Cardiovascular lumped parameter model specific parameter estimation method and device based on deep learning

Abstract

According to the method, the cardiovascular lumped parameter model specific parameter estimation method and device based on deep learning are realized. The method comprises the following steps: constructing a non-specific lumped parameter model, randomly sampling a specific parameter group, and setting and operating the non-specific lumped parameter model according to the randomly sampled specific parameter group to obtain an iconography index and arterial pressure; constructing a specific parameter estimation data set by taking the specific parameter group as a label and the iconography index and the arterial pressure as input data, and inputting the specific parameter estimation data set into the constructed specific parameter estimation model for training; and finally, inputting the to-be-tested iconography index and the arterial pressure into the trained specific parameter estimation model for processing to obtain a specific parameter group corresponding to the to-be-tested iconography index and the arterial pressure. According to the method, the inverse process of the non-specific lumped parameter model is realized, the method has the advantages of low calculation cost, high speed and relatively high precision, the specific parameter group can be obtained in real time, and the precision is ensured.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of biomedical engineering, and particularly relates to a cardiovascular lumped parameter model (LPM) specific parameter estimation method and device based on deep learning. BACKGROUND

[0002] The lumped parameter model (LPM) of the human cardiovascular system is a dimension reduction model, which mainly simulates the hemodynamic mechanism of chambers, blood vessels and the like by means of circuit elements. Specifically, a resistor is used to simulate the vascular resistance, a capacitor is used to simulate the vascular elasticity or the elasticity of the heart chamber, and an inductor is used to simulate the blood inertia. The LPM can simulate the hemodynamic phenomena of the real whole body cardiovascular system of a patient, and the accuracy is not lower than the traditional three-dimensional simulation calculation effect. At the same time, the dimension reduction model structure can significantly shorten the time required for traditional simulation calculation. Past researches are often directed to non-specific LPM, and disease simulation, surgery simulation or medical device performance testing are performed by changing the circuit structure or circuit parameters. However, since the non-specific LPM cannot reflect the individual differences of patients, the above tests or simulations are difficult to be truly popularized to individualized clinical applications.

[0003] At present, some researches have carried out related work on LPM specific parameter estimation. The specific parameter estimation of LPM refers to calculating the specific parameter group of LPM that can truly reflect the physiological state of a patient according to the clinical physiological indicators of the patient by means of algorithms or techniques. Some LPM specific researches model the whole blood circulation system (four-chamber heart + body circulation + pulmonary circulation), and simultaneously use the Levenberg-Marquardt algorithm for multi-parameter specific estimation. However, the Levenberg-Marquardt algorithm often significantly depends on the setting of initial values, and the optimization result is extremely easy to fall into a local minimum. Moreover, each iteration of the Levenberg-Marquardt algorithm needs to re-run the LPM, which significantly increases the calculation amount. In addition, most researches do not undergo model identifiability analysis when calculating the specific parameter group by means of the Levenberg-Marquardt algorithm or gradient descent algorithm through multiple iterations. This means that the specific parameter group calculated may not be the only parameter combination that can reflect the specific physiological indicators.

[0004] There are also some works that use physics-informed neural networks (PINN) to calculate the LPM-specific parameter set. However, the complexity of the fluid mechanics control equation embedded in the PINN greatly limits the complexity and completeness of the LPM model. Most LPM models optimized by the PINN algorithm have a structure with only six chambers (left heart, right heart, systemic artery, systemic vein, pulmonary artery, and pulmonary vein). This results in the inability to monitor important imaging indicators such as pulmonary capillary wedge pressure and arterial pressure, limiting the scope of its applicability. SUMMARY

[0005] To solve the problems in the background art, the present application provides a cardiovascular lumped parameter model-specific parameter estimation method and device based on deep learning, which solves the problem that the existing LPM-specific parameter set estimation technology is limited by the initial value setting and the completeness of the model, and ensures the uniqueness of the LPM model parameter set through recognizability analysis.

[0006] The technical solution adopted by the present application comprises:

[0007] I. A cardiovascular lumped parameter model-specific parameter estimation method based on deep learning:

[0008] S1, a non-specific lumped parameter model is constructed by circuit elements.

[0009] S2, a Latin hypercube sampling algorithm is used to randomly sample the pre-set specific parameter set several times, and the non-specific lumped parameter model is set by using the specific parameter set sampled each time, and the imaging indicators and arterial pressure corresponding to the specific parameter set sampled each time are collected by running the non-specific lumped parameter model.

[0010] S3, the specific parameter set sampled each time is taken as a label, and the imaging indicators and arterial pressure corresponding to the specific parameter set are taken as input data, so as to construct a specific parameter estimation data set.

[0011] S4, a specific parameter estimation model is constructed, the specific parameter estimation data set is input into the specific parameter estimation model for training, and a trained specific parameter estimation model is obtained.

