A flexible ultrasonic sensor array imaging distortion correction method and system

Abstract

The application is a flexible ultrasonic sensor array imaging distortion correction method and system. The application relates to the technical field of ultrasonic signal collection and imaging research. Two data sets are generated in advance: the imaging of error delay and (Delay-And-Sum, DAS) under different deformation curvatures is regarded as error samples; and the imaging of correct DAS under different deformation curvatures is regarded as correct samples. The application trains an image generation model by means of the Diffusion algorithm idea, and takes the error samples as input and the correct samples as output. Forward diffusion adds Gaussian noise to the sample image, and reverse diffusion uses a neural network to complete image reconstruction by predicting the noise distribution of the sample with added Gaussian noise, so as to generate an undistorted image, thereby effectively improving the image quality and diagnostic reliability.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic signal acquisition and imaging research technology, and is a method and system for correcting imaging distortion of a flexible ultrasonic sensor array. Background Technology

[0002] Flexible ultrasonic sensors, as an emerging testing device, offer significant advantages over traditional ultrasonic sensors. In terms of geometric adaptability, they can conform to complex surfaces and can be customized in shape and size to suit various testing scenarios. They exhibit excellent performance, possessing high sensitivity, wide bandwidth response, and low noise, accurately capturing signals, acquiring rich information, and ensuring testing accuracy. In terms of operation and application, they are lightweight and easy to operate, facilitating on-site and mobile testing. They are also easily integrated with other technologies to form multimodal systems, and their good biocompatibility makes them suitable for in vivo testing in the biomedical field. They also boast excellent durability and reliability, resisting vibration and shock, and exhibiting good long-term stability. They can operate stably in harsh environments, reducing maintenance costs and calibration workload, and providing strong support for testing in multiple fields.

[0003] In the application of flexible ultrasonic sensors, the diverse shapes of the measured objects can lead to unknown deformations in the flexible sensors. Traditional delay and beamforming methods calculate the transmission delay of radio frequency (RF) data using the distance from the focal point to the sensor. When the curvature of the flexible ultrasonic sensor is unknown, erroneous transmission delays will occur, affecting the beamforming effect. This error will result in significant defocusing and distortion in ultrasound images, especially B-mode images, impacting diagnostic accuracy.

[0004] The current mainstream solution is to integrate a strain sensor onto a flexible ultrasonic sensor. When the flexible sensor deforms, the curvature of the sensor is measured, and the transmission delay of each radio frequency data point is recalculated based on the curvature. However, integrating a strain sensor increases the overall manufacturing difficulty of the sensor. Besides adding an extra sensor, algorithms can be used for data optimization. To address this issue, the impact of sensor deformation on the beamforming process needs to be considered, and corresponding compensation strategies or improved beamforming algorithms should be proposed to improve image quality and diagnostic reliability. In the flexible ultrasonic sensor array system of this invention, to achieve high-resolution, high signal-to-noise ratio imaging echo acquisition, each piezoelectric element in the sensor array is driven by a precisely controlled excitation signal. The waveform of the excitation signal used references the existing single-element ultrasonic signal excitation method (see paragraphs

[0092] to

[0096] of CN112869773A, "A Flexible Ultrasonic Sensor and Its Arterial Blood Pressure Detection Method").

[0005] In this invention, the excitation of each piezoelectric element in the flexible ultrasonic sensor array adopts the pulse excitation method proposed in the reference patent "A Flexible Ultrasonic Sensor and its Arterial Blood Pressure Detection Method" (CN112869773A). A control circuit applies a pulse signal with a specific waveform to each element to drive it to emit ultrasonic waves. The excitation signals used include single exponential decay pulses, decaying oscillating pulses, and square wave modulated pulses. For example, single exponential decay pulses are suitable for scenarios requiring rapid excitation response; decaying oscillating pulses can improve frequency domain resolution while maintaining strong penetration; and square wave modulated pulses facilitate the synchronous operation of multiple elements and enhance timing control stability. In this invention, these waveforms can be generated in real time by an FPGA or programmable signal source according to set parameters and allocated to each unit in the array. Furthermore, different excitation time windows and repetition frequencies can be set. Frequency (PRF) enables sequential scanning of the array, imaging synchronization, and dynamic focusing of the target area, thereby improving the balance between spatial sampling density and imaging depth without increasing hardware complexity. This excitation method works in conjunction with the diffusion model distortion correction mechanism introduced in this invention to ensure high-quality raw signal input, providing clear and accurate basic data for subsequent reconstruction. Summary of the Invention

[0006] This invention addresses the issue of erroneous transmission delays caused by the diverse shapes of the measured objects during imaging in flexible ultrasonic sensors. It provides a system and method for optimizing imaging distortion in flexible array ultrasonic sensors. Using two pre-generated datasets, an image generation model is trained using a diffusion model. When the flexible ultrasonic sensor produces distorted images due to deformation, the distorted image is used to generate an undistorted image through the image generation model. This method utilizes algorithms to optimize the distorted image without adding any additional sensors.

