Cardiovascular disease medical image intelligent analysis method based on deep learning

By fusing multimodal data using deep learning methods and constructing the ST-RiskNet model, the problem of incomplete feature extraction in cardiovascular disease detection was solved. This enabled accurate identification of high-risk lesion areas and personalized risk prediction, forming a closed-loop management process that improves the accuracy and clinical application value of cardiovascular disease detection.

CN122290991APending Publication Date: 2026-06-26NANTONG COLLEGE OF SCIENCE & TECHNOLOGY
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202610300070.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-03-12
Publication Date
2026-06-26

AI Technical Summary

Technical Problem

Existing cardiovascular disease detection methods suffer from incomplete feature extraction, poor data fusion, insufficient risk prediction dimensions, and weak model generalization ability, failing to meet the needs of precise and personalized diagnosis and treatment.

Method used

We employ a deep learning-based intelligent analysis method for cardiovascular medical images. By fusing heterogeneous data of multimodal medical images and physiological signals, and combining deep learning networks for vascular structural feature extraction and quantitative analysis, we construct a personalized hemodynamic parameter set to identify and visualize high-risk lesion areas in three dimensions. We also build an ST-RiskNet fusion prediction model for two-dimensional risk prediction and utilize federated learning to optimize the entire closed-loop management process.

Benefits of technology

It achieves comprehensive fusion and automated extraction of multi-dimensional features, improves the accuracy and integration of lesion identification and risk prediction, enhances the model's generalization ability and privacy protection, forms a closed-loop management process, provides personalized clinical intervention recommendations, and reduces the incidence of cardiovascular events in medium- and high-risk populations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122290991A_ABST
    Figure CN122290991A_ABST
Patent Text Reader

Abstract

This invention provides a deep learning-based intelligent analysis method for cardiovascular medical images, comprising the following steps: S1, heterogeneous data fusion and acquisition of multimodal medical images and physiological signals; S2, intelligent extraction and quantitative analysis of vascular structural features based on deep learning; S3, hemodynamic parameter modeling and risk region segmentation based on fused structural features; S4, texture feature mining and intelligent identification of lesion types in high-risk areas; S5, cardiovascular disease risk stratification prediction based on multi-dimensional feature fusion; S6, federated learning optimization and closed-loop update of the model. This invention's deep learning-based intelligent analysis method for cardiovascular medical images achieves, for the first time, heterogeneous fusion of medical images, millimeter-wave radar-PPG physiological signals, and clinical molecular data, and achieves precise spatiotemporal alignment through GPS-controlled crystal oscillators, overcoming the limitations of single image analysis and providing a more comprehensive feature foundation for subsequent feature extraction and risk prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of medical image analysis, deep learning, and cardiovascular disease detection, and particularly to a deep learning-based intelligent analysis method for cardiovascular medical images. Background Technology

[0002] Cardiovascular disease is one of the leading causes of death worldwide. Early and accurate detection and risk prediction of lesions such as atherosclerosis and vascular fibrosis are crucial for disease prevention and control.

[0003] The patent application with publication number CN121414757A in the prior art breaks through the limitations of traditional detection methods, providing a comprehensive and accurate analysis of the cardiovascular system from multiple dimensions. It goes beyond the detection of a single indicator, comprehensively considering multiple key elements such as vascular topology, motion trajectory, and hemodynamics. By acquiring medical image sequences of the target patient, it extracts the vascular region contours from each frame and generates an initial vascular topology map, clearly displaying the morphology and layout of the cardiovascular system, providing a solid foundation for subsequent analysis. Based on the displacement vectors and grayscale differences of vascular regions in adjacent frames, it constructs vascular motion trajectories and marks abnormal displacement nodes, achieving real-time tracking of dynamic changes in blood vessels and capturing subtle abnormalities in vascular movement. It spatially registers the initial vascular topology map with a standard cardiovascular model, calculates the curvature deviation and diameter variation coefficient of each vascular branch, and performs quantitative analysis of blood vessels from a geometric morphology perspective, accurately revealing vascular deformation and dilation. In terms of hemodynamics, it extracts the temporal fluctuation characteristics of blood flow signals in the image sequence and combines them with a vascular structural abnormality index to generate blood flow... This method utilizes a set of mechanical parameters to gain a deeper understanding of the flow characteristics and patterns of blood flow in the cardiovascular system. This multi-dimensional analysis allows doctors to comprehensively understand the cardiovascular condition, detect subtle changes in the cardiovascular system, and identify lesions that are easily missed by traditional methods. In detecting early coronary artery disease, traditional methods may only focus on the degree of coronary artery stenosis. However, this method, by comprehensively analyzing vascular topology, motion trajectory, and hemodynamics, can detect early atherosclerotic plaques, micro-damage to the vessel wall, and abnormal changes in hemodynamics, thereby enabling early diagnosis and intervention. Through spatial registration and parameter calculation techniques, this method can accurately quantify changes in blood vessels, providing accurate data for diagnosis. When assessing the degree of vascular lesions, traditional methods often rely on the doctor's subjective judgment, resulting in significant errors and uncertainties. This method, by calculating parameters such as curvature deviation and coefficient of variation in vessel diameter, can accurately quantify the degree of vascular tortuosity and changes in vessel diameter, providing doctors with objective and accurate diagnostic evidence, making the diagnostic results more reliable and scientific, and providing strong support for subsequent treatment planning.

