A dynamic microwave imaging stroke monitoring method and system based on time series deep learning denoising and wavelet pooling detection

Abstract

This invention belongs to the interdisciplinary field of microwave medical imaging and deep learning, specifically relating to a dynamic microwave imaging stroke monitoring method and system based on temporal deep learning denoising and wavelet pooling detection. It can achieve data domain clutter suppression, improve the signal-to-noise ratio of scattered signals, and effectively separate lesion signals from background interference; it enables continuous monitoring at multiple time points, capturing the evolutionary characteristics of lesions over time, such as hematoma expansion, edema spread, and reperfusion, providing dynamic pathological change trends for clinical use; it proposes a temporal self-supervised denoising framework to improve denoising consistency under multi-frame input and avoid inter-frame artifact jumps; it increases the equivalent density of spatial sampling with a limited number of antennas, reduces imaging artifacts, and enhances lesion localization stability; and it optimizes multi-scale lesion detection performance, capturing the edges and details of stroke lesions of different sizes and shapes in microwave images.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of microwave medical imaging and deep learning, specifically relating to a dynamic microwave imaging stroke monitoring method and system based on temporal deep learning denoising and wavelet pooling detection. Background Technology

[0002] Microwave brain imaging (MBI) utilizes the scattering changes caused by differences in the dielectric properties of different tissues to achieve non-invasive identification of brain abnormalities such as hemorrhage and ischemia / edema. To adapt to pre-hospital emergency care, transport, and bedside applications, various portable or semi-portable microwave brain detection devices and supporting algorithms have emerged in recent years. Existing technologies mainly include the following categories: (1) Traditional signal processing methods: such as reference cancellation, spatial filtering, singular value decomposition (SVD), independent component analysis (ICA), etc. These methods have certain effects under ideal conditions, but in practical applications, due to factors such as head shape, positional changes, and tissue heterogeneity, clutter and target signals overlap severely in the time and frequency domains, making it difficult for traditional methods to separate them effectively. For example, correlation imaging methods based on interpolation introduce virtual array elements through zero-filling interpolation of two-dimensional fast fourier transform (2D-FFT), which can improve spatial sampling density, but still rely on a fixed signal model and has limited adaptability to complex head environments.

[0003] (2) Application of deep learning methods in microwave imaging: In recent years, deep learning has been introduced into the field of microwave imaging. For example, generative adversarial networks (GANs) in the image domain are used for image enhancement, and autoencoders in the signal domain are used for denoising. However, most existing methods rely on supervised learning of paired data, such as requiring the simultaneous acquisition of data from the same scene with and without lesions, which is difficult to achieve in real clinical environments. At the same time, most methods only focus on a single aspect, such as denoising or imaging, and lack a systematic end-to-end solution.

[0004] (3) Imaging and detection separation methods: Existing microwave imaging systems typically reconstruct the dielectric constant distribution map first, and then identify lesions through manual or traditional image processing methods. This process is cumbersome, has poor real-time performance, and requires high image quality. Although some studies have attempted to apply deep learning to image detection, most of these studies are aimed at high-quality images such as magnetic resonance imaging (MRI) and computed tomography (CT), and are poorly adapted to low-resolution, high-noise microwave reconstructed images.

[0005] (4) In the prior art: some patents propose a method to directly map lesions based on changes in scattering parameters, which is simple to implement but sensitive to noise; other patents use three-dimensional tomographic inversion to obtain dielectric distribution, but the iterative solution is large and highly dependent on the model and calibration, making it difficult to achieve real-time updates in portable scenarios.

[0006] In summary, existing technologies for microwave stroke image detection still suffer from problems such as difficulty in clutter suppression, reliance on paired data, separation of imaging and detection, and poor adaptability to low-quality images. Therefore, there is an urgent need for a stroke image detection method that can achieve clutter suppression under conditions without paired data and designs a dedicated detection network for the characteristics of microwave reconstructed images. To this end, this invention provides a dynamic microwave imaging stroke monitoring method and system based on temporal deep learning denoising and wavelet pooling detection. Summary of the Invention

[0007] The purpose of this invention is to provide a dynamic microwave imaging stroke monitoring method and system based on temporal deep learning denoising and wavelet pooling detection. This method and system can suppress data domain clutter, improve the signal-to-noise ratio of scattered signals, and effectively separate lesion signals from background interference. It enables continuous monitoring at multiple time points, capturing the evolutionary characteristics of lesions over time, such as hematoma expansion, edema spread, and reperfusion, providing dynamic pathological change trends for clinical use. A temporal self-supervised denoising framework is proposed to improve denoising consistency under multi-frame input and avoid inter-frame artifact jumps. With a limited number of antennas, the spatial sampling equivalent density is increased, reducing imaging artifacts and enhancing lesion localization stability. The system optimizes multi-scale lesion detection performance, capturing the edges and details of stroke lesions of different sizes and shapes in microwave images.

