A single-probe real-time ultrasound imaging method based on disordered metasurface and AI

Abstract

The application discloses a single-probe real-time ultrasonic imaging method based on a disordered super surface and AI, which comprises the following steps: placing an object to be imaged in an imaging system, transmitting a broadband signal by using a single transducer, and obtaining an encoded imaging signal by using a single spatially fixed hydrophone at a receiving end; sequentially performing time-frequency conversion on the imaging signal to obtain spectrum information of the imaging signal; inputting the spectrum information of the imaging signal into a trained neural network to obtain an imaging result of the object to be imaged; wherein the imaging system comprises an ultrasonic transducer, a disordered super surface and a hydrophone, wherein the ultrasonic transducer and the disordered super surface are placed on the same straight line, and the hydrophone is placed behind the super surface; the ultrasonic transducer emits sound waves to the object to be imaged, the emitted sound waves are scattered when reaching the object to be imaged to obtain scattered waves, the scattered waves interact with the disordered super surface to obtain imaging signals, and the imaging signals are received by the hydrophone; the disordered super surface is a composite material with randomly distributed spatial acoustic parameters and dispersion in a working frequency band.

Description

Technical Field

[0001] This invention relates to the field of acoustics, specifically to a single-probe real-time ultrasonic imaging method based on disordered metasurfaces and AI. Background Technology

[0002] Due to its high biocompatibility and low invasiveness, ultrasound imaging plays an irreplaceable role in many fields, including medical ultrasound, industrial crack and fatigue testing. The idea of ​​making structures or specimens visible through ultrasound originated from the military applications of underwater sonar ranging and navigation, similar to echolocation by marine animals or bats. Ultrasonic mirroring and ultrasound angiography were the first techniques to use ultrasound to detect internal defects in metal castings and to visualize the ventricles of the human brain, respectively.

[0003] Since its inception, ultrasound imaging technology has made significant progress. Phased arrays are commonly used to detect complex geometries and various material defects, as well as to scan internal organs for health diagnosis. Furthermore, specialized robotic surgery utilizes real-time ultrasound imaging for positioning and visualization. Traditional ultrasound imaging techniques typically employ complex phased arrays at both the transmitter and receiver to rapidly acquire spatial information of objects. However, the unavoidable high manufacturing costs and large system size hinder the integration and miniaturization of imaging systems. In contrast, existing single-probe imaging technologies offer significant advantages such as compact physical size, simple system construction, and low manufacturing costs. However, existing single-probe imaging technologies often require spatial mechanical scanning or multi-mode mask switching, and necessitate multiple time-consuming measurements, making real-time imaging difficult to achieve. Although computational imaging algorithms such as compressed sensing can reduce the number of measurements required to some extent, image reconstruction still requires multiple measurements, and the subsequent post-processing of measurement data significantly increases the workload. Therefore, the latency problem remains fundamentally unresolved. In addition, compressed sensing imaging requires known mask encoding, a requirement that becomes even more difficult to meet in terms of manufacturing and accuracy at higher ultrasound frequencies.

[0004] Therefore, low-cost real-time single-probe imaging remains a problem that urgently needs to be solved in both the optical and acoustic fields. Summary of the Invention

[0005] Purpose of the invention: To address the challenges of high integration and miniaturization in traditional ultrasound imaging technology, the large number of measurements and post-processing workload in existing single-probe imaging technology, and the inherent contradiction between single-probe imaging and real-time imaging in existing technologies, this invention proposes a single-probe, single-measurement real-time ultrasound imaging method based on disordered metasurfaces and AI. This method features fast imaging speed, low manufacturing cost, compact system size, and high imaging quality.

[0006] Technical solution: A single-probe real-time ultrasound imaging method based on disordered metasurfaces and AI, comprising the following steps:

[0007] Step 1: Place the object to be imaged in the imaging system, use a single transducer to transmit a broadband signal, and use a single spatially fixed hydrophone at the receiving end to obtain the encoded imaging signal.

[0008] Step 2: Perform time-frequency conversion on the imaging signal sequentially to obtain the spectral information of the imaging signal;

[0009] Step 3: Input the spectral information of the imaging signal into the trained neural network to obtain the imaging result of the object to be imaged;

[0010] The imaging system includes a transducer, a disordered metasurface, and a hydrophone. The ultrasonic transducer and the disordered metasurface are placed in a straight line, and the hydrophone is placed in a fixed position behind the metasurface. The object to be imaged is placed between the transducer and the disordered metasurface, and is kept in a straight line with both. The transducer emits a signal to the object to be imaged. The emitted signal is scattered when it reaches the object to be imaged, forming a scattered wave. The scattered wave interacts with the disordered metasurface to obtain an encoded imaging signal, which is received by the hydrophone. The disordered metasurface is a composite material with randomly distributed spatial acoustic parameters and dispersive properties in the operating frequency band.