[0012] S5, the imaging indicators and arterial pressure to be measured are input into the trained specific parameter estimation model for processing, and the specific parameter set corresponding to the imaging indicators and arterial pressure to be measured is obtained.

[0013] The non-specific lumped parameter model comprises right atrium, right ventricle, left atrium, left ventricle, aorta, systemic artery, systemic capillary, systemic vein, pulmonary artery sinus, pulmonary artery, pulmonary capillary, pulmonary vein and left ventricular assist device simulated by circuit elements. The specific parameter set comprises pulmonary artery sinus resistance, systemic large artery resistance, systemic small artery resistance, pulmonary vein compliance, aortic sinus resistance, systemic small artery compliance, left ventricular elastic coefficient maximum, right ventricular elastic coefficient maximum, cardiac cycle and blood volume. The imaging index and arterial pressure comprise left ventricular flow, left ventricular volume, right ventricular flow, right ventricular volume and arterial pressure.

[0014] The specific parameter estimation model comprises left ventricular end-systolic pressure model, right ventricular end-systolic pressure model, left ventricular end-systolic pressure physical constraint model, right ventricular end-systolic pressure physical constraint model and lumped parameter prediction model; the left ventricular flow and left ventricular volume in the specific parameter estimation data set are input into the left ventricular end-systolic pressure model for processing to obtain left ventricular end-systolic pressure; the right ventricular flow and right ventricular volume in the specific parameter estimation data set are input into the right ventricular end-systolic pressure model for processing to obtain right ventricular end-systolic pressure; the obtained left ventricular end-systolic pressure is input into the left ventricular end-systolic pressure physical constraint model for constraint processing to obtain left ventricular elastic coefficient maximum; the obtained right ventricular end-systolic pressure is input into the right ventricular end-systolic pressure physical constraint model for constraint processing to obtain right ventricular elastic coefficient maximum; the obtained left ventricular elastic coefficient maximum, right ventricular elastic coefficient maximum and arterial pressure in the specific parameter estimation data set are input into the lumped parameter prediction model for processing to obtain predicted pulmonary artery sinus resistance, systemic large artery resistance, systemic small artery resistance, pulmonary vein compliance, aortic sinus resistance, systemic small artery compliance, cardiac cycle and blood volume.

[0015] The left ventricular end-systolic pressure model, right ventricular end-systolic pressure model and lumped parameter prediction model have the same structure, all comprising first convolution layer, first pooling layer, second convolution layer, second pooling layer, third convolution layer, third pooling layer, fourth convolution layer, fourth pooling layer, first full connection layer, second full connection layer, third full connection layer, fourth full connection layer and fifth full connection layer; the input end of the first convolution layer and the input end of the first full connection layer together serve as the input end of the model, the first convolution layer, first pooling layer, second convolution layer, second pooling layer, third convolution layer, third pooling layer, fourth convolution layer and fourth pooling layer are connected in series, the first full connection layer, second full connection layer and third full connection layer are connected in series, the result output by the fourth pooling layer is spliced with the result output by the third full connection layer and then input into the fourth full connection layer for processing, the processed result is input into the fifth full connection layer, and the output end of the fifth full connection layer serves as the output end of the model.

[0016] the left ventricular volume in the specific parameter estimation dataset is input to the input end of the first fully connected layer of the left ventricular end-systolic pressure model; the fifth fully connected layer of the left ventricular end-systolic pressure model outputs the left ventricular end-systolic pressure, which is input to the input end of the left ventricular end-systolic pressure physical constraint model.

[0017] the right ventricular volume in the specific parameter estimation dataset is input to the input end of the first fully connected layer of the left ventricular end-systolic pressure model; the fifth fully connected layer of the left ventricular end-systolic pressure model outputs the left ventricular end-systolic pressure, which is input to the input end of the left ventricular end-systolic pressure physical constraint model.

[0018] the arterial pressure in the specific parameter estimation dataset is input to the input end of the first convolutional layer of the lumped parameter prediction model; the left and right ventricular end-systolic pressure physical constraint models output the maximum left ventricular elastance and the maximum right ventricular elastance, which are input to the input end of the first fully connected layer of the left ventricular end-systolic pressure model; the fifth fully connected layer of the lumped parameter prediction model outputs the predicted pulmonary sinus resistance, systemic arterial resistance, systemic arteriolar resistance, pulmonary venous compliance, aortic sinus resistance, systemic arteriolar compliance, cardiac cycle, and blood volume.

[0019] The left and right ventricular end-systolic pressure physical constraint models are set according to the following formula:

[0020]

[0021] wherein, and respectively represent the maximum left ventricular elastance and the maximum right ventricular elastance; and respectively represent the left ventricular end-systolic pressure and the right ventricular end-systolic pressure; and respectively represent the left ventricular volume and the right ventricular volume; and respectively represent the left ventricular volume at zero pressure and the right ventricular volume at zero pressure.

