CN110223781B

CN110223781B - A multi-dimensional plaque rupture risk early warning system

Info

Publication number: CN110223781B
Application number: CN201910476097.6A
Authority: CN
Inventors: 刘婷; 金士琪; 霍怀璧; 李思邈
Original assignee: First Hospital of China Medical University
Current assignee: First Hospital of China Medical University
Priority date: 2019-06-03
Filing date: 2019-06-03
Publication date: 2021-06-04
Anticipated expiration: 2039-06-03
Also published as: CN110223781A

Abstract

The multi-dimensional plaque rupture risk early warning system provided by the present invention integrates the obtained CCTA vulnerable plaque automatic identification method with the CT real-time hemodynamic analysis technology based on deep learning to create a more efficient, sensitive and accurate The non-invasive plaque rupture risk early warning model provides a comprehensive and reliable basis and key technical means for intelligent non-invasive accurate detection and early warning of coronary plaque rupture.

Description

Multidimensional plaque rupture risk early warning system

Technical Field

The invention relates to a multi-dimensional plaque rupture risk early warning system.

Background

Coronary heart disease is the first killer seriously harming human health, and a large number of studies indicate that thrombosis secondary to vulnerable plaque rupture is a major factor causing acute cardiovascular events (ACS). Although the understanding of the atherosclerotic plaque is continuously improved at present, the Vulnerable plaque (Vulnerable plaque) has the characteristics of hidden symptoms and burst rupture, so that the early accurate diagnosis is difficult clinically.

In recent years, Coronary Angiography (CAG) examination has been used as a gold standard for assessing coronary heart disease. However, it provides only lumen information and cannot distinguish in detail the coronary artery vessel wall and the inside of the plaque. Intravascular ultrasound (IVUS) and Optical Coherence Tomography (OCT) have been increasingly used as intravascular imaging detection techniques, and have shown their advantages in the field of coronary intervention.

At present, although intravascular imaging examination methods such as IVUS and OCT are excellent in identifying vulnerable plaque characteristics, their invasive examination methods limit their wide clinical application. CT Coronary Angiography (CCTA) is used as a non-invasive examination method to be widely applied in the diagnosis of Coronary heart disease, and is a main imaging method for non-invasive assessment of Coronary atherosclerotic plaque at present. Relevant studies have shown that vulnerable plaque features detected by CCTA significantly increase the likelihood of ACS occurring. At present, a great deal of evidence and diagnosis standards about vulnerable plaques mainly come from analysis of local morphological characteristics, risk early warning of coronary heart disease is mostly limited to assessment indexes of clinical risk factors, serum biochemistry and the like, and an assessment system based on accurate noninvasive anatomical function combination and an effective early warning model and a monitoring means are still lacking. If the vulnerable plaque can be accurately identified by a non-invasive CTA (computed tomography angiography) method or an MRI method and by combining with artificial intelligence means such as neural network deep learning, the knowledge of the vulnerable plaque can be expanded by monitoring the dynamic evolution process of the plaque from stability to instability or even rupture in vivo, the trigger mechanism of the vulnerability of the plaque in an acute cardiovascular event can be further deeply known, and early warning and risk stratification of the acute coronary event can be realized.

The convolutional neural network is a network structure containing a multilayer network, and a convolutional neural network model structure is established by an input layer, a convolutional layer, a downsampling layer and an output layer. Since the feature detection layer of the system is learned by training data, explicit feature extraction is avoided during use, and learning is performed implicitly from the training data, thereby achieving remarkable performance in the pattern recognition field.

Based on the convolutional neural network artificial intelligence technology, the medical image data can be deeply mined, the potential pathophysiology quantification related information in the medical image can be automatically extracted, the method is not limited to artificially designed iconography characteristics, and the method is expected to be capable of reducing the dimensions of the high-dimensional characteristics of the image data, constructing an efficient coronary plaque form recognition model, accurately predicting, recognizing and judging vulnerable plaques, and obtaining a diagnosis effect consistent with invasive intra-cavity examination, so that clinical decisions can be guided in an early personalized manner. Therefore, how to integrate the noninvasive imaging method and the artificial intelligent means such as neural network deep learning and the like to realize quantitative determination of the whole-heart coronary plaque and high-precision automatic detection of vulnerable plaque, establish a new imaging evaluation index and perfect a risk assessment system of the acute cardiovascular event is an important scientific problem to be solved urgently.

