System for detecting volume in medical instrument

Abstract

The invention relates to the technical field of medical instrument volume detection, and discloses a medical instrument internal volume detection system. A continuous wave and a pulse modulation signal are simultaneously transmitted through an annular phased array microwave antenna array, three-dimensional dielectric constant distribution is reconstructed by adopting a compressed sensing algorithm, and Doppler time-frequency domain characteristics are extracted through short-time Fourier transform; and inputting the phase gradient field and Doppler velocity field features into a convolutional neural network to generate a four-dimensional space-time volume evolution model, separating a macroscopic volume trend and a microcosmic pulsation component by adopting an adaptive Kalman filtering algorithm, and predicting an abnormal hemodynamic event through a long-short-term memory network. The technical problem of real-time precise detection of the multi-dimensional volume parameters of the blood circulation medical instrument under the complex pulsating flow condition is solved, and accurate evaluation of the instantaneous volume value, the pulsation amount and the pulsating frequency spectrum and early warning of abnormal events are achieved.

Description

Technical Field

[0001] This invention relates to the field of medical device volume detection technology, and more specifically, to a medical device internal volume detection system. Background Technology

[0002] Blood circulation medical devices require real-time, precise monitoring of internal blood volume during operation to assess device status and hemodynamic parameters. These devices have a multi-chamber structure, and the internal blood exhibits complex pulsating flow characteristics during operation, accompanied by rapid volume changes, multi-chamber coupled vibrations, and nonlinear fluid dynamic effects.

[0003] Traditional volumetric detection methods mainly rely on single phase coherence detection or Doppler shift detection. While phase coherence detection can obtain the spatial distribution of dielectric constant inside the device, it lacks sufficient temporal resolution and cannot capture rapid pulsation changes; Doppler shift detection can sense blood flow motion, but lacks the ability to spatially reconstruct macroscopic volumetric distribution.

[0004] The limitations of this single detection mechanism make it impossible to simultaneously obtain macroscopic volume changes and microscopic pulsation details, thus making it impossible to accurately assess key hemodynamic parameters such as instantaneous pulse volume, ejection fraction, and pulsation spectrum, affecting the precise monitoring of the operating status of medical devices and the timely early warning of abnormal events. Summary of the Invention

[0005] This invention provides a medical device internal volume detection system, which solves the technical problem in related technologies where a single detection mechanism cannot simultaneously obtain macroscopic volume changes and microscopic pulsation details, resulting in the inability to accurately assess key hemodynamic parameters and provide timely warnings of abnormal events.

[0006] This invention discloses a method for detecting the internal volume of a medical device, comprising the following steps: acquiring the multi-chamber structural parameters and hemodynamic model data of the blood circulation device to be tested; calculating the phase response and Doppler frequency shift response of a microwave signal in the multi-chamber structure using the finite element method, and generating a phase-Doppler coupling response prediction matrix; controlling a ring phased array microwave antenna to simultaneously transmit a continuous wave signal and a pulse modulation signal; performing orthogonal demodulation on the received signal to extract the phase data matrix; performing mixing and Fourier transform on the pulse signal to extract the Doppler frequency shift time series; using the phase data matrix as observation data, reconstructing the three-dimensional dielectric constant distribution using a compressed sensing algorithm, specifically obtaining sparse coefficients by solving the L1 norm regularization optimization problem using an iterative soft thresholding algorithm, and reconstructing the dielectric constant distribution through sparse transform; extracting time-frequency domain pulsation features, including instantaneous dominant frequency, spectral bandwidth, and time-frequency energy distribution, from the Doppler frequency shift time series using a short-time Fourier transform; calculating the spatial gradient field of the three-dimensional dielectric constant distribution as the phase gradient field, based on the Doppler frequency shift... Doppler velocity field features are constructed by inverting the blood flow velocity field and calculating velocity divergence and velocity curl. The phase gradient field and Doppler velocity field features are input into a convolutional neural network, and a four-dimensional spatiotemporal volume evolution model is generated through a cross-channel attention fusion module and a spatiotemporal convolutional coding module. The total volume time series is obtained by spatial integration of the four-dimensional spatiotemporal volume evolution model, and a state-space model containing macroscopic volume trends and microscopic pulsation components is established. An adaptive Kalman filter algorithm is used to adaptively estimate the process noise covariance and observation noise variance through a sliding window method based on the innovation sequence, separating the macroscopic volume trend and microscopic pulsation components, and calculating the instantaneous volume value, pulsation volume, and pulsation spectrum. The feature sequence containing the instantaneous total volume, macroscopic trend, pulsation components, pulsation volume, instantaneous dominant frequency, spectral bandwidth, and time-frequency energy is input into a long short-term memory network, which outputs the probability of abnormal hemodynamic events. Based on the probability threshold, a multi-level early warning decision is generated, and a multi-parameter monitoring report containing instantaneous volume value, volume trend curve, pulsation volume, pulsation spectrum, and abnormal warning level is generated.

[0007] This invention discloses a medical device internal volume detection system for performing the aforementioned medical device internal volume detection method. The system includes a ring phased array microwave antenna array, a signal transmission and reception unit, a data processing unit, and an early warning output unit. The ring phased array microwave antenna array is arranged around the multi-chamber structure of a blood circulation device. The signal transmission and reception unit controls the antenna array to transmit continuous wave signals and pulse modulation signals, and collects phase data and Doppler frequency shift data. The data processing unit performs compressed sensing reconstruction, feature extraction, neural network fusion, Kalman filtering, and anomaly prediction. The early warning output unit generates a multi-parameter monitoring report and issues an early warning signal based on the anomaly probability.