[0004] Current methods for detecting cardiovascular diseases largely rely on single medical image analysis, which suffers from incomplete feature extraction and strong subjectivity. The integration of physiological signals and clinical data is low, and the spatiotemporal alignment is poor, making it difficult to achieve comprehensive analysis of multi-dimensional features. Risk prediction models are mostly focused on a single short-term or long-term dimension, lacking an integrated design for acute event warning and long-term risk stratification. Furthermore, model training is limited by the amount of data from a single center, resulting in insufficient generalization ability. At the same time, a closed-loop management system covering image analysis, risk prediction, clinical intervention, and model optimization has not been established, which fails to meet the clinical demand for precise and personalized diagnosis and treatment.

[0005] Therefore, it is necessary to provide a deep learning-based intelligent analysis method for cardiovascular medical images to solve the above-mentioned technical problems. Summary of the Invention

[0006] This invention provides a deep learning-based intelligent analysis method for cardiovascular medical images, which solves the problems of single feature extraction, poor data fusion, insufficient risk prediction dimensions, and weak model generalization ability in existing technologies.

[0007] To address the aforementioned technical problems, this invention provides a deep learning-based intelligent analysis method for cardiovascular medical images, comprising the following steps: S1. Heterogeneous data fusion acquisition of multimodal medical images and physiological signals, completing the acquisition, preprocessing, construction and spatiotemporal alignment of multi-source data; S2. Intelligent extraction and quantitative analysis of vascular structural features based on deep learning, realizing vascular region segmentation, structural feature extraction and vascular structural abnormality index output; S3. Hemodynamic parameter modeling and risk area delineation by integrating structural features, constructing a personalized hemodynamic parameter set and completing the identification, expansion and three-dimensional visualization of high-risk lesion areas; S4. High-risk area texture feature mining and intelligent lesion type identification: Extract texture features and generate lesion probability prediction matrix to achieve intelligent lesion type determination. S5. Multi-dimensional feature fusion for cardiovascular disease risk stratification prediction: Construct the ST-RiskNet fusion prediction model to complete short-term and long-term risk dual-dimensional prediction and personalized risk subtype stratification. S6. Federated learning optimization and full-process closed-loop update of the model: Construct a federated learning framework, complete model robustness testing and continuous learning optimization, and realize full-process closed-loop management.

[0008] Preferably, S1 includes S11 multi-source data acquisition, S12 multi-source data preprocessing, and S13 heterogeneous dataset construction and spatiotemporal alignment; The cardiovascular medical image sequence acquired in S11 is continuous frame image data of CTA angiography or MRI angiography. The patient's subcutaneous arterial vibration signal and photoplethysmography signal are acquired by a millimeter-wave radar-PPG fusion sensor. The millimeter-wave radar operates at a frequency of 60.5-64.5 GHz with a displacement resolution of 0.02 mm. The PPG module uses 530 nm green light + 880 nm infrared light with a sampling rate of 1 kHz and a signal-to-noise ratio >90 dB. At the same time, the patient's clinical data are acquired through a dynamic questionnaire engine and OCR recognition technology.