Disclosure of Invention

The technical problem to be solved by the invention is as follows: the system solves the technical problems that an assessment system based on accurate noninvasive anatomical function combination, an effective early warning model and a monitoring means are lacked in the prior art, and provides a multi-dimensional plaque rupture risk early warning system.

The technical scheme adopted by the invention for solving the technical problems is as follows:

a multidimensional plaque rupture risk forewarning system, the system comprising:

a model building module: establishing a multi-dimensional noninvasive vulnerable plaque rupture risk early warning model;

a detection module: carrying out early warning based on the established early warning model;

the multi-dimensional noninvasive vulnerable plaque rupture risk early warning model is established based on a neural network deep learning coronary artery vulnerable plaque automatic identification system and a real-time hemodynamic evaluation method.

The model building module specifically comprises: and establishing an early warning module based on quantitative analysis of the CT on the plaque and the qualitative characteristics of the CT plaque.

The parameters used by quantitative analysis of the plaque and the CT plaque qualitative characteristics are obtained by a neural network deep learning coronary artery vulnerable plaque automatic identification system; the quantitative analysis of the plaque by the CT comprises the total load of the whole heart plaque, the plaque position, the plaque range, the stenosis rate, the plaque volumes with different densities and a reconstruction index; the CT plaque qualitative characteristics comprise positive reconstruction, napkin ring characteristics, punctate calcification and low-density plaque.

Wherein the model is established further comprising CT-FFR, CT-ESS based on real-time hemodynamic assessment.

Wherein the CT-FFR, CT-ESS based on real-time hemodynamic assessment comprises: and calculating CT-FFR and CT-ESS from the computer virtual coronary artery anatomy database model through training learning simulation.

Wherein the CT-ESS is calculated from the computer virtual coronary artery anatomy database model by training a learning simulation, the CT-ESS comprising:

1) establishing a coronary artery anatomical parameter computer virtual training database;

2) completing off-line training on a coronary artery anatomical structure virtual database by using a deep learning algorithm, and taking the simulated CFD-FFRct at the corresponding position as a target value;

3) finally, the algorithm model calculates FFR and ESS at any position of the coronary vessel tree, and performs color coding on the FFR and ESS to perform visual display.

The risk assessment method comprises the steps of establishing connection between multidimensional exposure factors and end point events through a Cox presentation-hazards risk model, and creating a risk assessment model based on a CT noninvasive imaging and deep learning system.

Wherein the establishment of the model is further based on clinical risk factors, novel serum biomarkers.

The multi-dimensional plaque rupture risk early warning system integrates the obtained CCTA vulnerable plaque automatic identification method with the CT real-time hemodynamic analysis technology based on deep learning, creates a more efficient, sensitive and accurate non-invasive plaque rupture risk early warning model, and provides comprehensive and reliable basis and key technical means for intelligent non-invasive accurate detection and early warning of coronary plaque rupture.

Drawings

The invention is further illustrated with reference to the following figures and examples.

FIG. 1 is a flow chart of a preferred embodiment of the present invention;

FIG. 2 is a multi-dimensional risk pre-warning model for non-invasive vulnerable plaque automation, tissue characterization, and functionality, according to a preferred embodiment of the present invention;

FIG. 3 is an automatic identification system of a vulnerable coronary plaque of a preferred embodiment of the present invention;

FIG. 4 is a convolutional 3D Convolutional Neural Network (CNN) training process of a preferred embodiment of the present invention;

FIG. 5 is an interpretation of vulnerable plaque signs of a preferred embodiment of the present invention;

FIG. 6 is a coronary plaque lesion localization and lesion segmentation model based on a deep convolutional neural network according to a preferred embodiment of the present invention;

FIG. 7 shows the DL-FFRct and DL-ESSct training and application based on deep learning network algorithm in the preferred embodiment of the present invention.