[0022] II. An apparatus for implementing the specific parameter estimation method of the cardiovascular lumped parameter model:

[0023] a first model module for constructing and storing a non-specific lumped parameter model;

[0024] The input module is configured to receive a plurality of sets of specific parameters sampled from the first model module and corresponding imaging indexes and arterial pressure;

[0025] The second model module is configured to construct and store a specific parameter estimation model, and input the imaging indexes and arterial pressure into the second model module to obtain a predicted set of specific parameters.

[0026] The output module is configured to output the predicted set of specific parameters from the second model module or visualize the predicted set of specific parameters.

[0027] Three, a computer device comprising a memory and a processor, the memory stores a computer program, and the processor implements the steps of the above method when executing the computer program.

[0028] Four, a computer readable storage medium having a computer program stored thereon, the computer program is executed by a processor to implement the steps of the above method.

[0029] The beneficial effects of the present application are:

[0030] 1. The method of the present application realizes the inverse process of the non-specific lumped parameter model, and by inputting the imaging indexes and arterial pressure into the trained specific parameter estimation model for processing, the corresponding set of specific parameters of the imaging indexes and arterial pressure can be obtained.

[0031] 2. The method of the present application realizes the beneficial effect of obtaining hemodynamic indexes by non-invasive method by inputting the non-invasive ultrasound data (ventricular flow and volume) and arterial pressure data into the trained specific parameter estimation model for processing to obtain a set of specific parameters, and then inputting the set of specific parameters into the non-specific lumped parameter model again to obtain hemodynamic indexes.

[0032] 3. The method of the present application has the advantages of low calculation cost, fast speed and high precision, and can obtain a set of specific parameters in real time while ensuring the accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 It is a non-specific lumped parameter model diagram of the method of the present application.

[0034] Figure 2 It is a pressure waveform diagram of left atrium, left ventricle and aorta (non-heart failure working condition) in the embodiment.

[0035] Figure 3 It is a left heart myocardial perfusion pressure waveform diagram (non-heart failure working condition) in the embodiment.

[0036] Figure 4Figure of aortic flow waveform in the embodiment (no heart failure condition).

[0037] Figure 5 Figure of left ventricular pressure-volume loop in the embodiment as a function of LVAD rotational speed (heart failure condition).

[0038] Figure 6 Figure of TSF results in the embodiment.

[0039] Figure 7 Figure of GSF results in the embodiment.

[0040] Figure 8 Figure of the structure of the specific parameter estimation model of the method of the present application.

[0041] Figure 9 Figure of the structure of the left ventricular end-systolic pressure model, the right ventricular end-systolic pressure model and the lumped parameter prediction model of the method of the present application.

[0042] Figure 10 Figure of the comparison of the predicted values and the true values of the hemodynamic indexes obtained in the first group in the embodiment.

[0043] Figure 11 Figure of the comparison of the predicted values and the true values of the hemodynamic indexes obtained in the second group in the embodiment. DETAILED DESCRIPTION

[0044] The present application will be described in more detail below with reference to the drawings and embodiments, but the present application is not limited thereto, and those skilled in the art can make several improvements and refinements without departing from the principles of the present application, and these improvements and refinements are also considered to be within the scope of protection of the present application. The contents not described in detail in the present specification are the prior art known to those skilled in the art.

[0045] The specific parameter estimation method of the cardiovascular lumped parameter model of the present embodiment comprises the following steps:

[0046] S1, constructing a non-specific lumped parameter model (non-specific LPM model) through circuit elements.

[0047] As shown in Figure 1 , the non-specific lumped parameter model comprises right atrium (RA), right ventricle (RV), left atrium (LA), left ventricle (LV), aorta (AO), systemic arterial (SAT), systemic capillary (SAR), systemic venous (SVN), pulmonary arterial sinus (PAS), pulmonary artery (PAT), pulmonary capillary (PCP), pulmonary venous (PVN) and left ventricular assist device (LVAD) simulated by circuit elements.

[0048] As shown in Figure 1As shown, the non-specific lumped parameter model of the present application includes a heart section, a systemic circulation section, a pulmonary circulation section and a LVAD, which are simulated by circuit elements.

[0049] The heart section consists of four chambers, namely RA, RV, LA and LV; each chamber is represented by a capacitor, and these capacitors are connected by diodes to simulate the one-way flow characteristics of the heart valves; RA is located at the upper right corner of the heart and is connected to RV through a diode, and the RA chamber is simulated by a capacitor to simulate its filling and contraction process; RV is located at the lower right corner of the heart and is connected to PAS through a diode, and the RV chamber is simulated by a capacitor to simulate its filling and contraction process; LA is located at the upper left corner of the heart and is connected to LV through a diode, and the LA chamber is simulated by a capacitor to simulate its filling and contraction process; LV is located at the lower left corner of the heart and is connected to AO through a diode, and the LV chamber is simulated by a capacitor to simulate its filling and contraction process.