[0008] The specific technical solution adopted by this invention is as follows: A dynamic microwave imaging stroke monitoring method and system based on temporal deep learning denoising and wavelet pooling detection, characterized by the following steps: S100: Collect the sequence of original frequency domain scattering parameter matrices of a loop antenna array at multiple time points. The original frequency domain scattering parameters are S-parameters. Construct a time-series complex scattering tensor. S200: Perform inverse discrete Fourier transform, modulation enhancement and normalization processing on the time-series complex scattering tensor to reconstruct it into a standard time-series input tensor; S300: Input the time-series input tensor into the time-series CycleGAN deep learning network to complete clutter and artifact removal, and output the denoised time-series scattering data tensor. S400: Based on the denoised scattering data tensor, the Born approximation algorithm with regularization is used to complete the joint inversion of multiple frequency points and reconstruct the time-series microwave dielectric constant image sequence. S500: The reconstructed temporal microwave dielectric constant image sequence is input into a multi-scale lesion detection network based on wavelet pooling and attention mechanisms to achieve lesion identification.

[0009] Preferably, specifically, a data acquisition module is constructed, which uses a ring multi-antenna array to perform body-domain microwave illumination and echo acquisition on the tested head to obtain the original frequency-domain scattering parameter matrix. ,in For antenna index, The frequency point is defined as follows: the antenna array consists of eight microwave antennas evenly distributed in a ring, and the operating frequency band can cover 0.5-2.5 GHz. To achieve dynamic monitoring, the system continuously collects multiple frames of scattering data at fixed time intervals of 3 minutes to form a sequence of original scattering parameter matrices in the frequency domain, and finally completes the construction of the temporal complex scattering tensor.

[0010] Preferably, in step S200, specifically, a data preprocessing module is constructed to perform an inverse discrete Fourier transform (IFT) on the original frequency domain scattering parameters, extract the time domain signal, and take its absolute value to obtain the time domain response signal; to enhance the signal-to-noise ratio, a modulated inverse discrete Fourier transform method is used; subsequently, normalization processing is performed to map the signal amplitude to the [0,1] interval, and an input feature map of size 128×128 is constructed based on the number of sampling frequency points and the number of antenna indices; the T consecutively acquired feature maps are stacked into a standard time-series input tensor. This serves as the input to the deep learning model for temporal clutter removal.

[0011] Preferably, in step S300, specifically, a data domain deep learning clutter removal module is constructed. This module builds a clutter suppression model based on self-supervised learning, employing the CycleGAN (CycleGenerativeAdversarial Network) framework, including two time-series generators (G and F) and two discriminators (D_X and D_Y). Generator G is used to map the clutter-containing mixed signal to a clean target signal, while generator F is used for reverse reconstruction, ensuring the reversibility of the mapping. Through cyclic consistency loss and adversarial loss, the network can learn clutter suppression under unpaired data conditions. The generator network employs an improved 3DU-Net structure or a hybrid structure of ConvLSTM and U-Net to model inter-frame dependencies. Taking ConvLSTM-U-Net as an example: the encoder part shares 2D convolutional weights at each time step to extract spatial features, and then captures temporal dynamics through ConvLSTM layers. The decoder part is symmetrical, with residual blocks introduced in the middle to enhance information transmission. To improve training stability, spectral normalization and attention gate mechanisms are introduced. The discriminator adopts a temporal PatchGAN structure to judge the authenticity of spatiotemporal local regions of the input signal.

[0012] Loss functions include: To counteract the loss, a least-squares GAN approach is used. ; Similarly, for F,D X .

[0013] Cyclic consistency loss: ; Temporal smoothing loss: Constrains the denoising results between adjacent frames to be smooth, avoiding artifact jumps; ; Optical flow consistency loss: The pre-trained optical flow network is used to estimate inter-frame motion and constrain the motion consistency of the denoised image. ; Total loss: ; in , , Used to balance various losses; the denoised temporal scattering data tensor As input to the subsequent microwave image reconstruction module.