[0008] Furthermore, the step of generating the phase-Doppler coupling response prediction matrix includes: dividing the multi-chamber structure into three-dimensional mesh elements, setting the dielectric constant and conductivity according to the material properties of each element; solving Maxwell's equations on the finite element mesh to calculate the electric field distribution and magnetic field distribution of the microwave signal, calculating the phase response based on the electric field distribution, and calculating the Doppler frequency shift response based on the blood flow velocity field and the microwave propagation direction; discretizing the volume change range and blood flow velocity range into multiple sets of simulation scenarios, calculating the phase response vector and Doppler frequency shift vector for each set of scenarios, and organizing the response vectors into a phase-Doppler coupling response prediction matrix, where each row of the matrix corresponds to a specific volume-flow velocity state.

[0009] Furthermore, the step of reconstructing the three-dimensional dielectric constant distribution using the compressed sensing algorithm includes: dividing the internal space of the multi-cavity chamber into three-dimensional voxels, defining the dielectric constant distribution vector as the unknown quantity; establishing a linear relationship between phase data and dielectric constant distribution based on electromagnetic wave propagation theory, and introducing a sparse transformation matrix to represent the dielectric constant distribution as a linear combination of sparse coefficients; constructing an optimization problem with the L1 norm of the sparse coefficients as the objective function and the L2 norm of the phase data fitting error as the constraint condition, and solving it using an iterative soft thresholding algorithm, adding a gradient correction term to the current estimate in each iteration and then applying the soft thresholding operator until the difference between the estimates of adjacent iterations is less than the convergence threshold; converting the obtained sparse coefficients into a three-dimensional dielectric constant distribution through sparse transformation, and segmenting and identifying blood regions and non-blood regions based on the dielectric constant threshold.

[0010] Furthermore, the convolutional neural network includes: a dual-channel input layer, which receives phase gradient field and Doppler velocity field features respectively; a feature extraction module, which applies three-dimensional convolutional layers to the two input channels to extract structural boundary features and flow pattern features respectively; a cross-channel attention fusion module, which performs global average pooling on the feature maps to obtain channel feature descriptors, calculates the attention weights of the two channels through fully connected layers and activation functions, applies the attention weights to the corresponding feature maps and concatenates them to form fused features; a spatiotemporal convolutional coding module, which stacks the multi-frame fused features from the time series into a four-dimensional tensor, and applies cascaded spatial and temporal convolutional kernels to extract spatiotemporal evolution features; and a volume prediction output layer, which maps the spatiotemporal features to the blood occupancy probability of each voxel at each time step through a fully convolutional network, and converts the probability into a binary label through a thresholding operation to obtain a four-dimensional spatiotemporal volume evolution model.

[0011] Furthermore, the training steps of the convolutional neural network include: generating a simulation scene with known volume distribution and flow velocity field based on a physical model, calculating the corresponding phase gradient and Doppler velocity field as training input, and using the real volume distribution as training label; using a weighted combination of Dice loss and cross-entropy loss as the loss function, which simultaneously considers the overlap between the predicted probability and the real label and the classification error; and using an optimizer that combines momentum method and adaptive learning rate adjustment for training, with the learning rate gradually reduced using an exponential decay strategy.

[0012] Furthermore, the adaptive Kalman filter algorithm includes: defining a state vector containing macroscopic volume trend and volume trend change rate; establishing a state transition equation and an observation equation, with the observation value being the total volume time series; performing Kalman filter iterations, including a prediction step and an update step; the prediction step calculates the state prior estimate and prior covariance; the update step calculates the Kalman gain and updates the state estimate and covariance based on the observation value; calculating the sample covariance of the innovation within the window using a sliding window method based on the innovation sequence; adaptively estimating the observation noise variance and process noise covariance; and smoothing the noise parameters to adapt the Kalman filter to the pulsation mode changes; the first component of the filtered state estimate is the macroscopic volume trend, and the difference between the total volume and the macroscopic trend is the microscopic pulsation component.

[0013] Furthermore, the step of calculating the pulsation quantity includes: performing autocorrelation analysis on the micro-pulsation components, calculating the autocorrelation function and finding the time delay corresponding to the first peak as the pulsation period; and calculating the difference between the maximum and minimum values ​​of the pulsation components within each pulsation period as the pulsation quantity of that period.

[0014] Furthermore, the input layer of the Long Short-Term Memory Network receives a time series, in which the feature vector at each moment includes the instantaneous total volume, macroscopic trend, pulsation component, pulsation quantity, instantaneous dominant frequency, spectral bandwidth, and time-frequency energy; the output layer of the Long Short-Term Memory Network maps the hidden state at the last moment to a scalar through a linear mapping layer, and outputs the probability of an abnormal event after passing through an activation function.

[0015] Furthermore, the step of converting the probability of an abnormal event into an early warning decision includes: determining a low-risk warning when the probability of an abnormal event is greater than or equal to a first threshold and less than a second threshold; determining a medium-risk warning when the probability of an abnormal event is greater than or equal to a second threshold and less than a third threshold; and determining a high-risk warning when the probability of an abnormal event is greater than or equal to a third threshold; wherein the first threshold is less than the second threshold, and the second threshold is less than the third threshold.

[0016] This invention acquires complementary information through phase-Doppler dual-modal coupling detection, achieves high-resolution three-dimensional reconstruction through compressed sensing algorithm, realizes deep fusion of multi-source information through convolutional neural network, achieves accurate separation of macroscopic and microscopic components through adaptive Kalman filtering, and achieves intelligent recognition of abnormal patterns through long short-term memory network. It solves the technical problem of real-time and precise detection of multi-dimensional volume parameters under complex pulsatile flow conditions in blood circulation medical devices, and achieves the technical effects of accurate evaluation of key hemodynamic parameters such as instantaneous volume value, pulsatility, ejection fraction, and pulsation spectrum, and early warning of abnormal hemodynamic events. Attached Figure Description

[0017] Figure 1 This is the main flowchart of the medical device internal volume detection method of the present invention; Figure 2 This is a flowchart of the sub-steps of step 300 of the present invention; Figure 3 This is a flowchart of the sub-steps of step 400 of the present invention; Figure 4 This is a flowchart of a sub-step of step 500 of the present invention. Detailed Implementation Detailed Implementation Method 1 Blood circulation medical devices (such as novel artificial heart pumps, extracorporeal membrane oxygenation (ECMO) devices, and intelligent hemodialysis machines) require real-time and precise monitoring of internal blood volume during operation to assess device status and hemodynamic parameters. These medical devices have a multi-chamber structure, and the internal blood exhibits complex pulsating flow characteristics during operation, accompanied by rapid volume changes, multi-chamber coupled vibrations, and nonlinear fluid dynamic effects.

