Microwave-induced thermoacoustic transcranial quantitative imaging method based on REAGU-Net network

Abstract

The invention relates to a microwave-induced thermoacoustic transcranial quantitative imaging method based on a REAGU-Net network, and the method achieves the simultaneous quantitative reconstruction of dielectric constant, conductivity and sound velocity in brain tissue through the MITAT technology, and provides a new methodology for the characteristic analysis of the brain tissue. A REAGU-Net novel network is developed, an attention guiding block and a residual error inducing block are combined to process the influence of acoustic non-uniformity on reconstruction quality, and effective application of deep learning in quantitative reconstruction is shown; the technology not only shows a good result in cerebral hemorrhage detection, but also can be potentially applied to detection and diagnosis of various diseases including breast cancer, liver diseases and thyroid diseases, and a new quantitative analysis tool is provided for clinic; the DL-MITAT method provided by the invention can provide quantized physical characteristic data, and the data not only is helpful for imaging, but also can provide support for physiological information prediction and the like in subsequent electromagnetic treatment; compared with a traditional TCD or ultrasonic imaging method, the transcranial MITAT has the remarkable advantages that due to the fact that ultrasonic waves of the transcranial MITAT only need to be transmitted with the skull for one time, the possibility of image attenuation and degeneration is reduced, and the imaging quality is improved.

Description

Technical Field

[0001] This invention relates to the fields of medical devices, information technology, biomedical engineering, and medical and health care technology, and in particular to a microwave-induced thermoacoustic transcranial quantitative imaging method based on the REAGU-Net network. Background Technology

[0002] Over the past two decades, intracerebral hemorrhage (ICH) has been considered one of the deadliest diseases globally. With an aging population and increasing life stress, the number of deaths from stroke has steadily increased over the past decade. Statistics show that ICH has become the most common type of stroke, accounting for more than a quarter of all stroke cases, with a one-month mortality rate as high as approximately 40%, and is accompanied by a severe risk of disability. Therefore, timely and accurate diagnosis of ICH is crucial for reducing mortality and disability rates.

[0003] Currently, the transcranial imaging techniques used in clinical practice mainly include computed tomography (CT), magnetic resonance imaging (MRI), and transcranial Doppler (TCD) ultrasound. Each of these techniques has its advantages, but also significant drawbacks in certain aspects. While CT can rapidly provide imaging information in acute situations, its use of ionizing radiation limits its use as a routine examination, especially for children. On the other hand, MRI is expensive and noisy, causing discomfort to patients, particularly children. Furthermore, TCD ultrasound suffers from severe distortion due to sound wave reflection, attenuation, and phase shift when penetrating the skull, affecting the accuracy and reliability of the imaging. These factors collectively drive the need for new imaging technologies.

[0004] Against this backdrop, microwave-induced thermoacoustic imaging (MITAT), as an emerging imaging technology, has shown great promise for biomedical applications. The basic principle of MITAT technology is to irradiate tissue with microwaves to induce ultrasound signals, and then capture these ultrasound signals to achieve imaging. This technology effectively combines the high resolution of ultrasound imaging with the high contrast of microwave imaging, and is non-invasive, low-noise, and economical. Therefore, it shows promising application prospects in multiple fields such as cancer detection, foreign body detection, dielectric property reconstruction, and transcranial brain imaging.

[0005] While MITAT technology demonstrates advantages over traditional imaging techniques in transcranial brain imaging, it still faces several challenges in its application. One key issue is the quantitative recovery of electromagnetic acoustic parameters such as dielectric constant, conductivity, and sound velocity. Quantitative recovery of these parameters not only improves imaging accuracy but also provides richer information for biomedical and clinical testing.

[0006] The dielectric constant is a crucial parameter describing a material's response to an electric field and is closely related to biological characteristics such as tissue water content and cell density. Electrical conductivity reflects a material's ability to conduct electric current and is typically closely related to the electrolyte content and structure of the tissue. In brain tissue imaging, changes in dielectric constant and conductivity are often closely associated with pathological changes, making them important biomarkers. Furthermore, the speed of sound, as a parameter related to the propagation speed of sound waves in a medium, can reflect changes in the physical properties of tissues and their variations under different pathological states. Therefore, if these three parameters can be quantitatively recovered simultaneously, it will provide a more accurate basis for the early diagnosis and treatment planning of brain diseases.