[0007] This invention provides the following technical solutions: A method for optimizing imaging distortion of a flexible array ultrasonic sensor, the method comprising the following steps: Step 1: Generating an incorrect DAS imaging dataset; Step 2: Generate the correct DAS imaging dataset; Step 3: The dataset samples are processed by the Encoder module of the Variational Autoencoder (VAE) to convert them into low-dimensional Latent features, and Gaussian noise is added step by step; Step 4: Add noise to the Latent of the erroneous samples and feed them into the Self-Attention and U-Net networks for training to predict the noise required to generate undistorted images; Step 5: The feature distribution is denoised sequentially using the noise predicted by the neural network. After complete denoising, the undistorted image is reconstructed using the Decoder module of the Variational Autoencoder (VAE).

[0008] Preferably, step 1 specifically comprises: When measuring objects with surfaces of varying curvature, the curvature of the array sensor is treated as 0, and the DAS of each RF is calculated using a predefined sensor model, calculated by the following formula:

[0009] in,( x f , y f , z f ) is the location of the focal point, ( x c , y c , z c ) is the reference center point of the focal point on the aperture of the sensor. x i , y i , z i () is an arbitrary chip of the sensor i Spatial center coordinates, c It's the speed of sound. t i It is the calculated DAS; The default curvature of the calculated DAS is 0. The DAS obtained at this time is erroneous data. Using this DAS for imaging will produce phenomena such as defocus and distortion. This dataset is regarded as an erroneous sample.

[0010] Preferably, step 2 specifically comprises: For each ultrasound image in the DAS imaging dataset that generates an error, there must be a corresponding correct DAS image in this dataset. For objects under test with different surface curvatures, the surface curvature of the object under test is measured by strain sensors. A flexible ultrasound sensor array model is established in the Field II simulation tool, and the curvature of each sensor element is set to the curvature of the object under test measured by the strain sensors. The spatial center coordinates of each sensor element can be obtained through the established sensor array model. The correct DAS is calculated, and this DAS is used for imaging. This imaging will not produce phenomena such as defocusing or distortion, and the generated dataset is used as the correct sample.

[0011] Preferably, step 3 specifically comprises: The graph-to-image model is trained based on the idea of ​​Diffusion algorithm, with erroneous samples as input images and correct samples as output images. The encoder of the dataset images is converted into low-dimensional latent features using the variational autoencoder (VAE). In the forward diffusion process, Gaussian noise is gradually added to the low-dimensional latent features of both datasets. x cond For the Latent distribution of erroneous data, x target For the correct data latent distribution, x cond The input used for the denoising model. x target The noise-adding process, used for calculating the loss function, is a Markov process:

[0012] Among them, the distribution of Latent features is x ~ q ( x ), x t For the first t Distribution of time steps t The hyperparameter for adding noise controls the amount of noise added at each step, where I is the variance of the added Gaussian noise.

[0013] Preferably, when T When the value is infinite, the distribution is complete Gaussian noise.

[0014] Preferably, step 4 specifically comprises: x cond The noise distribution after adding noise is then used for noise prediction via a neural network before denoising:

[0015] in, The mean of the noise predicted by the neural network. The variance of the noise predicted by the neural network. θ For neural network parameters; Will x cond The noisy distribution is then processed by Self-Attention and fed into the U-Net network. The loss function used for training the neural network is L2 loss, expressed by the following formula:

[0016] in, For noise predicted by a neural network, for x target Noise generated during the noise generation process.

[0017] Preferably, during reverse diffusion, the distribution of the distorted image after adding noise is processed by Self-Attention and then fed into the U-Net network to predict the noise required to reconstruct the image. The noise predicted by the neural network and the noise added to the undistorted image are used to calculate the loss function to update the neural network parameters. The undistorted image is reconstructed by predicting the noise required at each step through the neural network.