[0009] Preferably, in step S12, the medical image sequence is sequentially subjected to Gaussian filtering and nonlocal mean denoising; the arterial vibration signal and PPG signal are denoised in real time using a generative adversarial network, and motion interference is removed by motion component characterization, signal-to-noise separation denoising, and adaptive fusion correction motion artifact elimination algorithm; the clinical data are standardized and missing data are filled in using multiple interpolation.

[0010] Preferably, in step S13, clock synchronization of all data is achieved by using GPS-disciplined crystal oscillators, and time axis calibration of physiological signals and clinical data is performed based on the timestamp of medical image acquisition, so as to achieve precise spatiotemporal alignment of multi-source data.

[0011] Preferably, in step S21, the preprocessed medical image sequence is input into an improved nnUNet network to achieve end-to-end automatic segmentation of the vascular region. After performing morphological erosion on the output binarized mask, a connected component label map is generated. The spatial coordinates of the vascular centerline are extracted by the Zhang-Suen thinning algorithm. After fitting with cubic spline interpolation, a three-dimensional vascular topology map is constructed. In step S22, the deep features of the vascular system are extracted by the 3DResNet-50 network and the curvature deviation and diameter variation coefficient are calculated. The temporal motion features of the vascular system are captured by the LSTM-Att network and abnormal displacement nodes are marked. After deriving the local Young's modulus, a vascular elastic feature matrix is ​​generated. In step S23, the feature set is decomposed by wavelet and then convolution operation and weight optimization are completed. After fusion, a vascular structure abnormality index with a value range of [0, 1] is output.

[0012] Preferably, in step S31, an adaptive boundary layer mesh is constructed based on the transient Navier-Stokes equations improved by OpenFOAM to simulate blood flow and generate a personalized hemodynamic parameter set including blood flow velocity, wall shear stress, pressure gradient, and fluid shear force. In step S32, the 95th percentile of the Laplacian operator value is used as the initial dynamic threshold, and high-risk lesion areas are determined by linear regression trend analysis of the vascular structural abnormality index. In step S33, a region growth algorithm is used to expand the high-risk region, with the fluid shear force mutation point as the growth stopping boundary. The high-risk region is visualized in three dimensions by using the UnityHDRP engine combined with the medical DICOM plugin.

[0013] Preferably, in step S41, texture co-occurrence matrix analysis is performed on pixel clusters within high-risk lesion areas. Multiple sets of pixel grayscale sequences are extracted along the normal direction of the blood vessel wall and processed using the Gabor filter bank to extract the texture response spectrum. In step S42, PCA dimensionality reduction is performed on the texture response spectrum, and principal components with a cumulative variance contribution rate greater than 90% are retained to generate texture fingerprint codes. The coupling relationship between texture fingerprint codes and vascular structural abnormality indices is captured through a TransformerEncoder network, and a joint probability density function is constructed using a non-parametric kernel density estimation method. In step S43, a lesion probability prediction matrix is ​​generated through Monte Carlo sampling, and the DBSCAN spatial clustering algorithm is used to identify peak region features and determine the type of atherosclerotic or vascular fibrosis lesions.

[0014] Preferably, the input layer of the ST-RiskNet fusion prediction model in S51 includes three types of features: radiomics features, physiological signal features, and clinical molecular features. The model constructs two parallel branches: short-term acute event prediction and long-term risk prediction. In S52, the short-term branch adopts a lightweight CNN-LSTM hybrid model. After structured pruning and 8-bit quantization compression, the model size is controlled within 1.2MB. It outputs the probability curve of cardiovascular acute event risk in the next 6 hours. When the predicted probability is >15%, it automatically triggers the GPS positioning and audible and visual alarm of the emergency system that conforms to the IEC60601-1-8 standard. The long-term branch uses the XGBoost+SHAP interpretation model to construct a 5-year risk score model and outputs a long-term morbidity risk score of 0-100. In S53, four types of indicators of patient inflammation, metabolism, vascular aging, and genetic susceptibility are collected. Patients are divided into four personalized risk subtypes by t-SNE dimensionality reduction and DBSCAN clustering algorithm.

[0015] Preferably, the classification criteria for the four types of personalized risk subtypes are as follows: inflammation-dominant type: hs-CRP>3mg / L, metabolic disorder type: HOMA-IR>2.5, vascular aging type: PWV>10m / s, and genetic susceptibility type: PRS>90th percentile.

