Method for identifying mitral regurgitation spectrum envelope and measuring left atrial pressure

Abstract

The present invention discloses a method for identifying the envelope of the mitral regurgitation spectrum and a method for measuring left atrial pressure, belonging to the field of curve correction technology. The method includes building a BCD-Unet deep learning model to complete the preliminary envelope of the Doppler spectrum curve; then post-processing the BCD-Unet deep learning model, designing a curve correction module to deal with the problem of curve misalignment, and designing a curve filling module to deal with the problem of curve breakage caused by the failure of some curves to be detected. In order to better integrate the information generated by the filling curve and the correction curve, a curve fusion module is also designed to further fuse the features of the two curves, combine the advantages of the two curves and obtain the final edge curve to meet the requirements of accurate measurement; iteratively train the model using data collected by the hospital, thereby continuously optimizing the model. The trained deep learning model has the ability to automatically track the envelope.

Description

Technical Field

[0001] The present invention relates to the technical field of curve correction, and in particular to a method for identifying a mitral valve regurgitation spectrum envelope and a method for measuring left atrial pressure. Background Art

[0002] Left atrial pressure (LAP), a measure of left ventricular preload, is a highly sought-after hemodynamic parameter in clinical practice and a key parameter for adjusting treatment strategies. Currently, left atrial pressure is measured using a catheter, which requires direct insertion of a catheter into the left atrium. In clinical practice, pulmonary artery wedge pressure is often used instead of LAP, but this is still an invasive test and cannot be performed routinely and repeatedly. Currently, clinical assessment of left atrial pressure changes can only be based on central venous pressure, pulmonary artery systolic pressure, and clinical experience. This method cannot accurately diagnose the severity of heart disease or guide precise clinical treatment. Over the years, researchers have explored various noninvasive methods for assessing LAP, such as the "Doppler plus sphygmomanometer" method. However, these methods have not been effectively implemented in clinical practice due to limitations such as high error rates and complexity. The development of a noninvasive, convenient, and accurate method for quantitatively measuring left atrial pressure would undoubtedly advance the precision diagnosis and treatment of heart disease and possess significant clinical value.

[0003] In a study using the mitral regurgitation spectrum to estimate the left ventricular relaxation time constant (τ) (Sun Jing. Animal Experimental Study on Estimating the Left Ventricular Relaxation Time Constant Using Continuous Doppler Spectra of Mitral Regurgitation [D]. PLA General Hospital; PLA Medical College; PLA Continuing Medical School; PLA General Hospital, 2012), a new "two-variable linear equation system" was derived using mathematical methods using three classic formulas. The system contains two variables, τ and LAP. By substituting two time points of the mitral regurgitation spectrum into the equation system and eliminating LAP, τ can be calculated, which has been experimentally verified. Conversely, this "two-variable linear equation system" should also be able to estimate LAP. This method was used to estimate LAP, and a series of animal and clinical validation experiments were conducted using catheterization as the "gold standard" to verify the feasibility of the method.

[0004] However, the experiment found that precise identification of the edges of the mitral regurgitation spectrum is key to accurate LAP measurement. Using Doppler ultrasound to measure left atrial pressure still requires physicians with extensive practical experience to accurately analyze and calculate the spectrum. This undoubtedly raises the technical barriers to using this method, making it difficult to promote its use in the medical field. Furthermore, manual calculations based on images can lead to miscalculations or errors, requiring physicians to maintain a high level of concentration at all times, which is also a test of medical staff's mental state. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for identifying the envelope of the mitral regurgitation spectrum. The method is based on deep learning technology and automatically extracts key feature point data from the continuous Doppler spectrum image of mitral regurgitation to analyze and calculate the left atrial pressure value. At the same time, the patient's information and related data are stored to achieve the management and maintenance functions of the patient information and patient images. A visual graphical interface for desktop applications used daily by doctors is designed and implemented to provide an interactive method that is more in line with the doctors' daily use. The various values of the patients are used to provide doctors with a more intuitive analysis of the patient's condition, thereby reducing the doctors' mastery of unnecessary skills and greatly improving the doctors' work efficiency. At the same time, it provides a practical application for the use of the Doppler spectrum method to quantitatively measure left atrial pressure as a routine left atrial pressure monitoring method.

