Method and device for monitoring cerebral blood flow autoregulation function and training target model

Abstract

The present application discloses a method and device for monitoring cerebral blood flow autoregulation function and training a target model, belonging to the field of intelligent medical technology. The method comprises: based on an arterial blood pressure signal and a cerebral blood flow signal, determining a first transfer function analysis TFA variable for representing the current state of the cerebral blood flow autoregulation CA function to be monitored, wherein the arterial blood pressure signal is the input signal of the CA function, and the cerebral blood flow signal is the output signal of the CA function. Based on the first TFA variable, determining whether the CA function is damaged through a target model, the target model can monitor whether any CA function is damaged based on the TFA variable corresponding to any CA function, the target model comprises a first branch and a second branch, the first branch is used to determine the clustering loss of the first TFA variable, and the second branch is used to determine the reconstruction loss of the first TFA variable. The target model is obtained by training the model based on the reconstruction loss and the clustering loss, and the monitoring accuracy is improved compared with the logistic regression model.

Description

Technical Field

[0001] The present application relates to the field of intelligent medical technology, and in particular to a method and device for monitoring the automatic regulation function of cerebral blood flow and training a target model. Background Art

[0002] Cerebral autoregulation (CA) function is an important hemodynamic physiological protection mechanism that can maintain the cerebral blood flow velocity (CBFV) at a relatively stable level when the arterial blood pressure (ABP) fluctuates. Since impaired CA function increases the risk of secondary brain injury, there is usually a need to monitor CA function to determine whether it is damaged.

[0003] In the related art, the mean of the transfer function analysis (TFA) variables (generally including phase, gain, and correlation information) within each fixed frequency band is generally determined based on the input signal (ABP) and output signal (CBFV) of the CA function, so as to characterize the state of the CA function through the mean of the TFA variables within each fixed frequency band. Then, the mean of the TFA variables within each fixed frequency band is fitted with a logistic regression model to determine the difference of the TFA variables, so as to monitor the CA function based on the difference of the TFA variables.

[0004] However, due to the lack of complexity of the logistic regression model, it is prone to underfitting, resulting in inaccurate monitoring results. Summary of the invention

[0005] The embodiment of the present application provides a method and device for monitoring the automatic regulation function of cerebral blood flow and training a target model, which can solve the problem of inaccurate monitoring results in related technologies. The technical solution is as follows:

[0006] In a first aspect, a method for monitoring cerebral blood flow autoregulation function is provided, the method comprising:

[0007] Determine a first transfer function analysis TFA variable for representing a current state of a cerebral blood flow autoregulation (CA) function to be monitored based on an arterial blood pressure signal and a cerebral blood flow signal, wherein the arterial blood pressure signal is an input signal of the CA function and the cerebral blood flow signal is an output signal of the CA function;

[0008] Based on the first TFA variable, whether the CA function is damaged is determined by a target model. The target model can monitor whether any CA function is damaged based on the TFA variable corresponding to any CA function. The target model includes a first branch and a second branch. The first branch is used to determine the clustering loss of the first TFA variable, and the second branch is used to determine the reconstruction loss of the first TFA variable.

[0009] In a second aspect, a method for training a target model is provided, wherein the target model can monitor whether any CA function is damaged based on a TFA variable corresponding to any cerebral blood flow autoregulation CA function, the target model comprising a first branch and a second branch, the first branch being used to determine a clustering loss of the TFA variable, and the second branch being used to determine a reconstruction loss of the TFA variable; the method comprising:

[0010] Acquire multiple groups of sample data, each group of sample data in the multiple groups of sample data includes a sample arterial blood pressure signal and a sample cerebral blood flow signal;

[0011] Based on each group of sample data, determine the second TFA variable corresponding to each group of sample data to obtain multiple groups of second TFA variables, each group of second TFA variables is used to describe the state of the corresponding CA function;

[0012] Based on the multiple groups of second TFA variables, the deep convolutional embedded clustering model is iteratively trained to obtain the target model, wherein the deep convolutional embedded clustering model includes the first branch and the second branch.

[0013] In a third aspect, a device for monitoring the autoregulation function of cerebral blood flow is provided, the device comprising:

[0014] A first determination module is used to determine a first transfer function analysis TFA variable for representing a current state of a cerebral blood flow autoregulation (CA) function to be monitored based on an arterial blood pressure signal and a cerebral blood flow signal, wherein the arterial blood pressure signal is an input signal of the CA function and the cerebral blood flow signal is an output signal of the CA function;

[0015] The second determination module is used to determine whether the CA function is damaged based on the first TFA variable through a target model. The target model can monitor whether any CA function is damaged based on the TFA variable corresponding to any CA function. The target model includes a first branch and a second branch. The first branch is used to determine the clustering loss of the first TFA variable, and the second branch is used to determine the reconstruction loss of the first TFA variable.

[0016] In a fourth aspect, a device for training a target model is provided, wherein the target model can monitor whether any CA function is damaged based on a TFA variable corresponding to any cerebral blood flow autoregulation CA function, the target model comprises a first branch and a second branch, the first branch is used to determine a clustering loss of the TFA variable, and the second branch is used to determine a reconstruction loss of the TFA variable, the device comprising:

[0017] An acquisition module, used for acquiring multiple groups of sample data, each group of sample data in the multiple groups of sample data includes a sample arterial blood pressure signal and a sample cerebral blood flow signal;

[0018] a variable determination module, configured to determine, based on each set of sample data, a second TFA variable corresponding to each set of sample data, to obtain multiple sets of second TFA variables, each set of second TFA variables being used to describe a state of a corresponding CA function;

[0019] A training module is used to iteratively train the deep convolution embedded clustering model based on the multiple groups of second TFA variables to obtain the target model, wherein the deep convolution embedded clustering model includes the first branch and the second branch.

[0020] In a fifth aspect, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method described in the first aspect or the second aspect is implemented.

[0021] In a sixth aspect, a computer-readable storage medium is provided, on which instructions are stored. When the instructions are executed by a processor, the method described in the first aspect or the second aspect is implemented.

