A vascular lesion risk identification system based on a machine learning model

Abstract

Vascular lesion risk identification system based on machine learning model of the present invention: A control host acquires the blood oxygen value, pH value, lactic acid value, pressure value in the blood of the evaluator's blood vessels and the image of the inner wall structure of the blood vessels; the blood vessels are venous vessels. Based on the structural image, the diameter, the thickness of the inner, middle and outer membranes of the blood vessels are analyzed. The blood oxygen value, pH value, lactic acid value, pressure value, diameter, and the thickness of the inner, middle and outer membranes of the blood vessels of the evaluator are input into the machine learning model of the target venous blood vessels for training to output the identification result of the risk of venous blood vessel lesions; the blood vessels are arterial blood vessels. Based on the structural image, the diameter, the thickness of the inner, middle and outer membranes of the blood vessels are analyzed, and when there are arterial plaques in the arterial blood vessels, their composition and morphology are identified and the diameter stenosis degree value is calculated. The blood oxygen value, pH value, lactic acid value, pressure value, diameter, the thickness of the inner, middle and outer membranes of the blood vessels, the composition and morphology when there are arterial plaques, and the diameter stenosis degree value of the evaluator are input into the machine learning model of the target arterial blood vessels for training to output the identification result of the risk of arterial blood vessel lesions.

Description

Technical Field

[0001] The present invention relates to the field of machine learning and disease risk identification technology, and in particular to a vascular lesion risk identification system based on a machine learning model. This invention technology comes from the project: National Key R&D Program: "Research on Pan-vascular Disease Screening, System Evaluation and Prevention and Treatment System", No.: 2021YFC2500500. Background Art

[0002] Faced with the increasing burden of vascular diseases, how to use existing diagnosis and treatment resources to effectively prevent and control the risk of vascular diseases has become an important issue that needs to be urgently addressed in the current medical field. Existing intravascular ultrasound catheters can measure vascular structure images and obtain some vascular data based on vascular structure images. Doctors can determine whether patients have vascular lesion risks and which type of vascular lesion risks are involved based on vascular data and their own experience. This determination method has low accuracy and low real-time performance. Based on the existing defects, the present invention designs a vascular lesion risk identification system based on a machine learning model. Summary of the invention

[0003] In view of the problems and shortcomings of the prior art, the present invention provides a vascular lesion risk identification system based on a machine learning model.

[0004] The present invention solves the above technical problems through the following technical solutions:

[0005] The present invention provides a vascular lesion risk identification system based on a machine learning model, which includes an intravascular ultrasonic catheter and a control host, and is characterized in that a blood oxygen sensor, a pH sensor, a lactate sensor and a pressure sensor are installed on the outer surface of the intravascular ultrasonic catheter, and an ultrasonic sensor is installed on the outer surface of the front end of the intravascular ultrasonic catheter;

[0006] The control host is used to obtain the blood oxygen value, pH value, lactate value and pressure value in the blood vessels of the assessor through the blood oxygen sensor, pH sensor, lactate sensor and pressure sensor respectively, and obtain a high-resolution structural image of the inner wall of the assessor's blood vessels through the ultrasonic sensor and intravascular ultrasound technology;

[0007] The control host is also used to analyze whether the blood vessel is a venous blood vessel or an arterial blood vessel based on the blood oxygen value and the lactic acid value. When it is analyzed that the blood vessel is a venous blood vessel, image processing is performed on the structural image, and then the blood vessel diameter, the thickness of the vascular intima, the thickness of the vascular media, and the thickness of the vascular adventitia are analyzed. When it is analyzed that the blood vessel is an arterial blood vessel, image processing is performed on the structural image, and then the blood vessel diameter, the thickness of the vascular intima, the thickness of the vascular media, the thickness of the vascular adventitia, and whether there are arterial plaques in the arterial blood vessel are analyzed. And when it is analyzed that there are arterial plaques in the arterial blood vessel, the composition and morphology of the arterial plaques are identified, and the value of the diameter stenosis degree is calculated;

[0008] The control host is also used to, when the blood vessel is a venous blood vessel, input the blood oxygen value, pH value, lactic acid value, pressure value, blood vessel diameter, thickness of the vascular intima, thickness of the vascular media, and thickness of the vascular adventitia of the evaluator into the target venous blood vessel machine learning model for training and learning, and output the identification result of the venous blood vessel lesion risk. The identification result of the venous blood vessel lesion risk is no risk of venous blood vessel lesion, risk of deep vein valve insufficiency, and risk of deep vein thrombosis;

[0009] The control host is also used to, when the blood vessel is an arterial blood vessel, input the blood oxygen value, pH value, lactic acid value, pressure value, blood vessel diameter, thickness of the vascular intima, thickness of the vascular media, thickness of the vascular adventitia, the composition and morphology of the arterial plaques when there are arterial plaques, and the value of the diameter stenosis degree of the evaluator into the target arterial blood vessel machine learning model for training and learning, and output the identification result of the arterial blood vessel lesion risk. The identification result of the arterial blood vessel lesion risk is no risk of arterial blood vessel lesion, risk of atherosclerosis, risk of aneurysm, risk of thromboangiitis obliterans, and risk of arterial dissection.