[0050] The systemic circulation section includes AO, SAT, S AR and SVN; AO flows out from LV, and the resistance and inertia of blood flow are simulated by a resistor and an inductor, and the elasticity of the blood vessel wall is simulated by a capacitor; SAT branches from AO, and the flow of blood in the systemic artery is simulated by a resistor and an inductor, and the elasticity of the blood vessel wall is simulated by a capacitor; S AR branches from SAT, and the flow and exchange of blood in the capillary are simulated by a resistor and a capacitor, and the elasticity of the capillary wall is simulated by a capacitor; SVN converges from S AR, and the process of blood returning to RA is simulated by a resistor and an inductor, and the elasticity of the vein wall is simulated by a capacitor.

[0051] The pulmonary circulation section includes PAS, PAT, PCP and PVN; PAS flows out from RV, and the resistance and inertia of blood flow are simulated by a resistor and an inductor, and the elasticity of the blood vessel wall is simulated by a capacitor; PAT branches from PAS, and the flow of blood in the pulmonary artery is simulated by a resistor and an inductor, and the elasticity of the blood vessel wall is simulated by a capacitor; PCP branches from PAT, and the flow and gas exchange of blood in the pulmonary capillary are simulated by a resistor and a capacitor, and the elasticity of the capillary wall is simulated by a capacitor; PVN converges from PCP, and the process of blood returning to LA is simulated by a resistor and an inductor, and the elasticity of the vein wall is simulated by a capacitor.

[0052] The LVAD is located on the right side of LV, and the resistance and inertia effects encountered by blood entering and leaving the LVAD are simulated by a resistor and an inductor; the LVAD pump is equivalent to a voltage source, representing a certain pressure difference before and after the pump to push the blood flow. In addition, the resistance and blood inertia of the LVAD inflow section and outflow section are also simulated by a resistor and an inductor, respectively.

[0053] The parameters of the non-specific lumped parameter model are fine-tuned according to the reference parameters in the reference Fu Y, Qiao A, Yang Y and Fan X (2020) Numerical Simulation of the Effect of Pulmonary Vascular Resistance on the Hemodynamics of Reoperation After Failure of One and a Half Ventricle Repair. Front. Physiol. 11:207. doi: 10.3389 / fphys.2020.00207. For the no heart failure working condition (LVAD not started), the pressure curves of the left atrium, left ventricle and aorta simulated by the non-specific lumped parameter model constructed by the present application are as shown in Figure 2 , the aortic flow curve is as shown in Figure 3 , and the left ventricular myocardial perfusion pressure is as shown in Figure 4 . These pressure and flow waveforms are consistent with the normal hemodynamic curves of no heart failure.

[0054] Heart failure is usually manifested as a decrease in left ventricular contraction ability. The simulation of the heart failure working condition is realized by reducing the maximum value of the left ventricular elastic coefficient, i.e. . At the same time, the LVAD is started and the speed is adjusted to evaluate the coupling effect of the LVAD and the cardiovascular circulation system. Figure 5 The effects of different LVAD speeds on the pressure-volume loop of the left ventricle of heart failure are shown: as the speed increases, the pressure-volume loop gradually moves to the lower left corner, and the shape gradually changes from a parallelogram to a triangle.

[0055] It is fully proved by the above operations that the non-specific lumped parameter model constructed in the present embodiment fully meets the needs of the subsequent steps of the method of the present application.

[0056] S2, the Latin hypercube sampling algorithm is used to randomly sample the pre-set specific parameter groups for several times, and the specific parameter group of each random sampling is used to set the non-specific lumped parameter model, and the non-specific lumped parameter model is run to collect the imaging indexes and arterial pressure corresponding to each sampled specific parameter group.

[0057] The specific parameter group includes pulmonary artery sinus resistance (Rpas), systemic arterial resistance (Rsar), systemic arteriolar resistance (Rsat), pulmonary vein compliance (Cpvn), aortic sinus resistance (Rao), systemic arteriolar compliance (Csat), maximum left ventricular elastic coefficient, maximum right ventricular elastic coefficient, cardiac cycle and blood volume.

[0058] Imaging parameters and arterial pressure include left ventricular flow waveform, left ventricular volume (end-systole and end-diastole), right ventricular flow waveform, right ventricular volume (end-systole and end-diastole), and arterial pressure.

[0059] S3. Use the specific parameter set from each sampling as a label, and use the corresponding imaging indicators and arterial pressure as input data to construct a specific parameter estimation dataset.