[0019] Traditional volumetric detection methods primarily rely on either phase-coherent detection or Doppler shift detection. While phase-coherent detection can acquire the spatial distribution of the dielectric constant within the device, its temporal resolution is insufficient, making it unable to capture rapid pulsation changes. Doppler shift detection can sense blood flow motion but lacks the ability to spatially reconstruct macroscopic volumetric distribution. These limitations of a single detection mechanism prevent the simultaneous acquisition of macroscopic volumetric changes and microscopic pulsation details, thus hindering the accurate assessment of key hemodynamic parameters such as instantaneous pulse volume, ejection fraction, and pulsation spectrum. This impacts the precise monitoring of medical device operation and the timely warning of abnormal events.

[0020] The medical device internal volume detection method provided in this embodiment is applied to the real-time monitoring scenario of blood circulation medical devices. Before the method is executed, the blood circulation medical device has been equipped with a ring phased array microwave antenna array, which is arranged around the multi-chamber structure of the device for transmitting and receiving microwave signals.

[0021] Step 100: Obtain multi-chamber structural parameters and hemodynamic model data, and generate a phase-Doppler coupling response prediction matrix.

[0022] Obtain the multi-chamber structural parameters of the blood circulation device under test, including the geometric dimensions, wall thickness, material dielectric constant, and inter-chamber connectivity of each chamber. Simultaneously, obtain the hemodynamic model data of the device, including the blood dielectric constant, expected flow velocity range, pulsation frequency range, and pressure-volume relationship curves for each chamber.

[0023] Based on the above parameters, the propagation characteristics of microwave signals in a multi-chamber structure were calculated using the finite element method. Specifically, the multi-chamber structure was divided into three-dimensional mesh elements, and the dielectric constant of each element was set. and conductivity For the blood-filled chamber region, the dielectric constant is set to the blood dielectric constant. (Typically 60-80); For instrument wall structures, the dielectric constant is set to the material dielectric constant. .

[0024] Solve Maxwell's equations on a finite element mesh to calculate the electric field distribution of a 2.4 GHz microwave signal. and magnetic field distribution ,in This is a spatial position vector. The phase response of each antenna receiving point is calculated based on the electric field distribution. According to the blood flow velocity field Calculate the Doppler frequency shift response ,in The transmission frequency is 2.4 GHz. At the speed of light, This is the unit vector for the direction of microwave propagation.

[0025] Different volume states (volume variation range discretized into several representative points) and different blood flow velocity fields (velocity range discretized into several representative points) are combined to form multiple simulation scenarios. The phase response vector is calculated for each scenario. and Doppler frequency shift vector ,in This represents the number of phase detection channels. This represents the number of Doppler detection channels. These response vectors are organized into a phase-Doppler coupled response prediction matrix: ; in The total number of simulation scenarios is represented by a matrix, with each row corresponding to a specific volume-flow rate state. This matrix establishes a mapping relationship between the physical state of the instrument and its microwave response, providing prior information for subsequent volume reconstruction.

[0026] Step 200: Simultaneously transmit continuous wave and pulse modulation signals using a ring phased array microwave antenna, and acquire phase coherence data and Doppler frequency shift time series.

[0027] The antenna elements in the circular phased array microwave antenna are controlled to transmit 2.4 GHz continuous wave signals in some elements and pulse-modulated signals in others. The continuous wave signals are used for phase coherence detection, and the pulse-modulated signals are used for Doppler frequency shift detection.

[0028] For continuous wave signal transmission, the transmission power is set to 10-20 dBm, and a time-division multiplexing strategy is used to activate each transmitting antenna sequentially. When the... When one transmitting antenna is activated, the remaining antennas act as receiving antennas to receive transmitted or reflected signals. The received signals are then quadrature demodulated to extract the in-phase components. and orthogonal components Calculate phase data Traverse all transmitting antennas to obtain... 3D phase data matrix , of which Line 1 Column elements represent the first The antenna transmits, the first Phase value when received by each antenna.

[0029] For pulse-modulated signal transmission, the pulse repetition frequency is set to 1-10 kHz, and the pulse width is set to 1-10 μs. After the transmitting antenna emits the pulse signal, the receiving antenna receives the Doppler frequency shift signal caused by blood flow. The received signal is mixed to down-convert the radio frequency signal to baseband. The baseband signal is sampled, with the sampling rate set to at least 5 times the maximum value of the Doppler frequency shift (according to the Nyquist sampling theorem). The sampled signal is then subjected to a Fourier transform to extract the Doppler frequency shift spectrum. ,in For frequency variables, The time variable is recorded. The Doppler frequency shift time series of each receiving channel is recorded. ,in This represents the number of receive channels.

[0030] Continuous wave and pulse signals are transmitted using frequency division or time division multiplexing to avoid mutual interference. The acquisition process is continuous, and the sampling time interval is determined according to the blood flow pulsation frequency, usually set to less than 1 / 10 of the pulsation period to capture pulsation details.

[0031] Step 300: Reconstruct the three-dimensional dielectric constant distribution based on the compressed sensing algorithm, and extract the Doppler time-frequency domain pulsation features through short-time Fourier transform.

[0032] Step 310: Process phase data to reconstruct the three-dimensional dielectric constant distribution.

[0033] The phase data matrix acquired in step 200 As observation data. Due to the number of measurement channels of the loop antenna array. The number of voxels is much smaller than the total number of voxels after discretization in three-dimensional space, making this an underdetermined inverse problem. A compressed sensing algorithm is used for solving it.