[0014] Preferably, in step S400, a microwave image reconstruction module is constructed to convert the denoised temporal scattering data tensor... The microwave imaging algorithm is input to reconstruct the spatial dielectric constant distribution map; imaging is performed using a quadratic programming optimization algorithm based on the Born approximation (BIM); the microwave image reconstruction module employs a quadratic programming optimization framework to transform the inverse scattering problem into a constrained quadratic programming problem: ; Combining Tikhonov regularization and gradient regularization, where α1=10 is the Tikhonov regularization weight and β1=1e-6 is the gradient regularization weight, and introducing physical upper and lower bounds of the dielectric constant as constraints, improves image edge sharpness and target contrast. Jointly solving multi-frequency data at 0.8, 1.0, 1.2, 1.5, and 2.0 GHz further improves ill-conditioned images. The BIM imaging algorithm flow includes: initializing the dielectric constant distribution → forward modeling the scattering field → quadratic programming to solve for the contrast correction → iterative updating → outputting the image while meeting the convergence condition. By comparing the imaging results before and after denoising, the improvement effect of deep learning data augmentation on image noise suppression, target visibility sharpness, and artifact suppression is evaluated. The reconstructed time-series microwave dielectric constant image sequence... This serves as the input to the multi-scale lesion detection network based on wavelet pooling and attention mechanisms in the S500.

[0015] Preferably, in step S500, a microwave image domain lesion detection module is constructed, which builds a multi-scale lesion detection network based on wavelet pooling and attention mechanisms, denoted as the WAU-Net detection network. This network addresses the problems of sparse features and blurred edges in low-resolution microwave images, improving the representation ability of the target region's texture structure and spatial location. The network has a 2D structure, performing detection independently on each frame of the reconstructed image, with all frames sharing the same network weights; the input is a time-series microwave dielectric constant image sequence. The single-frame image in the image is output as the corresponding detection map. By processing all frames sequentially, a multi-frame detection result sequence is obtained. .

[0016] Preferably, the principle of the wavelet pooling module in S500 is as follows: Wavelet transform uses multi-resolution analysis to decompose a signal into sub-bands of different frequency bands; for two-dimensional images, discrete wavelet transform (DWT) is implemented through a set of orthogonal mirror filters; using the Haar wavelet basis, four two-dimensional convolutional kernel filters are defined, with the following elements: low-pass filter (approximate component): ; High-pass filter (detail components): ; The above filters satisfy orthogonality and have a fixed step size of 2, ensuring that the image size is halved after transformation; Wavelet pooling operation process: For the input feature map (H and W are spatial dimensions, and C is the number of channels), treating it as C independent single-channel images, and performing two-dimensional discrete wavelet transform on each. Specifically, four filters are applied to each channel for convolution, generating four sub-band feature maps: Low-frequency approximate subband: ; Horizontal detail sub-band: ; Vertical detail sub-band: ; Diagonal detail sub-band: ; in Represents two-dimensional convolution. This indicates a downsampling step of 2. The spatial dimensions of the four sub-bands mentioned above are all H / 2×W / 2, and the number of channels remains C.

[0017] At the same time, a traditional max pooling operation is performed on the input feature map to obtain... This is used to preserve the most salient local response.

[0018] The five sub-bands mentioned above are concatenated along the channel dimension to form the output of the wavelet pooling module: ; If the input is a single-channel image, the wavelet pooling output has 5×1=5 channels. This output retains all the information of the original image: low-frequency components represent smooth areas of the image, high-frequency components capture edge details in the horizontal, vertical, and diagonal directions, respectively, and max pooling enhances the most salient local features. By encoding spatial information into the channel dimension, lossless downsampling is achieved. In the S500, the input to the WAU-Net detection network is a 128×128×1 single-channel microwave image reconstructed from microwave brain imaging BIM, i.e., a sequence of microwave dielectric constant images. After the first convolutional residual module, the feature map size becomes 128×128×64. During the first wavelet pooling, the input feature map... The output after wavelet pooling is: ; That is, the spatial size is halved (128→64), and the number of channels changes from 64 to 320 (5×64); the subsequent downsampling stages follow the same pattern until a feature map with a minimum resolution of 8×8 is obtained.

[0019] If the input is a single-channel image, such as the network input layer, the wavelet pooling output channel count is 5 × 1 = 5 channels, halving the spatial size. This design ensures that high-frequency detail information is always preserved in the channel dimension and is not discarded during the entire network forward propagation.

[0020] Preferably, the network structure of the WAU-Net detection network is as follows: Encoding path: Each downsampling stage includes two convolutions, a wavelet pooling module, a channel attention module, and a pyramid-dilated coordinate attention module. The wavelet pooling module replaces the traditional pooling layer, preserving high-frequency edge information through two-dimensional discrete wavelet decomposition; the channel attention module redirects channel features; the pyramid-dilated coordinate attention module captures multi-scale contextual information through dilated convolutions and coordinate attention with different dilation rates of 1, 2, 3, and 4.

[0021] Decoding path: Each upsampling stage includes upsampling, full-scale skip connections, a spatial attention module, and two convolutions. The spatial attention module helps the network focus on the lesion area and suppress background interference; the full-scale skip connections unify the features of each layer of the encoder to the same resolution through pooling or interpolation, stack them with the features of the current layer of the decoder in the channel dimension, and fuse them through 3×3 convolution.

