A stroke risk scoring system based on carotid HR-MRI intraplaque hemorrhage imaging features and a storage medium
Patent Information
- Application Number
- CN202611218389.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-12
- Publication Date
- 2026-10-30
AI Technical Summary
[0009]针对现有技术的不足，本发明提供了一种基于颈动脉HR-MRI斑块内出血影像组学特征的卒中风险评分系统及存储介质，解决传统 IPH 评估主观性强、量化不足、缺乏人群特异性校正、预测效能有限的问题，实现 IPH 自动分割、影像组学量化提取、人群分层 Z分数校正、逻辑回归风险建模与三级智能分层，同时具备临床随访闭环优化能力，提升卒中早期筛查与辅助诊断的精准性与标准化
[0027]（1）实现斑块内出血区域自动识别。本发明采用基于3D U-Net深度学习模型的自动分割方式，对HR-MRI图像中的IPH区域进行自动检测和轮廓提取，降低人工勾画误差，提高处理效率和一致性。
Smart Images
Figure CN123017493A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical software, specifically to a stroke risk scoring system and storage medium based on radiomics characteristics of intraplaque hemorrhage in carotid artery HR-MRI. Background Technology
[0002] Vulnerable atherosclerotic plaques in the carotid arteries are a significant cause of ischemic stroke. Numerous studies have shown that intraplaque hemorrhage (IPH) is one of the strongest predictors of plaque vulnerability, with patients possessing IPH having a 3-5 times higher stroke risk compared to those without. High-resolution MRI (HR-MRI) T1-weighted black-blood sequences (T1WI-BB) and MP-RAGE sequences can detect IPH with high sensitivity and are currently the gold standard for IPH imaging identification. However, existing IPH assessment techniques have the following drawbacks:
[0003] Qualitative assessment is the primary method, lacking quantitative indicators: In clinical practice, radiologists only report "presence or absence of IPH" or "IPH range (small / medium / large)," failing to accurately quantify the heterogeneity of the IPH region (such as hemorrhage volume, morphological irregularity, spatial relationship with the lipid core, etc.), resulting in the loss of a large amount of microscopic information.
[0004] The predictive power of solely relying on the presence or absence of an IPH is limited: risk stratification based solely on the presence or absence of an IPH typically yields an AUC of only 0.70-0.75, which is insufficient for accurate screening. Some patients may have an IPH, but its area may be small, and it may not necessarily lead to short-term stroke; some patients without an IPH are also at high risk due to rupture of the fibrous cap.
[0005] Lack of population-specific correction: There are significant differences in carotid wall thickness and IPH volume distribution among different age groups, sexes, and those with a history of hypertension. Directly using a uniform threshold to judge "large IPH" can lead to misclassification.
[0006] IPH segmentation relies on manual labor, which is inefficient: manually delineating the IPH region is time-consuming and has poor repeatability, and there is a lack of automated segmentation algorithms for IPH.
[0007] Lack of closed-loop optimization capability: The existing IPH assessment model is static and cannot continuously optimize the risk threshold based on follow-up stroke events.
[0008] Therefore, there is an urgent need for a stroke risk scoring system based on IPH radiomics characteristics that is quantifiable and has population-specific correction and self-learning capabilities. Summary of the Invention
[0009] To address the shortcomings of existing technologies, this invention provides a stroke risk scoring system and storage medium based on carotid artery HR-MRI intraplaque hemorrhage radiomics features. It solves the problems of traditional IPH assessment, such as high subjectivity, insufficient quantification, lack of population-specific correction, and limited predictive efficacy. It achieves automatic IPH segmentation, radiomics quantitative extraction, population stratification Z-score correction, logistic regression risk modeling, and three-level intelligent stratification. At the same time, it has the ability to optimize clinical follow-up closed loop, improving the accuracy and standardization of early stroke screening and auxiliary diagnosis.