[0034] First, the internal space of the multi-cavity chamber is divided into There are three-dimensional voxels, and the dielectric constant of each voxel is the unknown quantity. Define the dielectric constant distribution vector. According to electromagnetic wave propagation theory, there is a linear relationship between phase data and dielectric constant distribution (under the Born approximation): ; in The sensitivity matrix has the following elements. Indicates the first The change in dielectric constant of an individual element affects the first The influence coefficient of the phase of each measurement channel can be obtained through finite element simulation in step 100; To measure the noise vector.

[0035] Because the dielectric constant of biological tissues and blood exhibits sparsity in certain transform domains (such as the discrete cosine transform domain or wavelet transform domain), a sparse transform matrix is ​​introduced. , making ,in Let be a sparse coefficient vector. The compressed sensing reconstruction problem is transformed into solving the following optimization problem: ; in It uses the L1 norm to promote sparsity; The L2 norm is used to constrain the data fitting error. This is the noise tolerance threshold, set based on the measured signal-to-noise ratio.

[0036] An iterative soft thresholding algorithm is used to solve the above optimization problem. Initialization The iterative formula is: ; in The soft threshold operator is defined as follows: ; The step size parameter can be set either through line search or by fixing. This is the regularization parameter, determined based on cross-validation. The iteration termination condition is... ,in This is the convergence threshold.

[0037] Obtaining sparsity coefficients Afterwards, through Reconstruct the three-dimensional dielectric constant distribution. Based on the dielectric constant threshold segmentation, identify blood regions (dielectric constant approximately 60-80) and non-blood regions (dielectric constant less than 10) to obtain the three-dimensional volume distribution.

[0038] Step 320: Extract Doppler time-frequency domain pulsation features.

[0039] The Doppler shift time series acquired in step 200 As input, since blood flow pulsation has time-varying characteristics, short-time Fourier transform (STFT) is used to analyze the time-frequency domain features.

[0040] For each receiving channel Choose the Hanning window function Window length Set to 1 / 5 to 1 / 3 of the expected pulsation period. The short-time Fourier transform is defined as: ; In the discrete implementation, the time series is divided into frames, with a frame length of... Each sampling point, frame shift is Sampling points (usually taken) (To ensure time-frequency resolution balance). Perform a Fast Fourier Transform (FFT) on each frame of data to obtain the spectrum at that moment. ,in For the first The center moment of the frame.

[0041] An energy-weighted average of the time spectra of all receiving channels is performed to obtain a comprehensive time-frequency diagram: ; Extracting pulsation feature parameters from the time-frequency plot: - Instantaneous dominant frequency At every moment Find the frequency corresponding to the peak value in the spectrum, i.e. - Spectrum bandwidth Calculate the 3dB bandwidth of the spectral energy - time-frequency energy distribution. : Indicates the intensity of the pulsation at each moment. These time-frequency domain features quantify the dynamic change patterns of blood flow pulsation, providing input for subsequent fusion analysis.

[0042] Step 400: Input the phase gradient field and Doppler velocity field into the convolutional neural network to generate a fused four-dimensional spatiotemporal volume evolution model.

[0043] Step 410: Calculate the phase gradient field and the Doppler velocity field.

[0044] Three-dimensional dielectric constant distribution reconstructed based on step 310 Calculate the spatial gradient field. Apply the three-dimensional gradient operator to the dielectric constant distribution: ; The magnitude of the gradient A peak appears at the blood-tissue interface to identify the volume boundary. The gradient direction indicates the interface normal.

[0045] Since phase measurement is directly related to the dielectric constant distribution, the phase gradient field is defined as: ; in Obtained through the reconstruction results of interpolation step 310.

[0046] Based on the Doppler frequency shift data extracted in step 320, the blood flow velocity field is calculated. The Doppler effect formula is then used. (in (The angle between the velocity direction and the microwave propagation direction), inverting the velocity components: ; in For the first Velocity components in each detection direction This is the angle between the direction and the velocity vector. Since there are multiple detection directions, the velocity components of each direction are fused using the least squares method to reconstruct the three-dimensional velocity field. .

[0047] Constructing the Doppler velocity field characteristic tensor ,in It is the velocity divergence (characterizing fluid compression or expansion). It is the velocity curl (characterizing fluid rotation).

[0048] Step 420: Generate a four-dimensional spatiotemporal volume evolution model using a convolutional neural network.

[0049] The phase gradient field obtained in step 410 Doppler velocity field characteristics Data preprocessing is performed. Since the phase gradient field (measured in radians per meter) and velocity field features (including velocity components measured in meters per second, and divergence and curl measured in seconds) have different dimensions and numerical ranges, directly inputting them into the neural network would lead to an uneven impact of different features on network training. Therefore, Z-score standardization is applied to both input channels to convert the feature values ​​of each channel into a standard distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of dimensional differences on network training.

[0050] The preprocessed phase gradient field and Doppler velocity field features are used as inputs to a convolutional neural network. This convolutional neural network is of very low scale and is used to fuse spatial structure information and dynamic flow information.

[0051] The network structure includes the following components: (1) Dual-channel input layer: receives phase gradient fields respectively Doppler velocity field characteristics (Including the three components of velocity, the six components of divergence and curl).

[0052] (2) Feature extraction module: Three-dimensional convolutional layers are applied to both input channels. The phase channel uses a convolutional kernel size of [missing value]. The filter count is 64, the activation function is ReLU, and structural boundary features are extracted. The velocity channel uses the kernel size. The filter count is 64, the activation function is ReLU, and the flow pattern features are extracted. .

[0053] (3) Cross-channel attention fusion module: Calculates the attention weights between two feature maps. First, global average pooling is performed on the feature maps to obtain channel feature descriptors. and Attention weights are calculated using a fully connected layer: ; ; in This indicates a splicing operation. It is the ReLU activation function. It is the Sigmoid activation function. This is a trainable weight matrix. The attention weights are applied to the feature map: ; in This represents element-wise multiplication (broadcast mechanism). It involves concatenating weighted feature maps. .