[0016] Preferably, in S61, a federated learning framework for joint training of multiple medical institutions is constructed based on the Paillier homomorphic encryption algorithm and the improved FedProx algorithm. In S62, the robustness of the model is tested by generating perturbation data using the fast gradient symbol method, storing difficult examples in a shared sample library and using a weighted sampling method, and retraining the model quarterly using the elastic weight consolidation continuous learning method. In S63, a hybrid reward function is constructed based on reinforcement learning recommendation units to generate personalized clinical intervention suggestions, and follow-up data and treatment results are fed back to the heterogeneous dataset to achieve closed-loop management of the entire process of image analysis, risk prediction, clinical intervention and model optimization.

[0017] Compared with related technologies, the intelligent analysis method for cardiovascular medical images based on deep learning provided by this invention has the following beneficial effects: This invention provides a deep learning-based intelligent analysis method for cardiovascular medical images, with comprehensive feature dimensions: it achieves heterogeneous fusion of medical images, millimeter-wave radar-PPG physiological signals, and clinical molecular data for the first time, and achieves precise spatiotemporal alignment through GPS-dominated crystal oscillators, breaking through the limitations of single image analysis and providing a more comprehensive and accurate feature foundation for subsequent feature extraction and risk prediction.

[0018] Deep learning drives automated feature extraction with high objectivity and accuracy: Improved deep learning networks such as nnUNet, 3DResNet-50, and LSTM-Att enable automatic segmentation of vascular regions and automated extraction of vascular structural and motion features, replacing traditional manual quantitative analysis, eliminating subjective judgment errors by doctors, and improving the calculation accuracy of parameters such as curvature deviation and diameter variation coefficient compared to traditional methods.

[0019] The system integrates lesion identification and risk prediction with comprehensive coverage: it integrates vascular structural features, hemodynamic parameters, and texture features to achieve accurate division of high-risk lesion areas and intelligent identification of lesion types; it constructs the ST-RiskNet multi-branch fusion model to achieve dual-dimensional prediction of future acute myocardial infarction warning and long-term morbidity risk score, while completing the stratification of multiple risk subtypes to meet the needs of different clinical scenarios.

[0020] Balancing model optimization and privacy protection with strong generalization ability: A federated learning framework is built based on Paillier homomorphic encryption and an improved FedProx algorithm to achieve joint modeling of multi-center data without data leaving the domain, solving the problems of data privacy and data silos; through FGSM attack testing and elastic weight consolidation for continuous learning, the robustness and generalization ability of the model are improved, and the prediction accuracy of the model on multi-center datasets increases.

[0021] The entire process is closed-loop management with strong clinical applicability: It forms a closed loop of image analysis, risk prediction, clinical intervention and model optimization. The diagnosis and treatment results are continuously fed back to the dataset and the model is optimized. At the same time, it provides personalized intervention suggestions for clinicians, realizing a full-chain service from "detection-prediction" to "intervention-optimization". It can reduce the incidence of major cardiovascular events in medium and high-risk groups and has important clinical application value. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of a preferred embodiment of the intelligent analysis method for cardiovascular medical images based on deep learning provided by the present invention. Figure 2 A schematic diagram illustrating the heterogeneous data fusion acquisition of multimodal medical images and physiological signals; Figure 3 This is a schematic diagram of intelligent extraction and quantitative analysis of vascular and structural features based on deep learning. Figure 4 A schematic diagram of hemodynamic parameter modeling and risk region delineation for fused structural features; Figure 5 A schematic diagram illustrating texture feature mining and intelligent lesion type identification in high-risk areas; Figure 6 A schematic diagram for predicting cardiovascular disease risk stratification by fusing multi-dimensional features; Figure 7 This diagram illustrates the federated learning optimization and closed-loop update process of the model. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Please refer to the following: Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7 ,in, Figure 1 This is a schematic diagram of a preferred embodiment of the intelligent analysis method for cardiovascular medical images based on deep learning provided by the present invention. Figure 2 A schematic diagram illustrating the heterogeneous data fusion acquisition of multimodal medical images and physiological signals; Figure 3 This is a schematic diagram of intelligent extraction and quantitative analysis of vascular and structural features based on deep learning. Figure 4 A schematic diagram of hemodynamic parameter modeling and risk region delineation for fused structural features; Figure 5 A schematic diagram illustrating texture feature mining and intelligent lesion type identification in high-risk areas; Figure 6A schematic diagram for predicting cardiovascular disease risk stratification by fusing multi-dimensional features; Figure 7 This diagram illustrates the federated learning optimization and closed-loop update process of the model. The deep learning-based intelligent analysis method for cardiovascular medical images includes the following steps: S1. Heterogeneous data fusion acquisition of multimodal medical images and physiological signals, completing the acquisition, preprocessing, construction and spatiotemporal alignment of multi-source data; S2. Intelligent extraction and quantitative analysis of vascular structural features based on deep learning, realizing vascular region segmentation, structural feature extraction and vascular structural abnormality index output; S3. Hemodynamic parameter modeling and risk area delineation by integrating structural features, constructing a personalized hemodynamic parameter set and completing the identification, expansion and three-dimensional visualization of high-risk lesion areas; S4. High-risk area texture feature mining and intelligent lesion type identification: Extract texture features and generate lesion probability prediction matrix to achieve intelligent lesion type determination. S5. Multi-dimensional feature fusion for cardiovascular disease risk stratification prediction: Construct the ST-RiskNet fusion prediction model to complete short-term and long-term risk dual-dimensional prediction and personalized risk subtype stratification. S6. Federated learning optimization and full-process closed-loop update of the model: Construct a federated learning framework, complete model robustness testing and continuous learning optimization, and realize full-process closed-loop management.