[0022] In a seventh aspect, a computer program product comprising instructions is provided, which, when executed on a computer, enables the computer to execute the method described in the first or second aspect above.

[0023] The beneficial effects of the technical solution provided by the embodiment of the present application are:

[0024] Based on the arterial blood pressure signal and the cerebral blood flow signal, the first transfer function analysis TFA variable for representing the current state of the cerebral blood flow autoregulation CA function to be monitored is determined, the arterial blood pressure signal is the input signal of the CA function, and the cerebral blood flow signal is the output signal of the CA function. Based on the first TFA variable, whether the CA function is damaged is determined by the target model. The target model can monitor whether any CA function is damaged based on the TFA variable corresponding to any CA function. The target model includes a first branch and a second branch. The first branch is used to determine the clustering loss of the first TFA variable, and the second branch is used to determine the reconstruction loss of the first TFA variable. The target model is obtained by training the model based on the reconstruction loss and the clustering loss. Compared with the logistic regression model, the monitoring accuracy is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0026] Figure 1 is a flow chart of a method for training a target model according to an exemplary embodiment;

[0027] Figure 2 is a schematic diagram of mean values ​​of TFA variables according to an exemplary embodiment;

[0028] Figure 3 is a schematic diagram of a structure of a deep convolutional embedded clustering model according to an exemplary embodiment;

[0029] Figure 4 is a schematic diagram of a deep convolutional embedded clustering model with two branches according to an exemplary embodiment;

[0030] Figure 5 is a schematic diagram showing a clustering result according to an exemplary embodiment;

[0031] Figure 6 is a flow chart of a method for monitoring CA function according to an exemplary embodiment;

[0032] Figure 7 is a schematic structural diagram of a device for monitoring the automatic regulation function of cerebral blood flow according to an exemplary embodiment;

[0033] Figure 8 is a schematic structural diagram of a device for training a target model according to an exemplary embodiment;

[0034] Fig. 9 The diagram is a schematic structural diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0035] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0036] It should be understood that the "multiple" mentioned in this application refers to two or more. In the description of this application, unless otherwise specified, " / " means or, for example, A / B can mean A or B; "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in order to facilitate the clear description of the technical solution of this application, the words "first" and "second" are used to distinguish between the same items or similar items with basically the same functions and effects. Those skilled in the art can understand that the words "first" and "second" do not limit the quantity and execution order, and the words "first" and "second" do not limit them to be different.

[0037] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0038] Before introducing the method provided in the embodiment of the present application in detail, the execution subject involved in the embodiment of the present application is introduced. The method provided in the embodiment of the present application can be executed by an electronic device. In one example, the electronic device can be a terminal such as a laptop, a desktop, a portable computer, a tablet computer, etc., which is not limited in the embodiment of the present application.

[0039] Next, the application scenarios involved in the embodiments of the present application are introduced.

[0040] In the medical field, CA function refers to the intrinsic ability of the brain to maintain adequate cerebral perfusion when arterial blood pressure changes. In patients with severe hypertension, brain damage, or undergoing surgery, CA function plays an important role in maintaining stable cerebral perfusion. However, it is currently known that a variety of diseases can damage CA function, and impaired CA function increases the risk of secondary brain damage. Therefore, monitoring CA function has become a clinical need to provide patients with useful clinical information. To this end, the embodiments of the present application provide a CA function monitoring method based on unsupervised learning, which can monitor in real time and accurately whether an individual's CA function is impaired.

[0041] In the embodiment of the present application, the target model is used to predict whether the CA function is impaired. The target model is obtained through training. Next, the method for training the target model provided in the embodiment of the present application is described in detail with reference to the accompanying drawings. Figure 1 , Figure 1 This is a flow chart of a method for training a target model according to an exemplary embodiment. As an example but not a limitation, the method can be applied to the above electronic device, and specifically may include part or all of the following contents:

[0042] Step 101: Acquire multiple groups of sample data, each group of sample data in the multiple groups of sample data includes a sample arterial blood pressure signal and a sample cerebral blood flow signal.

[0043] In order to train the target model, sample data must be collected first. In one example, 61 healthy individuals, 61 individuals with severe symptoms, and 2,247 individuals with mild to severe stenosis and no symptoms can be included. Among them, healthy individuals are collected from healthy people with no previous history of cardiovascular and cerebrovascular disease and no relevant risk factors. Healthy individuals are judged to have intact CA function; individuals with severe symptoms refer to subjects with severe intracranial vascular stenosis and stroke symptoms. Individuals with severe symptoms are identified as having impaired CA function; the other 2,247 individuals are individuals who cannot be judged as having impaired CA.

[0044] During the collection of sample data, each included individual can be asked to rest in a supine position with the head raised for at least 10 minutes. Professionals (such as medical staff) use a 2MHz transcranial Doppler probe head frame to continuously record the left and right CBFV of the bilateral middle cerebral arteries at a depth of 45-55mm distal to the M1 stenosis. At the same time, vascular unloading technology is used to non-invasively measure ABP from the finger artery. Healthy individuals only retain unilateral CBFV, individuals with severe symptoms retain the diseased CBFV, and individuals with mild to severe symptoms retain bilateral CBFV. In one example, in order to ensure the stability of the sample data, the CBFV signal and ABP signal collected synchronously after 10 minutes can be retained.

[0045] It is not difficult to understand that for individuals who have been diagnosed, the sample data collected from them are labeled, for example, the label is used to indicate that the CA function is impaired or not impaired. In this case, the collected sample data can be called labeled sample data; and for individuals who cannot be diagnosed, the sample data collected from them are unlabeled, and in this case, the collected sample data can be called unlabeled sample data. In this way, a total of 122 groups of labeled sample data (including ABP signals and CBFV signals of labeled individuals) and 4494 groups of unlabeled sample data (including ABP signals and CBFV signals of unlabeled individuals) can be collected, among which the unlabeled sample data are used for model training, and the labeled sample data are used to test the trained model.