[0007] However, current MITAT techniques still face some difficulties in quantifying and recovering these electromagnetic acoustic parameters. First, traditional MITAT imaging methods primarily reconstruct images proportional to the specific absorption rate (SAR) distribution, while direct SAR quantification presents practical challenges. The acoustic signals measured in MITAT experiments are often not measures of absolute pressure; therefore, SAR or power absorption distribution cannot be directly used as quantitative results. Although some studies treat the image as a conductivity distribution, this approach is only an approximation because the SAR distribution is also affected by electric field inhomogeneities.

[0008] Furthermore, addressing registration issues that may arise during imaging is crucial for achieving quantitative recovery of multiple parameters. Traditional imaging techniques often face challenges in aligning and registering images, especially under conditions of patient movement and varying imaging device parameter settings, which can lead to inconsistencies between images. Therefore, developing a novel algorithm to automate the registration process will significantly improve the clinical applicability of the method. Summary of the Invention

[0009] To address the aforementioned issues, a microwave-induced thermoacoustic transcranial quantitative imaging method based on the REAGU-Net network is proposed. By overcoming the limitations of traditional imaging techniques, MITAT combines the advantages of sound waves and electromagnetic waves, providing a more reliable means for the early diagnosis of brain diseases such as cerebral hemorrhage.

[0010] The technical solution of this invention is as follows: a microwave-induced transcranial quantitative imaging method based on the REAGU-Net network. This method utilizes the specific absorption rate distribution of simulated samples from a CST microwave studio, combines a pseudospectral time-domain algorithm to simulate the generation and propagation of thermoacoustic signals, reconstructs the initial brain image using the ultrasound back-projection (BP) algorithm, and normalizes the image to generate training and validation datasets. The REAGU-Net network is constructed and initialized, inputting the BP image dataset used for training, and outputting target dielectric constant, conductivity, and sound velocity distribution maps. Training is completed by optimizing the loss function. The performance of the REAGU-Net network is evaluated using a validation set, and the results are quantified using a structural similarity index. Hyperparameters are adjusted to complete network optimization. The trained network is then applied to test new simulated and experimental data, achieving quantitative output of parameters.

[0011] The microwave-induced thermoacoustic transcranial quantitative imaging method based on the REAGU-Net network specifically includes the following steps: 1) Create a two-dimensional brain model for simulation in CST Microwave Studio software, and set the dielectric constant and conductivity of each tissue. By changing the model's geometry and conductivity, dielectric constant, and sound velocity, construct N... t Different two-dimensional brain models were used, and the obtained conductivity, dielectric constant, and sound velocity parameter diagrams were normalized with the absorbed microwave power of the samples and used as subsequent reference data. 2) Based on the N constructed in step 1), t Different two-dimensional brain models are used to set up the entire simulated acoustic environment in a time-domain pseudospectral algorithm. This includes setting three acoustic parameters: sound velocity, density, and attenuation coefficient. The two-dimensional distribution of these acoustic parameters is consistent with the size of the entire simulation area. An initial two-dimensional sound pressure distribution is set as the sound source, generating sound waves. M ultrasound probes are arranged in a ring around the sound source, corresponding to the size of the initial two-dimensional sound pressure distribution, to receive ultrasound signals. N models are then constructed using data augmentation. t A set of independent two-dimensional simulation samples were used to obtain rich training data by varying the shape, size, thickness, sound velocity, and attenuation coefficient of the skull, as well as the location, shape, size, and microwave parameters of brain disease lesions within a set range. 3) For N t Using independent two-dimensional simulation samples, the time-domain ultrasound signals collected by the ultrasound probe were obtained through computer numerical calculations. The total signal dimension is M×L×N. t L represents the number of time points of the signal; 4) Preprocess the simulation signal by filtering, adding Gaussian noise, and incorporating the probe's receiving angle characteristics to improve the consistency between the simulation signal obtained by the simulation system and the actual experimental signal. 5) Simulation experiments were conducted using a calibrated simulated acoustic environment. It was assumed that the acoustic parameters in the entire environment were uniform. Preprocessed ultrasonic signals were used, and compressed sensing imaging technology was used to reconstruct the ultrasonic signals to obtain a preliminary two-dimensional image. The size and position of the imaging area were consistent with the initial two-dimensional sound pressure distribution. 6) Input the preliminary two-dimensional image obtained in step 5) into the network as training data, train the network, and obtain output data that is consistent with the size of the initial two-dimensional sound pressure distribution after passing through the network. 7) The conductivity, dielectric constant, and sound velocity distribution obtained from the electromagnetic simulation in step 1) are used as the artifact-free ground truth images and fed into the network as the benchmark data for training. The network described in steps 6) and 7) is a REAGU-Net network structure. The network input image in step 6) and the ground truth image in step 7) together form the training set. When the REAGU-Net network is trained, iteratively reduces the mean squared error loss function between the defined input data and the ground truth data, and updates many parameters of the REAGU-Net network to finally obtain the trained REAGU-Net network. The trained REAGU-Net network extracts the features of brain tumors and generates dielectric constant, conductivity and sound velocity distribution maps inside the brain tissue. 8) Build an experimental system, collect ultrasound data, and feed it into the previously trained REAGU-Net network for deep learning to obtain and restore the distribution of the three electromagnetic acoustic parameters of the human brain.