[0018] A flexible ultrasonic sensor array imaging distortion optimization system, the system comprising: An error dataset generation module generates an error DAS imaging dataset; A correct dataset generation module generates a correct DAS imaging dataset; The preprocessing module converts the dataset samples into low-dimensional Latent features using a variational autoencoder (VAE) and gradually adds Gaussian noise. The training module adds noise to the Latent of the erroneous samples and feeds them into the Self-Attention and U-Net networks for training to predict the noise required to generate undistorted images. The reconstruction module successively denoises the feature distribution using noise predicted by the neural network, and after complete denoising, reconstructs the undistorted image through the Decoder module of the Variational Autoencoder (VAE).

[0019] A computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement a method for correcting imaging distortion of a flexible ultrasonic sensor array.

[0020] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a method for correcting imaging distortion of a flexible ultrasonic sensor array.

[0021] The present invention has the following beneficial effects: This invention pre-generates two datasets: images of erroneous DAS images under different deformation curvatures, considered as erroneous samples; and images of correct DAS images under different deformation curvatures, considered as correct samples. This invention uses the Diffusion algorithm to train an image generation model, taking erroneous samples as input and correct samples as output. Forward diffusion adds Gaussian noise to the sample images, and backward diffusion uses a neural network to predict the noise distribution of the Gaussian noise-added samples to complete image reconstruction, thereby generating undistorted images and effectively improving image quality and diagnostic reliability. Attached Figure Description

[0022] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 The diagram shows a flowchart of the flexible array ultrasonic sensor imaging distortion optimization method of the present invention. Figure 2 The diagram shown is a schematic diagram of the DAS calculation of the present invention. Figure 3 The image shown is a bent sensor model of the present invention; Figure 4 The image shown is an unbent sensor model of the present invention. Detailed Implementation

[0024] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] The present invention will be described in detail below with reference to specific embodiments. Specific Implementation Example 1: according to Figures 1 to 4 As shown, the specific optimization technical solution adopted by the present invention to solve the above-mentioned technical problems is: The present invention relates to a flexible array ultrasonic sensor imaging distortion optimization system and method.

[0027] This invention provides a method for optimizing imaging distortion of a flexible array ultrasonic sensor, the method comprising the following steps: Step 1: Generating an incorrect DAS imaging dataset; Step 2: Generate the correct DAS imaging dataset; Step 3: The dataset samples are processed by a variational autoencoder (VAE) to convert them into low-dimensional latent features, and Gaussian noise is added progressively. Step 4: Add noise to the Latent of the erroneous samples and feed them into the Self-Attention and U-Net networks for training to predict the noise required to generate undistorted images; Step 5: The feature distribution is denoised sequentially using the noise predicted by the neural network. After complete denoising, the undistorted image is reconstructed using the Decoder module of the Variational Autoencoder (VAE). Specific Implementation Example 2: The only difference between Embodiment 2 and Embodiment 1 of this application is that: Step 1 specifically involves: When measuring objects with surfaces of varying curvature, the curvature of the array sensor is treated as 0, and the DAS of each RF is calculated using a predefined sensor model, calculated by the following formula:

[0029] in,( x f , y f , z f ) is the location of the focal point, ( x c , y c , z c ) is the reference center point of the focal point on the aperture of the sensor. x i , y i , z i () is an arbitrary chip of the sensor i Spatial center coordinates, c It's the speed of sound. t i It is the calculated DAS; The default curvature of the calculated DAS is 0. The DAS obtained at this time is erroneous data. Using this DAS for imaging will produce phenomena such as defocus and distortion. This dataset is regarded as an erroneous sample. Specific Implementation Example 3: The only difference between Embodiment 3 and Embodiment 2 of this application is that: Step 2 specifically involves: For each ultrasound image in the DAS imaging dataset that generates an error, there must be a corresponding correct DAS image in this dataset. For objects under test with different surface curvatures, the surface curvature of the object under test is measured by strain sensors. A flexible ultrasound sensor model is established in the Field II simulation tool, and the curvature of each sensor array element is set to the curvature of the object under test measured by the strain sensors. The spatial center coordinates of each sensor array element can be obtained through the established sensor array model. The correct DAS is calculated, and this DAS is used for imaging. This imaging will not produce phenomena such as defocusing or distortion, and the generated dataset is used as the correct sample. Specific Implementation Example 4: The only difference between Embodiment 4 and Embodiment 3 of this application is that: Step 3 specifically involves: The graph-to-image model is trained based on the idea of ​​Diffusion algorithm, with erroneous samples as input images and correct samples as output images. The Encoder module of Variational Autoencoder (VAE) is used to convert the dataset images into low-dimensional Latent features. In the forward diffusion process, Gaussian noise is gradually added to the low-dimensional latent features of both datasets. x cond For the Latent distribution of erroneous data, x target For the correct data latent distribution, x cond The input used for the denoising model. x target The noise-adding process, used for calculating the loss function, is a Markov process:

[0032] Among them, the distribution of Latent features is x ~ q ( x ), x t For the first t Distribution of time steps t The hyperparameter for adding noise controls the amount of noise added at each step, where I is the variance of the added Gaussian noise. Specific Implementation Example 5: The difference between Embodiment 5 and Embodiment 4 of the present invention lies only in: when T When the value is infinite, the distribution is complete Gaussian noise. Specific Implementation Example Six: The difference between Embodiment Six and Embodiment Five of the present invention lies only in: Step 4 specifically involves: x cond The noise distribution after adding noise is then used for noise prediction via a neural network before denoising:

[0035] in, The mean of the noise predicted by the neural network. The variance of the noise predicted by the neural network. θ For neural network parameters; Will x cond The noisy distribution is then processed by Self-Attention and fed into the U-Net network. The loss function used for training the neural network is L2 loss, expressed by the following formula:

[0036] in, For noise predicted by a neural network, for x target Noise generated during the noise addition process. Specific Implementation Example 7: The difference between Embodiment Seven and Embodiment Six of the present invention lies only in: During back-diffusion, the distribution of the distorted image after adding noise is processed by Self-Attention and then fed into the U-Net network to predict the noise required to reconstruct the image. The noise predicted by the neural network and the noise added to the distorted image are used to calculate the loss function to update the neural network parameters. The distorted image is reconstructed by predicting the noise required at each step through the neural network. Specific Implementation Example 8: The difference between Embodiment 8 and Embodiment 7 of the present invention lies only in: This invention provides a system for optimizing imaging distortion of a flexible array ultrasonic sensor, the system comprising: An error dataset generation module generates an error DAS imaging dataset; A correct dataset generation module generates a correct DAS imaging dataset; The preprocessing module converts the dataset samples into low-dimensional Latent features using a variational autoencoder (VAE) and gradually adds Gaussian noise. The training module adds noise to the Latent of the erroneous samples and feeds them into the Self-Attention and U-Net networks for training to predict the noise required to generate undistorted images. The reconstruction module successively denoises the feature distribution using noise predicted by the neural network, and after complete denoising, reconstructs the undistorted image through the Decoder module of the Variational Autoencoder (VAE). Specific Implementation Example Nine: The difference between Embodiment Nine and Embodiment Eight of the present invention lies only in: The present invention provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement a method for optimizing imaging distortion of a flexible array ultrasonic sensor. Specific Implementation Example 10: The only difference between Embodiment 10 and Embodiment 9 of the present invention is that: The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a method for optimizing imaging distortion of a flexible array ultrasonic sensor.

[0041] To address the issue of erroneous transmission delays caused by unknown deformations of the measured object due to the diverse shapes of the object during imaging in flexible ultrasonic sensors, this invention discloses a method for optimizing imaging distortion in flexible array ultrasonic sensors. Using two pre-generated datasets, an image generation model is trained using a diffusion model. When the flexible ultrasonic sensor produces distorted images due to deformation, the distorted image is used to generate an undistorted image through the image generation model. This method does not require the addition of an additional sensor and uses an algorithm to optimize the distorted image. The specific process is as follows: Figure 1 As shown.

[0042] This invention requires the pre-generation of two datasets: (1) Imaging of DAS under different deformation curvatures. When training the model, the objects to be tested should be of the same type, differing only in surface curvature (e.g., the wrists of different people). The step size for different curvatures should be sufficiently small and the coverage area sufficiently large. When generating this dataset, the curvature of the array sensor is treated as 0 when measuring objects with different surface curvatures, and the DAS of each RF is calculated using a predefined sensor model. The calculation formula is:

[0043] in( x f , y f , z f ) is the location of the focal point, ( x c , y c , z c) is the reference center point of the focal point on the aperture of the sensor. x i , y i , z i () is an arbitrary chip of the sensor i Spatial center coordinates, c It's the speed of sound. t i It is the calculated DAS.