[0046] It is worth mentioning that each set of sample data in the multiple sets of sample data is unlabeled sample data, which makes the method more universal.

[0047] Step 102: Based on each set of sample data, determine the second TFA variable corresponding to each set of sample data to obtain multiple sets of second TFA variables, each set of second TFA variables is used to describe the state of the corresponding CA function.

[0048] The electronic device can post-process the sample data using Matlab software. For each set of sample data (each set of sample data corresponds to an individual), two 90-second windows are slid simultaneously with 10-second increments of the ABP signal and the CBFV signal to obtain frequency segments with a resolution of 1 / 90 Hz, that is, each window includes 90 seconds of collected signals. Then, based on the sample data obtained after post-processing, the TFA variable is determined by the TFA method. For all the data, the average value of the TFA variable extracted from all its frequency segments represents the state of the individual's CA function.

[0049] In implementation, the specific implementation for determining the sample data by the TFA method may include the following:

[0050] A finite-impulse-response (FIR) filter (N=12) is used to estimate the impulse in each window, as well as the frequency and step response, where N is the number of taps of the FIR filter. In implementation, the finite impulse response can be estimated using the least mean square method, such as calculated by formula (1):

[0051]

[0052] Where h(n) represents the finite impulse response, p represents the ABP signal, v represents the CBFV signal, represents the inverse matrix of the ABP signal autocorrelation matrix, R pv(n) represents the cross-correlation matrix between ABP signal and CBFV signal.

[0053] Afterwards, discrete time Fourier transform (DTFT) can be performed by the following formula (2), specifically:

[0054]

[0055] Wherein, fs represents the sampling frequency.

[0056] Afterwards, the gain and phase of the DTFT are determined by formula (3) and formula (4) respectively:

[0057]

[0058]

[0059] Among them, H R (f) represents the real part of H(f), H I (f) represents the imaginary part of H(f). The phase value range is [-180°, 180°].

[0060] Then, DTFT is applied to autocorrelation and cross-correlation, and correlation information is calculated through auto-spectra and cross-spectrum, as shown in formula (5):

[0061]

[0062] Among them, MSC(f) represents the correlation information, S pp (f) and S w (f) represents the self-spectrum, which can be based on R pp (n) and R vv (n) get, S pv (f) represents the cross spectrum, which can be based on R pv (n) Determine.

[0063] Since the length of the frequency segment with a sampling frequency of 1 Hz is 90 seconds, according to the Nyquist theorem, there are 45 frequency points containing phase, gain and correlation information in the range of 0-0.5 Hz to represent the frequency domain features. The mean value of each frequency domain feature at each frequency point is as follows: Figure 2 As shown, Figure 2 Figure (a) is a schematic diagram of the mean value of the phase at each frequency point. Figure 2 Figure (b) is a schematic diagram of the mean gain at each frequency point. Figure 2Figure (c) is a schematic diagram of the mean value of the correlation information at each frequency point. The solid line represents the sampled data of healthy individuals, the dotted line represents the sampled data of individuals with severe symptoms, and the line parallel to the vertical axis represents the average standard deviation of the TFA variable along the frequency point. In an embodiment of the present application, for unlabeled samples, the phase, gain, and correlation information corresponding to each frequency point are taken as a set of sample data, where the phase, gain, and correlation information can be collectively referred to as TFA variables.

[0064] In order to make each dimension of the same frequency domain feature (such as phase, gain, or correlation information) comparable, each frequency domain feature can be preprocessed separately. In implementation, each frequency domain feature can be standardized separately by the following formula (6):

[0065]

[0066] Among them, x represents the original value of the frequency domain feature, x * represents the standardized value, μ represents the mean of the frequency domain feature, and σ represents the variance of the frequency domain feature.

[0067] Afterwards, the standardized frequency domain features can be normalized to within a preset numerical range. In one example, the extreme value normalization method can be used for processing. Among them, the preset numerical range can be set according to actual needs, usually referring to a data range near the zero point, such as the preset numerical range of [-0.5, 0.5]. In this way, since the activation function will obtain the maximum activation value near the zero point during the subsequent training process, and the larger the activation value, the faster the model learning speed can be, so the standardized frequency domain features are normalized to within the preset numerical range, which can speed up the learning of the model.

[0068] In addition, by processing each labeled sample data in the above manner, taking the phase, gain, correlation information, and true label corresponding to each frequency point as a set of test data, a test data set can be obtained.

[0069] Step 103: Based on multiple groups of second TFA variables, iteratively train the deep convolutional embedded clustering model to obtain a target model, wherein the deep convolutional embedded clustering model includes a first branch and a second branch.

[0070] The first branch has the ability to cluster input features, and the second branch has the ability to reconstruct input features.

[0071] For an example, see Figure 3 , Figure 3It is a structural diagram of a deep convolution embedded clustering model according to an exemplary embodiment. The first branch includes a coding module and a deep embedded clustering module, the second branch includes a coding module and a decoding module, and the second branch and the first branch can share the coding module.

[0072] In an embodiment of the present application, the specific implementation of iterative training of the deep convolutional embedded clustering model based on multiple groups of second TFA variables may include the following sub-steps:

[0073] (1) In each iteration, a specified number threshold groups of second TFA variables from multiple groups of second TFA variables are used as input features of the deep convolutional embedded clustering model.

[0074] The specified quantity threshold can be set according to actual needs. For example, the specified quantity threshold is 32, that is, the small batch size in the training process can be set to 32.

[0075] In each iteration of training, a small batch of second TFA variables is selected, and a corresponding matrix is ​​generated based on each selected group of second TFA variables. For example, since each group of second TFA variables includes three frequency domain features, namely phase, gain, and correlation information, and each frequency domain feature corresponds to 45 frequency point values, a 3*45 matrix can be generated based on each group of second TFA variables. The electronic device uses all the generated 3*45 matrices as input features and inputs them into the deep convolution embedded clustering model.

[0076] (2) The input features are clustered through the first branch, and the gradient of the first branch is back-propagated according to the clustering results.