Detailed Description

The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic views illustrating only the basic structure of the present invention in a schematic manner, and thus show only the constitution related to the present invention.

In order to overcome the technical problems of lack of an assessment system based on accurate noninvasive anatomical function combination and an effective early warning model and monitoring means in the prior art, the invention provides a multidimensional plaque rupture risk early warning system, which comprises:

the multi-dimensional noninvasive vulnerable plaque rupture risk early warning model is established based on a neural network deep learning coronary artery vulnerable plaque automatic identification system and a real-time hemodynamic evaluation method. The specific flow is shown in figure 1.

As shown in fig. 2: the establishment of the early warning model comprises a plurality of quantitative clinical risk factors and blood biochemical indexes. The model is added with a neural network plaque automatic identification system, comprises quantitative analysis of the plaque by CT (computed tomography) (including indexes such as total load of the whole heart plaque, plaque position, range, stenosis rate, plaque volume with different densities, reconstruction index and the like) and qualitative characteristics of the CT plaque (positive reconstruction, napkin ring characterization, punctate calcification and low-density plaque), and is also added with hemodynamic indexes such as CT-FFR (computed tomography-flow regression), CT-ESS and the like based on real-time hemodynamic evaluation and evaluation indexes of plaque development, so that quantified plaque information and external blood flow information are fused into the model, and the requirement of accurate medical treatment is met. Establishing a relation between multi-dimensional exposure factors and an end point event through a Cox reporting-hazards risk model, and creating a risk assessment model based on a CT noninvasive imaging and deep learning system; and comparing the model with efficacy of the us COMFIRMs score, european euro score, evaluating the superiority of the model.

The early noninvasive risk early warning model for the vulnerable plaque rupture is established by combining various quantitative indexes of patient risk factors, novel biomarkers, hemodynamic CT-FFR, CT-ESS, plaque progress and the like and comprehensively integrating morphology and functional multi-dimensional.

As shown in fig. 3, the automatic neural network plaque identification system first takes the coronary artery plaque of the coronary artery CTA image and the coronary artery plaque of the OCT image of the same sample as a region of interest (ROI). And inputting the CTA image and the OCT image of the same plaque into the CNN as different channels (channels), and automatically fusing the image characteristics of the CTA and the OCT in the subsequent CNN full-connection layer part, thereby establishing a model. And the validity of the model is verified in the test data set, thereby realizing the test of the convolutional neural network data training.

To ensure the accuracy of the identification system, the subject selection and the selected blood circulation biomarkers are now described:

1) study subject selection:

on the basis of 300 coronary CTA and IVUS data sets, 300 prospective stable angina pectoris cases are included, and 300 non-ST elevation ACS cases are included; all patients underwent coronary CTA examination, clinical risk factor assessment, and novel blood circulation marker indices, and non-ST elevation ACS patients underwent IVUS or OCT examinations simultaneously.

The specific grouping criteria are as follows: the patient with the typical angina pectoris symptom or the clinically confirmed non-ST elevation myocardial infarction is: determining that the patient has no definite history of myocardial infarction; ② the coronary artery blood vessel stenosis rate is more than 30 percent; ③ more than 1 risk factors of atherosclerosis, such as hypertension, diabetes, lipid metabolism disorder, smoking, etc.

Exclusion criteria: PCI or coronary step-and-bridge postoperative patients have been treated; second, there is history of other cardiovascular diseases such as congenital heart disease, cardiomyopathy, etc.; ③ patients allergic to iodine contrast agents; ST-segment elevation type myocardial infarction patients; poor quality of scanned image, unable to carry out data measurement and analysis.