[0022] Output layer: A pixel-level detection probability map is generated by sigmoid activation, a binary detection map is obtained by thresholding, and the coordinates of the lesion center and the area of ​​the region are calculated.

[0023] The loss function uses a hybrid loss: ; Where α is set to 0.5 to balance the two types of losses, BCE is the binary cross-entropy loss, and Dice is the Dice loss, with the specific formulas as follows: ; in This is a smoothing factor used to prevent division by zero errors.

[0024] In dynamic monitoring scenarios, this network is applied independently to each reconstructed image frame to obtain frame-by-frame detection results. Because each frame is processed independently, it can be computed in parallel, facilitating real-time implementation.

[0025] A dynamic microwave imaging stroke monitoring system based on temporal deep learning denoising and wavelet pooling detection includes: Multi-antenna array module: 8 microwave antennas arranged in a ring for transmitting and receiving microwave signals; Signal acquisition and preprocessing module: connected to the antenna array module, used to continuously acquire multi-frequency scattering parameters, and perform inverse Fourier transform, modulus taking and normalization processing on each frame of acquired signal to construct a time-series input tensor; The first deep learning module: deployed in an embedded processor, is used to perform data domain clutter suppression; Image reconstruction module: connected to the first deep learning module, used to reconstruct 128×128 microwave image sequences based on the Born quadratic programming algorithm; The second deep learning module is connected to the image reconstruction module and is used to independently perform lesion detection on each frame of the reconstructed image and output the frame-by-frame detection results. Portable housing and power supply module: used to encapsulate the above components and provide power. The power supply module uses a lithium battery and is suitable for mobile scenarios.

[0026] The technical effects achieved by this invention are as follows: Temporal self-supervised denoising: No paired labeled data required, highly adaptable, and suitable for continuous monitoring of microwave brain imaging in real clinical environments; Two-stage collaborative optimization: First, perform temporal denoising in the data domain, and then perform frame-by-frame detection in the image domain to avoid the accumulation of errors in a single stage; Lossless information downsampling: wavelet pooling preserves edge details and improves the accuracy of lesion detection in microwave brain imaging; Dynamic monitoring capability: Enables continuous monitoring at multiple time points, outputs detection results frame by frame, supports the analysis of lesion evolution trends in microwave brain imaging, and provides dynamic pathological change information for clinical reference by medical staff. Attached Figure Description

[0027] Figure 1 This is a block diagram of a dynamic microwave imaging stroke monitoring method and system based on temporal deep learning denoising and wavelet pooling detection according to the present invention. Figure 2 This is a physical image of the 8-element ring antenna array of the present invention; Figure 3 This is a schematic diagram of the microwave brain imaging detection method for stroke according to the present invention; Figure 4 This is a diagram of the self-supervised clutter suppression network in step S400 of the present invention; Figure 5 This is a schematic diagram of the wavelet pooling module in step S500 of the present invention; Figure 6 This is the structure diagram of the multi-scale lesion detection network based on wavelet pooling and attention mechanism in step S500 of the present invention, namely the WAU-Net detection network structure diagram. Figure 7 This is a comparison image of the standard time-series input tensor before and after denoising in step S300 of this invention. Figure 7 'a' represents the standard time-series input tensor before denoising. Figure 7 b is the temporal scattering data tensor after denoising. Detailed Implementation

[0028] To make the objectives and advantages of this invention clearer, the invention will be specifically described below with reference to embodiments. It should be understood that the following text is merely used to describe one or more specific embodiments of the invention and does not strictly limit the scope of protection specifically claimed by the invention.

[0029] like Figures 1-7 As shown, a dynamic microwave imaging stroke monitoring method and system based on temporal deep learning denoising and wavelet pooling detection includes the following steps: Data Acquisition and Preprocessing An 8-antenna ring array is used, such as Figure 2 As shown, the antennas are uniformly distributed, operating in the 0.5–2.5 GHz frequency band. A vector network analyzer was used to continuously acquire multiple frames of S-parameters, with each frame collecting 8 × 8 × 10¹ complex data points (101 frequency points). The frame interval was set to 3 minutes, and a total of T = 10 frames were collected to simulate the evolution of stroke. Each frame signal underwent an IFT transform, and the modulus was taken to obtain a time-domain signal with a length of 256 points. All channel signals were arranged into a 128 × 128 matrix according to antenna order and normalized to the [0,1] interval. The 10-frame matrices were stacked to form a standard time-series input tensor. .

[0030] A self-supervised denoising network is trained to denoise a standard temporal input tensor and output a denoised temporal scattering data tensor, such as... Figure 7 As shown.