[0077] In one example, the implementation of clustering the input features through the first branch may include: encoding the input features through the encoding module, and clustering the features obtained after the encoding process by the encoding module through the deep embedded clustering module.

[0078] The encoding module includes a convolution layer and a pooling layer. The convolution layer performs convolution processing to achieve feature extraction, and the pooling layer performs pooling processing to reduce the dimension of the feature to reduce the amount of computation.

[0079] The number of convolutional layers included in the encoding module can be one or more, and the number of pooling layers can be one or more.

[0080] In one example, the encoding module includes N branches of convolution kernels with different receptive field sizes, where N is an integer greater than or equal to 2. At this time, the specific implementation of encoding the input features through the encoding module may include: performing convolution processing on the input features through each of the multiple branches in the encoding module, respectively, to achieve encoding of the input features.

[0081] For example, the value of N is 2. Figure 4 , Figure 4 It is a schematic diagram of a deep convolution embedded clustering model with two branches according to an exemplary embodiment, and the convolution kernel size on each branch is different. Each branch includes a convolution layer, a pooling layer, a convolution layer, a pooling layer, a convolution layer, a deconvolution layer, an upsampling layer, a deconvolution layer, an upsampling layer, and a deconvolution layer. The convolution layer, pooling layer, convolution layer, pooling layer, and convolution layer included in each branch belong to the encoding module, and the deconvolution layer, upsampling layer, deconvolution layer, upsampling layer, and deconvolution layer included in each branch belong to the decoding module. That is, the encoding module includes two branches, and the decoding module includes two branches. In addition, from Figure 4 It is not difficult to see that the output of the middle convolutional layer in each branch is also connected to the deep embedded clustering module.

[0082] In an example, the convolution kernel sizes of the two branches can be set to 3 and 7 respectively, the number of channels of each branch changes in sequence: [3, 8, 4, 2, 4, 8, 8, 16, 3], the size and step size of the pooling layer are both 3, and the upsampling layer corresponding to the pooling layer is the nearest neighbor upsampling with a step size of 3.

[0083] In each training, after the input features are input into the encoding module, the encoding module performs encoding processing through each of the two branches to achieve feature extraction. Afterwards, the features extracted by each branch are respectively input into the deep embedded clustering module to be clustered by the deep embedded clustering module. The electronic device performs gradient back propagation on the first branch according to the clustering result to update the parameters of the first branch.

[0084] In one example, the deep embedded clustering module can use the K-means algorithm to cluster the features extracted by the encoding module and maximize the category separation ability. Assuming that there are only two types of CA states, healthy and severe symptoms, the K value is set to 2, and the DEC (Deep Embedded Clustering) self-training strategy is adopted. For the i-th sample data and the j-th cluster, the Student's t-distribution can be used to measure the data representation Z in the latent space m The centroid μ of the TFA variable with healthy CA or damaged CA k, as the soft assigned cluster distribution Q, as shown in formula (7). Then Q is raised to the quadratic power and normalized according to the frequency of each cluster to calculate the auxiliary target distribution P, as shown in formula (8). The clustering loss is then calculated by minimizing the two distributions using KL divergence, as shown in formula (9). The target distribution P is used as pseudo-supervisory information to optimize the data representation by learning a high-confidence auxiliary value with supervisory effect, which in turn supervises the update of the distribution Q, so that the data of each cluster cluster is closer to the cluster center, thereby improving the cohesion of the cluster.

[0085]

[0086]

[0087]

[0088] Here, α represents the degree of freedom.

[0089] It is worth mentioning that electronic devices extract rich local features of each variable under different receptive fields through multiple branches. From the frequency domain perspective, the local information under different receptive fields actually represents the local information in frequency bands of different widths. Then, features at different scales are fused so that the information contained in the intermediate layer features used for clustering can better represent the original features.

[0090] (3) The input features are reconstructed through the second branch to obtain feature reconstruction results, and the gradient backpropagation of the second branch is performed based on the feature reconstruction results.

[0091] The implementation of reconstructing the input features through the second branch to obtain the feature reconstruction result may include: reconstructing the features obtained after the encoding process by the encoding module through the decoding module to obtain the feature reconstruction result.

[0092] The decoding module is symmetrical to the encoding module. The decoding module includes a deconvolution layer and an upsampling layer. The deconvolution layer is used for decoding operations. The upsampling layer can reconstruct features of the same dimension as the input features to achieve feature reconstruction. The number of deconvolutions is the same as the number of convolution layers in the encoding module, and the number of upsampling layers is the same as the number of pooling layers in the encoding module.

[0093] It is not difficult to understand that since the first branch and the second branch use the encoding module together, one copy of the features output by the encoding module is transmitted to the deep embedded clustering module for clustering processing, and the other copy is output to the decoding module.

[0094] In one example, the decoding module includes N branches with convolution kernels of different receptive field sizes, where N is an integer greater than or equal to 2. At this time, the features obtained after the encoding processing of the encoding module are reconstructed by the decoding module, and the specific implementation of obtaining the feature reconstruction result may include: performing convolution processing on the features obtained after the encoding processing of the encoding module through each of the N branches in the decoding module. On each branch in the second branch, a jump connection method is used to combine the convolution result of a specified convolution layer in the encoding module with the convolution result of the same channel in the decoding module. The convolution results output by each branch included in the decoding module are fused to obtain a feature reconstruction result.

[0095] The specified convolution layer may include each convolution layer in each branch.

[0096] Please continue to refer to Figure 4 , when N is equal to 2, after the features are extracted through each convolutional layer in the encoding module, in order to retain the original features as much as possible, the features extracted by each convolutional layer can be passed to the corresponding deconvolution layer in the decoding module using a skip connection, such as Figure 4 As shown by the dotted line in . In this way, when the decoding operation is performed by the decoding module, the features decoded by each deconvolution layer are merged with the features transmitted by the encoding module along the channel. The decoding module adds the features of the final output of each branch feature by feature to achieve feature reconstruction. Then, the electronic device performs gradient backpropagation based on the feature reconstruction result to update the parameters in the second branch. In order to quantify the reconstruction loss, the loss function used by the second branch can be the minimum mean square error loss, for example, as shown in formula (10):

[0097]

[0098] Among them, y i represents the TFA variable of sample i, y' i Represents the reconstructed value.