2) Blood circulation biomarker detection

(1) LpPLA2 (lipoprotein-associated phospholipase A2) is involved in the development process of atherosclerosis, and has risk prediction value. (2) The relevance of MMP-9 (matrix protein metalloenzyme 9) to cardiovascular disease and prognosis has been demonstrated in a number of studies. (3) The NGAL/MMP-9 complex (neutrophil gelatinase-associated apolipoprotein/matrix protein metalloenzyme 9) is a cardiovascular event risk predictor associated with bleeding plaques. (4) OGN (glycoproteoglycan), the expression of which is significantly increased in the area of coronary calcification. The above four markers identify the presence of high risk plaque, mainly from different perspectives of body inflammation, plaque progression and plaque calcification. (5) vWF (von willebrand factor) is a sensitive indicator of the extent of damage and hypercoagulability of vascular endothelial cells. (6) CRP (C-reactive protein) belongs to an acute phase protein and is a pro-inflammatory factor related to plaque generation and development. (7) FIB (fibrinogen), which is involved in intravascular thrombosis, plays an important role in blood coagulation and thrombosis. (8) D-dimer (D-dimer) can be used as one of molecular markers of thrombus formation in vivo. (9) MPV (mean platelet volume) is closely related to platelet count and embolic events, etc. The above markers identify the presence of high-risk plaques mainly from the different perspectives of vascular endothelium, coagulation system and platelet status.

The 3D-CNN modeling process includes: inputting CTA and OCT images, fusing the images through 3D convolution, and extracting classification features by using a supervised learning method; and extracting more structural edge information by using a down-sampling layer, and removing redundant information and noise.

3D-CNN modeling Process as shown in FIG. 4: the original input is images of CTA and OCT, and the image characteristics of the CTA and the OCT are fused through 3D convolution; and the supervised learning method is utilized to more effectively extract the classification characteristics of the patients; the down-sampling layer enables the features to extract more structural edge information and simultaneously eliminates redundant information and noise; the multi-modal common input enables the raw input to require less domain information to provide vulnerable plaque feature identification accuracy.

Meanwhile, an abnormal characteristic map of a vulnerable plaque patient is constructed by CCTA-based whole-heart coronary artery vessel tree data, and the change of plaque morphological structure related to diseases is reflected; and performing plaque feature recognition on the patient coronary artery imaging data of the training set according to a method of a convolutional neural network to construct an intelligent classification model of the coronary atherosclerotic plaque.

Preferably, the system adopts a data expansion technology to improve the over-fitting problem caused by insufficient sample size, improves the classification performance by using transfer learning, and improves the deep learning effect; finally, the automated diagnostic model is validated through a training set. Preferably, the verification index includes: the method is characterized by comprising the steps of true positive, false positive, true negative, false negative, identification accuracy, horse maintenance correlation coefficient and running time.

Specifically, 500 cases of CCTA data are randomly selected as input training set data of the convolutional neural network, and the rest 100 cases are used as a test set to carry out validity verification on the constructed automatic diagnosis model. The method comprises the steps of firstly constructing an abnormal characteristic map of a vulnerable plaque patient based on CCTA whole-heart coronary artery vascular tree data of unstable angina and NSTE-ACS patients, and reflecting the change of plaque morphological structures related to diseases. Secondly, performing plaque feature recognition on the patient coronary artery imaging data of the training set according to a method of a convolutional neural network, and constructing an intelligent classification model of the coronary atherosclerotic plaque; on the basis of 500 cases of image data, data expansion is applied to improve the over-fitting problem caused by insufficient sample size, and the classification performance and the deep learning effect are improved by transfer learning. Finally, the automated diagnostic model is validated through a training set. The verification index includes: the method is characterized by comprising the steps of true positive, false positive, true negative, false negative, identification accuracy, horse maintenance correlation coefficient and running time.

The identification system adopts data amplification and transfer learning: in order to improve the effect of deep learning, on the basis of sample image data used for training, a data expansion method such as image rotation, scale transformation, noise addition and the like is carried out on an original data set, so that the overfitting problem caused by insufficient sample amount is improved. The method comprises the steps of using a transfer learning technology, namely using the existing public data set for pre-training, using weight parameters obtained by training to initialize a network, and then carrying out fine adjustment on network parameters through medical images acquired by a training project, thereby realizing classification of the medical images and improving classification performance.