[0011] The scattered waves interact with the disordered metasurface to obtain an encoded imaging signal. Specifically, the disordered metasurface encodes the scattered waves, encoding the shape information of the object into the spectrum of any point in the metasurface space.

[0012] Furthermore, the size of the transducer is determined by the size of the object to be imaged, and its diameter should be larger than the size of the object to be imaged.

[0013] Furthermore, the size of the transducer's emitting surface is such that the emitted signal can cover the object to be imaged.

[0014] Furthermore, the size of the disordered metasurface is chosen to ensure that it can receive most of the information from the scattered sound field of the object being imaged. This ensures that the metasurface interacts with most of the cross-section of the scattered sound field from the object. Further, a neural network decodes the spectral information of the imaging signal.

[0015] Furthermore, the center frequency and bandwidth of the transmitted signal are determined by the imaging resolution requirements.

[0016] Furthermore, the transducer transmits an ultrasonic pulse with a bandwidth of 60% or more of the central operating frequency to the object to be imaged, and the wavelength of the central operating frequency is less than one-quarter of the required imaging resolution.

[0017] Furthermore, the disordered metasurface is a metasurface obtained by randomly doping two materials with different acoustic parameters, satisfying a random spatial parameter distribution and exhibiting dispersion in the operating frequency band.

[0018] Furthermore, the disordered metasurface is a composite material with randomly distributed spatial acoustic parameters and dispersive properties in the operating frequency band, obtained by doping agar and randomly distributed steel balls.

[0019] Furthermore, the doping of agar and randomly distributed steel balls specifically includes the following operations:

[0020] Take agar powder, steel balls and water in a mass ratio of 1:8:9, mix the agar powder and water thoroughly to obtain a mixed solution, and heat the mixed solution to boiling.

[0021] When the mixed solution becomes viscous, steel balls are added, and after thorough stirring, it is poured into a container to cool and solidify, resulting in a composite material with randomly distributed spatial acoustic parameters and dispersive properties in the operating frequency band.

[0022] Furthermore, the spectrum is a 3MHz spectrum within a 2.5MHz bandwidth range close to the center operating frequency.

[0023] Furthermore, the neural network is a multilayer perceptron with two hidden layers.

[0024] Furthermore, the neural network is a convolutional neural network or a cascaded neural network.

[0025] Furthermore, the trained neural network is obtained by training according to the following training steps:

[0026] S10: Collect a certain number of objects of the same type as the object to be imaged, forming an object set; there are differences between the objects in the object set; place the objects in the object set one by one into the imaging system to obtain the corresponding encoded imaging signal, and perform time-frequency conversion on the imaging signal in sequence to obtain the corresponding time-frequency conversion spectrum information; thus constructing the training set.

[0027] S20: The test set is constructed according to S10, and the number of samples in the test set is less than the number of samples in the training set.

[0028] S30: The neural network is trained using the spectral information in the test set, and the trained neural network is tested using the test set to obtain a neural network that meets the imaging quality requirements.

[0029] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0030] (1) The method of the present invention breaks the contradiction between single-probe imaging and real-time imaging in the prior art. It utilizes low-cost disordered metasurfaces, and only requires a single measurement data to achieve real-time high-precision imaging of objects with a single probe. It also has the advantage of compact system size.

[0031] (2) The imaging quality obtained by the method of the present invention does not depend on the spatial resolution of the sensor, so there is no need for a high-cost high-resolution probe, and it truly realizes low-cost, low-system-complexity, and high-quality real-time object imaging with a single ultrasonic probe.

[0032] (3) The hydrophone in the method of the present invention has a fixed position and only needs to be measured once to achieve real-time imaging, without the need for mechanical scanning and multiple measurements;

[0033] (4) The imaging architecture of the method of the present invention is not limited to transmission imaging, and the imaging architecture based on echo scattering is also within the scope of protection of this patent. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the process of a single-probe real-time ultrasonic imaging method based on disordered metasurfaces and AI proposed in this invention.

[0035] Figure 2 A schematic diagram of typical imaging results of handwritten digits in the test set and training set;

[0036] Figure 3 To train the overall graph, where, Figure 3 (a) in the figure shows the change of the average SSIM in the training and test sets with the number of training iterations. Figure 3 (b) in the figure represents the SSIM distribution of 600 handwritten digits in the training set. Figure 3 (c) in the test set represents the SSIM distribution of 100 handwritten digits. Detailed Implementation

[0037] The technical solution of the present invention will now be further described in conjunction with the accompanying drawings and embodiments.