[0006] To achieve the above object, the technical solutions adopted by the present invention are as follows:

[0007] A method for identifying the envelope of a mitral regurgitation spectrum is provided. The method is to identify the envelope of a mitral regurgitation continuous Doppler spectrum. The method specifically comprises the following steps:

[0008] (1) Build a deep learning network model to perform edge detection on the mitral regurgitation spectrum; since the edges predicted by the network model are misaligned and broken, a post-processing module is used to correct the edges; the post-processing module includes a curve correction module, a curve fusion module, and a curve fusion module, wherein: the curve correction module is used to deal with the curve misalignment problem, the curve filling module is used to deal with the curve breakage problem, and the curve fusion module is used to further fuse the features of the two curves (the curve processed by the curve correction module and the curve processed by the curve filling module), combining the advantages of the two curves to obtain a final continuous and smooth edge curve;

[0009] (2) iteratively training the model described in step (1) using data collected by the hospital, and the trained deep learning model has the ability to automatically track the envelope;

[0010] (3) Using the model trained in step (2), the mitral regurgitation spectrum is used as input, and the output is the edge curve (envelope) and the left atrial pressure value.

[0011] In the above step (1), the network model is a BCD-Unet deep learning model, and is trained using a cross entropy loss function, which is shown in formula (1):

[0012]

[0013] In formula (1), y i Indicates whether the pixel i belongs to the pixel of the envelope curve, if it does, it is 1, otherwise it is 0; p i is the probability that pixel i is predicted to belong to the envelope curve pixel; N is the number of pixels.

[0014] The training process of BCD-Unet is as follows: after an image is input to the model, the model first performs a convolution operation on the image to extract features. The resulting feature map is then downsampled, reducing its size to half, allowing the convolution operation to extract features over a larger image range. After several rounds of convolution-downsampling operations, the feature map is gradually upsampled and convolved to restore it to its original image size. A loss function is then used to evaluate the difference between each pixel in the model's output and the true label, and backpropagation is also used to adjust the model parameters. Finally, the model is trained and has the ability to roughly detect curves in the spectrogram. During the experiment, the training parameters of the BCD-Unet model were set to epochs = 50 and batch size = 8, and the Adam optimizer with a learning rate of 0.004 was used to accelerate the training process.

[0015] In the above step (1), the curve correction module consists of two parts: curve fitting and one-dimensional unet (1D-Unet). The purpose of curve fitting is to obtain a continuous and smooth curve using polynomial fitting, while 1D-Unet corrects the fitted curve through deep regression to make it closer to the edge curve formed by the doctor's manual marking of edge key points; the curve processed by the curve correction module can effectively solve the problem of curve misalignment. The input of 1D-Unet is the fitted curve, and the output is the corrected curve. These curves are all stored in the form of a one-dimensional array.

[0016] In the above step (1), the curve correction module is trained using the L2 loss function, which is shown in formula (2):

[0017]

[0018] In formula (2), represents the true value of point i, represents the curve correction module's predicted value for point i, and N represents the number of elements in the array. The L2 loss function is a commonly used loss function in regression tasks, with advantages such as fast computation and convenient differentiation. The training process of the curve correction module is similar to that of BCD-Unet, still using backpropagation to continuously adjust the model by evaluating the difference between the model output and the true value. The training parameters of the curve correction module are set to: epochs = 50, batch_size = 16.

[0019] In step (1) above, the curve filling module processes the curve using a one-dimensional generative adversarial network (1D-GAN), training it in a discriminator-generator adversarial manner. After training, the curve filling module can repair the missing parts of the curve, thereby obtaining a smooth filled curve. The input of 1D-GAS is the curve with the missing part, and the output is the repaired curve, which is also stored in the form of a one-dimensional array.

[0020] The curve filling module is based on the idea of generative adversarial networks and uses two loss functions to train the discriminator and generator respectively. The loss functions of the discriminator and generator are shown in formula (3) and formula (4) respectively:

[0021]

[0022]

[0023] Where: E represents mathematical expectation, C defect represents the rough detection result obtained by BCD-Unet, C truth Represents the true curve, z is the randomly added noise, G represents the generator, D represents the discriminator, G() and D() represent the output results of the generator and discriminator respectively.