[0012] Furthermore, the brain model in step 1) includes the skull, brain tissue, cerebrospinal fluid, cerebral hemorrhage area and coupling oil layer to simulate the propagation characteristics of electromagnetic waves in the head structure. The absorbed microwave power of the sample is derived by performing electromagnetic simulation in CST.

[0013] Furthermore, the REAGU-Net network consists of a backbone network and multiple alternately connected residual excitation blocks (RE) and attention-guided blocks (AG), used to simultaneously reconstruct dielectric constant, conductivity, and sound velocity from back-projected thermoacoustic images. The input is a grayscale thermoacoustic image obtained by the ultrasonic back-projection BP algorithm. The image first enters the feature extraction layer of the network backbone, and multi-scale features are extracted through multiple convolutions, batch normalization, and ReLU activation functions. To achieve effective fusion of features at different levels, skip connection blocks are set between adjacent layers of the feature extraction layer in the backbone. Each skip connection block consists of alternating RE and AG modules. In the RE module, features are reconstructed through residual paths, and in the AG... In this module, an attention weight matrix is ​​obtained by calculating the similarity between high- and low-level features. The fused features are weighted to highlight key organizational structures and suppress artifact effects. The fused features output by all skip connection blocks are upsampled and decoded step by step to restore the original resolution and are then input to the feature decomposition layer at the end of the network. The feature decomposition layer generates three-channel outputs by setting a 1×1 convolutional layer at the output of the network feature extraction. Each channel corresponds to a quantitative prediction result of a physical parameter. During the training phase, measured data of dielectric constant, conductivity, and sound velocity are used as supervision signals and trained with the three channel outputs respectively, so that the network can automatically learn the coupling relationship and differences between the three physical quantities.

[0014] Furthermore, each skip connection block input includes a feature map f from a higher layer. hi and lower-level feature maps f l , where f hi Upsampled to f through deconvolution l With the same spatial dimensions, and f l The common input module generates the fused feature f oi In the RE module, the input features are first channel aligned and compressed through 1×1 convolution, then dot product operation is performed to enhance the spatial location information carried by the low-level features, and features are reconstructed through improved residual paths to enhance structural details.