[0038] Example 1:

[0039] This embodiment discloses a single-probe real-time ultrasonic imaging method based on disordered metasurfaces and AI. This method utilizes the strong randomness of disordered media and extracts rich imaging-related information through artificial intelligence (AI). The main steps include:

[0040] Step 1: Construct an imaging system comprising an ultrasonic transducer, a disordered metasurface, and a hydrophone. The ultrasonic transducer emits a broadband ultrasonic signal towards the object to be imaged. Upon reaching the object, the emitted broadband ultrasonic signal is scattered, forming a scattered wave. This scattered wave interacts with the disordered metasurface located behind the object. The disordered metasurface encodes the scattered wave, encoding the object's shape information into the spectrum of any point behind the metasurface, thus obtaining the encoded imaging signal. The hydrophone, located behind the disordered metasurface, receives the imaging signal. In this embodiment, the ultrasonic transducer's emitted signal is determined to be a broadband ultrasonic pulse based on the required imaging resolution. Ideally, the bandwidth should be 60% or more of the center frequency, and the wavelength of the center operating frequency should generally be less than one-quarter of the required imaging resolution. The disordered metasurface used in this embodiment only needs to have randomly distributed spatial acoustic parameters and dispersive properties in its operating frequency band. Therefore, the simplest approach is to randomly dope materials with two different acoustic parameters, such as doping agar blocks with steel balls. When using it, it is necessary to ensure that the ultrasonic transducer, the object under study, the disordered metasurface, and the hydrophone are placed in the same propagation path in sequence.

[0041] Step 2: Perform time-frequency conversion on the encoded imaging signal to obtain the spectrum information of the imaging signal; input the spectrum information of the imaging signal into the trained neural network, and the neural network decodes the spectrum information of the imaging signal. In this embodiment, the neural network is used to extract the shape information of the object under study from the spectrum of the signal.

[0042] The neural network used in this embodiment is determined based on the specific imaging object. For complex objects, convolutional neural networks or cascaded neural networks can be used; for simple imaging objects, simple neural networks such as multilayer perceptrons can be used. With only limited data for training, the neural network can effectively construct the computational relationship between the input signal and the output object shape without requiring any prior knowledge of the metasurface.

[0043] The neural network training process of this embodiment will now be further explained:

[0044] Constructing the training set: Based on actual imaging requirements, a certain number of objects belonging to the same category as the object to be imaged are collected to form a set. The objects in this set should possess certain differences, such as different shapes. An object from this set is placed between an ultrasonic transducer and a disordered metasurface. The ultrasonic transducer emits a broadband ultrasonic signal to the object, and a hydrophone receives the encoded imaging signal. Each received imaging signal corresponds one-to-one with a placed object. Each received imaging signal is zero-padded before undergoing a Fast Fourier Transform (FFT) to obtain the spectral information of the imaging signal.

[0045] Construct a test set. The requirements for constructing a test set are the same as those for the training set, but the number of test sets is generally about one-fifth of that of the training set.

[0046] Neural network training: The signals from the training and test sets are used as inputs to train the neural network. The label of the neural network is the shape of the object. This embodiment uses the structural similarity (SSIM) metric to measure image quality. By continuously adjusting various hyperparameters of the neural network, the optimal SSIM is achieved as much as possible. Generally, an SSIM > 0.6 in the test set is considered to indicate that the image has been reconstructed well.

[0047] This embodiment utilizes the different scattered waves generated when ultrasound encounters objects of different shapes. By leveraging the strong spatial parameter randomness and dispersive properties of the disordered metasurface, the scattered waves can be effectively encoded. The randomness of the spatial acoustic parameter distribution facilitates encoding in the spatial domain, while the dispersive properties allow for multiplexing of the encoding pattern at different frequency points. Therefore, the spatial information of the scattered waves emitted from the object can be effectively reflected in the spectrum of the sound wave at any location behind the metasurface. Finally, using a neural network, the object shape corresponding to the input spectrum can be obtained without any prior knowledge of the disordered metasurface.