[0024] The training process of the one-dimensional generative adversarial network (1D-GAN) is as follows: first, two deep learning models are set up, namely the generator and the discriminator; the role of the generator is to output a smooth and continuous generated curve by inputting a given rough detection curve, and the role of the discriminator is to judge the difference between the generated curve and the true curve. The closer the two curves are, the better the generation effect of the generator is, and the closer the output value of the discriminator is to 1. Conversely, the generation effect is poor, and the discriminator output value is close to 0; therefore, after the generator generates a curve, its result will be input into the discriminator for judgment, and then the discriminator parameters will be updated by back propagation. Then, the generator adjusts its own parameters according to the result given by the discriminator to improve the accuracy of the generated result. The two models are trained adversarially to improve the performance together. Finally, the result generated by the generator can be very close to the true curve, that is, the ability to fill the missing part of the curve is completed.

[0025] The overall objective function of the curve filling module is shown in formula (5):

[0026]

[0027] Where λ represents the weight; the training parameters of the curve filling module are set as: epochs = 50, batch_size = 16.

[0028] In step (1), the input of the curve fusion module is the result of the curve correction module and the curve filling module, and the output is the fused curve. These curves are also stored in a one-dimensional array. The L1 loss function is used as its objective function. The L1 loss function is shown in formula (6):

[0029]

[0030] In formula (6), represents the true value of point i, represents the predicted value of the curve correction module for point i, and N represents the number of elements in the array. The L1 loss function is robust and fast in calculation, and can better integrate the results of the curve correction module and the curve filling module. This module also uses backpropagation for training to evaluate the difference between the model output value and the true value and continuously adjust the model.

[0031] In step (2) above, the specific method of iterative training is as follows: the model takes the mitral regurgitation spectrum as input and outputs the edge curve; there are a total of 44 spectrum images in the data set. The 44 spectrum images are cropped and divided into 88 slices, and the size is adjusted to 256*256. 71 slices are used for training and 17 slices are used for testing. To overcome the problem of small data volume, data enhancement is performed on the 71 training slices, and after enhancement, 710 slices are obtained, which can meet the data volume requirements of the model.

[0032] The left atrial pressure is measured using the mitral regurgitation spectrum envelope identified by the method. The measurement method is as follows: the mitral regurgitation spectrum envelope identified by the model is mapped to the coordinate system given by the spectrum diagram, and the equations listed in formula (7) and formula (8) are solved to calculate the value of the left atrial pressure LAP.

[0033] τ=(t1-t2) / ln((LAP+16) / LAP+4) (7);

[0034] τ=(t1-t3) / ln((LAP+36) / LAP+4) (8);

[0035] In formulas (7) and (8), τ is the left ventricular relaxation time constant, t1, t2, and t3 are the times when the blood flow velocities in the descending branch of mitral regurgitation continuous Doppler spectrum are 1 m / s, 2 m / s, and 3 m / s, respectively. The left ventricular relaxation time constant τ = P / (-dP / dt), where P is the left ventricular pressure and t is the time from -dp / dtmax.

[0036] The advantages and beneficial effects of the present invention are as follows:

[0037] 1. The present invention derives a new formula for non-invasive quantitative measurement of left atrial pressure using ultrasound;

[0038] 2. The present invention collects the original audio signal of the mitral regurgitation spectrum, establishes a spectrum envelope mapping method based on the deep learning model framework, and post-processes it into a "wide" spectrum composed of more "spectral lines" to improve the accuracy of time-velocity curve measurement;

[0039] 3. The accuracy of the new method was verified by comparing it with the "gold standard" catheter method through animal and clinical experiments;

[0040] 4. The present invention has been compared with the existing "Doppler + sphygmomanometer" method through clinical trials to verify its accuracy, superiority and practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of continuous Doppler spectrum of mitral regurgitation.

[0042] Figure 2 This is the overall diagram of the deep learning network model proposed in this invention.