[0010] To achieve the above objectives, the present invention is implemented through the following technical solution: a stroke risk scoring system and storage medium based on carotid artery HR-MRI intraplaque hemorrhage radiomics features, including a data acquisition module, an IPH radiomics database, a population-specific reference database, an intelligent segmentation and feature extraction engine, a risk scoring and stratification module, a report generation module, and a feedback optimization module.
[0011] Among them, the data acquisition module is configured to acquire basic clinical information of the subject's carotid artery HR-MRI images.
[0012] The IPH radiomics database pre-stores subgroup reference medians and standard deviations of IPH regional radiomics characteristics based on local populations.
[0013] Population-specific reference database: Pre-stored specific reference data on carotid artery bifurcation wall thickness, but without baseline incidence data of IPH in the population.
[0014] Intelligent segmentation and feature extraction engine: including IPH automatic segmentation unit and radiomics feature extraction unit.
[0015] Risk scoring and stratification module: includes feature screening unit, Z-score correction unit, comprehensive risk index calculation unit, and risk stratification unit.
[0016] The report generation module is configured to generate a structured screening report based on the relationship between the comprehensive risk index and a preset threshold.
[0017] The feedback optimization module is configured to continuously update the reference range and risk thresholds based on clinical follow-up data.
[0018] Preferably, the IPH automatic segmentation unit uses a 3D U-Net deep learning model to automatically segment the IPH region. The 3D U-Net deep learning model takes T1WI black blood sequence and MP-RAGE sequence as input, and the Dice coefficient of the model test set is not less than 0.85. The radiomics feature extraction unit extracts multi-dimensional radiomics features of IPH.
[0019] Preferably, the feature screening unit uses LASSO regression to screen core radiomics features; the Z-score correction unit retrieves population stratification reference values to calculate feature-corrected Z-scores; the comprehensive risk index calculation unit constructs the IPH-RI stroke risk index through a logistic regression model; and the risk stratification unit divides stroke risk into three levels based on preset thresholds.
[0020] Preferably, the Z-score correction unit is configured according to the formula... ) calculate, where and Based on the subject's age, gender, and history of hypertension, data are dynamically retrieved from the IPH radiomics database.
[0021] Preferably, the formula for calculating the comprehensive risk index is as follows: Weight It was obtained through logistic regression training.
[0022] Preferably, the risk stratification unit defines IPH-RI < 0.15 as low risk, 0.15-0.45 as medium risk, and ≥0.45 as high risk.
[0023] Preferably, the IPH radiomics database and the population-specific reference database are fitted with a continuous reference curve using LOESS local weighted regression, with fitting parameters span=0.5 and degree=1.
[0024] Preferably, the feedback optimization module refits the reference median, standard deviation, and logistic regression coefficient every 6 months, and synchronously updates the database parameters, logistic regression weights, and risk thresholds every 6 months, using the Youden index maximization criterion to select the optimal boundary threshold.
[0025] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the function of the stroke risk scoring system described in any of the above embodiments.
[0026] This invention provides a stroke risk scoring system and storage medium based on radiomics characteristics of intraplaque hemorrhage on carotid artery HR-MRI. It has the following beneficial effects:
[0027] (1) Automatic identification of intraplaque hemorrhage areas. This invention adopts an automatic segmentation method based on the 3D U-Net deep learning model to automatically detect and extract the contours of IPH regions in HR-MRI images, reducing manual delineation errors and improving processing efficiency and consistency.
[0028] (2) Achieve quantitative evaluation of IPH heterogeneity. Unlike the traditional method of only judging "whether there is an IPH", this invention achieves quantitative analysis of the microstructure of IPH by extracting multi-dimensional image omics parameters such as volume features, morphological features, texture features and wavelet features.
[0029] (3) Improve the accuracy of stroke risk prediction. By screening core radiomics features through LASSO and combining them with a logistic regression model to establish the IPH-RI comprehensive risk index, the accuracy of stroke risk prediction was improved, and the AUC of the test set reached 0.91.