Preferably, the system further comprises labeling and interpreting vulnerable plaque signs, and plaque quantification analysis. When marking and judging vulnerable plaque signs, carrying out characteristic explanation, namely CT plaque qualitative characteristics (positive reconstruction, napkin ring characteristics, punctate calcification and low-density plaque); and (3) napkin ring marking: a high density of annular rings of non-calcified portions of coronary plaque surrounding a low density of centers; low density plaque: mean CT number <30HU over three regions of interest; spot calcification: the diameter of the micro calcification in any direction is less than 3 mm; positive reconstitution: the ratio of the lumen of the lesion to the lumen of the adjacent reference vessel is greater than 1.1. As shown in fig. 5 (interpretation of vulnerable plaque signs), where a-non-calcified plaque was accompanied by positive remodeling (arrows); b-partially calcified plaque in the middle right coronary artery where low HU plaque occurs; c-plaque napkin ring sign; d-calcification occurred (< 3mm diameter in all directions).

Quantitative plaque analysis, namely quantitative analysis of plaque by CT (including indexes such as total plaque load, plaque position, range, stenosis rate, plaque volume with different densities, reconstruction index and the like): the quantitative analysis of the coronary artery plaque in the blood vessel centerline is automatically adjusted to be carried out on the whole heart coronary artery. If multiple plaques are found within a coronary segment, quantitative measurements of all plaques in the full heart coronary artery are aggregated. The reconstruction index is calculated by dividing the diameter of the blood vessel at the position of the minimum lumen diameter by the ratio of the average lumen diameter of the proximal and distal reference points. The plaque length calculation method is the centerline distance from the proximal end to the distal end of the plaque. The diametric stenosis is calculated as the minimum lumen diameter divided by the mean lumen diameter of the proximal and distal reference points.

IVUS (intravascular ultrasound) and OCT examinations: reference is made to the IVUS test guidelines of the american college of cardiology. Qualitative analysis included dividing the plaques into plaques (lipid plaques) and hard plaques (fibrous plaques and calcified plaques). IVUS diagnoses plaque rupture, thrombus. The results of OCT diagnosis were vulnerable plaque (TCFA), plaque rupture, plaque erosion.

Preferably, the vulnerable plaque quantification index is extracted on the basis of the total volume of all coronary branch plaques (TP), the non-calcified volume (NCP), the calcified volume (CP), the low density plaque volume (LDP), the total plaque load, the plaque density and the maximum Remodeling Index (RI).

The coronary artery plaque automatic identification system based on the neural network deep learning improves the accuracy of coronary artery central line identification, automatically extracts full-heart coronary artery lesions through the RCNN convolutional neural network, and establishes accurate positioning and fine segmentation of a blood vessel tree and plaques so as to realize automatic quantitative identification of the full-heart coronary artery plaques; establishing a novel convolution network algorithm; on the basis of indexes such as total volume (TP) of all coronary artery branch plaques, non-calcified volume (NCP), calcified volume (CP), low density plaque volume (LDP), total plaque load, plaque density and maximum Remodeling Index (RI) by adopting a standardized measurement method, a novel potential vulnerable plaque quantification index is extracted. Therefore, a novel convolution network method for optimizing an automatic identification scheme for full-heart coronary plaque extraction and accurate quantitative analysis is realized.

Preferably, the system can also realize the step of extracting the plaque lesion of the whole heart coronary artery: firstly, obtaining potential target lesion areas through sliding of windows with different widths and heights, and then performing normalization operation to serve as standard input of a convolutional neural network; then carrying out convolution pooling operation according to the input to obtain feature vector output with fixed dimensionality; and finally, classifying according to the feature vectors output in the last step, and detecting an accurate lesion target area through boundary regression.

As shown in fig. 6, the model for coronary plaque lesion localization and lesion segmentation based on the deep convolutional neural network is mainly used for automatically localizing the lesion position and range from a CTA anatomical image; the segmentation model mainly realizes automatic segmentation of the outline of the focus; that is, the aorta region and each main coronary branch are automatically segmented from the CCTA image, and the region of coronary atherosclerotic lesion is automatically detected. And establishing accurate positioning and fine segmentation of the blood vessel tree and the plaque. Automatically extracting the whole-heart coronary artery lesion through an RCNN convolutional neural network, firstly, obtaining a potential target lesion area through sliding of windows with different widths and heights, and then performing normalization operation to be used as standard input of the convolutional neural network; then carrying out convolution pooling operation according to the input to obtain feature vector output with fixed dimensionality; and finally, classifying according to the feature vectors output in the last step, and detecting an accurate lesion target area through boundary regression.