[0043] Figure 3 This is a schematic diagram of the continuous Doppler spectrum of mitral regurgitation corrected by the method of the present invention. DETAILED DESCRIPTION

[0044] In order to further understand the present invention, the present invention is described below with reference to examples. However, the examples are only for further elaboration of the features and advantages of the present invention, rather than for limitation of the claims of the present invention.

[0045] The maximum rate of decrease in left ventricular pressure after systole (-dP / dt) is exponentially distributed. Based on the Weiss exponential equation and the simplified Bernoulli equation, a "two-variable linear equation system" containing the left ventricular relaxation time constant (τ) and LAP can be derived. Figure 1. The calculation of τ has been verified by animal experiments and applied to the built-in measurement software of the ultrasound machine. Although both are excellent indicators for accurately evaluating left ventricular function and hemodynamic status, they are rarely used in clinical diagnosis and treatment. Mainly because: 1. Each step of the measurement needs to be very precise, and a small error may result in a large bias; 2. The mitral valve must have an appropriate amount of regurgitation to obtain the accurate outline of the regurgitation spectrum for measurement. However, with the advent of the big data era, the continuous improvement of algorithms and the introduction of machine deep learning, the image processing of the regurgitation spectrum can be made more accurate, so that the derivation formula in the theory may be applied in clinical practice.

[0046] The present invention compares the accuracy of measuring left atrial pressure using different automatic envelope methods of continuous Doppler spectrum and establishes a spectrum envelope line drawing method based on a deep learning model framework.

[0047] The first step is to build a deep learning network model. The basic network model used in this paper is the BCD-Unet deep learning model. This model is used to perform edge prediction on the mitral regurgitation spectrum, demonstrating excellent edge detection capabilities.

[0048] However, due to the presence of misalignment and breakage in the edges predicted by BCD-Unet, using it alone for detection is far from sufficient. Therefore, the present invention proposes using a post-processing module to perform fine-tuning operations on the edges. The post-processing module consists of three parts: a curve correction module, a curve fusion module, and a curve fusion module.

[0049] In order to deal with the problem of curve misalignment, the present invention designs a curve correction module. This module consists of two parts: curve fitting and one-dimensional unet (1D-Unet). The purpose of curve fitting is to obtain a continuous and smooth curve using polynomial fitting, while 1D-Unet corrects the fitted curve through deep regression to make it closer to the edge curve formed by the doctor's manual marking of edge key points. The curve processed by this curve correction module can effectively solve the problem of curve misalignment. The input of 1D-Unet is the fitted curve, and the output is the corrected curve; these curves are all stored in the form of a one-dimensional array.

[0050] In addition to the curve correction module, we also designed a curve filling module to deal with the problem of curve breakage. Generally, due to inconsistent edge features in the spectrum graph, some curves will not be detected, resulting in curve breakage. Therefore, the defective part can be regarded as a masked area, and our goal is to repair it. Masked area repair has always been one of the classic problems of Generative Adversarial Networks (GAN), so we designed a curve filling module based on the one-dimensional generative adversarial network (1D-GAN) and trained it in an adversarial manner between the discriminator and the generator. After training, the module can repair the missing parts on the curve to obtain a smooth filled curve.

[0051] To better integrate the information generated by the filling and correction curves, the present invention also designs a curve fusion module. This module further fuses the features of the curve processed by the curve correction module with those processed by the curve filling module, combining the advantages of both curves to produce a final continuous, smooth edge curve. Compared to the edge curve predicted by BCD-Unet, the resulting edge curve after this series of processing is more continuous and smooth, meeting the requirements. Figure 2 This is the overall diagram of the model.

[0052] For BCD-Unet, the present invention uses the cross entropy loss function to train it. The cross entropy loss function is shown in formula (1):

[0053]

[0054] In formula (1), y i Indicates whether the pixel i belongs to the pixel of the envelope curve, if it does, it is 1, otherwise it is 0; p iis the probability that pixel i is predicted to belong to the envelope curve; N is the number of pixels. The advantage of using the cross-entropy loss function is that the model remains stable and converges quickly during training. The BCD-Unet training process is as follows: After an image is input to the model, the model first performs a convolution operation on the image to extract features. The resulting feature map is then downsampled, reducing its size to half, allowing the convolution operation to extract features over a larger image area. After several rounds of convolution-downsampling operations, the feature map is gradually upsampled and then convolved to restore it to its original size. A loss function is then used to evaluate the difference between each pixel in the model's output and the true label, and backpropagation is also used to adjust the model parameters. Finally, the model is trained and has the ability to roughly detect curves in spectrograms. During the experiment, the BCD-Unet model training parameters were set to 50 epochs and 8 batch sizes, and the Adam optimizer with a learning rate of 0.004 was used to accelerate the training process.