[0031] A real-world dataset containing 608 target signal sequences (including lesion evolution) and 548 background signal sequences (without lesions) was used, with each sequence consisting of 10 frames. The training and test sets were divided in a 7:3 ratio. Background signals were collected from measurements taken at different time points from head models at different positions and rotation angles to ensure that the background in the test set did not overlap with the training set. The temporal CycleGAN network training parameters were: epochs=100, batch size=4, initial learning rate 1e-5, using the Adam optimizer (β1=0.5, β2=0.999). , , The generator employs a ConvLSTM-U-Net architecture: the encoder contains four 3D convolutional blocks (kernel size 3×3×3, with a temporal dimension of 1, i.e., frame-by-frame 2D convolution), which are then fed into a ConvLSTM layer (hidden dimension 64, kernel 3×3). The decoder is symmetrical, with nine residual blocks inserted in the middle, and an attention gate is introduced. The discriminator uses a temporal PatchGAN (spatiotemporal patch 70×70×5). After training, the model is saved for subsequent denoising.

[0032] Microwave image reconstruction like Figure 4 As shown, the imaging area is divided into a 128×128 grid, and the background medium is set to the equivalent dielectric constant of brain tissue (εr≈45). Multi-frequency (0.8, 1.0, 1.2, 1.5, 2.0 GHz) joint inversion is employed, with each BIM iteration solving a quadratic programming problem and setting regularization parameters. The upper and lower bounds of the dielectric constant were set to [2, 80] based on the physiological range. After 10 iterations, a 128×128 dielectric constant distribution map was output. Comparing the imaging results before and after denoising, the denoised image showed reduced artifacts, improved lesion contrast, and better inter-frame continuity.

[0033] like Figure 5 as well as Figure 6 As shown, the training of the multi-scale lesion detection network based on wavelet pooling and attention mechanisms is denoted as WAU-Net detection network training. Training and validation were performed using a self-built simulation dataset (containing 128×128 microwave images reconstructed from BIM and their corresponding lesion detection labels). The dataset contains single-frame images from different time points, which can be considered as independent samples. The input is a single-frame reconstructed image (size 128×128×1), and the output is a binary detection image of the same size. Data augmentation includes random rotation (±10°), translation, scaling, and flipping. Network training parameters: epochs=300, batch size=8, optimizer Adam (lr=1e-4), loss function is BCE-Dice hybrid loss (α=0.5). Network structure details are as follows: Residual module: Each convolutional block contains two 3×3 convolutions, skip connections, and the ReLU activation function.

[0034] like Figure 5 as well as Figure 6 As shown, the wavelet pooling module uses Haar wavelets to perform two-dimensional discrete wavelet decomposition on the input feature map, generating four sub-bands: LL, LH, HL, and HH (each size is halved, while the number of channels remains the same). These sub-bands are then stacked with the max pooling results in the channel dimension. If the number of input channels is C, the number of output channels is 5C, and the spatial size is halved. For example, the first wavelet pooling layer has an input of 128×128×64 and an output of 64×64×320.

[0035] Channel attention: Squeeze operation with global average pooling, activation operation with two fully connected layers (dimensionality reduction coefficient r=16), sigmoid activation.

[0036] Pyramid-expanded coordinate attention: parallel dilated convolution (dilation rates 1, 2, 3, 4), 1×1 convolution fusion and coordinate attention. Coordinate attention is pooled along the horizontal and vertical directions respectively, concatenated and then subjected to 1×1 convolution, then separated and activated by sigmoid.

[0037] Spatial attention: average pooling and max pooling along the channel dimension, 7×7 convolution, sigmoid.

[0038] Full-scale skip connections: Features from each layer of the encoder (resolutions 128, 64, 32, 16, 8) are pooled or interpolated to be unified to the same resolution as the current layer of the decoder, stacked in the channel dimension, and fused by 3×3 convolution.

[0039] Physical phantom experiment verification A PVP brain tissue equivalent phantom was prepared, and ischemic and hemorrhagic lesions with different dielectric properties were created by adjusting the water-to-PVP ratio. A microwave measurement platform was constructed, including a vector network analyzer, a custom broadband antenna array, and a phantom fixation device. Real-world data was collected, input into a pre-trained denoising model and a detection model, and lesion detection results were output.

[0040] This invention enables data domain clutter suppression, improves the signal-to-noise ratio of scattered signals, and effectively separates lesion signals from background interference; it achieves continuous monitoring at multiple time points, capturing the evolutionary characteristics of lesions over time, such as hematoma expansion, edema spread, and reperfusion, providing dynamic pathological change trends for clinical use; it proposes a temporal self-supervised denoising framework, improving denoising consistency under multi-frame input and avoiding inter-frame artifact jumps; it increases the equivalent density of spatial sampling with a limited number of antennas, reduces imaging artifacts, and enhances the stability of lesion localization; and it optimizes multi-scale lesion detection performance, capturing the edges and details of stroke lesions of different sizes and shapes in microwave images.