[0025] S1 includes S11 multi-source data acquisition, S12 multi-source data preprocessing, and S13 heterogeneous dataset construction and spatiotemporal alignment. The cardiovascular medical image sequence acquired in S11 is continuous frame image data of CTA angiography or MRI angiography. The patient's subcutaneous arterial vibration signal and photoplethysmography signal are acquired by a millimeter-wave radar-PPG fusion sensor. The millimeter-wave radar operates at a frequency of 60.5-64.5 GHz with a displacement resolution of 0.02 mm. The PPG module uses 530 nm green light + 880 nm infrared light with a sampling rate of 1 kHz and a signal-to-noise ratio >90 dB. At the same time, the patient's clinical data are acquired through a dynamic questionnaire engine and OCR recognition technology.

[0026] In step S12, the medical image sequence is sequentially subjected to Gaussian filtering and nonlocal mean denoising; the arterial vibration signal and PPG signal are denoised in real time using generative adversarial network, and motion interference is removed by motion component characterization, signal-to-noise separation denoising, and adaptive fusion correction motion artifact elimination algorithm; the clinical data is standardized and missing data is filled in using multiple interpolation method.

[0027] In S13, the clock synchronization of all data is achieved by using GPS to discipline the crystal oscillator, and the time axis calibration of physiological signals and clinical data is performed based on the timestamp of medical image acquisition, so as to achieve precise spatiotemporal alignment of multi-source data.

[0028] In step S21, the preprocessed medical image sequence is input into the improved nnUNet network to achieve end-to-end automatic segmentation of the vascular region. After morphological erosion of the output binarized mask, a connected component label map is generated. The spatial coordinates of the vascular centerline are extracted by the Zhang-Suen thinning algorithm. After cubic spline interpolation fitting, a three-dimensional vascular topology map is constructed. In step S22, the deep features of the vascular system are extracted by the 3DResNet-50 network and the curvature deviation and diameter variation coefficient are calculated. The temporal motion features of the vascular system are captured by the LSTM-Att network and abnormal displacement nodes are marked. After deriving the local Young's modulus, a vascular elastic feature matrix is ​​generated. In step S23, the feature set is decomposed by wavelet and then convolution operation and weight optimization are completed. After fusion, a vascular structure abnormality index with a value range of [0, 1] is output.

[0029] In step S31, an adaptive boundary layer mesh is constructed based on the transient Navier-Stokes equations improved by OpenFOAM to simulate blood flow and generate a personalized hemodynamic parameter set including blood flow velocity, wall shear stress, pressure gradient, and fluid shear force. In step S32, the 95th percentile of the Laplacian operator value is used as the initial dynamic threshold, and high-risk lesion areas are determined by linear regression trend analysis of the vascular structural abnormality index. In step S33, a region growth algorithm is used to expand the high-risk region, with the fluid shear force mutation point as the growth stopping boundary. The three-dimensional visualization of the high-risk region is achieved by combining the UnityHDRP engine with the medical DICOM plugin.