[0055] For the curve correction module, the present invention uses the L2 loss function to train it. The L2 loss function is shown in formula (2):

[0056]

[0057] In formula (2), represents the true value of point i, represents the curve correction module's predicted value for point i, and N represents the number of elements in the array. The L2 loss function is a commonly used loss function in regression tasks, with advantages such as fast computation and convenient differentiation. The training process of the curve correction module is similar to that of BCD-Unet, still using backpropagation to continuously adjust the model by evaluating the difference between the model output and the true value. The training parameters of the curve correction module are set to: epochs = 50, batch_size = 16.

[0058] The curve filling module uses a one-dimensional generative adversarial network (1D-GAN) to process the curve, training it in a discriminator-generator adversarial manner. After training, the curve filling module can repair missing parts on the curve, resulting in a smooth filled curve. 1D-GAS takes a curve with missing parts as input and outputs the repaired curve, which is also stored as a one-dimensional array.

[0059] For the curve filling module, based on the idea of generative adversarial networks, two loss functions are used to train the discriminator and generator respectively. The loss functions of the discriminator and generator are shown in formula (3) and formula (4) respectively:

[0060]

[0061]

[0062] Where: E represents mathematical expectation, C defect represents the rough detection result obtained by BCD-Unet, C truth Represents the true curve, z is the randomly added noise, G represents the generator, D represents the discriminator, G() and D() represent the output results of the generator and discriminator respectively.

[0063] The training process of the one-dimensional generative adversarial network (1D-GAN) is as follows: first, two deep learning models are set up, namely the generator and the discriminator; the role of the generator is to output a smooth and continuous generated curve by inputting a given rough detection curve, and the role of the discriminator is to judge the difference between the generated curve and the true curve. The closer the two curves are, the better the generation effect of the generator is, and the closer the output value of the discriminator is to 1. Conversely, the generation effect is poor, and the discriminator output value is close to 0; therefore, after the generator generates a curve, its result will be input into the discriminator for judgment, and then the discriminator parameters will be updated through back propagation. The generator then adjusts its own parameters according to the result given by the discriminator to improve the accuracy of the generated result. The two models are trained adversarially to improve performance together. Finally, the result generated by the generator can be very close to the true curve, that is, it has completed the ability to fill in the missing part of the curve.

[0064] The overall objective function of the curve filling module is shown in formula (5):

[0065]

[0066] Where λ represents the weight; the training parameters of the curve filling module are set as: epochs = 50, batch_size = 16.

[0067] The input of the curve fusion module is the result of the curve correction module and the curve filling module, and the output is the fused curve. These curves are also stored in a one-dimensional array. The L1 loss function is used as its objective function. The L1 loss function is shown in formula (6):

[0068]

[0069] In formula (6), represents the true value of point i, represents the predicted value of the curve correction module for point i, and N represents the number of elements in the array. The L1 loss function is robust and fast in calculation, and can better integrate the results of the curve correction module and the curve filling module. This module also uses backpropagation for training to evaluate the difference between the model output value and the true value and continuously adjust the model.

[0070] The second step is to iteratively train the model using data collected by the hospital, thereby continuously optimizing the model. The trained deep learning model has the ability to automatically track the envelope. The specific method of the iterative training is: the model takes the mitral regurgitation spectrum as input and outputs the edge curve; there are a total of 44 spectrum images in the data set. The 44 spectrum images are cropped and divided into 88 slices, and the size is adjusted to 256*256. 71 slices are used for training and 17 slices are used for testing; in order to overcome the problem of small data volume, the 71 training slices are enhanced, and 710 slices are obtained after enhancement, which can meet the model's requirements for data volume.