[0041] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained in this invention are implemented according to conventional methods in the art unless otherwise specified or limited.

Claims

1. A dynamic microwave imaging stroke monitoring method based on temporal deep learning denoising and wavelet pooling detection, characterized in that: Includes the following steps: S100: Collect the frequency domain original scattering parameter matrix sequence of the loop antenna array at multiple time points and construct the time-series complex scattering tensor; S200: Perform inverse discrete Fourier transform, modulation enhancement and normalization processing on the time-series complex scattering tensor to reconstruct it into a standard time-series input tensor; S300: Input the standard time-series input tensor into the time-series CycleGAN deep learning network to complete clutter and artifact removal, and output the denoised time-series scattering data tensor. S400: Based on the denoised scattering data tensor, the Born approximation algorithm with regularization is used to complete the joint inversion of multiple frequency points and reconstruct the time-series microwave dielectric constant image sequence. S500: The reconstructed temporal microwave dielectric constant image sequence is input into a multi-scale lesion detection network based on wavelet pooling and attention mechanisms to achieve lesion identification.

2. The dynamic microwave imaging stroke monitoring method based on temporal deep learning denoising and wavelet pooling detection according to claim 1, characterized in that: In step S100, specifically, a data acquisition module is constructed, which uses a ring multi-antenna array to perform body-domain microwave illumination and echo acquisition on the tested head to obtain the original frequency-domain scattering parameter matrix. ,in For antenna index, The frequency point is defined as follows: the antenna array consists of eight microwave antennas evenly distributed in a ring, and the operating frequency band can cover 0.5-2.5 GHz. To achieve dynamic monitoring, the system continuously collects multiple frames of scattering data at fixed time intervals of 3 minutes to form a sequence of original scattering parameter matrices in the frequency domain, and finally completes the construction of the temporal complex scattering tensor.

3. The dynamic microwave imaging stroke monitoring method based on temporal deep learning denoising and wavelet pooling detection according to claim 1, characterized in that: Specifically, in S200, a data preprocessing module is constructed to perform inverse discrete Fourier transform on the original frequency domain scattering parameters, extract the time domain signal, and take its absolute value to obtain the time domain response signal. To enhance the signal-to-noise ratio, a modulation inverse discrete Fourier transform method is adopted. Subsequently, normalization processing is performed to map the signal amplitude to the [0,1] interval. Based on the number of sampling frequency points and the number of antenna indices, an input feature map with a size of 128×128 is constructed. The T consecutively acquired feature maps are stacked into a standard time-series input tensor. This serves as the input to the deep learning model for temporal clutter removal.