And establishing the early warning model, wherein the early warning model further comprises CT-FFR and CT-ESS based on real-time hemodynamic evaluation. The CT-FFR, CT-ESS based on real-time hemodynamic assessment comprises: and calculating CT-FFR and CT-ESS from the computer virtual coronary artery anatomy database model through training learning simulation. Deep learning-based hemodynamic algorithm training: the core of the algorithm is that FFRct and ESSct are calculated from a computer virtual coronary artery anatomy database model through training learning simulation; the method specifically comprises the following steps:

1) establishing a coronary artery anatomical parameter computer virtual training database; (including coronary artery vessel tree branch conditions, radius of normal vessel position, tip angle, branch length, number and degree of stenosis of whole blood vessel tree, parameters of lesion proximal and distal vessels, and other comprehensive anatomical parameters).

As shown in fig. 7, a CCTA-based real-time blood flow dynamics analysis model is constructed by a deep learning model. The core of the algorithm is to calculate FFRct and ESSct from a computer virtual coronary artery anatomy database model through training learning simulation. Firstly, automatically extracting features from a reconstructed coronary artery anatomical model, and sending the features serving as input into a deep learning model trained in advance. Then applying the constructed model to perform individual arterial hemodynamic evaluation on the selected patient based on CCTA data, selecting the patient in which invasive FFR examination is performed, and performing contrast verification on DL-FFRct, DL-ESSct, CFD-FFRct, CFD-ESSct and invasive FFR calculated based on a deep learning algorithm; the hemodynamic differences of OCT-confirmed stable plaque versus vulnerable plaque local were also determined by lateral studies; longitudinal studies compare changes in hemodynamic indices before and after plaque progression over one to two years of follow-up.

Wherein DL in DL-FFRct refers to deep learning of deep learning, namely the fractional flow reserve based on deep learning, and DL-ESSct is the vascular endothelial shearing force based on deep learning; CFD-FFRct is the fractional flow reserve based on hydrodynamics (traditional method), and CFD-ESSct is the vascular endothelial shear force based on hydrodynamics (traditional method).

Establishing a relation between multidimensional exposure factors and an end point event through a Cox presentation-hazards risk model, and creating a risk evaluation model based on a CT noninvasive imaging and deep learning system; the algorithm is applied to DL-FFR and DL-ESS at all positions of the centerline of the coronary anatomical model.

The full-automatic accurate quantitative identification technology of the full-heart coronary artery plaque is applied to the regular follow-up tracking of the coronary artery atherosclerotic plaque patient. By comparing CCTA data of follow-up two years, accurate quantification results of plaques between two years are obtained, quantitative characteristics of all-heart coronary artery plaques and changes of hemodynamics are obtained through analysis of an automatic vulnerable plaque identification system and a hemodynamics evaluation system, the stability of plaques and future rupture risks are evaluated, the judgment results of the automatic vulnerable plaque identification system of the coronary artery are deeply learned based on a neural network, the automatic vulnerable characteristic identification system of the coronary artery is applied to a noninvasive risk evaluation early warning model of multi-factor integration analysis, and prognosis information (main end-point events are all-cause death, non-fatal myocardial infarction and unstable angina hospitalization) is obtained by combining quantitative indexes of patient risk factors, novel biomarkers, hemodynamics CT-FFR, CT-ESS, plaque progress and the like to establish a comprehensive morphology and functional multi-dimensionality noninvasive rupture early warning model of plaques and follow-up 2 years of follow-up visits Molding;

the follow-up visit for 2 years specifically relates to: follow-up visits were performed three months, half a year, one year, two years after the first examination, respectively. The follow-up end-point events were MACE events, including cardiac death, myocardial infarction, or revascularization of target lesions resulting from ischemia, etc. And in addition, one year and two years are selected as longitudinal comparison nodes, and an automatic identification system based on CCTA neural network deep learning is selected twice before and after the comparison to evaluate the form of the whole-heart coronary artery plaque and the development condition of the hemodynamic index, so that the stability of the plaque and the future rupture risk are predicted.