[0010] The positive and progressive effects of the present invention are as follows:

[0011] The blood vessel lesion risk identification system based on the machine learning model designed by the present invention can accurately analyze the identification result of the blood vessel lesion risk of the patient. The identification result of the blood vessel lesion risk includes the identification result of the venous blood vessel lesion risk and the identification result of the arterial blood vessel lesion risk. The present invention has high accuracy and high real-time performance, and improves the detection efficiency. Description of the Drawings

[0012] Figure 1 It is a structural block diagram of the blood vessel lesion risk identification system based on the machine learning model of the preferred embodiment of the present invention. Detailed Embodiment

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0014] As Figure 1 shown, this embodiment provides a risk identification system for vascular lesions based on a machine learning model, which includes an intravascular ultrasound catheter 100 and a control host 200. A blood oxygen sensor 10, a pH sensor 20, a lactic acid sensor 30, and a pressure sensor 40 are installed on the outer surface of the intravascular ultrasound catheter 100, and an ultrasound sensor 50 is installed on the outer surface of the foremost end of the intravascular ultrasound catheter 100.

[0015] The control host 200 is used to obtain the blood oxygen value, pH value, lactic acid value, and pressure value in the blood vessels of the evaluator through the blood oxygen sensor 10, pH sensor 20, lactic acid sensor 30, and pressure sensor 40 respectively, and obtain a high-resolution structural image of the inner wall of the evaluator's blood vessels through the ultrasound sensor 50 and intravascular ultrasound technology.

[0016] The control host 200 is also used to analyze whether the blood vessel is a venous blood vessel or an arterial blood vessel based on the blood oxygen value and lactic acid value. When it is analyzed that the blood vessel is a venous blood vessel, after image processing of the structural image, analyze the blood vessel diameter (pipe diameter), the thickness of the vascular intima, the thickness of the vascular media, and the thickness of the vascular adventitia; when it is analyzed that the blood vessel is an arterial blood vessel, after image processing of the structural image, analyze the blood vessel diameter, the thickness of the vascular intima, the thickness of the vascular media, the thickness of the vascular adventitia, and whether there is an arterial plaque in the arterial blood vessel, and when it is analyzed that there is an arterial plaque in the arterial blood vessel, identify the composition and morphology of the arterial plaque and calculate the pipe diameter stenosis degree value = (blood vessel diameter - arterial plaque cross-sectional area) / blood vessel diameter.

[0017] In this embodiment, based on the characteristics of the blood oxygen value and lactic acid value in venous blood vessels and arterial blood vessels, in venous blood vessels, the blood oxygen value is low and the lactic acid value is high, and in arterial blood vessels, the blood oxygen value is high and the lactic acid value is low. Accordingly, it can be analyzed whether the blood vessel is a venous blood vessel or an arterial blood vessel. For venous blood vessels and arterial blood vessels, different data are collected.

[0018] Among them, identifying the composition of the arterial plaque specifically includes constructing an arterial plaque composition recognition model and identifying the composition of the arterial plaque based on the constructed arterial plaque composition recognition model.

[0019] Constructing an arterial plaque composition recognition model: High-resolution structural images of the inner walls of arterial blood vessels of A historical patients are collected through an ultrasonic sensor 50 and intravascular ultrasound technology. After image processing of the structural images of each historical patient, plaque image segmentation is performed to obtain the plaque segmentation images of each historical patient. Feature extraction is performed on each plaque segmentation image to obtain corresponding multiple plaque composition features, and specific composition labels of the arterial plaques of each historical patient are marked. These specific composition labels of the arterial plaques include lipid core composition labels, necrotic substance composition labels, fibrous cap composition labels, inflammatory cell composition labels, calcium deposition composition labels, and smooth muscle cell composition labels. A plaque composition feature set for each historical patient is constructed. The plaque composition feature set consists of multiple plaque composition features and corresponding specific composition labels of the arterial plaques. A plaque composition training sample set and a plaque composition verification sample set are constructed. The plaque composition training sample set includes the plaque composition feature sets of A1 historical patients, and the plaque composition verification sample set includes the plaque composition feature sets of A2 historical patients. A = A1 + A2, where A, A1, and A2 are all positive integers and A1 is much larger than A2. A random forest regression model is selected to construct the arterial plaque composition recognition model. The plaque composition training sample set is input into the arterial plaque composition recognition model for model training, and the plaque composition verification sample set is input into the trained arterial plaque composition recognition model for model verification to construct the target arterial plaque composition recognition model.

[0020] For example: The plaque composition feature sets of 80% of the historical patients constitute the plaque composition training sample set, and the plaque composition feature sets of 20% of the historical patients constitute the plaque composition verification sample set.

[0021] Identifying the composition of arterial plaques based on the constructed arterial plaque composition recognition model: When it is analyzed that there are arterial plaques in the arterial blood vessels, plaque image segmentation is performed on the image after image processing by the evaluator to obtain the plaque segmentation image of the evaluator. Feature extraction is performed on the plaque segmentation image to obtain multiple plaque composition features of the evaluator. The multiple plaque composition features of the evaluator are input into the target arterial plaque composition recognition model to identify the composition of the arterial plaques.

[0022] Identifying the morphology of arterial plaques, specifically including constructing an arterial plaque morphology recognition model and identifying the morphology of arterial plaques based on the constructed arterial plaque morphology recognition model.

[0023] Construct an arterial plaque morphology recognition model: Extract features from the plaque segmentation images of each historical patient to obtain multiple corresponding plaque morphology features, and label the specific morphological markers of the arterial plaques of each historical patient. These specific morphological markers of arterial plaques include stable plaque markers, unstable plaque markers, vulnerable plaque markers, calcified plaque markers, fibrous plaque markers, and thrombotic plaque markers. Construct a plaque morphology feature set for each historical patient. The plaque morphology feature set consists of multiple plaque morphology features and the corresponding specific morphological markers of arterial plaques. Construct a plaque morphology training sample set and a plaque morphology verification sample set. The plaque morphology training sample set includes the plaque morphology feature sets of A1 historical patients, and the plaque morphology verification sample set includes the plaque morphology feature sets of A2 historical patients. Select a random forest regression model to construct an arterial plaque morphology recognition model. Input the plaque morphology training sample set into the arterial plaque morphology recognition model for model training, and input the plaque morphology verification sample set into the trained arterial plaque morphology recognition model for model verification to construct the target arterial plaque morphology recognition model.