[0054] (4) Spatiotemporal convolutional coding module: converts feature maps from multiple frames in the time series. Stacked as a four-dimensional tensor The spatiotemporal evolution features are extracted by applying 3D convolution plus temporal convolution (i.e., 3D+1D decomposition convolution). Specifically, using... Spatial convolution kernel and Concatenated temporal convolution kernels capture both spatial correlations and temporal dynamics.

[0055] (5) Volume Prediction Output Layer: A fully convolutional network maps spatiotemporal features to volume prediction probabilities. The network output is the probability of blood occupancy for each voxel at each time step. The probability is converted into a binary label through a thresholding operation: when If a region is identified as blood-occupied (marked as 1), it is identified as a non-blood region (marked as 0), thus obtaining a four-dimensional spatiotemporal volume evolution model. .

[0056] The input layer of this convolutional neural network is a dual-channel input layer that receives the phase gradient field. Doppler velocity field characteristics The output layer is a volume prediction layer, which outputs the blood occupancy probability of each voxel. After thresholding, a four-dimensional spatiotemporal volume evolution model is obtained. , where 1 represents blood-occupied areas and 0 represents non-blood-occupied areas.

[0057] The convolutional neural network is trained using a supervised learning model. Training data is generated through simulation: based on the physical model from step 100, a scene with known volume distribution and flow velocity field is generated, and the corresponding phase gradient and Doppler velocity field are calculated as inputs, with the real volume distribution serving as the label. The loss function is a weighted combination of Dice loss and cross-entropy loss. ; in To predict probabilities, The labels are real labels. The optimization strategy uses the Adam optimizer, which combines momentum and adaptive learning rate adjustment. The learning rate is initially set to 0.001 and uses an exponential decay strategy to decrease to 0.5 times the original value every 20 epochs. The training lasts for a total of 100 epochs, and the batch size is set to 8.

[0058] This convolutional neural network achieves deep fusion of phase spatial information and Doppler dynamic information, generating a four-dimensional spatiotemporal volume evolution model that simultaneously includes the spatial distribution and temporal evolution of volume.

[0059] The cross-channel attention fusion module of the aforementioned convolutional neural network can also employ other fusion methods. For example, the two feature maps can be directly concatenated and then convolved, i.e. This method omits the attention weight calculation process. While this direct concatenation method has lower computational complexity, its fusion effect is not as good as the adaptive allocation of attention weights.

[0060] The 3D+1D decomposition convolution in the aforementioned spatiotemporal convolutional coding module can also be replaced with a complete four-dimensional convolution. That is, directly using... The convolution kernel performs convolution operations in four dimensions of space and time. This complete four-dimensional convolution can capture spatiotemporal coupling patterns, but the number of parameters increases significantly, requiring a larger training dataset.

[0061] Step 500: Use the adaptive Kalman filter algorithm to separate the macroscopic volume trend and the microscopic pulsation component, and calculate the instantaneous volume value, pulsation amount and pulsation spectrum.

[0062] The four-dimensional spatiotemporal volume evolution model generated in step 400 It contains complete spatiotemporal information about the volume. Spatial integration of this model yields the total volume time series: ; in The volume of a single voxel.

[0063] The total volume time series It comprises two components: macroscopic volumetric trends (such as the average filling volume of the device) and microscopic pulsation components (periodic fluctuations caused by heartbeats, pumping, etc.). An adaptive Kalman filter algorithm is used for separation.

[0064] The adaptive Kalman filter algorithm is an unconventional algorithm that adds an online adaptive estimation process for the noise covariance to the standard Kalman filter. The algorithm includes the following steps: Step 510: Establish the state-space model.

[0065] Define state vector ,in For the first Macroscopic volume trend at any given moment Let be the rate of change of the volume trend. The state transition equation is: ; Where the state transition matrix , The sampling time interval, The process noise has a mean of zero and a covariance matrix of... The Gaussian distribution.

[0066] The observation equation is: ; Among the observed values Observation matrix , To observe the noise, it follows a pattern with a mean of zero and a variance of . The Gaussian distribution. Observation noise corresponds to the microscopic pulsation component.

[0067] Step 520: Perform Kalman filter iteration.

[0068] Initial state estimation Initial covariance matrix (Identity matrix).

[0069] For time Perform the following iterations: (a) Prediction steps: ; ; (b) Update steps: ; ; ; in This is the Kalman gain.

[0070] Step 530: Adaptively estimate the noise covariance.

[0071] In standard Kalman filtering, the process noise covariance and observation noise variance This is a fixed value. In adaptive Kalman filtering, it is based on the innovation sequence. Adaptive estimation of noise parameters.

[0072] The sample covariance of the innovation within the window is calculated using the sliding window method: ; in This is the window length (typically 10-50 sampling points). Update the observation noise variance. ,in It is a smoothing coefficient (usually taken as 0.9-0.95).

[0073] Similarly, the noise covariance during the update process: ; By adaptively adjusting noise parameters, Kalman filtering can adapt to changes in pulsation patterns and improve separation accuracy.

[0074] Step 540: Calculate the instantaneous volume, pulsation volume, and pulsation spectrum.

[0075] Filtered state estimation The first component is the macroscopic volume trend. The micro-pulsation components are: ; The instantaneous volume value is the total volume. It also outputs macro trends and pulsation components for subsequent analysis.

[0076] Pulsatility is defined as the difference between the maximum and minimum values ​​of a pulsating component within one pulse cycle. First, the pulse cycle is detected: for the pulsating component... Autocorrelation analysis is performed, which takes the pulsating component time series as input and calculates the autocorrelation function. Find the time delay corresponding to the first peak of the autocorrelation function. Output pulsation period Calculate within each period: ; in For the first The amount of pulse in each cycle.

[0077] Pulsating spectrum through Fourier transform yields: ; in The duration of observation is given. The pulsation spectrum reflects the distribution of pulsation energy at different frequencies, with the main peak frequency corresponding to the fundamental pulsation frequency and the harmonic peaks reflecting the non-sinusoidal characteristics of the pulsation waveform.