The plaque risk assessment is not limited to morphology, functional science, biochemical indexes, clinical risk factors and plaque development conditions are integrated, a multi-dimensional noninvasive whole-heart coronary plaque rupture risk early warning model is constructed, and a comprehensive and reliable means is provided for understanding the vulnerable plaque generation and rupture mechanism and early accurate early warning.

The multi-dimensional plaque rupture risk early warning system integrates the obtained CCTA vulnerable plaque automatic identification method with the CT real-time hemodynamic analysis technology based on deep learning, creates a more efficient, sensitive and accurate non-invasive plaque rupture risk early warning model, and provides a comprehensive and reliable basis and a key technical means for intelligent non-invasive accurate detection and early warning of future coronary plaque rupture.

In light of the foregoing description of the preferred embodiment of the present invention, many modifications and variations will be apparent to those skilled in the art without departing from the spirit and scope of the invention. The technical scope of the present invention is not limited to the content of the specification, and must be determined according to the scope of the claims.

Claims

1. A coronary artery multi-dimensional plaque rupture risk early warning system which characterized in that: the system comprises: a model building module: establishing a coronary artery multi-dimensional noninvasive vulnerable plaque rupture risk early warning model, which comprises an early warning module established based on CT quantitative analysis of plaque and CT plaque qualitative characteristics;

wherein the multidimensional noninvasive vulnerable plaque rupture risk early warning model is established based on a neural network deep learning coronary artery vulnerable plaque automatic identification system and a real-time hemodynamic evaluation method,

the automatic identification system for the coronary vulnerable plaque for deep learning of the neural network is characterized in that a coronary artery plaque of a coronary artery CTA image and a coronary artery plaque of an OCT image which are selected from the same sample are taken as regions of interest, the CTA image and the OCT image of the same plaque are input into a convolution 3D convolution neural network as different channels, image features of the CTA and the OCT are automatically fused in a full-connection layer part of the subsequent 3D convolution neural network, and classification features are extracted by using a supervised learning method; the system adopts a data expansion technology to improve the overfitting problem caused by insufficient sample size, thereby realizing the test of the convolutional neural network data training.

2. The system of claim 1, wherein: the model building module specifically comprises: and establishing an early warning module based on quantitative analysis of the CT on the plaque and the qualitative characteristics of the CT plaque.

3. The system of claim 2, wherein: parameters used by quantitative analysis of the plaque and CT plaque qualitative characteristics are obtained by a neural network deep learning coronary vulnerable plaque automatic identification system; the quantitative analysis of the plaque by the CT comprises the total load of the whole heart plaque, the plaque position, the plaque range, the stenosis rate, the plaque volumes with different densities and a reconstruction index; the CT plaque qualitative characteristics comprise positive reconstruction, napkin ring characteristics, punctate calcification and low-density plaque.

4. The system of claim 1, wherein: the model building also includes CT-FFR, CT-ESS based on real-time hemodynamic assessment.

5. The system of claim 4, wherein: the CT-FFR, CT-ESS based on real-time hemodynamic assessment comprises: and calculating CT-FFR and CT-ESS from the computer virtual coronary artery anatomy database model through training learning simulation.

6. The system of claim 5, wherein: the CT-FFR is calculated from the computer virtual coronary artery anatomy database model through training learning simulation, and the CT-ESS comprises:

3) finally, the algorithm calculates FFR and ESS at any position of the coronary vessel tree, and performs color coding on the FFR and ESS to perform visual display.

7. The system of claim 1, wherein: and establishing a relation between the multidimensional exposure factors and the end point events through a Cox reporting-hazards risk model, and creating a risk assessment model based on a CT noninvasive imaging and deep learning system.

8. The system of any of the preceding claims 1 to 7, wherein: the model is also established based on clinical risk factors and novel serum biomarkers.