[0044] The calculated DAS (Discrete Aperture Sensor) is used to image objects with different surface curvatures. Multiple imaging sessions should be conducted on objects with the same curvature to ensure better generalization. However, since the curvature of the flexible ultrasonic sensor has changed, but the calculated DAS defaults to a curvature of 0, the DAS obtained in this step is erroneous data. Imaging with this DAS will result in defocusing and distortion. This dataset is considered an erroneous sample.

[0045] (2) Imaging of correct DAS under different deformation curvatures. For each ultrasound image in the erroneous dataset, there must be a corresponding correct DAS image in this dataset. For the test object with different surface curvatures, the surface curvature of the test object is measured by strain sensors. A flexible ultrasound sensor array model is established in the Field II simulation tool, and the curvature of each sensor element is set to the curvature of the test object measured by the strain sensors. The spatial center coordinates of each sensor element can be obtained through the established sensor array model. The correct DAS is calculated using the DAS calculation formula used in the incorrect dataset. Imaging with this DAS will not produce defocus, distortion, or other issues. The dataset generated in this step is considered the correct sample.

[0046] This invention trains a graph-based image model based on the Diffusion algorithm, with incorrect samples serving as input images and correct samples as output images. To improve the algorithm's speed, a Variational Autoencoder (VAE) encoder is used to convert the dataset images into low-dimensional latent features.

[0047] During the forward diffusion process, low-dimensional latent features are gradually applied to both datasets. x cond For the Latent distribution of erroneous data, x target Gaussian noise is gradually added to the correct data (Latent distribution). x cond The input used for the denoising model. x target Used for calculating the loss function. The noise-adding process is a Markov process. Its formula is expressed as:

[0048] Where the Latent feature distribution is: x ~ q ( x ), x t For the first t Distribution of time steps t The hyperparameter for adding noise controls the amount of noise added at each step, where I is the variance of the added Gaussian noise. T When the value is infinite, the distribution is complete Gaussian noise.

[0049] x cond The noise distribution after adding noise is then used for noise prediction via a neural network before denoising. The denoising formula is:

[0050] in The mean of the noise predicted by the neural network. The variance of the noise predicted by the neural network. θ These are the parameters of the neural network.

[0051] The specific steps for implementing a neural network are as follows: x cond The noise-added distribution is then processed by Self-Attention and fed into the U-Net network. This structure predicts the noise at each step to reconstruct the image. The loss function used for training the neural network is L2 loss. Its formula is:

[0052] in For noise predicted by a neural network, for x target Noise generated during the noise generation process.

[0053] Low-dimensional latent features are reconstructed from the predicted noise, and then the VAE's Decoder module reconstructs a reconstructed image with the same dimensions as the input image. This output image is the corrected, undistorted image. Specific Implementation Example Eleven: The only difference between Embodiment Eleven and Embodiment Ten of this invention is that: In the fields of sports medicine and rehabilitation, accurately acquiring ultrasound images of human arm muscles is crucial for assessing muscle condition, diagnosing muscle injuries, and developing rehabilitation plans. However, the irregular shape of the human arm and the changing morphology and position of muscles during different movements present numerous challenges for flexible ultrasound sensors when imaging arm muscles.

[0055] When using flexible ultrasound sensors to image the muscles of a human arm, the sensor undergoes unknown deformation due to the arm's bending and stretching movements. This causes traditional delay and direct-scan (DAS) beamformers to apply incorrect time delays to the radio frequency (RF) data, resulting in defocused and distorted B-mode images. Doctors find it difficult to accurately determine the muscle structure and health from these images.

[0056] To address this issue, this invention employs advanced flexible array ultrasonic sensor imaging distortion optimization technology. First, two key datasets are generated. In generating the imaging dataset for erroneous DAS (Distortion Assay) under different deformation curvatures, the arms of different individuals are used as the research object. Although the arms belong to the same type of measurement target, the thickness, muscle development, and surface curvature of each person's arm differ. When calculating the DAS, the curvature of the array sensor is considered to be 0, and the DAS of each RF (Radial Frequency) is calculated using a predefined sensor model. Using the DAS calculated in this way to image the muscles of different arms, because the sensor actually deforms while the default curvature is 0 during calculation, the resulting images will exhibit defocusing and distortion problems. These images are considered erroneous samples.