[0071] Figure 3 Figure 1 is a schematic diagram of the corrected continuous Doppler spectrum of mitral regurgitation, where: τ is the left ventricular relaxation time constant, t1, t2, and t3 are the times when the blood flow velocity of the descending branch of the continuous Doppler spectrum of mitral regurgitation is 1 m / s, 2 m / s, and 3 m / s respectively, and the left ventricular relaxation time constant (τ): τ = P / (-dP / dt), where P is the left ventricular pressure and t is the time from -dp / dtmax

[0072] After training the model, it has the ability to detect envelope curves. The model's detection results are mapped to the coordinate system given by the spectrum graph, and the left atrial pressure (LAP) value is calculated using formulas (7) and (8):

[0073] τ=(t1-t2) / ln((LAP+16) / LAP+4) (7)

[0074] τ=(t1-t3) / ln((LAP+36) / LAP+4) (8)

[0075] After solving the equations listed in formula (7) and formula (8), the system can calculate the value of LAP and display it.

[0076] Factors that influence the accurate tracing of the mitral regurgitation spectral envelope include the amount of mitral regurgitation, the smoothness of the envelope, and its ability to accurately reflect mitral regurgitation. More test cases can be collected as needed to construct a richer dataset for model training, allowing this method to accurately traverse the envelope even in cases with less regurgitation.

Claims

1. A method for identifying the envelope of a mitral regurgitation spectrum, characterized by: The method is to identify the envelope of the continuous Doppler spectrum of mitral regurgitation; the method specifically includes the following steps: (1) Build a deep learning network model to perform edge detection on the mitral regurgitation spectrum. Since the edges predicted by the network model are misaligned and broken, a post-processing module is used to correct the edges. The post-processing module includes a curve correction module, a curve filling module, and a curve fusion module. The curve correction module is used to handle the curve misalignment problem, the curve filling module is used to handle the curve breakage problem, and the curve fusion module is used to further fuse the features of the two curves, combining the advantages of the two curves to obtain a final continuous and smooth edge curve. The network model is a BCD-Unet deep learning model, and is trained using a cross entropy loss function, which is shown in formula (1): ； In formula (1), y i Indicates whether the pixel i belongs to the pixel of the envelope curve, if it does, it is 1, otherwise it is 0; p i is the probability that pixel i is predicted to belong to the envelope curve pixel; N is the number of pixels; The curve correction module consists of two parts: curve fitting and one-dimensional unet (1D-Unet). The purpose of curve fitting is to obtain a continuous and smooth curve using polynomial fitting, while 1D-Unet corrects the fitted curve through deep regression to make it closer to the edge curve formed by manually marking edge key points by the doctor; the curve processed by the curve correction module can effectively solve the problem of curve misalignment; the input of 1D-Unet is the fitted curve, and the output is the corrected curve; these curves are all stored in the form of one-dimensional arrays; The curve correction module is trained using the L2 loss function, which is shown in formula (2): ； In formula (2), represents the true value of point i, represents the predicted value of the curve correction module for point i, and N represents the number of elements in the array. The L2 loss function is one of the loss functions commonly used in regression tasks. It has the advantages of fast calculation speed and convenient differentiation. The training process of the curve correction module is similar to that of BCD-Unet. It still uses backpropagation for training, and continuously adjusts the model by evaluating the difference between the model output value and the true value. The training parameters of the curve correction module are set as: epochs=50, batch_size=16. The curve filling module processes the curve using a one-dimensional generative adversarial network (1D-GAN), training it in a discriminator-generator adversarial manner. After training, the curve filling module can repair the missing parts of the curve to obtain a smooth filled curve. The input of the 1D-GAN is a curve with missing parts, and the output is the repaired curve, which is also stored in a one-dimensional array. The curve filling module is based on the idea of generative adversarial networks and uses two loss functions to train the discriminator and generator respectively. The loss functions of the discriminator and generator are shown in formula (3) and formula (4) respectively: ； Where: E represents the mathematical expectation, Represents the rough detection result obtained by BCD-Unet, Represents the true curve, z is the randomly added noise, G represents the generator, D represents the discriminator, G() and D() represent the output results of the generator and discriminator respectively; The input of the curve fusion module is the result of the curve correction module and the curve filling module, and the output is the fused curve; these curves are also stored in a one-dimensional array; the L1 loss function is used as its objective function, and the L1 loss function is shown in formula (6): ； In formula (6), represents the true value of point i, represents the predicted value of point i by the curve correction module, and N represents the number of elements in the array. The L1 loss function is robust and fast, and can better integrate the results of the curve correction module and the curve filling module. This module also uses backpropagation for training, evaluating the difference between the model output value and the true value and continuously adjusting the model. (2) Iteratively training the model described in step (1) using data collected by the hospital, and the trained deep learning model has the ability to automatically track the envelope; (3) Using the model trained in step (2), the mitral regurgitation spectrum is used as input, and the output is the edge curve and the left atrial pressure value LAP, where the edge curve is the envelope line.