[0078] The adaptive estimation of noise covariance in the aforementioned adaptive Kalman filter algorithm can also employ other adaptive methods. For example, the variance matching method based on the residual sequence can be used to calculate the theoretical residual covariance. Adjust the variance by minimizing the difference between the actual residual sample covariance and the actual residual sample covariance. and This method exhibits better robustness when noise statistics change drastically.

[0079] To improve time resolution, the aforementioned pulsation spectrum calculation can employ a short-time Fourier transform instead of a global Fourier transform. Specifically, a Hanning window with a window length of 2-3 pulsation periods is selected. Perform sliding window processing and calculate the power spectral density for each window. This allows us to observe the evolution of the pulsation spectrum over time, promptly detect abnormal changes in the spectrum, and provide more detailed input features for subsequent anomaly prediction.

[0080] Step 600: Analyze the volumetric pulsation pattern sequence based on long short-term memory network to predict abnormal hemodynamic events and generate a multi-parameter monitoring report.

[0081] The multiple time-series parameters calculated in step 500 are organized into a pulsating pattern feature sequence. At each time step... Define the eigenvector: ; The components are instantaneous total volume, macroscopic trend, pulsation component, pulsation volume, instantaneous main frequency extracted in step 320, spectral bandwidth, and time-frequency energy, respectively.

[0082] Because the dimensions and numerical ranges of the components in the feature vector differ significantly (volume parameter is measured in milliliters, frequency parameter in Hertz, and energy parameter in energy units), directly inputting them into the neural network would lead to features with larger numerical values ​​dominating the training process. Therefore, mean normalization based on the range is applied to each component of the feature vector, mapping each feature component to the interval between 0 and 1, thus eliminating the influence of differences in dimensions and numerical ranges.

[0083] Time window The normalized feature vectors within are organized into sequences. ,in The sequence length (usually the length containing 3-5 pulse cycles).

[0084] The preprocessed feature sequences are input into a Long Short-Term Memory (LSTM) network for analysis. LSTM is a conventional model that includes three gating units: a forget gate, an input gate, and an output gate, and is capable of capturing long-term dependencies in time series.

[0085] The input layer of LSTM is defined as: receiving time series , where the feature vector at each time step It includes seven components: instantaneous total volume, macroscopic trend, pulsation component, pulse volume, instantaneous dominant frequency, spectral bandwidth, and time-frequency energy, with a sequence length of [missing information]. Set the duration to include 3-5 pulse cycles.

[0086] The output layer of LSTM is defined as: the hidden state at the last time step is mapped through a linear mapping layer. (The number of hidden units is set to 128) Mapped to scalar anomaly probability, the weight matrix of this linear mapping layer is: , bias is After passing through the Sigmoid activation function, the output is the probability of abnormal events occurring. .

[0087] The LSTM was trained using a supervised learning model. Training data was derived from a clinical database, containing time-series samples of normal and abnormal hemodynamic events. The loss function used was binary cross-entropy. ; in The labels are real numbers. The optimization strategy uses the Adam optimizer with a learning rate of 0.0001, a training period of 50 epochs, and a batch size of 32.

[0088] LSTM output probability of abnormal events This needs to be translated into specific early warning decisions. A threshold-based judgment method is used: when... When the system determines that an abnormal state has occurred, it issues a warning signal; when The system is then judged to be in a normal state. To improve the reliability of the early warning, a three-level early warning mechanism is further established: when... When the risk level is low, At the time, a medium-risk warning was issued. When a high-risk warning is issued, different risk levels correspond to different alarm intensity and intervention recommendations.

[0089] Based on the LSTM output and the calculation results from step 500, a multi-parameter monitoring report is generated. The report includes the following: - The instantaneous volume value at the current moment. - Macroscopic volume trend curve and its rate of change - the pulsation volume of the most recent cycle Mean Pulse Quantity - Pulse Spectrum and main frequency - Probability of abnormal events And the corresponding warning level (normal, low risk, medium risk or high risk) - 3D visualization of multi-chamber volume distribution (based on the 4D model in step 400 at the current moment). The monitoring report is updated every pulse cycle, allowing medical staff to monitor the operating status of medical devices in real time.

[0090] To improve the interpretability of anomaly predictions, the aforementioned LSTM network can have an attention weight calculation module added before the output layer. Specifically, this module calculates the contribution weight of each time step in the sequence to the final prediction. For each time step... Hidden state The weights are calculated using the attention scoring function: ; in , and These are trainable parameters. Weighted summation of hidden states: ; Will The anomaly probability is calculated using the input linear mapping layer. Attention weights. It reflects the importance of abnormality identification at each moment, can be visualized in the monitoring report, and helps medical staff understand the cause of abnormalities.

[0091] Technical effects of this embodiment: This implementation overcomes the limitations of single detection methods by employing a phase-Doppler dual-mode coupling detection method. Phase coherence detection, based on the phase delay difference of electromagnetic waves in media with different dielectric constants, can accurately reconstruct the three-dimensional dielectric constant distribution inside the blood circulation device, thereby obtaining spatial structural information of the macroscopic volume. Doppler frequency shift detection, based on the frequency modulation effect of moving blood on microwaves, can capture the dynamic pulsation characteristics of blood flow, and quantify the instantaneous frequency, bandwidth, and energy evolution of the pulsation through time-frequency domain features extracted by short-time Fourier transform. The synchronous acquisition of the two detection methods ensures the temporal alignment of spatial and temporal information, providing a complementary data foundation for subsequent fusion analysis.

[0092] The compressed sensing reconstruction algorithm leverages the sparsity prior of the dielectric constant of biological tissue in the sparse transform domain and promotes sparse solutions through L1 norm regularization, achieving accurate reconstruction of the three-dimensional dielectric constant distribution under underdetermined inverse problem conditions. This algorithm overcomes the ill-posedness problem of traditional least squares methods when the number of measurement channels is much smaller than the number of unknowns, significantly improving spatial resolution. Simultaneously, the iterative soft thresholding process achieves a balance between data fitting and sparsity through alternating gradient descent and threshold shrinkage, avoiding overfitting and underfitting.