[0015] The beneficial effects of this invention are as follows: This invention utilizes the REAGU-Net network-based microwave-induced thermoacoustic transcranial quantitative imaging method to simultaneously and quantitatively reconstruct the dielectric constant, conductivity, and speed of sound (SOS) of brain tissue using MITAT technology, providing a new methodology for the characteristic analysis of brain tissue. A novel network called REAGU-Net was developed, combining attention-guided (AG) blocks and residual evoked (RE) blocks to address the impact of acoustic inhomogeneities on reconstruction quality, demonstrating the effective application of deep learning in quantitative reconstruction. This technology not only shows good results in the detection of cerebral hemorrhage but also has potential applications in the detection and diagnosis of various diseases, including breast cancer, liver disease, and thyroid disease, providing new quantitative analysis tools for clinical practice. The proposed DL-MITAT method can provide quantitative physical characteristic data, which not only helps in imaging but also supports the prediction of physiological information in subsequent electromagnetic therapy. Compared with traditional TCD or ultrasound imaging methods, transcranial MITAT has significant advantages because its ultrasound waves only need to propagate through the skull once, reducing the possibility of image attenuation and degradation, thereby improving imaging quality. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the REAGU-Net network structure proposed in this invention; Figure 2 This is the overall implementation framework for the electromagnetic acoustic three-parameter quantitative recovery method using deep learning-enabled microwave-induced thermoacoustic imaging proposed in this invention. Figure 3 This is a schematic diagram of the simulation model of the present invention; Figure 4 The figures show the simulation test results of two simplified brain models of the present invention. Figure 5 The results of simulation tests using two real brain models are shown in this invention. Figure 6 The results of sound velocity reconstruction using conventional methods on brain models with and without skulls; Figure 7 A comparison chart showing the dielectric property reconstruction results using different networks; Figure 8 This is a schematic diagram of the experimental setup for the in vitro test of the present invention; Figure 9 The results of in vitro experiments using three different brain tissue samples are presented in this invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0018] like Figure 1 As shown, the proposed REAGU-Net network consists of a backbone network and multiple alternating residual excitation blocks (RE blocks) and attention-guided blocks (AG blocks) to simultaneously reconstruct dielectric constant, conductivity, and sound velocity from back-projected thermoacoustic images. The input is a grayscale thermoacoustic image obtained by the ultrasonic back-projection algorithm (BP), with a size of 160×160 and 1 channel. This image first enters the feature extraction layer of the network backbone, where multi-scale features are extracted through multiple convolutions, batch normalization, and the ReLU activation function. The feature map resolutions are 160×160, 80×80, 40×40, 20×20, and 10×10, corresponding to 64, 128, 256, 512, and 1024 channels, respectively. To achieve effective fusion of features at different levels, skip connection blocks are set between adjacent layers of the backbone, each consisting of alternating RE and AG modules. Taking the i-th level skip connection block as an example, its input includes feature maps f from higher layers. hi and lower-level feature maps f l , where f hi Upsampled to f through deconvolution l With the same spatial dimensions, and f l The common input module generates the fused feature f oi In the RE module, the input features are first channel-aligned and compressed using 1×1 convolutions, then dot product operations are performed to enhance the spatial location information carried by low-level features, and feature reconstruction is performed through improved residual paths to enhance structural details. In the AG module, an attention weight matrix is ​​obtained by calculating the similarity between high- and low-level features, and the fused features are weighted to highlight key organizational structures and suppress artifact effects. The fused features f output by all skip connection blocks are... oi After stepwise upsampling and decoding, the original resolution (160×160) is restored and input to the feature decomposition layer at the end of the backbone network.

[0019] This feature decomposition layer generates three-channel outputs through a 1×1 convolutional layer at the output of the network's feature extraction. Specifically, the fused feature map obtained after all skip connections and decoding processes has a size of 160×160×64, containing comprehensive features from different scales. The 1×1 convolutional layer takes this feature map as input and performs a linear weighted mapping on the 64-dimensional feature vector at each pixel location, with a kernel size of 3, thereby generating three independent channels. Since the 1×1 convolution only operates on the channel dimension and does not change the spatial resolution, the output remains a 160×160 image, with each channel corresponding to a quantitative prediction of a physical parameter. During the training phase, this invention uses measured data of dielectric constant, conductivity, and sound velocity as supervision signals, training them corresponding to the three channel outputs respectively, enabling the network to automatically learn the coupling relationships and differences between the three physical quantities. Specifically, the first channel primarily responds to the electric field energy storage characteristics of the medium, thereby learning the dielectric constant distribution; the second channel focuses on energy dissipation behavior, reflecting the conductivity distribution; and the third channel reconstructs the sound velocity by capturing differences in acoustic propagation paths. The three convolutional kernels form independent parameter sets during training, ensuring that the three output channels are physically related yet remain independent. The 1×1 convolutional layer here not only achieves the mapping from the high-dimensional feature space to the physical parameter space but also preserves the structural information in the backbone network without increasing computational complexity. Finally, the network outputs a fused image of size 160×160×3, with the three channels representing the spatial distribution of dielectric constant, conductivity, and sound velocity, respectively, achieving simultaneous quantitative reconstruction of the tissue's electrical and acoustic properties based on a single thermoacoustic backprojection image. This quantitative output enables the simultaneous reconstruction of dielectric constant, conductivity, and sound velocity.