[0060] S4. Construct a specific parameter estimation model. Input the specific parameter estimation dataset into the specific parameter estimation model for training to obtain a trained specific parameter estimation model.

[0061] like Figure 8 As shown, the specific parameter estimation models include the left ventricular end-systolic pressure model (neural network A), the right ventricular end-systolic pressure model (neural network B), the left ventricular end-systolic pressure physical constraint model, the right ventricular end-systolic pressure physical constraint model, and the lumped parameter prediction model (neural network C).

[0062] Left ventricular flow and volume from the specific parameter estimation dataset are input into the left ventricular end-systolic pressure model for processing to obtain left ventricular end-systolic pressure. Right ventricular flow and volume from the specific parameter estimation dataset are input into the right ventricular end-systolic pressure model for processing to obtain right ventricular end-systolic pressure. The obtained left ventricular end-systolic pressure is input into the left ventricular end-systolic pressure physical constraint model and, after constraint processing, the maximum left ventricular elasticity coefficient is obtained. The obtained right ventricular end-systolic pressure is input into the right ventricular end-systolic pressure physical constraint model and, after constraint processing, the maximum right ventricular elasticity coefficient is obtained. The obtained maximum left ventricular elasticity coefficient, maximum right ventricular elasticity coefficient, and arterial pressure from the specific parameter estimation dataset are input into the lumped parameter prediction model for processing to obtain predicted pulmonary sinus resistance, large systemic artery resistance, small systemic artery resistance, pulmonary vein compliance, aortic sinus resistance, small systemic artery compliance, cardiac cycle, and blood volume.

[0063] Pulmonary sinus resistance, large systemic artery resistance, small systemic artery resistance, pulmonary vein compliance, aortic sinus resistance, small systemic artery compliance, cardiac cycle, and blood volume all contribute to this. Figure 8 and Figure 9 LPM-specific parameters in the model.

[0064] like Figure 9 As shown, the left figure is a structural diagram of the left ventricular end-systolic pressure model and the right ventricular end-systolic pressure model, and the right figure is a structural diagram of the lumped parameter prediction model.

[0065] The left ventricular end-systolic pressure model, the right ventricular end-systolic pressure model and the lumped parameter prediction model have the same structure, and all include a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, a fourth pooling layer, a first full connection layer, a second full connection layer, a third full connection layer, a fourth full connection layer and a fifth full connection layer. The input end of the first convolutional layer and the input end of the first full connection layer together serve as the input end of the model, the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the third pooling layer, the fourth convolutional layer and the fourth pooling layer are connected in series, the first full connection layer, the second full connection layer and the third full connection layer are connected in series, the result output by the fourth pooling layer and the result output by the third full connection layer are spliced and input into the fourth full connection layer for processing, the processed result is input into the fifth full connection layer, and the output end of the fifth full connection layer serves as the output end of the model.

[0066] The left ventricular flow in the specific parameter estimation dataset is input into the input end of the first convolutional layer of the left ventricular end-systolic pressure model; the left ventricular volume in the specific parameter estimation dataset is input into the input end of the first full connection layer of the left ventricular end-systolic pressure model; the fifth full connection layer of the left ventricular end-systolic pressure model outputs the left ventricular end-systolic pressure, which is input into the input end of the left ventricular end-systolic pressure physical constraint model.

[0067] The right ventricular flow in the specific parameter estimation dataset is input into the input end of the first convolutional layer of the right ventricular end-systolic pressure model; the right ventricular volume in the specific parameter estimation dataset is input into the input end of the first full connection layer of the left ventricular end-systolic pressure model; the fifth full connection layer of the right ventricular end-systolic pressure model outputs the right ventricular end-systolic pressure; and the right ventricular end-systolic pressure is input into the input end of the right ventricular end-systolic pressure physical constraint model.

[0068] The arterial pressure in the specific parameter estimation dataset is input into the input end of the first convolutional layer of the lumped parameter prediction model; the left ventricular elastic coefficient maximum and the right ventricular elastic coefficient maximum output by the left and right ventricular end-systolic pressure physical constraint model are input together into the input end of the first full connection layer of the left ventricular end-systolic pressure model; and the fifth full connection layer of the lumped parameter prediction model outputs the predicted pulmonary sinus resistance, systemic arterial resistance, systemic arteriolar resistance, pulmonary vein compliance, aortic sinus resistance, systemic arteriolar compliance, a cardiac cycle and a blood volume.

[0069] The left and right ventricular end-systolic pressure physical constraint models are set according to the following formula:

[0070]

[0071] wherein, and respectively represent the maximum left ventricular elastic coefficient and the maximum right ventricular elastic coefficient; and respectively represent the left ventricular end-systolic pressure and the right ventricular end-systolic pressure; and respectively represent the left ventricular volume and the right ventricular volume; and respectively represent the left ventricular volume at zero pressure and the right ventricular volume at zero pressure.