[0030] (4) Achieve population-specific correction. Establish a reference database based on factors such as age, gender, and history of hypertension, and use a Z-score correction mechanism to eliminate individual differences and improve the reliability of risk assessment results.
[0031] (5) Achieve dynamic self-learning optimization. The system can continuously update database parameters and model weights based on subsequent clinical follow-up results, achieving closed-loop optimization of the risk assessment model. Attached Figure Description
[0032] Figure 1 This is a system flowchart of the present invention. Detailed Implementation
[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0034] Please see Figure 1 This invention provides a technical solution: a stroke risk scoring system based on carotid artery HR-MRI plaque hemorrhage radiomics features, comprising a data acquisition module, an IPH radiomics database, a population-specific reference database, an intelligent segmentation and feature extraction engine, a risk scoring and stratification module, a report generation module, and a feedback optimization module. Wherein:
[0035] 1. Data Acquisition Module
[0036] The system is configured to acquire the subject's carotid artery HR-MRI images (containing at least T1WI black blood sequences and MP-RAGE sequences), basic clinical information (age, sex, history of hypertension, history of diabetes, history of dyslipidemia, smoking history), and history of stroke / TIA.
[0037] 2. IPH Radiomics Database
[0038] Pre-stored reference ranges for IPH regional radiomics features based on local populations, including:
[0039] Subgroup median and standard deviation of IPH volume (mm³)
[0040] Subgroup median and standard deviation of IPH surface irregularity
[0041] Subgroup median and standard deviation of IPH to total plaque volume ratio
[0042] Subgroup median and standard deviation of IPH region gray-level co-occurrence matrix (GLCM) contrast
[0043] The above characteristics were stratified by age (≤65 years / >65 years), gender, and history of hypertension.
[0044] The construction method is as follows:
[0045] S1: Construction of the IPH radiomics database
[0046] We collected data from 800 patients with carotid artery plaques confirmed by HR-MRI and pathology (post-CEA specimens) or clinical follow-up (of whom at least 200 had ipsilateral stroke within 6 months).
[0047] Two neuroradiologists manually delineated the IPH region on T1WI black blood sequences (included with Dice coefficient ≥0.85).
[0048] Stratified by age (≤65 years / >65 years), sex, and history of hypertension, the median and standard deviation of IPH volume, surface irregularity index, and GLCM contrast were calculated for each subgroup.
[0049] Local weighted regression (LOESS, parameters span=0.5, degree=1) is used to fit continuous feature-reference median and feature-reference standard deviation curves to solve the boundary jump problem caused by hierarchical discretization.
[0050] S2: Construction of Population-Specific Reference Database
[0051] We collected 500 healthy controls with no history of stroke and no significant plaque in the ipsilateral carotid artery (IMT < 1.0 mm), stratified by age and sex, and calculated the median and standard deviation of the carotid artery bifurcation wall thickness.
[0052] S3: Feature Selection and Weight Training
[0053] Using ipsilateral stroke events within 6 months as the gold standard, the radiomics features of the training set (accounting for 70% of the total) were screened by LASSO.
[0054] Features with non-zero coefficients are retained as the core feature set (expected to be 5-10).
[0055] Using the adjusted Z-scores of the core features and clinical variables (age, history of hypertension) as independent variables and stroke events as dependent variables, multivariate logistic regression was performed to obtain the regression coefficients β_i for each feature.
[0056] Normalized weights Preserve the sign direction (positive / negative risk).
[0057] S4: Risk Threshold Determination
[0058] On the validation set (15% of the total), sensitivity and specificity were calculated at different IPH-RI thresholds, and the point that maximized the Youden index (sensitivity + specificity - 1) was used as the default risk threshold. Meanwhile, a low-risk threshold of 0.15 (90% sensitivity) and a high-risk threshold of 0.45 (80% specificity) were preset for flexible clinical selection.