[0024] Identify the morphology of the arterial plaque based on the constructed arterial plaque morphology recognition model: When it is analyzed that there are arterial plaques in the arterial blood vessels, perform plaque image segmentation on the image processed by the evaluator to obtain the plaque segmentation image of the evaluator, extract features from the plaque segmentation image to obtain multiple plaque morphology features of the evaluator, and input the multiple plaque morphology features of the evaluator into the target arterial plaque morphology recognition model to identify the morphology of the arterial plaque.

[0025] The control host 200 is also used to input the blood oxygen value, pH value, lactic acid value, pressure value, blood vessel diameter, intima thickness, media thickness, and adventitia thickness of the evaluator into the target venous blood vessel machine learning model for training and learning when the blood vessel is a venous blood vessel, and output the venous blood vessel lesion risk recognition result. The venous blood vessel lesion risk recognition result includes no venous blood vessel lesion risk, deep vein valve insufficiency risk, and deep vein thrombosis risk.

[0026] Among them, constructing the target venous blood vessel machine learning model includes: Obtain the blood oxygen value, pH value, lactic acid value, and pressure value in the blood vessels of B historical venous testers through a blood oxygen sensor, a pH sensor, a lactic acid sensor, and a pressure sensor respectively, obtain high-resolution structural images of the inner walls of the blood vessels of B historical venous testers through an ultrasonic sensor and intravascular ultrasound technology, perform image processing on the structural images of each historical venous tester and analyze the blood vessel diameter, intima thickness, media thickness, and adventitia thickness, and label the venous blood vessel lesion markers of each historical venous tester. These venous blood vessel lesion markers include no venous blood vessel lesion markers, deep vein valve insufficiency markers, and deep vein thrombosis markers.

[0027] Construct the venous vessel feature set for each historical venous detector. The venous vessel feature set consists of blood oxygen value, pH value, lactic acid value, pressure value, vessel diameter, intima thickness of the vessel, media thickness of the vessel, adventitia thickness of the vessel, and the corresponding venous vessel lesion markers; construct the venous vessel sample set, which includes the venous vessel feature sets of B historical venous detectors, where B is a positive integer.

[0028] Construct a venous vessel machine learning model based on a convolutional neural network, use the particle swarm optimization algorithm and genetic algorithm to optimize the hyperparameters of the venous vessel machine learning model, and use the venous vessel sample set to train the venous vessel machine learning model. Obtain the optimal hyperparameters through continuous iteration, and use the optimal hyperparameters to construct the target venous vessel machine learning model.

[0029] Specifically: construct a venous vessel machine learning model based on a convolutional neural network, initialize the hyperparameters of the venous vessel machine learning model, set the number of particles L in the particle swarm, the maximum number of iterations T, and the dimension D of the particle search space corresponding one-to-one to the hyperparameters. Each particle represents a set of hyperparameters, and initialize the velocity and position of each particle.

[0030] Construct the corresponding venous vessel machine learning model based on any particle as a hyperparameter, use the venous vessel sample set to train the venous vessel machine learning model corresponding to the particle, and calculate the corresponding fitness value using the fitness function. Calculate the fitness values of all particles based on this. The maximum fitness value among all particles is used as the optimal fitness, the current position of each particle is used as the individual optimal position of each particle, and the current position of the particle corresponding to the optimal fitness is used as the global optimal position of the particle swarm. Among them, the fitness function MSE(X) represents the mean squared error of the deviation.

[0031] Update the velocity and position of each particle:

[0032]

[0033] X ld (t + 1) = X ld (t) + V ld (t + 1)

[0034]

[0035] where, V ld (t) and X ld (t) respectively represent the velocity and position of the current particle, V ld (t + 1) and X ld(t + 1) represent the velocity and position of the updated particle respectively, l represents the l-th particle, d represents the d-th dimension of the search space, t represents the current iteration number, W(t) represents the adaptive inertia weight, W min1 represents the minimum value of the inertia weight, W max1 represents the maximum value of the inertia weight, C 1 and C 2 represent the acceleration factors, Xpb ld represents the historical individual best position, and its corresponding fitness value is the historical individual best fitness value, Xgb d represents the historical global best position, and its corresponding fitness value is the historical global best fitness value.

[0036] Update the historical individual best position and the historical global best position: Calculate the fitness values of each updated particle using the fitness function, compare the updated fitness values of each particle with the corresponding historical individual best fitness values. If the updated fitness value of any particle is greater than the corresponding historical individual best fitness value, then update the historical individual best fitness value of this particle = the updated fitness value of this particle, and update the historical individual best position of this particle = the updated position of this particle, otherwise do not update; then take the maximum value among the current historical individual best fitness values of each particle as the current best fitness value. If the current best fitness value is greater than the historical global best fitness value, then update the historical global best fitness value = the current best fitness value, and update the historical global best position = the current position of the particle corresponding to the current best fitness value, otherwise do not update.