[0030] In step S41, texture co-occurrence matrix analysis is performed on pixel clusters within high-risk lesion areas. Multiple sets of pixel grayscale sequences are extracted along the normal direction of the blood vessel wall and processed using the Gabor filter bank to extract the texture response spectrum. In step S42, PCA dimensionality reduction is performed on the texture response spectrum, and principal components with a cumulative variance contribution rate greater than 90% are retained to generate texture fingerprint codes. The coupling relationship between texture fingerprint codes and vascular structural abnormality indices is captured through a TransformerEncoder network, and a joint probability density function is constructed using a non-parametric kernel density estimation method. In step S43, a lesion probability prediction matrix is ​​generated through Monte Carlo sampling, and the DBSCAN spatial clustering algorithm is used to identify peak region features to determine the type of atherosclerotic or vascular fibrosis lesions.

[0031] The input layer of the ST-RiskNet fusion prediction model in S51 includes three types of features: radiomics features, physiological signal features, and clinical molecular features. The model constructs two parallel branches for short-term acute event prediction and long-term risk prediction. In S52, the short-term branch adopts a lightweight CNN-LSTM hybrid model. After structured pruning and 8-bit quantization compression, the model size is controlled within 1.2MB. It outputs the probability curve of cardiovascular acute event risk in the next 6 hours. When the predicted probability is >15%, it automatically triggers the GPS positioning and audible and visual alarm of the emergency system in accordance with the IEC60601-1-8 standard. The long-term branch uses the XGBoost+SHAP interpretation model to construct a 5-year risk score model and outputs a long-term morbidity risk score of 0-100. In S53, four types of indicators of patient inflammation, metabolism, vascular aging, and genetic susceptibility are collected. Patients are divided into four personalized risk subtypes by t-SNE dimensionality reduction and DBSCAN clustering algorithm.

[0032] The criteria for classifying the four personalized risk subtypes are as follows: Inflammation-dominant type: hs-CRP>3mg / L, Metabolic disorder type: HOMA-IR>2.5, Vascular aging type: PWV>10m / s, Genetic susceptibility type: PRS>90th percentile.

[0033] In S61, a federated learning framework for joint training among multiple medical institutions is constructed based on the Paillier homomorphic encryption algorithm and the improved FedProx algorithm. In S62, the rapid gradient symbol method is used to generate perturbation data to test the robustness of the model. Difficult sample samples are stored in a shared sample library and a weighted sampling method is used. The model is retrained quarterly using the elastic weight consolidation continuous learning method. In S63, a hybrid reward function is constructed based on reinforcement learning recommendation units to generate personalized clinical intervention suggestions. Follow-up data and treatment results are fed back to the heterogeneous dataset to realize the closed-loop management of the entire process of image analysis, risk prediction, clinical intervention and model optimization.

[0034] Compared with related technologies, the intelligent analysis method for cardiovascular medical images based on deep learning provided by this invention has the following beneficial effects: This invention provides a deep learning-based intelligent analysis method for cardiovascular medical images, which deeply integrates multi-source data and provides comprehensive feature dimensions. It is the first to achieve heterogeneous fusion of medical images, millimeter-wave radar-PPG physiological signals, and clinical molecular data, and achieves precise spatiotemporal alignment through GPS-controlled crystal oscillators. This breaks through the limitations of single image analysis and provides a more comprehensive and accurate feature foundation for subsequent feature extraction and risk prediction.

[0035] Deep learning drives automated feature extraction with high objectivity and accuracy: Improved deep learning networks such as nnUNet, 3DResNet-50, and LSTM-Att enable automatic segmentation of vascular regions and automated extraction of vascular structural and motion features, replacing traditional manual quantitative analysis, eliminating subjective judgment errors by doctors, and improving the calculation accuracy of parameters such as curvature deviation and diameter variation coefficient compared to traditional methods.

[0036] The system integrates lesion identification and risk prediction with comprehensive coverage: it integrates vascular structural features, hemodynamic parameters, and texture features to achieve accurate division of high-risk lesion areas and intelligent identification of lesion types; it constructs the ST-RiskNet multi-branch fusion model to achieve dual-dimensional prediction of future acute myocardial infarction warning and long-term morbidity risk score, while completing the stratification of multiple risk subtypes to meet the needs of different clinical scenarios.