[0059] 3. Population-Specific Reference Database
[0060] The following reference data is pre-stored:
[0061] Age / sex-specific reference median and standard deviation of ipsilateral carotid artery bifurcation wall thickness
[0062] Age / sex-specific baseline incidence of IPH in stroke-free populations
[0063] 4. Intelligent segmentation and feature extraction engine
[0064] include:
[0065] (a) Automatic IPH segmentation unit: A deep learning model based on 3D U-Net is used to automatically segment the IPH region on T1WI black blood sequences and MP-RAGE sequences. The model is trained on 1000 manually labeled IPH datasets, with a target Dice coefficient ≥ 0.85.
[0066] (b) Radiomics Feature Extraction Unit: Extract the following features from the segmented IPH region:
[0067] Feature categories Specific features quantity Volumetric morphological characteristics Absolute volume (mm³), relative volume (IPH volume / patch volume), surface irregularity index, sphericity 8 First-order statistical characteristics Mean, median, standard deviation, skewness, kurtosis, energy, and entropy of signal intensity in the IPH region. 12 Texture Features (GLCM) Contrast, Correlation, Energy, Homogeneity, Dissimilarity 24 Texture Features (GLRLM) Unevenness in journey length, unevenness in grayscale, emphasis on short journeys 16 wavelet features The above features were extracted after undergoing eight wavelet decomposition methods. Approximately 200 total Approximately 260
[0068] 5. Risk Scoring and Stratification Module
[0069] include:
[0070] (a) Feature selection unit: LASSO regression (10-fold cross-validation) was used to select 5-10 core radiomics features most relevant to stroke events from 260 features.
[0071] (b) Z-score correction unit: For each core feature, based on the subject's age, sex, and history of hypertension, the corresponding subgroup reference median M(h) and reference standard deviation SD(h) are retrieved from the IPH radiomics database, and the corrected Z-score is calculated using the following formula:
[0072] in, This represents the original value of the i-th radiomics feature of the subject.
[0073] (c) Integrated Risk Index (IPH-RI) Calculation Unit: A logistic regression model is used, with the diagnosis result (ipsilateral ischemic stroke within 6 months = 1, no stroke = 0) as the dependent variable, and the adjusted Z-scores of each core feature and clinical variables (age, history of hypertension) as independent variables. Normalized weights are obtained through training. The formula for calculating the comprehensive risk index is as follows:
[0074]
[0075] The output value is between 0 and 1.
[0076] (d) Risk Stratification Unit: Based on the relationship between the IPH-RI value and a preset threshold, subjects are divided into three risk levels:
[0077] IPH-RI range Risk level Recommended management strategies < 0.15 Low risk Routine follow-up and lifestyle intervention 0.15 - 0.45 Medium risk Intensive statin therapy, with a follow-up HR-MRI at 6 months. ≥ 0.45 High risk Consider carotid endarterectomy (CEA) or carotid stent implantation (CAS).
[0078] The preset threshold is determined in the training set by maximizing the Yoden exponent and is iteratively updated with the feedback optimization module.
[0079] 6. Report Generation Module
[0080] Configured to generate structured screening reports containing the following:
[0081] Basic information of the examinee (age, gender, history of hypertension)
[0082] Automatic IPH segmentation results (IPH volume, IPH / plaque volume ratio)
[0083] Core radiomics features and corrected Z-scores
[0084] IPH-RI Comprehensive Risk Index and Risk Level
[0085] Stroke risk probability (prediction of ipsilateral stroke incidence within 6 months)
[0086] Clinical management recommendations (based on risk level)
[0087] 7. Feedback Optimization Module
[0088] It is configured to: receive clinical follow-up diagnostic data (whether ipsilateral ischemic stroke occurred within 6 months), periodically (every 6 months) refit the IPH radiomics feature reference range (reference median, standard deviation), retrain the logistic regression model to optimize feature weights and risk thresholds, update the IPH radiomics database and population-specific reference database, and achieve closed-loop self-optimization of the system.