[0020] The improved RE block differs from traditional residual blocks in three ways. First, it uses 1×1 convolutions instead of directly inputting the data into the block. Second, instead of concatenating two feature maps, we multiply a low-level feature map with the corresponding high-level feature map using a dot product. This is because the low-level feature map contains additional positional information from the input image. By multiplying a feature map layer, the result establishes a more accurate relationship between the network's output image and the input data. Finally, we can save some computational resources by halving the original input channels of the 3×3 convolution part. The improved RE module reduces memory usage to some extent while improving accuracy.

[0021] The pyramid pooling parameters of the AG blocks were modified to obtain feature maps of sizes 1×1, 2×2, 6×6, and 9×9, respectively, making the corresponding sizes physically compatible with the input and output images. The output of the current network is set to any value between 0 and 1, rather than being limited to binary (0 and 1) results.

[0022] like Figure 2 The framework shown is for transcranial tissue property reconstruction using a quantitative recovery algorithm for electromagnetic acoustic three parameters based on microwave-induced thermoacoustic imaging, taking a two-dimensional case as an example. It includes the following steps: Step 1: A two-dimensional brain model for simulation was created in the CST Microwave Studio software, and the dielectric constant and conductivity of each tissue were set. The brain model included the skull, brain tissue, cerebrospinal fluid (CSF), the area of ​​cerebral hemorrhage, and the coupling oil layer to simulate the propagation characteristics of electromagnetic waves in the head structure. Two types of brain models were used: one was a simplified model with an irregular skull shape and uneven thickness; the other was a realistic brain model obtained from 3.0T MRI images, with complex tissue morphology, which could more accurately reflect the actual brain structure. Electromagnetic simulation was performed in CST to derive the absorbed microwave power of the sample. Furthermore, by changing the model's geometry and conductivity, dielectric constant, and sound velocity, an N0... t Different two-dimensional brain models were used, and the obtained conductivity, dielectric constant, and sound velocity parameter diagrams were normalized with the absorbed microwave power of the samples for subsequent reference data.

[0023] Step 2: Based on the N constructed in Step 1 t Different two-dimensional brain models are used to set up the entire simulated acoustic environment in PSTD (Pseudospectral Time-Domain Algorithm). The size of the entire simulation area is S. x × S y (If it is a three-dimensional case, then it is S) x ×S y ×S z (The same applies below). Three acoustic parameters, including sound velocity, density, and attenuation coefficient, need to be set. The two-dimensional distribution of these acoustic parameters should be consistent with the size of the entire simulation area. The size is I. x ×I y (If it is a three-dimensional case, then it is I) x ×I y ×I z The initial two-dimensional sound pressure distribution (proportional to the microwave power absorbed by the sample) is set as the sound source to generate sound waves. M ultrasonic probes are arranged in a ring around I (or a three-dimensional curved surface in the case of three-dimensionality, which can be part of a sphere or a special curved surface designed according to different application scenarios). x ×I y The area around the sound source of a certain size is used to receive ultrasonic signals. A total of N is established through data augmentation. t The training set data is enriched by using independent two-dimensional simulation samples, in which the shape, size, thickness, sound velocity, and attenuation coefficient of the skull, as well as the location, shape, size, and microwave parameters of brain disease lesions, vary within a certain range.

[0024] Step 3: For N tUsing independent two-dimensional simulation samples, the time-domain ultrasound signals collected by the ultrasound probe were obtained through computer numerical calculations. The total signal dimension is M×L×N. t L represents the number of time points of the signal.

[0025] Step 4: Preprocess the simulation signal by filtering, adding Gaussian noise, and incorporating the probe's receiving angle characteristics to improve the consistency between the simulation signal obtained by the simulation system and the actual experimental signal, thereby improving the final imaging quality.

[0026] Step 5: A simulation experiment is conducted using a calibrated simulated acoustic environment. Assuming the acoustic parameters are uniform throughout the environment, pre-processed ultrasonic signals are used, and compressed sensing imaging technology is employed to reconstruct the image, obtaining a preliminary two-dimensional image. The size and location of the imaging area are related to the initial sound pressure distribution. x ×I y Maintain consistency.

[0027] Step 6: Input the preliminary 2D image obtained in Step 5 into the network as training data, train the network, and obtain the output data (also of size I) after passing through the network. x ×I y ).