[0093] The cross-channel attention fusion module of the convolutional neural network adaptively assigns fusion weights to the two features by learning the correlation between the phase gradient field and the Doppler velocity field. When the device is in a static or slowly changing state, the weight of phase information increases; when blood flow pulsation is intense, the weight of Doppler information increases. This data-driven adaptive fusion strategy overcomes the performance instability problem of fixed-weight fusion methods under different operating conditions. The spatiotemporal convolutional coding module captures both spatial structural correlation and temporal evolution dynamics through 3D+1D decomposition convolution. The generated four-dimensional spatiotemporal volume evolution model includes both the volume distribution of each chamber and reflects the continuous trajectory of volume change over time, providing complete spatiotemporal information for macro-micro component separation.

[0094] The adaptive Kalman filter algorithm overcomes the performance degradation problem of the standard Kalman filter under conditions of unknown or time-varying noise statistics by online estimation and dynamic adjustment of the noise covariance. The sliding window covariance estimation based on the innovation sequence can track changes in pulsation patterns. When the pulsation frequency or amplitude changes, the observed noise variance is adaptively adjusted to ensure the filter always operates in an optimal state. This algorithm accurately separates the slowly changing macroscopic volume trend from the rapidly fluctuating microscopic pulsation component. The separated pulsation signal is used to calculate key hemodynamic parameters such as pulsatility and pulsation spectrum, providing quantitative indicators for the functional evaluation of medical devices.

[0095] Long Short-Term Memory (LSTM) networks selectively retain and forget historical information through gating mechanisms, enabling them to learn long-term dependencies in pulsatile pattern sequences. Under normal hemodynamic conditions, pulsatile patterns exhibit periodicity and stability; however, when abnormal events occur (such as air embolism, pump failure, or chamber obstruction), the frequency, amplitude, and waveform shape of the pulsatile patterns show characteristic changes. By learning the characteristic representations of these abnormal patterns from training data, LSTM can promptly identify early abnormal signs in real-time monitoring, issuing warnings before clinical symptoms appear and providing a window of opportunity for medical intervention.

[0096] In summary, this implementation method acquires complementary information through phase-Doppler dual-modal coupling detection, achieves high-resolution three-dimensional reconstruction through compressed sensing algorithms, realizes deep fusion of multi-source information through convolutional neural networks, achieves accurate separation of macroscopic and microscopic components through adaptive Kalman filtering, and achieves intelligent identification of abnormal patterns through long short-term memory networks. It solves the technical problem of real-time and precise detection of multi-dimensional volume parameters under complex pulsatile flow conditions in blood circulation medical devices, and realizes accurate assessment of key hemodynamic parameters such as instantaneous volume value, pulsatility, ejection fraction, and pulsation spectrum, as well as early warning of abnormal hemodynamic events.

Claims

1. A method for detecting the internal volume of a medical device, characterized in that, Includes the following steps: The multi-chamber structural parameters and hemodynamic model data of the blood circulation device to be tested are obtained. The phase response and Doppler frequency shift response of the microwave signal in the multi-chamber structure are calculated by the finite element method, and the phase-Doppler coupling response prediction matrix is ​​generated. The system controls a circular phased array microwave antenna array to simultaneously transmit continuous wave signals and pulse modulated signals, performs orthogonal demodulation on the received signals to extract the phase data matrix, and performs frequency mixing and Fourier transform on the pulse signals to extract the Doppler frequency shift time series. Using the phase data matrix as observation data, a compressed sensing algorithm is used to reconstruct the three-dimensional dielectric constant distribution. Specifically, the sparse coefficients are obtained by solving the L1 norm regularization optimization problem through an iterative soft thresholding algorithm, and the dielectric constant distribution is reconstructed through sparse transformation. The time-frequency domain pulsation features, including instantaneous dominant frequency, spectral bandwidth, and time-frequency energy distribution, are extracted from the Doppler frequency shift time series using short-time Fourier transform. The spatial gradient field of the three-dimensional dielectric constant distribution is calculated as the phase gradient field. The blood flow velocity field is inverted based on Doppler frequency shift data, and the velocity divergence and velocity curl are calculated to construct the Doppler velocity field features. The phase gradient field and Doppler velocity field features are input into a convolutional neural network, and a four-dimensional spatiotemporal volume evolution model is generated through a cross-channel attention fusion module and a spatiotemporal convolutional coding module. Spatial integration is performed on the four-dimensional spatiotemporal volume evolution model to obtain the total volume time series. A state-space model containing macroscopic volume trend and microscopic pulsation components is established. An adaptive Kalman filter algorithm is used to adaptively estimate the process noise covariance and observation noise variance through a sliding window method based on the innovation sequence, separating the macroscopic volume trend and microscopic pulsation components, and calculating the instantaneous volume value, pulsation amount and pulsation spectrum. The feature sequence, including instantaneous total volume, macroscopic trend, pulsation component, pulsatility, instantaneous dominant frequency, spectral bandwidth, and time-frequency energy, is input into a long short-term memory network. The network outputs the probability of abnormal hemodynamic events and generates multi-level early warning decisions based on probability thresholds. It also generates a multi-parameter monitoring report that includes instantaneous volume value, volume trend curve, pulsatility, pulsation spectrum, and abnormal early warning level.

2. The method for detecting the internal volume of a medical device according to claim 1, characterized in that, The step of generating the phase-Doppler coupled response prediction matrix includes: The multi-chamber structure is divided into three-dimensional mesh units, and the dielectric constant and conductivity are set according to the material properties of each unit. Solve Maxwell's equations on a finite element mesh to calculate the electric and magnetic field distributions of the microwave signal. Calculate the phase response based on the electric field distribution and the Doppler frequency shift response based on the blood flow velocity field and microwave propagation direction. The volume change range and blood flow velocity range are discretized into multiple simulation scenarios. For each scenario, the phase response vector and Doppler frequency shift vector are calculated. The response vectors are organized into a phase-Doppler coupled response prediction matrix, where each row of the matrix corresponds to a specific volume-velocity state.