[0089] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the functions described in the above system.
[0090] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A stroke risk scoring system based on radiomics characteristics of intraplaque hemorrhage on carotid artery HR-MRI, characterized in that: It includes a data acquisition module, an IPH radiomics database, a population-specific reference database, an intelligent segmentation and feature extraction engine, a risk scoring and stratification module, a report generation module, and a feedback optimization module; Among them: the data acquisition module is configured to acquire basic clinical information of the subject's carotid artery HR-MRI images; The IPH radiomics database pre-stores subgroup reference medians and standard deviations of IPH regional radiomics characteristics based on local populations. Population-specific reference database: Pre-stored specific reference data on carotid artery bifurcation wall thickness, but without baseline incidence data of IPH in the population; Intelligent segmentation and feature extraction engine: including IPH automatic segmentation unit and radiomics feature extraction unit; Risk scoring and stratification module: includes feature screening unit, Z-score correction unit, comprehensive risk index calculation unit, and risk stratification unit; The report generation module is configured to generate a structured screening report based on the relationship between the comprehensive risk index and a preset threshold; The feedback optimization module is configured to continuously update the reference range and risk thresholds based on clinical follow-up data.
2. The stroke risk scoring system based on carotid artery HR-MRI plaque hemorrhage radiomics characteristics according to claim 1, characterized in that: The IPH automatic segmentation unit uses a 3D U-Net deep learning model to automatically segment the IPH region. The 3D U-Net deep learning model takes T1WI black blood sequence and MP-RAGE sequence as input, and the Dice coefficient of the model test set is not less than 0.
85. The radiomics feature extraction unit extracts multi-dimensional radiomics features of IPH.
3. The stroke risk scoring system based on radiomics characteristics of intraplaque hemorrhage on carotid artery HR-MRI according to claim 1, characterized in that: The feature screening unit uses LASSO regression to screen core radiomics features; the Z-score correction unit retrieves population stratification reference values to calculate feature-corrected Z-scores; the comprehensive risk index calculation unit constructs the IPH-RI stroke risk index through a logistic regression model; and the risk stratification unit divides stroke risk into three levels based on preset thresholds.
4. A stroke risk scoring system based on radiomics characteristics of intraplaque hemorrhage on carotid artery HR-MRI according to claim 3, characterized in that: The Z-score correction unit is according to the formula ( ) Calculate, where and Based on the subject's age, gender, and history of hypertension, data are dynamically retrieved from the IPH radiomics database.
5. A stroke risk scoring system based on radiomics characteristics of intraplaque hemorrhage on carotid artery HR-MRI according to claim 1, characterized in that: The formula for calculating the comprehensive risk index is as follows: Weight It was obtained through logistic regression training.
6. A stroke risk scoring system based on radiomics characteristics of intraplaque hemorrhage on carotid artery HR-MRI according to claim 5, characterized in that: The risk stratification unit defines IPH-RI < 0.15 as low risk, 0.15-0.45 as medium risk, and ≥0.45 as high risk.
7. A stroke risk scoring system based on radiomics characteristics of intraplaque hemorrhage on carotid artery HR-MRI according to claim 1, characterized in that: The IPH radiomics database and the population-specific reference database were fitted with continuous reference curves using LOESS local weighted regression with fitting parameters span=0.5 and degree=1.
8. A stroke risk scoring system based on radiomics characteristics of intraplaque hemorrhage on carotid artery HR-MRI according to claim 5, characterized in that: The feedback optimization module refits the reference median, standard deviation, and logistic regression coefficients every 6 months, and synchronously updates the database parameters, logistic regression weights, and risk thresholds every 6 months. The optimal boundary threshold is selected by maximizing the Youden index.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the function of the stroke risk scoring system according to any one of claims 1 to 8.