[0037] Judge whether the historical global best fitness value reaches the first set fitness value or whether the iteration number reaches the maximum iteration number T. If so, obtain the optimal hyperparameters based on the particle at the historical global best position, and construct the target venous blood vessel machine learning model using the optimal hyperparameters. Otherwise, judge whether the current iteration number t is less than the set threshold T1. If it is, perform a hybridization operation on the updated particles using the genetic algorithm, and update the velocity and position of each particle in the particle population formed after hybridization again. If it is not, directly update the velocity and position again, where T / 2 ≤ T1 ≤ T.

[0038] At this time, a genetic algorithm is used to perform a hybridization operation on the updated particles, and the velocity and position of each particle in the particle population formed after hybridization are updated again: the current historical individual optimal fitness values of each particle are arranged in descending order, the L1 particles at the end of the sorting are eliminated, the particle with the optimal position in the historical population among the uneliminated particles is used as the parent generation, and L1 particles are randomly selected from the remaining uneliminated particles as the parent generation. Each selected parent generation is hybridized with the parent generation to generate L1 particles. The current positions of the L1 particles generated by hybridization are used as the positions of the corresponding particles after t iterations respectively. The current velocity of each particle among the L1 particles generated by hybridization is the average value of the velocities of the particles on both sides of the particle, so as to obtain the particle population formed after hybridization. The particle population contains L particles. The velocity and position of each particle in the particle population formed after hybridization are updated again. L1 is a positive integer and 1≤L1≤L / 2, and t = t + 1.

[0039] The control host 200 is further configured to, when the blood vessel is an arterial blood vessel, input the blood oxygen value, pH value, lactic acid value, pressure value, blood vessel diameter, blood vessel intima thickness, blood vessel media thickness, blood vessel adventitia thickness, the composition and morphology of the arterial plaque when there is an arterial plaque, and the value of the diameter stenosis degree of the evaluator into the target arterial blood vessel machine learning model for training and learning, and output the arterial blood vessel lesion risk identification result. The arterial blood vessel lesion risk identification result includes no arterial blood vessel lesion risk, atherosclerosis risk, aneurysm risk, thromboangiitis obliterans risk, and aortic dissection risk.

[0040] Among them, constructing the target arterial blood vessel machine learning model includes: respectively obtaining the blood oxygen value, pH value, lactic acid value, and pressure value in the blood vessels of C historical arterial detectors through a blood oxygen sensor, a pH sensor, a lactic acid sensor, and a pressure sensor, obtaining high-resolution structural images of the inner walls of the blood vessels of C historical arterial detectors through an ultrasonic sensor and intravascular ultrasound technology, performing image processing on the structural image of each historical arterial detector and then analyzing the blood vessel diameter, blood vessel intima thickness, blood vessel media thickness, blood vessel adventitia thickness, and whether there is an arterial plaque in the arterial blood vessel. And when it is analyzed that there is an arterial plaque in the arterial blood vessel, the composition and morphology of the arterial plaque are identified by using the arterial plaque composition identification model and the arterial plaque morphology identification model, and the value of the diameter stenosis degree is calculated, and the arterial blood vessel lesion label of each historical arterial detector is marked. These arterial blood vessel lesion labels include no arterial blood vessel lesion label, atherosclerosis label, aneurysm label, thromboangiitis obliterans label, and aortic dissection label.

[0041] Construct the arterial vessel feature set for each historical artery detector. The arterial vessel feature set consists of blood oxygen value, pH value, lactic acid value, pressure value, vessel diameter, intima thickness of the vessel, media thickness of the vessel, adventitia thickness of the vessel, composition and morphology of the arterial plaque and the value of the diameter stenosis degree when there is an arterial plaque, and the corresponding arterial vessel lesion markers. Construct the arterial vessel sample set, which includes the arterial vessel feature sets of C historical artery detectors, where C is a positive integer.

[0042] Construct an arterial vessel machine learning model based on a convolutional neural network. Use the particle swarm optimization algorithm and genetic algorithm to optimize the hyperparameters of the arterial vessel machine learning model, and use the arterial vessel sample set to train the arterial vessel machine learning model. Obtain the optimal hyperparameters through continuous iteration, and construct the target arterial vessel machine learning model using the optimal hyperparameters.

[0043] Specifically: construct an arterial vessel machine learning model based on a convolutional neural network, initialize the hyperparameters of the arterial vessel machine learning model, set the number of particles N in the particle swarm, the maximum number of iterations S, and the particle search space dimension K corresponding one-to-one to the hyperparameters. Each particle represents a set of hyperparameters, and initialize the velocity and position of each particle.

[0044] Construct the corresponding arterial vessel machine learning model based on any particle as a hyperparameter, use the arterial vessel sample set to train the arterial vessel machine learning model corresponding to the particle, and calculate the corresponding fitness value using the fitness function. Calculate the fitness values of each particle based on this. The maximum fitness value among each particle is used as the optimal fitness, the current position of each particle is used as the individual optimal position of each particle, and the current position of the particle corresponding to the optimal fitness is used as the global optimal position of the particle swarm. Among them, the fitness function MSE(X) represents the mean squared error of the deviation.

[0045] Update the velocity and position of each particle:

[0046]

[0047] X nk (s + 1) = X nk (s) + V nk (s + 1)

[0048]

[0049] Among them, V nk (s) and X nk (s) respectively represent the velocity and position of the current particle, V nk (s + 1) and X nk(s + 1) represent the velocity and position of the updated particle respectively, n represents the nth particle, k represents the kth dimension of the search space, s represents the current iteration number, W(s) represents the adaptive inertia weight, W min2 represents the minimum value of the inertia weight, W max2 represents the maximum value of the inertia weight, C 3 and C 4 represent the acceleration factors, Xpb nk represents the historical individual best position, and its corresponding fitness value is the historical individual best fitness value, Xgb k represents the historical global best position, and its corresponding fitness value is the historical global best fitness value.