3. The method for detecting the internal volume of a medical device according to claim 1, characterized in that, The steps for reconstructing the three-dimensional dielectric constant distribution using compressed sensing algorithm include: The internal space of the multi-cavity chamber is divided into three-dimensional voxels, and the dielectric constant distribution vector is defined as the quantity to be determined. Based on the electromagnetic wave propagation theory, a linear relationship between phase data and dielectric constant distribution is established, and a sparse transformation matrix is ​​introduced to represent the dielectric constant distribution as a linear combination of sparse coefficients. An optimization problem is constructed with the L1 norm of sparse coefficients as the objective function and the L2 norm of the phase data fitting error as the constraint. An iterative soft thresholding algorithm is used to solve the problem. In each iteration, a gradient correction term is added to the current estimate and then the soft thresholding operator is applied until the difference between the estimates of adjacent iterations is less than the convergence threshold. The obtained sparse coefficients are converted into a three-dimensional dielectric constant distribution through sparse transformation, and blood regions and non-blood regions are segmented and identified based on the threshold of the dielectric constant.

4. The method for detecting the internal volume of a medical device according to claim 1, characterized in that, The convolutional neural network includes: The dual-channel input layer receives phase gradient field and Doppler velocity field features respectively; The feature extraction module applies three-dimensional convolutional layers to the two input channels to extract structural boundary features and flow pattern features, respectively. The cross-channel attention fusion module performs global average pooling on the feature maps to obtain channel feature descriptors, calculates the attention weights of the two channels through fully connected layers and activation functions, applies the attention weights to the corresponding feature maps, and concatenates them to form fused features; The spatiotemporal convolutional coding module stacks multi-frame fused features from the time series into a four-dimensional tensor, and applies cascaded spatial and temporal convolutional kernels to extract spatiotemporal evolution features; The volume prediction output layer maps spatiotemporal features to the blood occupancy probability of each voxel at each time step through a fully convolutional network, and converts the probability into a binary label through a thresholding operation to obtain a four-dimensional spatiotemporal volume evolution model.

5. The method for detecting the internal volume of a medical device according to claim 4, characterized in that, The training steps of the convolutional neural network include: Simulation scenarios with known volume distribution and flow velocity field are generated based on physical models. The corresponding phase gradient and Doppler velocity field are calculated as training inputs, and the real volume distribution is used as training labels. A weighted combination of Dice loss and cross-entropy loss is used as the loss function, which simultaneously considers the overlap between the predicted probability and the true label and the classification error. An optimizer combining momentum and adaptive learning rate adjustment is used for training, with the learning rate gradually reduced using an exponential decay strategy.

6. The method for detecting the internal volume of a medical device according to claim 1, characterized in that, The adaptive Kalman filter algorithm includes: The state vector is defined to include the macroscopic volume trend and the rate of change of the volume trend. The state transition equation and the observation equation are established, and the observation value is the total volume time series. Perform Kalman filtering iterations, including a prediction step and an update step. The prediction step calculates the state prior estimate and prior covariance, and the update step calculates the Kalman gain and updates the state estimate and covariance based on the observations. Based on the innovation sequence, the sliding window method is used to calculate the sample covariance of the innovation within the window, adaptively estimate the observation noise variance and process noise covariance, and make the Kalman filter adapt to the pulsation mode change by smoothly updating the noise parameters. The first component of the filtered state estimate is the macroscopic volume trend, and the difference between the total volume and the macroscopic trend is the microscopic pulsation component.

7. The method for detecting the internal volume of a medical device according to claim 1, characterized in that, The steps for calculating the pulse volume include: Autocorrelation analysis was performed on the micro-pulsation components to calculate the autocorrelation function and find the time delay corresponding to the first peak as the pulsation period; The difference between the maximum and minimum values ​​of the pulsation component within each pulsation cycle is calculated as the pulsation amount for that cycle.

8. The method for detecting the internal volume of a medical device according to claim 1, characterized in that, The input layer of the Long Short-Term Memory Network receives a time series, in which the feature vector at each moment includes the instantaneous total volume, macroscopic trend, pulsation component, pulsation quantity, instantaneous main frequency, spectral bandwidth, and time-frequency energy. The output layer of the Long Short-Term Memory Network maps the hidden state at the last moment to a scalar through a linear mapping layer, and outputs the probability of an abnormal event after passing through an activation function.

9. The method for detecting the internal volume of a medical device according to claim 8, characterized in that, The steps for converting the probability of an abnormal event occurrence into an early warning decision include: When the probability of an abnormal event occurring is greater than or equal to the first threshold and less than the second threshold, it is judged as a low-risk warning. When the probability of an abnormal event occurring is greater than or equal to the second threshold and less than the third threshold, it is judged as a medium-risk warning. A high-risk warning is issued when the probability of an abnormal event occurring is greater than or equal to the third threshold. The first threshold is less than the second threshold, and the second threshold is less than the third threshold.

10. A medical device internal volume detection system, used to perform the medical device internal volume detection method according to any one of claims 1 to 9, characterized in that, The system includes a ring phased array microwave antenna array, a signal transmission and reception unit, a data processing unit, and an early warning output unit; the ring phased array microwave antenna array is arranged around the multi-chamber structure of the blood circulation device; The signal transmitting and receiving unit controls the antenna array to transmit continuous wave signals and pulse modulation signals, and collects phase data and Doppler frequency shift data; the data processing unit performs compressed sensing reconstruction, feature extraction, neural network fusion, Kalman filtering and anomaly prediction; the early warning output unit generates a multi-parameter monitoring report and issues an early warning signal based on the anomaly probability.

Images