[0072] The specific parameter estimation model uses RMSE as the loss function of the model, and the loss function is set according to the following formula:

[0073]

[0074] Wherein, RMSE represents the loss function; N represents the total number of samples in the specific parameter estimation data set; i represents the index; and respectively represent the true value and the predicted value of the specific parameter group corresponding to the i th sample.

[0075] S5, input the to-be-measured imaging index and arterial pressure into the trained specific parameter estimation model for processing to obtain the specific parameter group corresponding to the to-be-measured imaging index and arterial pressure.

[0076] Further, according to the specific parameter group corresponding to the to-be-measured imaging index and arterial pressure, the non-specific lumped parameter model constructed in step S1 is set and run to obtain the hemodynamic index (application).

[0077] The hemodynamic index includes arterial pressure, pulmonary artery pressure and capillary pressure.

[0078] The method of the present application realizes the beneficial effect of obtaining the hemodynamic index by a non-invasive method, by inputting the non-invasive ultrasound data (ventricular flow and volume) and arterial pressure data into the trained specific parameter estimation model for processing to obtain the specific parameter group, and then inputting the specific parameter group into the non-specific lumped parameter model and running to obtain the hemodynamic index.

[0079] The to-be-measured imaging index and arterial pressure include to-be-measured left ventricular flow, left ventricular volume, right ventricular flow, right ventricular volume and arterial pressure.

[0080] The specific parameter group corresponding to the to-be-measured imaging index and arterial pressure includes the predicted specific parameter group, including pulmonary artery sinus resistance, systemic arterial resistance, systemic arteriolar resistance, pulmonary vein compliance, aortic sinus resistance, systemic arteriolar compliance, maximum left ventricular elastic coefficient, maximum right ventricular elastic coefficient, cardiac cycle and blood volume.

[0081] The device of the cardiovascular lumped parameter model specific parameter estimation method of the embodiment comprises:

[0082] A first model module is configured to construct and store a non-specific lumped parameter model.

[0083] An input module is configured to receive a specific parameter group sampled several times from the first model module and corresponding imaging indexes and arterial pressure.

[0084] A second model module is configured to construct and store a specific parameter estimation model and input the imaging indexes and arterial pressure into the second model module to obtain a predicted specific parameter group.

[0085] An output module is configured to output the predicted specific parameter group from the second model module or visualize the predicted specific parameter group.

[0086] In order to present the uniqueness and importance of the preset specific parameter group of the application, the following experiments are further performed in the embodiment.

[0087] The parameters of the LPM specific parameter group required to be analyzed by the method of the application are of three types: (1) systemic circulation and pulmonary circulation parameters required to be specifically estimated; (2) heart parameters required to be specifically estimated; and (3) clinically directly non-invasively obtained but still required to be analyzed as variables. The (2) type is determined as the maximum elastic coefficients of the left and right ventricles in the application. Because the difference between heart failure and normal myocardial contractility is mainly reflected in these two parameters. The (3) type is determined as the cardiac cycle and blood volume in the application. Because the individual difference of the cardiac cycle is very significant and is crucial for the evaluation of hemodynamics. The individual difference of the blood volume is also very significant and is mainly related to the height and weight. The (1) type parameters are determined by the following contents.

[0088] The (1) type parameter estimation problem of the LPM is essentially an inverse problem, that is, the model parameters are estimated from the experimental observation values (non-invasive ultrasound data and cuff pressure detector data are used in the present study). If the model complexity is high, there will be a parameter non-identifiability problem, that is, a combination of multiple model parameters will lead to the same experimental observation value. The following three numerical methods can be used to quantitatively analyze the parameter identifiability:

[0089] (1) Traditional sensitivity function (TSF): evaluating the sensitivity of a certain model parameter to a certain observation value.

[0090] (2) Fisher information matrix (FIM): based on TSF, evaluating how accurately the model parameters can be estimated from the observation data.

[0091] (3) Generalized sensitivity function (GSF): based on FIM, assess how much the multiple model parameters affect the observation value at different times, which can be used to simplify the model or optimize the experimental design.

[0092] The three methods are in a progressive relationship. The specific process is often: first calculate the TSF to determine the top few parameters that are most sensitive to the observation value; then calculate the FIM and GSF to check whether there are problems of parameter non-identifiability among the parameters. According to this process, the TSF is first calculated, and the top 6 parameters with the largest values are taken; then the FIM and GSF of the 6 parameters are calculated, and the GSF curve is drawn to check whether there are problems of parameter non-identifiability among the parameters. For example, Figure 6 The TSF results are shown in the following table, Figure 7 The GSF curves of the 6 parameters are shown in the following figure, where R represents resistance, and C represents capacitance.