[0099] It is worth mentioning that in the above process, the skip connection method and multiple branches are used to extract rich local information of each TFA variable. In addition, the skip connection method improves the robustness of reconstructing the original input features based on the hidden layer features, ensures the effectiveness of the feature extraction process, and is more conducive to the deep learning of the second branch.

[0100] It should be noted that the above steps (3) and (4) are performed in parallel, that is, during the training process, the first branch and the second branch are iteratively trained at the same time.

[0101] The deep convolution embedded clustering model of the embodiment of the present application can be designed through the pytorch open source framework. The model parameters can be randomly initialized, and the activation function can be selected as the tanh function. From the overall perspective of the deep convolution embedded clustering model, the goal of the deep convolution embedded clustering model is to learn a feature mapping that can represent the input features and is suitable for clustering. Therefore, the overall loss function of the deep convolution embedded clustering model is shown in formula (11), including the reconstruction loss L r and clustering loss L c Two parts. Reconstruction loss L r Ensure that the second branch learns feature representation in an unsupervised manner while maintaining the intrinsic spatial structure of the data distribution; clustering loss L c The extracted feature space is manipulated to fit the embedding points.

[0102] L=L r +λL c (11)

[0103] Among them, L represents the loss function of the deep convolutional embedded clustering model; λ is the weight coefficient, which indicates the degree of deformation of the embedding space.

[0104] In one embodiment, the iterative training process may include two stages: pre-training and subsequent iterative training. In the pre-training process, a first specified number of iterative trainings are performed, and in the subsequent iterative trainings, the update interval is set to a second specified number of times, and the second specified number of times is less than the first specified number of times to avoid instability of the clustering centroid caused by frequent updates. Both the first specified number of times and the second specified number of times can be set according to actual needs, for example, the first specified number of times is 200 times, and the second specified number of times is 100 times.

[0105] As an example of the present application, during iterative training, the optimizer used for gradient back propagation in the training process before a preset number of iterations is the Adam optimizer, and the optimizer used for gradient back propagation in the training process after the preset number of iterations is the SGD (stochastic gradient descent) optimizer.

[0106] The preset number of iterations can be set according to actual needs. For example, the preset number of iterations can be 200. That is, the gradient back propagation in the first 200 iterations of training can use the Adam optimizer, and the gradient back propagation in the subsequent iterations of training can use the SGD optimizer.

[0107] In one example, the Adam optimizer is used in pre-training and the learning rate is set to 0.01, the SGD optimizer is used in subsequent training and the learning rate is set to 0.001, and the learning rate decays to 1 / 10 every 1000 iterations.

[0108] In this way, the Adam optimizer can be used in the early stage of training to make the model converge quickly, and the SGD optimizer can be used to slowly find the optimal solution in the later stage of training, thereby improving the efficiency of model training.

[0109] (4) When the iterative training meets the training end conditions, the obtained deep convolutional embedded clustering model is determined as the target model.

[0110] Among them, the target model can monitor whether any CA function is damaged based on the TFA variables corresponding to any CA function. The target model includes a first branch and a second branch. The second branch is used to determine the clustering loss of the TFA variables, and the first branch is used to determine the reconstruction loss of the TFA variables.

[0111] The training end condition can be set according to actual needs. For example, the training end condition can refer to the number of iterative training reaching the number threshold, and the number threshold can be set according to actual needs, for example, the number threshold can be set to 10000. In another example, the training end condition can also refer to the target distribution P between two consecutive iterations changing less than 0.1%.

[0112] Furthermore, before determining the obtained deep convolutional embedded clustering model as the target model, in order to verify the prediction performance of the obtained deep convolutional embedded clustering model, the above prediction dataset can also be used for testing. For example, the prediction dataset can be used as the input feature of the obtained deep convolutional embedded clustering model, and the Hungarian algorithm can be used to calculate the clustering distribution c covering all possible clusters. i With the true label l i There is a one-to-one mapping between them, as shown in formula (12).

[0113]

[0114] Among them, p(c i ) represents the cluster assignment c i probability.

[0115] In one embodiment, the electronic device verifies the target model. In the real data set, the clustering discrimination accuracy of the target model is about 81.8%. The clustering results are as follows: Figure 5 shown.

[0116] In an embodiment of the present application, multiple groups of sample data are obtained, and each group of sample data in the multiple groups of sample data includes a sample arterial blood pressure signal and a sample cerebral blood flow signal. Based on each group of sample data, the second TFA variable corresponding to each group of sample data is determined to obtain multiple groups of second TFA variables, and each group of second TFA variables is used to describe the state of the corresponding CA function. Based on the multiple groups of second TFA variables, the deep convolution embedded clustering model is iteratively trained to obtain a target model. The target model can automatically adjust the TFA variables corresponding to the CA function based on any cerebral blood flow to monitor whether any CA function is damaged. The target model includes a first branch and a second branch. The first branch is used to determine the clustering loss of the TFA variable, and the second branch is used to determine the reconstruction loss of the TFA variable. The target model is obtained by training the model based on the reconstruction loss and the clustering loss, which improves the accuracy of monitoring compared to the logistic regression model.

[0117] In addition, the method provided in the embodiment of the present application has an extremely low number of floating-point operations, and the predicted operation time is in the millisecond level, which can meet real-time requirements.

[0118] Figure 6 FIG. 1 is a flow chart of a method for monitoring a CA function according to an exemplary embodiment. The method can be applied to the above electronic device. The method may include the following steps:

[0119] Step 601: Based on the arterial blood pressure signal and the cerebral blood flow signal, determine the first TFA variable used to represent the current state of the cerebral blood flow autoregulation CA function to be monitored, the arterial blood pressure signal is the input signal of the CA function, and the cerebral blood flow signal is the output signal of the CA function.