4. The dynamic microwave imaging stroke monitoring method based on temporal deep learning denoising and wavelet pooling detection according to claim 1, characterized in that: Specifically, in S300, a data domain deep learning clutter removal module is constructed. This module builds a clutter suppression model based on self-supervised learning, using the CycleGAN framework, and includes two time series generators and two discriminators. The two discriminators are discriminator D_X and discriminator D_Y; the two time series generators are time series generator G and time series generator F. Generator G is used to map the clutter-containing mixed signal to a clean target signal, while generator F is used for reverse reconstruction to ensure the reversibility of the mapping; through cyclic consistency loss and adversarial loss, the network can learn clutter suppression under unpaired data conditions. The generator network employs an improved 3DU-Net architecture or a hybrid architecture of ConvLSTM and U-Net to model inter-frame dependencies; it introduces... Spectral normalization and attention gate mechanism; The discriminator uses a temporal PatchGAN structure to judge the authenticity of the spatiotemporal local regions of the input signal; Loss functions include: To counteract the loss, a least-squares GAN approach is used. ； in A generator that maps cluttered signals to clean signals. A discriminator to determine whether the input is a real, clean signal. This represents a source domain dataset containing clutter. This represents a clean target domain dataset; Represents the mathematical expectation. To from the source domain Sampled clutter-containing signal samples To from the target domain Clean signal samples from the middle sample; This indicates that the generator responds to the input. The generated denoised signal, The discriminator's judgment result on the generated signal. This represents the discriminator's judgment of the real clean signal; the loss function is used to make the generated signal approximate the real clean signal through adversarial training. Similarly, for F,D X ; ； in Represents a reverse generator and discriminator The adversarial loss function between them; It is an inverse generator, responsible for reconstructing a clean signal into a cluttered signal form; It is a source domain discriminator used to determine whether the input is a true source domain signal, i.e., a signal containing clutter. The reverse generator reconstructs a clean signal into a cluttered output. This represents the discriminator's judgment result on the reconstructed signal; The discriminator judges the real cluttered signal; this symmetrical structure ensures that the bidirectional mapping of CycleGAN can maintain good generation quality, so that both the forward denoising and backward reconstruction processes can be optimized through adversarial training, thereby improving the stability and reconstruction fidelity of the entire network. To maintain the integrity of the signal content, a cycle consistency loss is introduced: ； in It is a reverse generator, responsible for reconstructing the denoised signal back to its original form. This represents the result of reconstructing the original signal from the denoised signal. This represents the L1 norm, which is the sum of absolute errors; this loss ensures that the signal content remains consistent before and after denoising. Temporal smoothing loss: constrains the smoothing of denoising results between adjacent frames; ； in This represents the total number of frames in the time series. This is a time frame index, with values ​​ranging from 1 to T-1; and These represent the output signals or images after denoising processing for frame t and frame (t+1), respectively. This represents the L1 norm, which is used to calculate the sum of the absolute values ​​of the positional differences between corresponding pixels in two frames. This is a normalization factor used to average the differences between adjacent frame pairs; This means summing over all adjacent frame pairs; the loss function penalizes drastic changes between frames by calculating the L1 distance between adjacent denoised result frames, ensuring that the denoised result has smooth transition characteristics in the time dimension, which is especially important in dynamic microwave imaging such as real-time monitoring of stroke evolution. Optical flow consistency loss: The pre-trained optical flow network is used to estimate inter-frame motion and constrain the motion consistency of the denoised image. ； in and These represent the original noisy input signals or images of frame t and frame (t+1), respectively. and These represent the output signals or images after denoising processing for frame t and frame (t+1), respectively. This represents a pre-trained optical flow network used to estimate pixel-level motion vector fields between two frames; To estimate the optical flow field from the original noisy input, reflecting the real physical motion; The optical flow field estimated from the denoised output; The L2 norm, or Euclidean distance, is used to calculate the difference between two optical flow fields; symbol This represents the element-wise subtraction of two optical flow fields; Total loss: ； in To generate positive counter-loss, To generate adversarial loss in reverse, A discriminator for distinguishing source domain signals, , , This is used to balance various losses; Denoising temporal scattering data tensor As input to the subsequent microwave image reconstruction module.

5. The dynamic microwave imaging stroke monitoring method based on temporal deep learning denoising and wavelet pooling detection according to claim 1, characterized in that: In step S400, a microwave image reconstruction module is constructed to convert the denoised temporal scattering data tensor... The microwave imaging algorithm is input to reconstruct the spatial dielectric constant distribution map; imaging is performed using a quadratic programming optimization algorithm based on the Born approximation; the microwave image reconstruction module employs a quadratic programming optimization framework to transform the inverse scattering problem into a constrained quadratic programming problem: ； in This represents the summation over all receiving antennas, where This represents the total number of receiving antennas. Indicates the first The scattered field values ​​actually measured by each receiving antenna are known measured data. This represents the summation of all grid cells after discretization of the imaging region, where The total number of grids used to divide the imaging region; Let be the Green's function, describing the first... The electromagnetic wave propagation characteristics from the first grid to the mm-th antenna are known quantities calculated based on electromagnetic field theory. For the first The contrast of each grid represents the difference between the dielectric constant of that grid location and the dielectric constant of the background medium, and is an unknown quantity to be solved. For the first The total electric field within each grid is also an unknown quantity to be solved. This represents the square of the modulus, used to calculate the error power between the theoretical and measured scattered fields; Combining Tikhonov regularization and gradient regularization, where α1=10 is the Tikhonov regularization weight and β1=1e-6 is the gradient regularization weight, and introducing physical upper and lower bounds of the dielectric constant as constraints, the algorithm improves ill-conditioned performance by jointly solving multi-frequency data at 0.8, 1.0, 1.2, 1.5, and 2.0 GHz. The BIM imaging algorithm flow includes: initializing the dielectric constant distribution → forward modeling the scattering field → quadratic programming to solve for the contrast correction → iterative updating → outputting the reconstructed temporal microwave dielectric constant image sequence that meets the convergence condition. This serves as the input to the multi-scale lesion detection network based on wavelet pooling and attention mechanisms in the S500.

6. The dynamic microwave imaging stroke monitoring method based on temporal deep learning denoising and wavelet pooling detection according to claim 1, characterized in that: The S500 constructs a microwave image domain lesion detection module, and builds a multi-scale lesion detection network based on wavelet pooling and attention mechanisms, denoted as the WAU-Net detection network. This network addresses the problems of sparse features and blurred edges in low-resolution microwave images, improving the representation ability of the target region's texture structure and spatial location. The network has a 2D structure, independently detecting each frame of the reconstructed image, while all frames share the same network weights. The input is a time-series microwave dielectric constant image sequence. The single-frame image in the image is output as the corresponding detection map. By processing all frames sequentially, a multi-frame detection result sequence is obtained. .