[0037] Balancing model optimization and privacy protection with strong generalization ability: A federated learning framework is built based on Paillier homomorphic encryption and an improved FedProx algorithm to achieve joint modeling of multi-center data without data leaving the domain, solving the problems of data privacy and data silos; through FGSM attack testing and elastic weight consolidation for continuous learning, the robustness and generalization ability of the model are improved, and the prediction accuracy of the model on multi-center datasets increases.

[0038] The entire process is closed-loop management with strong clinical applicability: It forms a closed loop of image analysis, risk prediction, clinical intervention and model optimization. The diagnosis and treatment results are continuously fed back to the dataset and the model is optimized. At the same time, it provides personalized intervention suggestions for clinicians, realizing a full-chain service from "detection-prediction" to "intervention-optimization". It can reduce the incidence of major cardiovascular events in medium and high-risk groups and has important clinical application value.

[0039] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A deep learning-based intelligent analysis method for cardiovascular medical images, characterized in that, Includes the following steps: S1. Heterogeneous data fusion acquisition of multimodal medical images and physiological signals, completing the acquisition, preprocessing, construction and spatiotemporal alignment of multi-source data; S2. Intelligent extraction and quantitative analysis of vascular structural features based on deep learning, realizing vascular region segmentation, structural feature extraction and vascular structural abnormality index output; S3. Hemodynamic parameter modeling and risk area delineation by integrating structural features, constructing a personalized hemodynamic parameter set and completing the identification, expansion and three-dimensional visualization of high-risk lesion areas; S4. High-risk area texture feature mining and intelligent lesion type identification: Extract texture features and generate lesion probability prediction matrix to achieve intelligent lesion type determination. S5. Multi-dimensional feature fusion for cardiovascular disease risk stratification prediction: Construct the ST-RiskNet fusion prediction model to complete short-term and long-term risk dual-dimensional prediction and personalized risk subtype stratification. S6. Federated learning optimization and full-process closed-loop update of the model: Construct a federated learning framework, complete model robustness testing and continuous learning optimization, and realize full-process closed-loop management.

2. The intelligent analysis method for cardiovascular medical images based on deep learning according to claim 1, characterized in that, S1 includes S11 multi-source data acquisition, S12 multi-source data preprocessing, and S13 heterogeneous dataset construction and spatiotemporal alignment. The cardiovascular medical image sequence acquired in S11 is continuous frame image data of CTA angiography or MRI angiography. The patient's subcutaneous arterial vibration signal and photoplethysmography signal are acquired by a millimeter-wave radar-PPG fusion sensor. The millimeter-wave radar operates at a frequency of 60.5-64.5 GHz with a displacement resolution of 0.02 mm. The PPG module uses 530 nm green light + 880 nm infrared light with a sampling rate of 1 kHz and a signal-to-noise ratio >90 dB. At the same time, the patient's clinical data are acquired through a dynamic questionnaire engine and OCR recognition technology.

3. The intelligent analysis method for cardiovascular medical images based on deep learning according to claim 2, characterized in that, In S12, the medical image sequence is sequentially subjected to Gaussian filtering and nonlocal mean denoising; the arterial vibration signal and PPG signal are denoised in real time using generative adversarial network, and motion interference is removed by motion component characterization, signal-noise separation denoising, and adaptive fusion correction motion artifact elimination algorithm. Clinical data were standardized and missing data were filled using multiple imputation methods. 4.The deep learning-based intelligent analysis method for cardiovascular disease medical images according to claim 2, characterized in that, In S13, the clock synchronization of all data is achieved by using GPS to discipline the crystal oscillator, and the time axis calibration of physiological signals and clinical data is performed based on the timestamp of medical image acquisition, so as to achieve precise spatiotemporal alignment of multi-source data. 5.The deep learning-based intelligent analysis method for cardiovascular disease medical images according to claim 1, characterized in that, In step S21, the preprocessed medical image sequence is input into the improved nnUNet network to achieve end-to-end automatic segmentation of the vascular region. After morphological erosion of the output binarized mask, a connected component label map is generated. The spatial coordinates of the vascular centerline are extracted by the Zhang-Suen thinning algorithm. After cubic spline interpolation fitting, a three-dimensional vascular topology map is constructed. In step S22, the deep features of the vascular system are extracted by the 3DResNet-50 network and the curvature deviation and diameter variation coefficient are calculated. The temporal motion features of the vascular system are captured by the LSTM-Att network and abnormal displacement nodes are marked. After deriving the local Young's modulus, a vascular elastic feature matrix is ​​generated. In step S23, the feature set is decomposed by wavelet and then convolution operation and weight optimization are completed. After fusion, a vascular structure abnormality index with a value range of [0, 1] is output.