[0050] Update the historical individual best position and the historical global best position: Calculate the fitness values of each updated particle using the fitness function, compare the updated fitness values of each particle with the corresponding historical individual best fitness values. If the updated fitness value of any particle is greater than the corresponding historical individual best fitness value, then update the historical individual best fitness value of this particle = the updated fitness value of this particle, and update the historical individual best position of this particle = the updated position of this particle, otherwise do not update; then take the maximum value among the current historical individual best fitness values of each particle as the current best fitness value. If the current best fitness value is greater than the historical global best fitness value, then update the historical global best fitness value = the current best fitness value, and update the historical global best position = the current position of the particle corresponding to the current best fitness value, otherwise do not update.

[0051] Judge whether the historical global best fitness value reaches the second set fitness value or whether the iteration number reaches the maximum iteration number S. If so, obtain the optimal hyperparameters based on the particle at the historical global best position, and construct the target arterial blood vessel machine learning model using the optimal hyperparameters. Otherwise, judge whether the current iteration number s is less than the set threshold S1. If it is, perform the hybridization operation on the updated particles using the genetic algorithm, and update the velocity and position of each particle in the particle population formed after hybridization again. If it is not, directly update the velocity and position again, where S / 2 ≤ S1 ≤ S.

[0052] At this time, a genetic algorithm is used to perform a hybridization operation on the updated particles, and the velocity and position of each particle in the particle population formed after hybridization are updated again. Specifically: the current historical individual optimal fitness values of each particle are arranged in descending order, the N1 particles ranked at the end are eliminated, the particle with the optimal position in the historical population among the uneliminated particles is used as the mother generation, and N1 particles are randomly selected from the remaining uneliminated particles as the father generation. Each selected father generation is hybridized with the mother generation to generate N1 particles. The current positions of the N1 particles generated by hybridization are respectively used as the positions of the corresponding particles after s iterations. The current velocity of each particle among the N1 particles generated by hybridization is the average value of the velocities of the particles on both sides of the particle, so as to obtain the particle population formed after hybridization. The particle population contains N particles. The velocity and position of each particle in the particle population formed after hybridization are updated again. N1 is a positive integer and 1≤N1≤N / 2, s = s + 1.

[0053] In this embodiment, in the middle and early stages of iteration, the hybridization operation is used to prevent the particle swarm from falling into a local optimal solution, and in the later stage of iteration, the hybridization operation is reduced to ensure the stability and convergence of the solution; this embodiment combines the advantages of the particle swarm algorithm and the genetic algorithm, and can construct an optimized venous blood vessel machine learning model and an arterial blood vessel machine learning model.

[0054] In this embodiment, the ultrasonic sensor 50 and the intravascular ultrasound technology can monitor the structure of the blood vessel wall, and based on this, blood vessel parameters can be obtained. The blood oxygen sensor 10, pH sensor 20, lactate sensor 30, and pressure sensor 40 can simultaneously monitor the changes in blood oxygen and the local chemical environment, providing more parameters for the diagnosis of blood vessel lesions. Then, the external control host is built-in with software, and through the artificial intelligence machine learning algorithm, the risk of blood vessel lesions can be automatically identified in real-time data. The intelligent data analysis can greatly improve the efficiency and accuracy of diagnosis and reduce human errors.

[0055] Although the specific implementation manners of the present invention have been described above, those skilled in the art should understand that these are only examples, and the protection scope of the present invention is defined by the appended claims. Without departing from the principles and essence of the present invention, those skilled in the art can make various changes or modifications to these implementation manners, but these changes and modifications all fall within the protection scope of the present invention.

Claims

1. A vascular lesion risk identification system based on a machine learning model, comprising an intravascular ultrasound catheter and a control host, characterized in that: The outer surface of the intravascular ultrasonic catheter is equipped with a blood oxygen sensor, a pH sensor, a lactate sensor and a pressure sensor, and the front end outer surface of the intravascular ultrasonic catheter is equipped with an ultrasonic sensor; The control host is used to obtain the blood oxygen value, pH value, lactate value and pressure value in the blood vessels of the assessee through the blood oxygen sensor, pH sensor, lactate sensor and pressure sensor respectively, and obtain a high-resolution structural image of the inner wall of the assessee's blood vessels through the ultrasonic sensor and intravascular ultrasound technology; The control host is also used to analyze whether the blood vessel is a vein or an artery based on the blood oxygen value and the lactate value, and when the blood vessel is analyzed to be a vein, analyze the blood vessel diameter, the thickness of the blood vessel intima, the thickness of the blood vessel middle membrane and the thickness of the blood vessel adventitia after performing image processing on the structural image; when the blood vessel is analyzed to be an artery, analyze the blood vessel diameter, the thickness of the blood vessel intima, the thickness of the blood vessel middle membrane, the thickness of the blood vessel adventitia and whether there is an arterial plaque in the artery after performing image processing on the structural image, and when the arterial plaque is analyzed to be present in the arterial vessel, identify the composition and morphology of the arterial plaque and calculate the stenosis degree value of the vessel diameter; The control host is also used for inputting the evaluator's blood oxygen value, pH value, lactate value, pressure value, blood vessel diameter, intima thickness, media thickness and adventitia thickness into the target venous blood vessel machine learning model for training and learning, and outputting a venous vascular disease risk identification result, the venous vascular disease risk identification result being no venous vascular disease risk, deep vein valve insufficiency risk and deep vein thrombosis risk; The control host is also used to input the evaluator's blood oxygen value, pH value, lactate value, pressure value, blood vessel diameter, intima thickness, media thickness, adventitia thickness, composition and morphology of arterial plaques when arterial plaques are present, and degree of stenosis of the vessel diameter into the target arterial machine learning model for training and learning, and output an arterial vascular disease risk identification result, and the arterial vascular disease risk identification result is no arterial vascular disease risk, atherosclerosis risk, aneurysm risk, thromboangiitis obliterans risk and arterial dissection risk.