[0048] Example 2:

[0049] The technical solution of this invention will now be further illustrated using underwater ultrasonic handwritten digital imaging as an example. Specifically, it includes:

[0050] Step 1: Construct the training set for the object to be imaged; In this embodiment, the object to be imaged is a metal plate engraved with handwritten digits, measuring 2cm × 2cm and 6mm thick. The handwritten digits are etched in the central region (1.4cm × 1.4cm), while the outer parts remain unprocessed. 600 handwritten digits are selected from the MNIST dataset, including 60 instances of each digit from 0 to 9. These are then binarized, and laser cutting is used to cut off the portions with values ​​of 1, while the portions with values ​​of 0 are left unprocessed. Because each handwritten digit in the MNIST dataset contains 28 × 28 pixels, the pixel size on the metal plate is 0.5mm × 0.5mm.

[0051] Step 2: Construct a test set of the objects to be imaged: Select 100 handwritten digits from the MNIST dataset that are different from those in the training set, including 10 instances of each digit from 0 to 9. The other steps are the same as for the training set.

[0052] Step 3: Determine the emitted acoustic wave of the ultrasonic transducer according to the imaging resolution requirements: Based on the geometry of the object to be imaged, in this embodiment, the signal emitted by the ultrasonic transducer is a pulsed sinusoidal signal with a center frequency of 2.5MHz and a cycle number of 2, to ensure sufficient energy within a bandwidth of 1-4MHz. No wavefront shaping is performed on the emitted ultrasonic waves. There are no specific requirements for the wavefront of the ultrasonic waves emitted by the transducer in the experiment; it is sufficient that the waveform is consistent each time.

[0053] Step 4: Preparation of the disordered metasurface required for imaging: The disordered metasurface used in this embodiment consists of agar and randomly distributed steel beads. The steel beads have a density of approximately 7.77 g / cm³, a diameter of 0.8 mm, and correspond to the wavelength of the ultrasonic signal at its center frequency emitted by the ultrasonic transducer. The agar used is laboratory-grade agar powder. During preparation, the mass ratio of agar powder, steel beads, and water is 1:8:9. The agar powder and water are thoroughly mixed and heated to boiling. When the agar solution becomes viscous, the steel beads are added, and the mixture is thoroughly stirred before being poured into a container. Once the agar solution containing the steel beads cools and solidifies, the disordered metasurface required for imaging is complete. Finally, the material is cut into the desired shape and size. Since stirring is a process that results in the steel beads being randomly distributed in the agar solution, and the method of this embodiment only requires the steel beads to be randomly distributed without any other requirements, no precise or expensive measuring instruments or manufacturing methods are needed throughout the process.

[0054] Step 5: Constructing the Imaging System: This embodiment is conducted in a 30*30*30cm water tank. A frame housing the ultrasonic transducer, disordered metasurface, and hydrophone is fabricated using Computer Numerical Control (CNC) technology and fixed to an external, adjustable-size aluminum alloy frame. The frame for the object to be imaged (metal plate) is 3D printed and installed on the edge of the water tank for easy replacement during the experiment. The ultrasonic transducer, the object to be imaged, the disordered metasurface, and the hydrophone are placed on the same propagation path. The transmitted signal is provided by a signal generator and amplified by a connected power amplifier before being transmitted to the ultrasonic transducer.

[0055] Step 6: Training Set Signal Acquisition: The ultrasonic transducer emits a broadband signal, which is received by a hydrophone after passing through the objects in the training set and the disordered metasurface. The received signal is received and recorded by an oscilloscope. The sampling rate is set to 50MHz, the number of sampling points is set to 1000, and only 20µs of time-domain signal needs to be acquired.

[0056] Step 7: Simple signal processing: After importing the recorded signals sequentially into MATLAB software, each signal is artificially padded to 400µs by appending a sequence of zero values ​​to the end. Then, a Fast Fourier Transform (FFT) is performed to obtain the signal spectrum. In this embodiment, a 3MHz spectrum within a 2.5MHz bandwidth close to the center frequency is selected as the input to the subsequent neural network. The entire process avoids complex and time-consuming operations such as matrix inversion, matrix decomposition, and iteration, resulting in a very short operation time.

[0057] Step 8: Test set signal acquisition and processing: The acquisition and processing of test set signals are the same as those of the training set.

[0058] Step 9: Neural Network Training: The signals collected from the training and test sets are used as inputs to train the neural network. The labels of the neural network are the shapes of handwritten digits. In this embodiment, a multilayer perceptron with two hidden layers is constructed based on the Adaptive Moment Estimation (Adam) optimizer widely used in deep learning. The number of neurons in the four layers are 501, 600, 700, and 784, respectively. The activation function used in the first three layers is the tanh function, while the output layer uses the sigmoid function. Then, the output tensor is changed from 1×784 to 28×28, thereby calculating the loss function as shown below:

[0059]

[0060] Where l is the batch size of the data processed in one operation, SSIM(·) is the structural similarity operation, and y and Let l be the label tensor and l be the output tensor, both of shape m×n. In this training, l = 32, m = n = 28. k and These are the label in the data batch of a single operation and the k-th data point in the output tensor, respectively. and Perform a single operation on the label in the data batch and the i-th row and j-th column of the k-th data in the output tensor.