6. The intelligent analysis method for cardiovascular medical images based on deep learning according to claim 1, characterized in that, In step S31, an adaptive boundary layer mesh is constructed based on the transient Navier-Stokes equations improved by OpenFOAM to simulate blood flow and generate a personalized hemodynamic parameter set including blood flow velocity, wall shear stress, pressure gradient, and fluid shear force. In step S32, the 95th percentile of the Laplacian operator value is used as the initial dynamic threshold, and high-risk lesion areas are determined by linear regression trend analysis of the vascular structural abnormality index. In step S33, a region growth algorithm is used to expand the high-risk region, with the fluid shear force mutation point as the growth stopping boundary. The three-dimensional visualization of the high-risk region is achieved by combining the UnityHDRP engine with the medical DICOM plugin. 7.The deep learning-based intelligent analysis method for cardiovascular disease medical images according to claim 1, characterized in that, In step S41, texture co-occurrence matrix analysis is performed on pixel clusters within high-risk lesion areas. Multiple sets of pixel grayscale sequences are extracted along the normal direction of the blood vessel wall and processed using the Gabor filter bank to extract the texture response spectrum. In step S42, PCA dimensionality reduction is performed on the texture response spectrum, and principal components with a cumulative variance contribution rate greater than 90% are retained to generate texture fingerprint codes. The coupling relationship between texture fingerprint codes and vascular structural abnormality indices is captured through a TransformerEncoder network, and a joint probability density function is constructed using a non-parametric kernel density estimation method. In step S43, a lesion probability prediction matrix is ​​generated through Monte Carlo sampling, and the DBSCAN spatial clustering algorithm is used to identify peak region features to determine the type of atherosclerotic or vascular fibrosis lesions. 8.The deep learning-based intelligent analysis method for cardiovascular disease medical images according to claim 1, characterized in that, The input layer of the ST-RiskNet fusion prediction model in S51 includes three types of features: radiomics features, physiological signal features, and clinical molecular features. The model constructs two parallel branches for short-term acute event prediction and long-term risk prediction. In S52, the short-term branch adopts a lightweight CNN-LSTM hybrid model. After structured pruning and 8-bit quantization compression, the model size is controlled within 1.2MB. It outputs the probability curve of cardiovascular acute event risk in the next 6 hours. When the predicted probability is >15%, it automatically triggers the GPS positioning and audible and visual alarm of the emergency system in accordance with the IEC60601-1-8 standard. The long-term branch uses the XGBoost+SHAP interpretation model to construct a 5-year risk score model and outputs a long-term morbidity risk score of 0-100. In S53, four types of indicators of patient inflammation, metabolism, vascular aging, and genetic susceptibility are collected. Patients are divided into four personalized risk subtypes by t-SNE dimensionality reduction and DBSCAN clustering algorithm. 9.The deep learning-based intelligent analysis method for cardiovascular disease medical images according to claim 1, characterized in that, The criteria for classifying the four personalized risk subtypes are as follows: Inflammation-dominant type: hs-CRP>3mg / L, Metabolic disorder type: HOMA-IR>2.5, Vascular aging type: PWV>10m / s, Genetic susceptibility type: PRS>90th percentile. 10.The deep learning-based intelligent analysis method for cardiovascular disease medical images according to claim 1, characterized in that, In S61, a federated learning framework for joint training among multiple medical institutions is constructed based on the Paillier homomorphic encryption algorithm and the improved FedProx algorithm. In S62, the rapid gradient symbol method is used to generate perturbation data to test the robustness of the model. Difficult sample samples are stored in a shared sample library and a weighted sampling method is used. The model is retrained quarterly using the elastic weight consolidation continuous learning method. In S63, a hybrid reward function is constructed based on reinforcement learning recommendation units to generate personalized clinical intervention suggestions. Follow-up data and treatment results are fed back to the heterogeneous dataset to realize the closed-loop management of the entire process of image analysis, risk prediction, clinical intervention and model optimization.

Citation Information

Patent Citations

  • Cardiovascular disease detection method based on image analysis

    CN121414757A