2. The vascular lesion risk identification system based on machine learning model according to claim 1, characterized in that: The control host is also used to construct an arterial plaque composition recognition model: collect high-resolution structural images of the inner wall of the arterial blood vessels of A historical patients by using an ultrasonic sensor and intravascular ultrasound technology, perform image processing on the structural image of each historical patient and then perform plaque image segmentation to obtain a plaque segmentation image of each historical patient, perform feature extraction on each plaque segmentation image to obtain a corresponding plurality of plaque composition features, and mark specific composition markers of the arterial plaques of each historical patient, wherein the specific composition markers of the arterial plaques include a lipid core composition marker, a necrotic substance composition marker, a fibrous cap composition marker, an inflammatory cell composition marker, a calcium deposition composition marker, and a smooth muscle cell composition marker; Construct a plaque composition feature set for each historical patient. The plaque composition feature set consists of multiple plaque composition features and corresponding arterial plaque specific composition labels. Construct a plaque composition training sample set and a plaque composition verification sample set. The plaque composition training sample set includes the plaque composition feature sets of A1 historical patients, and the plaque composition verification sample set includes the plaque composition feature sets of A2 historical patients. A=A1+A2, A, A1 and A2 are all positive integers and A1 is much larger than A2. A random forest regression model was used to construct an arterial plaque composition recognition model, the plaque composition training sample set was input into the arterial plaque composition recognition model for model training, and the plaque composition verification sample set was input into the trained arterial plaque composition recognition model for model verification, so as to construct a target arterial plaque composition recognition model; The control host is also used to perform plaque image segmentation on the image processed by the evaluator to obtain the evaluator's plaque segmentation image when analyzing the existence of arterial plaque in the arterial blood vessel, perform feature extraction on the plaque segmentation image to obtain the evaluator's multiple plaque composition features, input the evaluator's multiple plaque composition features into the target arterial plaque composition recognition model, and identify the composition of the arterial plaque.

3. The vascular lesion risk identification system based on machine learning model as claimed in claim 2, characterized in that: The control host is also used to construct an arterial plaque morphology recognition model: feature extraction is performed on the plaque segmentation image of each historical patient to obtain a plurality of corresponding plaque morphology features, and specific morphological marks of the arterial plaque of each historical patient are marked, and the specific morphological marks of the arterial plaque include stable plaque marks, unstable plaque marks, vulnerable plaque marks, calcified plaque marks, fibrous plaque marks and thrombotic plaque marks; Constructing a plaque morphology feature set for each historical patient, the plaque morphology feature set is composed of multiple plaque morphology features and corresponding arterial plaque specific morphology marks, constructing a plaque morphology training sample set and a plaque morphology verification sample set, the plaque morphology training sample set includes the plaque morphology feature set of A1 historical patients, and the plaque morphology verification sample set includes the plaque morphology feature set of A2 historical patients; A random forest regression model was used to construct an arterial plaque morphology recognition model, the plaque morphology training sample set was input into the arterial plaque morphology recognition model for model training, the plaque morphology verification sample set was input into the trained arterial plaque morphology recognition model for model verification, and the target arterial plaque morphology recognition model was constructed; The control host is also used to perform plaque image segmentation on the image processed by the evaluator to obtain the evaluator's plaque segmentation image when analyzing the existence of arterial plaque in the arterial blood vessel, perform feature extraction on the plaque segmentation image to obtain the evaluator's multiple plaque morphological features, input the evaluator's multiple plaque morphological features into the target arterial plaque morphology recognition model, and identify the morphology of the arterial plaque.

4. The vascular lesion risk identification system based on machine learning model according to claim 1, characterized in that: The control host is also used to build a target venous blood vessel machine learning model: The blood oxygen value, pH value, lactate value and pressure value in the blood vessels of B historical vein testers are obtained through a blood oxygen sensor, a pH sensor, a lactate sensor and a pressure sensor respectively, and a high-resolution structural image of the inner wall of the blood vessels of the B historical vein testers is obtained through an ultrasonic sensor and intravascular ultrasound technology, and the structural image of each historical vein tester is processed and analyzed to obtain the blood vessel diameter, the thickness of the vascular intima, the thickness of the vascular middle membrane and the thickness of the vascular adventitia, and the venous vascular lesion markers of each historical vein tester are marked, and the venous vascular lesion markers include a venous vascular lesion-free marker, a deep vein valve insufficiency marker and a deep vein thrombosis marker; Construct a venous blood vessel feature set for each historical vein tester, the venous blood vessel feature set consists of blood oxygen value, pH value, lactate value, pressure value, blood vessel diameter, intima thickness, media thickness and adventitia thickness and corresponding venous vascular lesion markers, and construct a venous blood vessel sample set, the venous blood vessel sample set includes the venous blood vessel feature sets of B historical vein testers, B is a positive integer; A venous vascular machine learning model was constructed based on a convolutional neural network. The particle swarm algorithm and genetic algorithm were used to optimize the hyperparameters of the venous vascular machine learning model. The venous vascular sample set was used to train the venous vascular machine learning model. The optimal hyperparameters were obtained through continuous iteration, and the target venous vascular machine learning model was constructed using the optimal hyperparameters.