[0120] The process of collecting arterial blood pressure signals and cerebral blood flow signals can refer to the above embodiment. In addition, the process of determining the first TFA variable based on the arterial blood pressure signals and cerebral blood flow signals can also refer to the above embodiment, which will not be repeated here.

[0121] That is, when it is necessary to monitor the CA function of an individual, the arterial blood pressure signal and the cerebral blood flow signal of the individual may be collected, and then the first TFA variable may be determined based on the collected arterial blood pressure signal and the cerebral blood flow signal.

[0122] Step 602: Based on the first TFA variable, determine whether the CA function is damaged through the target model. The target model can monitor whether any CA function is damaged based on the TFA variable corresponding to any CA function. The target model includes a first branch and a second branch. The first branch is used to determine the clustering loss of the first TFA variable, and the second branch is used to determine the reconstruction loss of the first TFA variable.

[0123] The target model is obtained through the above Figure 1After obtaining the first TFA variable, the first TFA variable can be used as an input feature of the target model and input into the target model to perform prediction through the target model and output a prediction result, the prediction result including a result indicating that the CA function is impaired or a result indicating that the CA function is not impaired, that is, whether the CA function is impaired can be determined according to the prediction result.

[0124] In an embodiment of the present application, based on the arterial blood pressure signal and the cerebral blood flow signal, a first transfer function analysis TFA variable is determined to represent the current state of the cerebral blood flow autoregulation CA function to be monitored, the arterial blood pressure signal is the input signal of the CA function, and the cerebral blood flow signal is the output signal of the CA function. Based on the first TFA variable, whether the CA function is damaged is determined by the target model. The target model can monitor whether any CA function is damaged based on the TFA variable corresponding to any CA function. The target model includes a first branch and a second branch. The first branch is used to determine the clustering loss of the first TFA variable, and the second branch is used to determine the reconstruction loss of the first TFA variable. The target model is obtained by training the model based on the reconstruction loss and the clustering loss. Compared with the logistic regression model, the accuracy of monitoring is improved.

[0125] Figure 7 This is a schematic diagram of a device for monitoring the automatic regulation function of cerebral blood flow according to an exemplary embodiment. The device can be implemented by software, hardware or a combination of both, and is referred to as part or all of the above electronic device. The device may include:

[0126] An acquisition module 710, configured to acquire multiple sets of sample data, each set of sample data in the multiple sets of sample data including a sample arterial blood pressure signal and a sample cerebral blood flow signal;

[0127] A variable determination module 720, configured to determine, based on each set of sample data, a second TFA variable corresponding to each set of sample data, to obtain multiple sets of second TFA variables, each set of second TFA variables being used to describe a state of a corresponding CA function;

[0128] The training module 730 is used to iteratively train the deep convolution embedded clustering model based on the multiple groups of second TFA variables to obtain the target model, wherein the deep convolution embedded clustering model includes the first branch and the second branch.

[0129] As an example of the present application, the training module 730 is used to:

[0130] In each iteration, a specified number threshold groups of second TFA variables from the plurality of groups of second TFA variables are used as input features of the deep convolutional embedded clustering model;

[0131] Clustering the input features through the first branch, and performing gradient back propagation of the first branch according to the clustering result;

[0132] Reconstructing the input features through the second branch to obtain a feature reconstruction result, and performing gradient back propagation of the second branch according to the feature reconstruction result;

[0133] When the iterative training meets the training end condition, the obtained deep convolution embedded clustering model is determined as the target model.

[0134] As an example of the present application, the training module 730 is used to:

[0135] Performing encoding processing on the input features by the encoding module;

[0136] The deep embedded clustering module performs clustering processing on the features obtained after the encoding processing by the encoding module.

[0137] As an example of the present application, the training module 730 is used to:

[0138] The decoding module performs feature reconstruction on the features obtained after the encoding process by the encoding module to obtain the feature reconstruction result.

[0139] As an example of the present application, the encoding module includes N branches of convolution kernels with different receptive field sizes, where N is an integer greater than or equal to 2; the training module 730 is used to:

[0140] The input features are respectively convolved through each of the multiple branches in the encoding module to achieve encoding of the input features.

[0141] As an example of the present application, the decoding module includes N branches of convolution kernels with different receptive field sizes, where N is an integer greater than or equal to 2; the training module 730 is used to:

[0142] Performing convolution processing on the features obtained after the encoding processing of the encoding module through each of the N branches in the decoding module;

[0143] On each branch in the second branch, a convolution result of a specified convolution layer in the encoding module is combined with a convolution result of the same channel in the decoding module by using a skip connection method;

[0144] The convolution results output by each branch included in the decoding module are fused to obtain the feature reconstruction result.

[0145] As an example of the present application, during iterative training, the optimizer used for the gradient back propagation in the training process before a preset number of iterations is the Adam optimizer, and the optimizer used for the gradient back propagation in the training process after the preset number of iterations is the stochastic gradient descent SGD optimizer.

[0146] In an embodiment of the present application, multiple groups of sample data are obtained, and each group of sample data in the multiple groups of sample data includes a sample arterial blood pressure signal and a sample cerebral blood flow signal. Based on each group of sample data, the second TFA variable corresponding to each group of sample data is determined to obtain multiple groups of second TFA variables, and each group of second TFA variables is used to describe the state of the corresponding CA function. Based on the multiple groups of second TFA variables, the deep convolution embedded clustering model is iteratively trained to obtain a target model. The target model can automatically adjust the TFA variables corresponding to the CA function based on any cerebral blood flow to monitor whether any CA function is damaged. The target model includes a first branch and a second branch. The first branch is used to determine the clustering loss of the TFA variable, and the second branch is used to determine the reconstruction loss of the TFA variable. The target model is obtained by training the model based on the reconstruction loss and the clustering loss, which improves the accuracy of monitoring compared to the logistic regression model.