2. The method for identifying the mitral regurgitation spectrum envelope according to claim 1, characterized in that: In step (1), The training process of BCD-Unet is as follows: after the image is input into the model, the model first performs a convolution operation on the image to extract features, and then downsamples the obtained feature map to reduce the size of the feature map to half of the original size, so that the convolution operation can extract features on a larger image range; after several convolution-downsampling operations, the feature map is gradually upsampled and convolved so that the feature map can be restored to the original image size; then the loss function is used to evaluate the difference between each pixel in the output result of the model and the true label, and the parameters of the model are adjusted using the back propagation method; finally, the model training is completed, and the model has the ability to roughly detect curves in the spectrum graph; during the experiment, the training parameters of the BCD-Unet model are set to: epochs=50, batch_size=8, and the Adam optimizer with a learning rate of 0.004 is used to accelerate the training process.

3. The method for identifying the mitral regurgitation spectrum envelope according to claim 1, characterized in that: The training process of the one-dimensional generative adversarial network (1D-GAN) is as follows: first, two deep learning models are set up, namely the generator and the discriminator; the role of the generator is to output a smooth and continuous generated curve by inputting a given rough detection curve, and the role of the discriminator is to judge the difference between the generated curve and the true curve. The closer the two curves are, the better the generation effect of the generator is, and the closer the output value of the discriminator is to 1. Conversely, the generation effect is poor, and the discriminator output value is close to 0; therefore, after the generator generates a curve, its result will be input into the discriminator for judgment, and then the discriminator parameters will be updated through back propagation. The generator then adjusts its own parameters according to the result given by the discriminator to improve the accuracy of the generated result. The two models are trained adversarially to improve performance together. Finally, the result generated by the generator can be very close to the true curve, that is, it has completed the ability to fill in the missing part of the curve.

4. The method for identifying the envelope of the mitral regurgitation spectrum according to claim 1, wherein: The overall objective function of the curve filling module is shown in formula (5): ； Where λ represents the weight; the training parameters of the curve filling module are set to: epochs=50, batch_size=16.

5. The method for identifying the envelope of mitral regurgitation spectrum according to claim 1, characterized in that: In step (2), the specific method of iterative training is as follows: the model takes the mitral regurgitation spectrum as input and outputs the edge curve; there are a total of 44 spectrum images in the data set; the 44 spectrum images are cropped and divided into 88 slices, and their sizes are adjusted to 256*256; 71 slices are used for training and 17 slices are used for testing; in order to overcome the problem of small data volume, data enhancement is performed on the 71 training slices, and after enhancement, 710 slices are obtained, which can meet the model's requirements for data volume.

6. A method for measuring left atrial pressure using the mitral regurgitation spectrum envelope identified by the method of any one of claims 1 to 5, characterized in that: The mitral regurgitation spectrum envelope identified by the model is mapped to the coordinate system given by the spectrum diagram, and after solving the equations listed in formula (7) and formula (8), the value of left atrial pressure LAP can be calculated; ； In formulas (7) and (8), τ is the left ventricular relaxation time constant, t1, t2, and t3 are the times when the blood flow velocities in the descending branch of mitral regurgitation continuous Doppler spectrum are 1 m / s, 2 m / s, and 3 m / s, respectively. The left ventricular relaxation time constant τ = P / (-dP / dt), where P is the left ventricular pressure and t is the time from -dp / dtmax.

Images