5. The vascular lesion risk identification system based on machine learning model according to claim 4, characterized in that: The control host is also used to construct a venous blood vessel machine learning model based on a convolutional neural network, use a particle swarm algorithm and a genetic algorithm to optimize the hyperparameters of the venous blood vessel machine learning model, and use a venous blood vessel sample set to train the venous blood vessel machine learning model, obtain the optimal hyperparameters through continuous iteration, and use the optimal hyperparameters to construct a target venous blood vessel machine learning model: A venous machine learning model is constructed based on a convolutional neural network. The hyperparameters of the venous machine learning model are initialized. The number of particles L, the maximum number of iterations T, and the particle search space dimension D corresponding to the hyperparameters are set. Each particle represents a set of hyperparameters. The speed and position of each particle are initialized. Based on any particle used as a hyperparameter, a corresponding venous machine learning model is constructed, and the venous machine learning model corresponding to the particle is trained using a venous sample set, and a corresponding fitness value is calculated using a fitness function. The fitness function is a function that is inversely proportional to the mean square error of the model deviation. Based on this, the fitness value of each particle is calculated, the maximum fitness value among each particle is used as the optimal fitness, the current position of each particle is used as the individual optimal position of each particle, and the current position of the particle corresponding to the optimal fitness is used as the group optimal position of the particle population; Update the velocity and position of each particle: X ld (t+1)=X ld (t)+V ld (t+1) Among them, V ld (t) and X ld (t) represent the current particle speed and position, V ld (t+1) and X ld (t+1) represents the updated particle speed and position, l represents the lth particle, d represents the dth dimension of the search space, t represents the current iteration number, W(t) represents the adaptive inertia weight, W min1 Indicates the minimum inertia weight, W max1 Indicates the maximum inertia weight, C1 and C2 indicate the acceleration factor, Xpb ld Indicates the optimal position of the historical individual, and its corresponding fitness value is the optimal fitness value of the historical individual, Xgb d Indicates the optimal position of the historical group, and its corresponding fitness value is the optimal fitness value of the historical group; Update the historical individual optimal position and the historical group optimal position: use the fitness function to calculate the fitness value of each particle after the update, compare the updated fitness value of each particle with the corresponding historical individual optimal fitness value, if the updated fitness of any particle is greater than the corresponding historical individual optimal fitness value, then update the historical individual optimal fitness value of the particle = the updated fitness of the particle, and update the historical individual optimal position of the particle = the updated position of the particle, otherwise do not update; then take the maximum value of the historical individual optimal fitness values ​​of the current particles as the current optimal fitness value, if the current optimal fitness value is greater than the historical group optimal fitness value, then update the historical group optimal fitness value = the current optimal fitness value, and update the historical group optimal position = the current position of the particle corresponding to the current optimal fitness value, otherwise do not update; It is determined whether the optimal fitness value of the historical group reaches the first set fitness value or whether the number of iterations reaches the maximum number of iterations T. If so, the optimal hyperparameters are obtained based on the particles at the optimal position of the historical group, and the target venous machine learning model is constructed using the optimal hyperparameters. Otherwise, it is determined whether the current number of iterations t is less than the set threshold T1. If yes, the genetic algorithm is used to perform hybridization operations on the updated particles, and the speed and position of each particle in the particle population formed after hybridization are updated again. If no, the speed and position are directly updated again, T / 2≤T1≤T.

6. The vascular lesion risk identification system based on machine learning model according to claim 5, characterized in that: The control host is further used to perform a hybridization operation on the updated particles using a genetic algorithm when the answer is yes, and update the speed and position of each particle in the particle population formed after the hybridization again: arrange the current historical individual optimal fitness values ​​of each particle in descending order, eliminate L1 particles with the lowest order, use the particle with the optimal position in the historical group among the non-eliminated particles as the parent generation, and randomly select L1 particles from the remaining non-eliminated particles as the parent generation, each selected parent generation is hybridized with the parent generation to generate L1 particles, and the current positions of the L1 particles generated by hybridization are respectively used as the positions of the corresponding particles after t iterations, and the current speed of each particle in the L1 particles generated by hybridization is the average speed of the particles on both sides of the particle, so as to obtain the particle population formed after the hybridization, and update the speed and position of each particle in the particle population formed after the hybridization again, L1 is a positive integer and 1≤L1≤L / 2, t=t+1.

7. The vascular disease risk identification system based on machine learning model according to claim 1, characterized in that: The control host is also used to construct a machine learning model for the target artery: the blood oxygen value, pH value, lactate value and pressure value in the blood of C historical artery testers are obtained through a blood oxygen sensor, a pH sensor, a lactate sensor and a pressure sensor respectively; high-resolution structural images of the inner wall of the blood vessels of the C historical artery testers are obtained through an ultrasonic sensor and intravascular ultrasound technology; the structural image of each historical artery tester is processed and then the blood vessel diameter, the thickness of the vascular intima, the thickness of the vascular middle membrane, the thickness of the vascular adventitia and whether there is an arterial plaque in the artery are analyzed; when the presence of an arterial plaque in the artery is analyzed, the composition and morphology of the arterial plaque are identified and the degree of stenosis of the vessel diameter is calculated; and arterial vascular lesion markers of each historical arterial tester are marked, and the arterial vascular lesion markers include a marker for no arterial vascular lesion, a marker for atherosclerosis, a marker for aneurysm, a marker for thromboangiitis obliterans and a marker for arterial dissection; Construct an arterial blood vessel feature set for each historical arterial tester. The arterial blood vessel feature set consists of blood oxygen value, pH value, lactate value, pressure value, blood vessel diameter, blood vessel intima thickness, blood vessel media thickness, blood vessel adventitia thickness, composition and morphology of arterial plaques when arterial plaques exist, and the degree of stenosis of the vessel diameter, as well as corresponding arterial blood vessel lesion markers. Construct an arterial blood vessel sample set. The arterial blood vessel sample set includes the arterial blood vessel feature sets of C historical arterial testers, where C is a positive integer. An arterial vascular machine learning model was constructed based on a convolutional neural network. The particle swarm algorithm and genetic algorithm were used to optimize the hyperparameters of the arterial vascular machine learning model. The arterial vascular sample set was used to train the arterial vascular machine learning model. The optimal hyperparameters were obtained through continuous iteration, and the target arterial vascular machine learning model was constructed using the optimal hyperparameters.