[0147] Figure 8 : is a schematic diagram of a structure of a device for training a target model according to an exemplary embodiment. The device can be implemented by software, hardware, or a combination of both, and is referred to as part or all of the above electronic device. The device may include:

[0148] A first determination module 810 is used to determine a first transfer function analysis TFA variable for representing a current state of a cerebral blood flow autoregulation (CA) function to be monitored based on an arterial blood pressure signal and a cerebral blood flow signal, wherein the arterial blood pressure signal is an input signal of the CA function and the cerebral blood flow signal is an output signal of the CA function;

[0149] The second determination module 820 is used to determine whether the CA function is damaged based on the first TFA variable through a target model. The target model can monitor whether any CA function is damaged based on the TFA variable corresponding to any CA function. The target model includes a first branch and a second branch. The first branch is used to determine the clustering loss of the first TFA variable, and the second branch is used to determine the reconstruction loss of the first TFA variable.

[0150] In an embodiment of the present application, based on the arterial blood pressure signal and the cerebral blood flow signal, a first transfer function analysis TFA variable is determined to represent the current state of the cerebral blood flow autoregulation CA function to be monitored, the arterial blood pressure signal is the input signal of the CA function, and the cerebral blood flow signal is the output signal of the CA function. Based on the first TFA variable, whether the CA function is damaged is determined by the target model. The target model can monitor whether any CA function is damaged based on the TFA variable corresponding to any CA function. The target model includes a first branch and a second branch. The first branch is used to determine the clustering loss of the first TFA variable, and the second branch is used to determine the reconstruction loss of the first TFA variable. The target model is obtained by training the model based on the reconstruction loss and the clustering loss. Compared with the logistic regression model, the accuracy of monitoring is improved.

[0151] Fig. 9 This is a schematic diagram of the structure of an electronic device provided by an embodiment of the present application. Fig. 9 As shown, the electronic device 9 of this embodiment includes: at least one processor 90 ( Fig. 9 Only one is shown in the figure), a memory 91, and a computer program 92 stored in the memory 91 and executable on the at least one processor 90, wherein the processor 90 implements the steps of any of the above-mentioned method embodiments when executing the computer program 92.

[0152] The electronic device 9 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The electronic device may include, but is not limited to, a processor 90 and a memory 91. Those skilled in the art will appreciate that Fig. 9 It is only an example of the electronic device 9 and does not constitute a limitation on the electronic device 9. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, it may also include input and output devices, network access devices, etc.

[0153] The processor 90 may be a CPU (Central Processing Unit), or other general-purpose processors, DSP (Digital Signal Processor), ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0154] In some embodiments, the memory 91 may be an internal storage unit of the electronic device 9, such as a hard disk or memory of the electronic device 9. In other embodiments, the memory 91 may also be an external storage device of the electronic device 9, such as a plug-in hard disk, SMC (Smart Media Card), SD (Secure Digital) card, Flash Card, etc. equipped on the electronic device 9. Further, the memory 91 may also include both an internal storage unit of the electronic device 9 and an external storage device. The memory 91 is used to store an operating system, an application program, a boot loader, data, and other programs, such as the program code of the computer program. The memory 91 may also be used to temporarily store data that has been output or is to be output.

[0155] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0156] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0157] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A method for monitoring cerebral blood flow autoregulation function, characterized in that: The method comprises: Determine a first transfer function analysis TFA variable for representing a current state of a cerebral blood flow autoregulation (CA) function to be monitored based on an arterial blood pressure signal and a cerebral blood flow signal, wherein the arterial blood pressure signal is an input signal of the CA function and the cerebral blood flow signal is an output signal of the CA function; Based on the first TFA variable, determining whether the CA function is damaged by using a target model, wherein the target model can monitor whether any CA function is damaged based on the TFA variable corresponding to any CA function, and the target model includes a first branch and a second branch, wherein the first branch is used to determine a clustering loss of the first TFA variable, and the second branch is used to determine a reconstruction loss of the first TFA variable; The first branch includes an encoding module and a deep embedded clustering module, the encoding module includes N branches with convolution kernels of different receptive field sizes, and N is an integer greater than or equal to 2; the second branch includes the encoding module and a decoding module, and the decoding module includes N branches with convolution kernels of different receptive field sizes; The target model is obtained by iteratively training a deep convolution embedded clustering model based on a second TFA variable corresponding to each group of sample data in a plurality of groups of sample data, wherein each group of sample data includes a sample arterial blood pressure signal and a sample cerebral blood flow signal, and each group of second TFA variables is used to describe the state of the corresponding CA function; the method further includes: In each iteration, a specified number threshold groups of second TFA variables from the plurality of groups of second TFA variables are used as input features of the deep convolutional embedded clustering model; Performing convolution processing on the input features through each of the multiple branches in the encoding module to achieve encoding of the input features, clustering the features obtained after the encoding processing of the encoding module through the deep embedded clustering module, and performing gradient backpropagation of the first branch according to the clustering result; The features obtained after the encoding processing of the encoding module are convolved through each of the N branches in the decoding module; on each branch in the second branch, the convolution result of the specified convolution layer in the encoding module is combined with the convolution result of the same channel in the decoding module by using a skip connection method; the convolution results output by each branch included in the decoding module are fused to obtain a feature reconstruction result, and the gradient backpropagation of the second branch is performed according to the feature reconstruction result; When the iterative training meets the training end condition, the obtained deep convolution embedded clustering model is determined as the target model.