[0028] Step 7: The conductivity, dielectric constant, and sound velocity distribution obtained from the electromagnetic simulation in Step 1 are used as artifact-free ground truth images and fed into the network for training. For brain detection, the initial sound pressure at the tumor site differs from that of the surrounding brain tissue. The novel REAGU-Net network structure proposed in this invention can focus on extracting features of brain tumors and generating dielectric constant, conductivity, and sound velocity distribution maps within brain tissue, which helps improve the reliability of brain disease detection. The input images and ground truth images together constitute the final training set. During network training, the mean squared error loss function between the defined input data and ground truth data is continuously reduced through iteration, while simultaneously updating many parameters of the network.

[0029] In this invention, the REAGU-Net neural network is an improvement upon the traditional U-Net, integrating attention-guided modules (AGblocks) and residual activation modules (RE blocks) to address the following issues: input back-projection (BP) images often contain numerous artifacts, and directly stitching together high- and low-level features may introduce noise and redundant information. Simultaneous output of the dielectric constant, conductivity, and speed of sound (SOS) distribution of brain tissue is required, demanding higher accuracy and robustness from the network.

[0030] Taking a two-dimensional backpropagation (BP) image as input as an example, the feature map first passes through a feature extraction module consisting of standard convolutional layers. In each convolutional unit, the feature map undergoes 3×3 convolution, batch normalization, and non-linear activation functions (such as ReLU). The result of each convolution is passed layer by layer and further optimized through a residual block. After the residual block, the spatial dimension of the feature map remains unchanged, but the number of channels increases. Subsequently, the feature map is processed by pooling layers, halving its size (or rounding down if not divisible), becoming 80×80. This process continues through the encoder path on the left, where the feature map is downsampled layer by layer, decreasing the spatial resolution of each layer while gradually increasing the number of channels, until reaching the deepest layer of the network. At this point, the feature map size is reduced to 10×10, while the number of channels increases to 1024.

[0031] The AG module, set up within the attention-focused processes of cascaded operations and skip connections, prioritizes the extraction of key information from the original downsampling process. In other words, when using the AG module, we focus more on its assessment of feature regions in our model, such as the location and size of bleed points. The RE module, on the other hand, focuses on detailed adjustments to the overall position. Specifically, in the RE module, the use of convolutional structures and dot product operations with the original feature map ensures the reuse of the spatial locations of the image obtained from our BP.

[0032] AG module (Attention-Guided Block): First, input the high-level feature map (f hi Low-resolution feature maps (f) contain semantic information but have lower resolution. l (): High resolution, containing location and texture information. For high-level feature maps f hi Upsampling is performed: a new f is obtained using deconvolution, batch normalization (BN), and the ReLU activation function. hi For the low-level feature map f l Perform a 1×1 convolution to extract key information, while reducing the number of channels for better matching. Then adjust the f... hi and f l The feature map is fed into pyramid pooling to perform pooling operations at different scales (1 × 1, 2 × 2, 6 × 6, 9 × 9). The pooling results are then fused using matrix multiplication (*) to generate attention weights, which are normalized to the [0, 1] range using an activation function (such as sigmoid). The newly obtained high-level feature map f is then processed...hi The attention weight matrix is ​​multiplied pixel-by-pixel to generate fused features. Finally, the fused features are convolved with a 3×3 matrix to generate the output feature map f. oi This serves as the final output of the module.

[0033] RE Module (Residual Evoke Block): First, input the high-level feature map (f) hi ): Rich in semantic information. Low-level feature map (f l ): Contains edge and location information. For high-level feature maps f hi Perform deconvolution to make it convolve with f l With the same resolution, adding batch normalization (BN) and the activation function (ReLU) yields a new f. hi For the low-level feature map f l Perform a 1×1 convolution to extract key information, while reducing the number of channels for better matching. Then adjust the f... hi and f l The feature map is fused using dot products to better utilize the complementarity of high- and low-level features. The feature map after a 3×3 convolution operation (including batch normalization and ReLU activation function) is used as the final output f of the module. oi .

[0034] In the decoder branch, the feature maps are upsampled (deconvolutioned) layer by layer to restore resolution: First layer: Restored from 20×20 to 40×40; Second layer: Restored from 40×40 to 80×80; Third layer: Restored from 80×80 to 160×160.

[0035] At each layer, feature maps from the corresponding layer in the encoder branch are fused with decoder feature maps via skip connections. Alternating AG Block and RE Block processing extracts and fuses feature information from different levels, and an attention mechanism enhances the expressive power of important regions.