[0093] From the above, Figure 7 It can be seen that the 6 parameters do not have the same or opposite trends at the same phase, and are 6 independent curves that are complementary to each other. This indicates that the contributions of the 6 parameters to the observation data are independent of each other, and there is no problem of parameter non-identifiability. Therefore, the pulmonary sinus resistance, the systemic arterial resistance, the systemic arteriolar resistance, the pulmonary vein compliance, the aortic sinus resistance, and the systemic arteriolar compliance are selected.

[0094] In order to highlight the beneficial effects of the method of the present application, the following experiment is further carried out in this embodiment:

[0095] In this embodiment, the Latin hypercube sampling algorithm is again used to randomly sample the pre-set specificity parameter group twice to obtain two groups of specificity parameter groups for testing,

[0096] The two groups of specificity parameter groups for testing are used to set the non-specific lumped parameter model, and the non-specific lumped parameter model is run to collect the image indicators and hemodynamic indicators (true value GT) corresponding to each group of specificity parameters for testing.

[0097] The obtained image indicators and arterial pressure are input into the trained specificity parameter estimation model for processing to obtain the processed specificity parameter group. The processed specificity parameter group is used to set the specificity lumped parameter model and run to obtain the processed hemodynamic indicators (predicted value NN).

[0098] The two processed hemodynamic indicators (predicted value NN) are compared with the corresponding hemodynamic indicators (true value GT), respectively as Figure 10 and Figure 11As shown in the figure, wherein, AP is arterial pressure, PAP is pulmonary artery pressure, PCWP is capillary pressure. Here, the hemodynamic indicators are only selected as arterial pressure, pulmonary artery pressure and capillary pressure.

[0099] The method realizes the inverse process of the non-specific lumped parameter model, and the specific parameter group corresponding to the imaging index and the arterial pressure can be obtained by inputting the imaging index and the arterial pressure into the trained specific parameter estimation model for processing.

[0100] The above examples are only preferred embodiments for fully illustrating the present application, and the protection scope of the present application is not limited thereto. Any equivalent replacement or transformation made by the skilled in the art on the basis of the present application is within the protection scope of the present application. The protection scope of the present application is subject to the claims.

Claims

1. A deep learning-based cardiovascular lumped parameter model-specific parameter estimation method, characterized by, The method comprises the following steps: S1, constructing a non-specific lumped parameter model by circuit elements; S2, using a Latin hypercube sampling algorithm to randomly sample a plurality of specific parameter groups, using each randomly sampled specific parameter group to set the non-specific lumped parameter model, and running the non-specific lumped parameter model to collect the imaging indexes and arterial pressure corresponding to each sampled specific parameter group; S3, taking each sampled specific parameter group as a label and the imaging indexes and arterial pressure corresponding to the specific parameter group as input data, thereby constructing a specific parameter estimation dataset; S4, constructing a specific parameter estimation model, inputting the specific parameter estimation dataset into the specific parameter estimation model for training, and obtaining a trained specific parameter estimation model; S5, inputting the imaging indexes and arterial pressure to be measured into the trained specific parameter estimation model for processing, and obtaining the specific parameter group corresponding to the imaging indexes and arterial pressure to be measured.

2. The cardiovascular lumped parameter model specific parameter estimation method according to claim 1, wherein: the non-specific lumped parameter model comprises right atrium, right ventricle, left atrium, left ventricle, aorta, systemic artery, systemic capillary, systemic vein, pulmonary artery sinus, pulmonary artery, pulmonary capillary, pulmonary vein and left ventricular assist device simulated by circuit elements.

3. The cardiovascular lumped parameter model specific parameter estimation method according to claim 1, wherein: the specific parameter group comprises pulmonary artery sinus resistance, systemic aorta resistance, systemic arteriole resistance, pulmonary vein compliance, aortic sinus resistance, systemic arteriole compliance, left ventricular elastic coefficient maximum, right ventricular elastic coefficient maximum, cardiac cycle and blood volume; and the imaging indexes and arterial pressure comprise left ventricular flow, left ventricular volume, right ventricular flow, right ventricular volume and arterial pressure.