2. A method for training a target model, characterized in that: The target model can monitor whether any CA function is damaged based on a TFA variable corresponding to any cerebral blood flow autoregulation CA function, the target model includes a first branch and a second branch, the first branch is used to determine the clustering loss of the TFA variable, and the second branch is used to determine the reconstruction loss of the TFA variable; the method includes: Acquire multiple groups of sample data, each group of sample data in the multiple groups of sample data includes a sample arterial blood pressure signal and a sample cerebral blood flow signal; Based on each group of sample data, determine the second TFA variable corresponding to each group of sample data to obtain multiple groups of second TFA variables, each group of second TFA variables is used to describe the state of the corresponding CA function; Iteratively training a deep convolutional embedded clustering model based on the multiple groups of second TFA variables to obtain the target model, wherein the deep convolutional embedded clustering model includes the first branch and the second branch; The first branch includes an encoding module and a deep embedded clustering module, the encoding module includes N branches with convolution kernels of different receptive field sizes, and N is an integer greater than or equal to 2; the second branch includes the encoding module and a decoding module, and the decoding module includes N branches with convolution kernels of different receptive field sizes; the deep convolution embedded clustering model is iteratively trained based on the multiple groups of second TFA variables to obtain the target model, including: In each iteration, a specified number threshold groups of second TFA variables from the plurality of groups of second TFA variables are used as input features of the deep convolutional embedded clustering model; Performing convolution processing on the input features through each of the multiple branches in the encoding module to achieve encoding of the input features, clustering the features obtained after the encoding processing of the encoding module through the deep embedded clustering module, and performing gradient backpropagation of the first branch according to the clustering result; The features obtained after the encoding processing of the encoding module are convolved through each of the N branches in the decoding module; on each branch in the second branch, the convolution result of the specified convolution layer in the encoding module is combined with the convolution result of the same channel in the decoding module by using a skip connection method; the convolution results output by each branch included in the decoding module are fused to obtain a feature reconstruction result, and the gradient backpropagation of the second branch is performed according to the feature reconstruction result; When the iterative training meets the training end condition, the obtained deep convolution embedded clustering model is determined as the target model.

3. The method according to claim 2, characterized in that During iterative training, the optimizer used for the gradient back propagation during the training process before the preset number of iterations is the Adam optimizer, and the optimizer used for the gradient back propagation during the training process after the preset number of iterations is the stochastic gradient descent SGD optimizer.

4. A device for monitoring the autoregulation function of cerebral blood flow, characterized in that: The device comprises: A first determination module is used to determine a first transfer function analysis TFA variable for representing a current state of a cerebral blood flow autoregulation (CA) function to be monitored based on an arterial blood pressure signal and a cerebral blood flow signal, wherein the arterial blood pressure signal is an input signal of the CA function and the cerebral blood flow signal is an output signal of the CA function; a second determination module, configured to determine whether the CA function is damaged based on the first TFA variable by using a target model, wherein the target model can monitor whether any CA function is damaged based on the TFA variable corresponding to any CA function, wherein the target model includes a first branch and a second branch, wherein the first branch is used to determine a clustering loss of the first TFA variable, and the second branch is used to determine a reconstruction loss of the first TFA variable; The first branch includes an encoding module and a deep embedded clustering module, the encoding module includes N branches with convolution kernels of different receptive field sizes, and N is an integer greater than or equal to 2; the second branch includes the encoding module and a decoding module, and the decoding module includes N branches with convolution kernels of different receptive field sizes; The target model is obtained by iteratively training the deep convolution embedded clustering model based on the second TFA variable corresponding to each group of sample data in multiple groups of sample data, wherein each group of sample data includes a sample arterial blood pressure signal and a sample cerebral blood flow signal, and each group of second TFA variables is used to describe the state of the corresponding CA function; the device is also used for: In each iteration, a specified number threshold groups of second TFA variables from the plurality of groups of second TFA variables are used as input features of the deep convolutional embedded clustering model; Performing convolution processing on the input features through each of the multiple branches in the encoding module to achieve encoding of the input features, clustering the features obtained after the encoding processing of the encoding module through the deep embedded clustering module, and performing gradient backpropagation of the first branch according to the clustering result; The features obtained after the encoding processing of the encoding module are convolved through each of the N branches in the decoding module; on each branch in the second branch, the convolution result of the specified convolution layer in the encoding module is combined with the convolution result of the same channel in the decoding module by using a skip connection method; the convolution results output by each branch included in the decoding module are fused to obtain a feature reconstruction result, and the gradient backpropagation of the second branch is performed according to the feature reconstruction result; When the iterative training meets the training end condition, the obtained deep convolution embedded clustering model is determined as the target model.

5. A device for training a target model, characterized in that: The target model can monitor whether any CA function is damaged based on a TFA variable corresponding to any cerebral blood flow autoregulation CA function, the target model includes a first branch and a second branch, the first branch is used to determine the clustering loss of the TFA variable, and the second branch is used to determine the reconstruction loss of the TFA variable, the device includes: An acquisition module, used for acquiring multiple groups of sample data, each group of sample data in the multiple groups of sample data includes a sample arterial blood pressure signal and a sample cerebral blood flow signal; a variable determination module, configured to determine, based on each set of sample data, a second TFA variable corresponding to each set of sample data, to obtain multiple sets of second TFA variables, each set of second TFA variables being used to describe a state of a corresponding CA function; A training module, configured to iteratively train a deep convolutional embedded clustering model based on the multiple groups of second TFA variables to obtain the target model, wherein the deep convolutional embedded clustering model includes the first branch and the second branch; The first branch includes an encoding module and a deep embedded clustering module, the encoding module includes N branches with convolution kernels of different receptive field sizes, and N is an integer greater than or equal to 2; the second branch includes the encoding module and a decoding module, the decoding module includes N branches with convolution kernels of different receptive field sizes; the training module is used to: In each iteration, a specified number threshold groups of second TFA variables from the plurality of groups of second TFA variables are used as input features of the deep convolutional embedded clustering model; Performing convolution processing on the input features through each of the multiple branches in the encoding module to achieve encoding of the input features, clustering the features obtained after the encoding processing of the encoding module through the deep embedded clustering module, and performing gradient backpropagation of the first branch according to the clustering result; The features obtained after the encoding processing of the encoding module are convolved through each of the N branches in the decoding module; on each branch in the second branch, the convolution result of the specified convolution layer in the encoding module is combined with the convolution result of the same channel in the decoding module by using a skip connection method; the convolution results output by each branch included in the decoding module are fused to obtain a feature reconstruction result, and the gradient backpropagation of the second branch is performed according to the feature reconstruction result; When the iterative training meets the training end condition, the obtained deep convolution embedded clustering model is determined as the target model.

Images