[0061] Figure 2To illustrate typical imaging results from this embodiment, the original shapes of objects in the training and test sets were compared with their corresponding imaging results. The differences between the two sets of images in the training set are almost negligible, indicating near-complete reconstruction of shape information. The two sets of images in the test set also show high similarity, signifying the high fidelity and accuracy of the imaging mechanism proposed in this invention. Therefore, it can be concluded that, based on the training data provided by the training set, the neural network can effectively decode the scattered wave information encoded by disordered metasurfaces. Notably, once the neural network is trained, the entire imaging process can be completed within tens of microseconds, and the imaging speed can be further increased by increasing the bandwidth of the emitted signal. This demonstrates the potential of this invention for real-time imaging of dynamic objects (such as a beating heart), which is of great significance in practical medical applications.

[0062] Figure 3 (a) shows the training process of the neural network, demonstrating the excellent convergence of the model. The average SSIM on the training and test sets reaches 0.9870 and 0.7853, respectively. Figure 3 (b) and Figure 3 (c) shows the SSIM distribution bubble plots for each imaging result in the training and test sets, quantifying the similarity between the imaging result and its corresponding original object for each sample in both sets. In the training set, the SSIM for each sample exceeds 0.92, exhibiting significant clustering in the range of 0.96–1.00. In the test set, the imaging results for each sample show an SSIM exceeding the threshold of 0.6, with no instances falling below this threshold. This demonstrates the effectiveness of the proposed method in achieving efficient imaging of all samples, with no instances of poor imaging quality for any single sample. This consistency and reliability of imaging performance is crucial for practical applications, significantly improving diagnostic accuracy.

[0063] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0064] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and 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 this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A single-probe real-time ultrasound imaging method based on disordered metasurface and AI, characterized in that: Includes the following steps: Step 1: Place the object to be imaged in the imaging system, use a single transducer to transmit a broadband signal, and use a single spatially fixed hydrophone at the receiving end to obtain the encoded imaging signal. Step 2: Perform time-frequency conversion on the imaging signal to obtain the spectral information of the imaging signal; Step 3: Input the spectral information of the imaging signal into the trained neural network. The neural network decodes the spectral information of the imaging signal to obtain the imaging result of the object to be imaged. The imaging system includes a transducer, a disordered metasurface, and a hydrophone. The transducer and the disordered metasurface are placed on the same straight line, and the hydrophone is placed at a fixed position behind the metasurface. The object to be imaged is placed between the transducer and the disordered metasurface, and is kept on the same straight line as both. The transducer emits a signal to the object to be imaged. The emitted signal is scattered when it reaches the object to be imaged, forming a scattered wave. This scattered wave interacts with the disordered metasurface to obtain an encoded imaging signal, which is received by the hydrophone. The disordered metasurface is made of randomly doped different materials, has randomly distributed spatial acoustic parameters, and exhibits dispersivity in the operating frequency band. The scattered waves interact with the disordered metasurface to obtain an encoded imaging signal, specifically including: The disordered metasurface performs encoding of the scattered waves, encoding the shape information of the object into the spectrum of any point in the metasurface space; The trained neural network was obtained by following training steps: S10: Collect a certain number of objects of the same type as the object to be imaged, forming an object set; there are differences between the objects in the object set; place the objects in the object set one by one into the imaging system to obtain the corresponding encoded imaging signal, and perform time-frequency conversion on the imaging signal in sequence to obtain the corresponding time-frequency conversion spectrum information; thus constructing the training set. S20: The test set is constructed according to S10, and the number of samples in the test set is less than the number of samples in the training set. S30: The neural network is trained using the spectral information in the training set, and the trained neural network is tested using the test set to obtain a neural network that meets the imaging quality requirements.

2. The method of claim 1, wherein: The center frequency and bandwidth of the transmitted signal are determined by the imaging resolution requirements.

3. The method of claim 1, wherein: The transducer's emitting surface size is such that the emitted signal can cover the object to be imaged.

4. The method of claim 1, wherein: The size of the disordered metasurface is chosen to ensure that most of the information from the scattered sound field of the object to be imaged is received.

Images