8. The vascular lesion risk identification system based on machine learning model according to claim 7, characterized in that: The control host is also used to build an arterial machine learning model based on a convolutional neural network, use a particle swarm algorithm and a genetic algorithm to optimize the hyperparameters of the arterial machine learning model, and use an arterial sample set to train the arterial machine learning model, obtain the optimal hyperparameters through continuous iteration, and use the optimal hyperparameters to build a target arterial machine learning model: An arterial vascular machine learning model is constructed based on a convolutional neural network, and the hyperparameters of the arterial vascular machine learning model are initialized. The number of particles N in the particle population, the maximum number of iterations S, and the particle search space dimension K corresponding to the hyperparameters are set. Each particle represents a set of hyperparameters, and the speed and position of each particle are initialized. Based on any particle used as a hyperparameter, a corresponding arterial blood vessel machine learning model is constructed, and the arterial blood vessel sample set is used to train the arterial blood vessel machine learning model corresponding to the particle, and a corresponding fitness value is calculated using a fitness function, where the fitness function is a function that is inversely proportional to the mean square error of the model deviation, and the fitness value of each particle is calculated based on this, the maximum fitness value among each particle is used as the optimal fitness, the current position of each particle is used as the individual optimal position of each particle, and the current position of the particle corresponding to the optimal fitness is used as the group optimal position of the particle population; Update the velocity and position of each particle: X nk (s+1)=X nk (s)+V nk (s+1) Among them, V nk (s) and X nk (s) represent the current particle speed and position, V nk (s+1) and X nk (s+1) represents the updated particle speed and position, n represents the nth particle, k represents the kth dimension of the search space, s represents the current iteration number, W(s) represents the adaptive inertia weight, W min2 Indicates the minimum inertia weight, W max2 Indicates the maximum inertia weight, C3 and C4 indicate the acceleration factor, Xpb nk Indicates the optimal position of the historical individual, and its corresponding fitness value is the optimal fitness value of the historical individual, Xgb k Indicates the optimal position of the historical group, and its corresponding fitness value is the optimal fitness value of the historical group; Update the historical individual optimal position and the historical group optimal position: use the fitness function to calculate the fitness value of each particle after the update, compare the updated fitness value of each particle with the corresponding historical individual optimal fitness value, if the updated fitness of any particle is greater than the corresponding historical individual optimal fitness value, then update the historical individual optimal fitness value of the particle = the updated fitness of the particle, and update the historical individual optimal position of the particle = the updated position of the particle, otherwise do not update; then take the maximum value of the historical individual optimal fitness values ​​of the current particles as the current optimal fitness value, if the current optimal fitness value is greater than the historical group optimal fitness value, then update the historical group optimal fitness value = the current optimal fitness value, and update the historical group optimal position = the current position of the particle corresponding to the current optimal fitness value, otherwise do not update; It is determined whether the optimal fitness value of the historical group reaches the second set fitness value or whether the number of iterations reaches the maximum number of iterations S. If so, the optimal hyperparameters are obtained based on the particles at the optimal position of the historical group, and the target arterial vascular machine learning model is constructed using the optimal hyperparameters. Otherwise, it is determined whether the current number of iterations s is less than the set threshold S1. If yes, the genetic algorithm is used to perform hybridization operations on the updated particles, and the speed and position of each particle in the particle population formed after hybridization are updated again. If no, the speed and position are directly updated again, S / 2≤S1≤S.

9. The vascular lesion risk identification system based on machine learning model according to claim 8, characterized in that: The control host is further used to perform a hybridization operation on the updated particles using a genetic algorithm when the answer is yes, and update the speed and position of each particle in the particle population formed after the hybridization again: arrange the current historical individual optimal fitness values ​​of each particle in descending order, eliminate N1 particles with the lowest order, use the particle with the optimal position in the historical group among the non-eliminated particles as the parent generation, and randomly select N1 particles from the remaining non-eliminated particles as the parent generation, each selected parent generation is hybridized with the parent generation to generate N1 particles, the current positions of the N1 particles generated by the hybridization are respectively used as the positions of the corresponding particles after s iterations, the current speed of each particle in the N1 particles generated by the hybridization is the average speed of the particles on both sides of the particle, so as to obtain the particle population formed after the hybridization, and update the speed and position of each particle in the particle population formed after the hybridization again, N1 is a positive integer and 1≤N1≤N / 2, s=s+1.

10. The vascular disease risk identification system based on machine learning model according to claim 1, characterized in that: The control host is also used to calculate the value of vessel diameter stenosis degree = (vessel diameter - arterial plaque cross-sectional area) / vessel diameter.

Images