4. The cardiovascular lumped parameter model specific parameter estimation method according to claim 1, wherein: the specific parameter estimation model comprises left ventricular end-systolic pressure model, right ventricular end-systolic pressure model, left ventricular end-systolic pressure physical constraint model, right ventricular end-systolic pressure physical constraint model and lumped parameter prediction model; the left ventricular flow and left ventricular volume in the specific parameter estimation dataset are input into the left ventricular end-systolic pressure model for processing, and the left ventricular end-systolic pressure is obtained. The right ventricular flow and the right ventricular volume in the specificity parameter estimation dataset are input into the right ventricular end-systolic pressure model for processing to obtain the right ventricular end-systolic pressure; the obtained left ventricular end-systolic pressure is input into the left ventricular end-systolic pressure physical constraint model for constraint processing to obtain the left ventricular elasticity coefficient maximum value; the obtained right ventricular end-systolic pressure is input into the right ventricular end-systolic pressure physical constraint model for constraint processing to obtain the right ventricular elasticity coefficient maximum value; the obtained left ventricular elasticity coefficient maximum value, the right ventricular elasticity coefficient maximum value and the arterial pressure in the specificity parameter estimation dataset are input into the lumped parameter prediction model for processing to obtain the predicted pulmonary sinus resistance, systemic arterial resistance, systemic arteriolar resistance, pulmonary vein compliance, aortic sinus resistance, systemic arteriolar compliance, a cardiac cycle and a blood volume.

5. The cardiovascular lumped parameter model specificity parameter estimation method according to claim 4, characterized in that: the left ventricular end-systolic pressure model, the right ventricular end-systolic pressure model and the lumped parameter prediction model have the same structure and each include a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a third pooling layer, a fourth convolutional layer, a fourth pooling layer, a first full connection layer, a second full connection layer, a third full connection layer, a fourth full connection layer and a fifth full connection layer; the input end of the first convolutional layer and the input end of the first full connection layer together serve as the input end of the model, the first convolutional layer, the first pooling layer, the second convolutional layer, the second pooling layer, the third convolutional layer, the third pooling layer, the fourth convolutional layer and the fourth pooling layer are connected in series, the first full connection layer, the second full connection layer and the third full connection layer are connected in series, the result output by the fourth pooling layer is spliced with the result output by the third full connection layer and then input into the fourth full connection layer for processing, the processed result is input into the fifth full connection layer, and the output end of the fifth full connection layer serves as the output end of the model.

6. The cardiovascular lumped parameter model specificity parameter estimation method according to claim 5, characterized in that: the left ventricular flow in the specificity parameter estimation dataset is input into the input end of the first convolutional layer of the left ventricular end-systolic pressure model; the left ventricular volume in the specificity parameter estimation dataset is input into the input end of the first full connection layer of the left ventricular end-systolic pressure model; the fifth full connection layer of the left ventricular end-systolic pressure model outputs the left ventricular end-systolic pressure, and the left ventricular end-systolic pressure is input into the input end of the left ventricular end-systolic pressure physical constraint model; the right ventricular flow in the specificity parameter estimation dataset is input into the input end of the first convolutional layer of the right ventricular end-systolic pressure model; the right ventricular volume in the specificity parameter estimation dataset is input into the input end of the first full connection layer of the left ventricular end-systolic pressure model; the fifth full connection layer of the right ventricular end-systolic pressure model outputs the right ventricular end-systolic pressure; and the right ventricular end-systolic pressure is input into the input end of the right ventricular end-systolic pressure physical constraint model; The arterial pressure in the specific parameter estimation dataset is input to the input end of the first convolutional layer of the lumped parameter prediction model; the left and right ventricular end-systolic pressure physical constraint model outputs the maximum left ventricular elastic coefficient and the maximum right ventricular elastic coefficient, which are input to the input end of the first fully connected layer of the left ventricular end-systolic pressure model; the fifth fully connected layer of the lumped parameter prediction model outputs the predicted pulmonary sinus resistance, systemic arterial resistance, systemic arteriolar resistance, pulmonary vein compliance, aortic sinus resistance, systemic arteriolar compliance, a cardiac cycle, and a blood volume.

7. The cardiovascular lumped parameter model specific parameter estimation method according to claim 4, characterized in that: The left and right ventricular end-systolic pressure physical constraint model is set according to the following formula: wherein, and Ea and Eb represent the maximum left and right ventricular elastance, respectively; and Pd and Pr represent the end-systolic pressure of the left and right ventricle, respectively; and Vl and Vr represent the left and right ventricular volume, respectively; and Vzo and Vzr represent the left and right ventricular volume at zero pressure, respectively.

8. An apparatus for implementing the method of estimating model-specific parameters of a cardiovascular lumped parameter model according to any one of claims 1 to 7, characterized in that It comprises: A first model module for constructing and storing a non-specific lumped parameter model; An input module for receiving a specific parameter group sampled several times from the first model module and corresponding image indicators and arterial pressure; A second model module for constructing and storing a specific parameter estimation model and inputting the image indicators and arterial pressure to the second model module to obtain a predicted specific parameter group; An output module for outputting the predicted specific parameter group from the second model module or visualizing the predicted specific parameter group. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is characterized in that: The processor executes the computer program to implement the steps of the method of any one of claims 1 to 7.

10. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 7.

Images