[0036] The network output contains three channels, each representing the reconstruction result of one physical parameter: Channel 1: Dielectric constant distribution; Channel 2: Conductivity distribution; Channel 3: Speed ​​of sound (SOS) distribution.

[0037] This laid the foundation for subsequent physical analysis or medical applications.

[0038] Step 8: Build an experimental system, collect ultrasound data, and feed it into the previously trained REAGU-Net network for deep learning to obtain and restore the distribution of the three electromagnetic acoustic parameters of the human brain (dielectric constant, conductivity, and sound velocity).

[0039] Figure 3 The diagrams show the brain models used to train and validate the DL-MITAT technology. (a) and (b) include two two-dimensional brain models of different levels of complexity, and (c) is a random skull shape used for training.

[0040] Figure 4 The simulation test results using a simplified brain model are shown in the figure. The results of quantitative reconstruction of the simplified brain model using DL-MITAT technology are presented, including the distribution of dielectric constant, conductivity and speed of sound (SOS), and a comparison with the true values ​​(Ground Truth).

[0041] Figure 5 The figures show the simulation test results using a real brain model, illustrating the quantitative reconstruction results of the DL-MITAT technology on a real brain model, including reconstruction results under two different cases. The real brain model is closer to the structure of the actual human brain and has higher complexity, further validating the applicability and robustness of the method.

[0042] Figure 6 The images show the results of SOS reconstruction of brain models (with and without skulls) using conventional methods. They compare the results of SOS reconstruction of brain models using conventional methods (Bayesian reformulation), with the images showing the reconstruction effects in the skull-less and skull-containing cases, respectively.

[0043] Figure 7 The dielectric properties are reconstructed using different networks (where AGREU-net is the network used in this patent). Figure 6 and Figure 7 The comparison aims to highlight the superiority of DL-MITAT technology.

[0044] To ensure the practical usability and robustness of this technology in real-world devices, we conducted actual thermoacoustic experiments in the laboratory using a brain model, such as... Figure 8As shown, the samples were all immersed in an oil tank filled with cooking oil as the coupling fluid for microwave and ultrasound. Although water is a good ultrasound coupling agent, it is not suitable for this situation because of its strong absorption of microwaves. A microwave source with a peak power of 20 kW excited a waveguide (WR430) antenna, radiating microwave pulse signals from bottom to top toward the samples with a pulse width of 0.5 μs and a duty cycle of 0.05%. An ultrasound probe with a center frequency of 2.25 MHz was placed outside a bovine bone ring and rotated by a rotating motor to scan along a circular ring, scanning a total of 360 positions and acquiring ultrasound signals. The signals were amplified by 59 dB by a single-channel signal amplifier and then recorded using a data acquisition card. Figure 9 This image shows the test results of in vitro experiments using three different brain samples. It illustrates the reconstruction results of this technique in in vitro experiments, including three different brain samples (tissue phantoms and pig brain slices), along with the reconstruction performance and error analysis of dielectric constant, conductivity, and speed of sound (SOS).