7. The dynamic microwave imaging stroke monitoring method based on temporal deep learning denoising and wavelet pooling detection according to claim 6, characterized in that: In the S500, the input to the WAU-Net detection network is a 128×128×1 single-channel microwave image reconstructed from microwave brain imaging BIM, i.e., a sequence of microwave dielectric constant images. After the first convolutional residual module, the feature map size becomes 128×128×64; during the first wavelet pooling, the input feature map... The output after wavelet pooling is: ； In this formula, This represents the output feature map of the wavelet pooling module; This indicates that a splicing operation is performed along the channel dimension; This is a low-frequency approximate subband obtained after processing with a low-pass filter, containing information about the smooth regions of the image; This is the horizontal detail sub-band obtained after processing with a horizontal high-pass filter, which contains the vertical edge information of the image; The vertical detail sub-band is obtained after processing with a vertical high-pass filter and contains the horizontal edge information of the image. The diagonal detail subband is obtained by processing with a diagonal high-pass filter and contains diagonal edge information of the image; This is the result obtained after performing traditional max pooling on the input feature map, used to preserve the most salient local responses; The dimension of the output feature map is represented by the first two numbers, 64×64, which indicate the spatial size of the output feature map, i.e., the height and width are both 64 pixels. The third number, 320, indicates the number of channels in the output feature map. Taking the embodiment of the present invention as an example, when the dimension of the input feature map X is 128×128×64, after processing by the wavelet pooling module, each of the four wavelet subbands outputs a 64×64×64 feature map. The max pooling result is also 64×64×64. After concatenating the five parts in terms of channel dimension, a 64×64×320 output feature map is obtained. That is, the spatial size is halved (128→64), and the number of channels changes from 64 to 320 (5×64); the subsequent downsampling stages follow the same pattern until a feature map with a minimum resolution of 8×8 is obtained. If the input is a single-channel image, the wavelet pooling output channel number is 5×1=5 channels, and the spatial size is halved.

8. A dynamic microwave imaging stroke monitoring method based on temporal deep learning denoising and wavelet pooling detection according to claim 7, characterized in that: The network structure of the WAU-Net detection network: Encoding path: Each downsampling stage includes two convolutions, a wavelet pooling module, a channel attention module, and a pyramid dilatation coordinate attention module; the wavelet pooling module replaces the traditional pooling layer and retains high-frequency edge information through two-dimensional discrete wavelet decomposition; the channel attention module redirects channel features; the pyramid dilatation coordinate attention module captures multi-scale contextual information through dilated convolutions and coordinate attention with different dilation rates of 1, 2, 3, and 4; Decoding path: Each upsampling stage includes upsampling, full-scale skip connections, spatial attention module, and two convolutions; the spatial attention module helps the network focus on the lesion area and suppress background interference; Full-scale skip connections unify the features of each layer of the encoder to the same resolution through pooling or interpolation, stack them with the current layer features of the decoder in the channel dimension, and fuse them through 3×3 convolution. Output layer: Pixel-level detection probability map is generated by sigmoid activation, and a binary detection map is obtained by thresholding. The coordinates of the lesion center and the location information of the area are calculated. The loss function uses a hybrid loss: ； Where α is set to 0.5 to balance the two types of losses, BCE is the binary cross-entropy loss, and Dice is the Dice loss, with the specific formulas as follows: ； in This is a smoothing factor used to prevent division by zero errors.

9. A dynamic microwave imaging stroke monitoring system based on temporal deep learning denoising and wavelet pooling detection, used to execute the monitoring method according to any one of claims 1-8, characterized in that: include: Multi-antenna array module: 8 microwave antennas arranged in a ring for transmitting and receiving microwave signals; Signal acquisition and preprocessing module: connected to the antenna array module, used to continuously acquire multi-frequency scattering parameters, and perform inverse Fourier transform, modulus taking and normalization processing on each frame of acquired signal to construct a time-series input tensor; The first deep learning module: deployed in an embedded processor, is used to perform data domain clutter suppression; Image reconstruction module: connected to the first deep learning module, used to reconstruct 128×128 microwave image sequences based on the Born quadratic programming algorithm; The second deep learning module is connected to the image reconstruction module and is used to independently perform lesion detection on each frame of the reconstructed image and output the frame-by-frame detection results. Portable housing and power supply module: used to encapsulate the above components and provide power. The power supply module uses a lithium battery and is suitable for mobile scenarios.

Images