patients; Image analysis module, used for parameter extraction and quantitative analysis of calcification images; The model processing module includes a trained adaptive risk assessment model, which is used to receive parameter analysis results and output breast cancer risk values; The assessment report generation module is used to generate a breast cancer risk assessment report based on the risk value.

10. The breast cancer dynamic risk assessment system based on the adaptive model according to claim 9, characterized in that: The model processing module includes: The federated learning training submodule is used to receive encrypted data from multiple participants and conduct cross-center collaborative training.

Images