[0045] The embodiments described above merely illustrate specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for microwave-induced thermoacoustic transcranial quantitative imaging based on REAGU-Net network, characterized in that, The specific steps are as follows: 4) The simulation signal is preprocessed by filtering, adding Gaussian noise, adding probe receiving angle characteristics, etc., to improve the consistency of the simulation signal obtained by the simulation system and the actual experimental signal; 2. The method of claim 1, wherein the REAGU-Net network is used for microwave-induced thermoacoustic transcranial quantitative imaging. 5) The calibrated simulation acoustic environment is used for simulation experiment, assuming that the acoustic parameters in the whole environment are uniform, the preprocessed ultrasonic signal is used, and the image reconstruction of the ultrasonic signal is carried out by means of compressed sensing imaging technology to obtain a preliminary two-dimensional image; The size and position of the imaging area are consistent with the size of the two-dimensional initial sound pressure distribution; 1) Establish a two-dimensional brain model for simulation in CST Microwave Studio software, and set the dielectric constant and conductivity of each tissue. By changing the model geometry and conductivity, dielectric constant, and sound speed, construct N t different two-dimensional brain models, and standardize the resulting conductivity, dielectric constant, and sound speed parameter maps with the sample's microwave power absorption for use as subsequent reference data; 2) N t different two-dimensional brain models, in the time domain pseudospectral algorithm, the whole simulation acoustic environment is set, including the setting of three acoustic parameters of sound velocity, density, and attenuation coefficient, the two-dimensional distribution of acoustic parameters is consistent with the size of the whole simulation area; the two-dimensional initial sound pressure distribution is set as a sound source to generate sound waves; M ultrasonic probes are arranged in a ring around the sound source with the size of the two-dimensional initial sound pressure distribution to receive ultrasonic signals; N t independent two-dimensional simulation samples are established through data augmentation, the shape, size, thickness, sound velocity, and attenuation coefficient of the skull, and the position, shape, size, and microwave parameters of the brain disease lesion vary in the set range to obtain rich training set data; 3) for N t independent two-dimensional simulation samples, time-domain ultrasound signals collected by an ultrasound probe are obtained by computer numerical calculation, and the total signal dimension is M x L x N t , L represents the number of time points of the signal; 6) The preliminary two-dimensional image obtained in step 5) is input into the network as training data, and the network is trained, and the output data consistent with the size of the two-dimensional initial sound pressure distribution is obtained through the network; 7) The conductivity, permittivity and sound speed distribution obtained by electromagnetic simulation in step 1) are used as the true value image without artifacts as the benchmark data for network training and input into the network for training; The network in steps 6) and 7) is a REAGU-Net network structure, the network input image in step 6) and the true value image in step 7) together constitute a training set, and when the REAGU-Net network is trained, the mean square error loss function between the defined input data and the true value data is continuously iterated and reduced, and the parameters of the REAGU-Net network are updated at the same time, and finally the trained REAGU-Net network is obtained; The trained REAGU-Net network extracts the features of brain tumors and generates the permittivity, conductivity and sound speed distribution map inside the brain tissue; 8) Build an experimental system, collect ultrasonic data, and input it into the aforementioned trained REAGU-Net network to obtain the recovered brain electromagnetic acoustic three-parameter distribution through deep learning. The brain model in step 1) includes skull, brain tissue, cerebrospinal fluid, brain hemorrhage area and coupling oil layer to simulate the propagation characteristics of electromagnetic waves in the head structure, and the microwave power absorbed by the sample is derived through electromagnetic simulation in CST. ​ 3. The method of claim 2, wherein the REAGU-Net network is a residual network (ResNet) and a generative adversarial network (GAN). ​ 4. The method of claim 2, wherein the REAGU-Net network is a residual network. The REAGU-Net network is composed of a backbone network and a plurality of alternately connected residual excitation blocks RE and attention guiding blocks AG, and is used for simultaneously reconstructing permittivity, conductivity and sound speed from back projection thermoacoustic images; an input is a gray-scale thermoacoustic image obtained by an ultrasonic back projection BP algorithm, the image first enters a feature extraction layer of a network backbone part, and multi-scale features are extracted through a plurality of convolution, batch normalization and ReLU activation functions; in order to realize effective fusion of features at different levels, a skip connection block is arranged between adjacent layers of the feature extraction layer of the backbone, each skip connection block is alternately composed of an RE module and an AG module, in the RE module, feature reconstruction is performed through a residual path, in the AG module, an attention weight matrix is obtained by calculating the similarity of high-level and low-level features, and fused features are weighted to highlight key tissue structures and suppress the influence of artifacts; the fused features output by all the skip connection blocks are restored to the original resolution after being progressively up-sampled and decoded, and are input to a feature decomposition layer at the end of the network; The feature decomposition layer realizes generation of three-channel output through a 1×1 convolution layer arranged at the output end of network feature extraction, each channel corresponds to a quantitative prediction result of one physical parameter, and in the training stage, measured data of permittivity, conductivity and sound speed are taken as a supervision signal, and three-channel output is trained respectively, so that the network automatically learns the coupling relationship and difference between the three physical quantities.

5. The method of claim 4, wherein the REAGU-Net network is a residual network. Each of the jump connection block inputs comprises a feature map f hi from a higher layer and a feature map f l from a lower layer, wherein f hi is up-sampled to the same spatial size as f l by deconvolution, and a fusion feature f l is generated by a joint input module with f oi ; in the RE module, the input features are first aligned and compressed in channels by 1x1 convolution, then a dot product operation is performed to strengthen the spatial position information carried by the low-level features, and the features are reconstructed by an improved residual path to enhance the structural details.

Images