[0057] Next, imaging datasets of the correct DAS under different deformation curvatures are generated. A 128-channel flexible ultrasound sensor is used in this scheme. The surface curvature of different arms was measured using a high-precision strain sensor, and then a flexible ultrasonic sensor model was built in the Field II simulation tool. Figure 3 , Figure 4 (This involves sensor models that are bent and unbent, respectively, with the sensor curvature set to the actual measured arm curvature.) Based on this model, the correct DAS (Discrete Aspect Ratio) is calculated and images are generated. The resulting images are clear, without defocus or distortion, and these images serve as correct samples.

[0058] With these two datasets, this invention utilizes Diffusion to train the image-generating model. When imaging a patient's arm muscles, distorted images resulting from deformation caused by the flexible ultrasound sensor's contact with the arm are used as erroneous samples input into the model. During forward diffusion, the model first processes the sample images using a variational autoencoder (VAE) to convert the input images into low-dimensional latent features, then progressively adds Gaussian noise, controlling the amount of noise added at each step through hyperparameters. During backward diffusion, the distribution of the distorted image after adding noise is processed by Self-Attention and then fed into a U-Net network to predict the noise required for image reconstruction. The noise predicted by the neural network and the noise added to the undistorted image are used to calculate the loss function to update the neural network parameters. The undistorted image is reconstructed by predicting the noise required at each step through the neural network.

[0059] The optimized images obtained through this technology allow doctors to clearly see the texture and fiber direction of the patient's arm muscles, as well as whether there are any abnormalities such as damage or inflammation within the muscles. For example, it can accurately determine the location and severity of muscle strains, providing a strong basis for developing personalized rehabilitation treatment plans and greatly improving the accuracy of diagnosis and the effectiveness of treatment. The successful application of this technology in human arm muscle imaging has brought a more advanced and reliable detection method to the fields of sports medicine and rehabilitation therapy.

[0060] The above description is merely a preferred embodiment of a flexible array ultrasonic sensor imaging distortion optimization system and method. The scope of protection for this system and method is not limited to the above embodiments; all technical solutions falling within this conceptual framework are within the scope of protection of this invention. It should be noted that for those skilled in the art, any improvements and variations made without departing from the principles of this invention should also be considered within the scope of protection of this invention.

Claims

1. A method for correcting imaging distortion of a flexible ultrasonic sensor array, characterized in that: The method includes the following steps: Step 1: Generating an incorrect DAS imaging dataset; Step 2: Generate the correct DAS imaging dataset; Step 3: The dataset samples are processed by the Encoder module of the Variational Autoencoder (VAE) to convert them into low-dimensional Latent features, and Gaussian noise is added step by step; Step 3 specifically involves: The graph-to-image model is trained based on the idea of ​​Diffusion algorithm, with erroneous samples as input images and correct samples as output images. The encoder of the dataset images is converted into low-dimensional latent features using the variational autoencoder (VAE). In the forward diffusion process, Gaussian noise is gradually added to the low-dimensional latent features of both datasets. x cond For the Latent distribution of erroneous data, x target For the correct data latent distribution, x cond The input used for the denoising model. x target The noise-adding process, used for calculating the loss function, is a Markov process: Among them, the distribution of Latent features is x ~ q ( x ), x t For the first t Distribution of time steps t The hyperparameter for adding noise controls the amount of noise added at each step, where I is the variance of the added Gaussian noise. Step 4: Add noise to the Latent of the erroneous samples and feed them into the Self-Attention and U-Net networks for training to predict the noise required to generate undistorted images; Step 4 specifically involves: x cond The noise distribution after adding noise is then used for noise prediction via a neural network before denoising: in, The mean of the noise predicted by the neural network. The variance of the noise predicted by the neural network. θ For neural network parameters; Will x cond The noisy distribution is then processed by Self-Attention and fed into the U-Net network. The loss function used for training the neural network is L2 loss, expressed by the following formula: in, For noise predicted by a neural network, for x target Noise during the noise addition process; Step 5: The feature distribution is denoised sequentially using the noise predicted by the neural network. After complete denoising, the undistorted image is reconstructed using the Decoder module of the Variational Autoencoder (VAE).

2. The method according to claim 1, characterized in that: Step 1 specifically involves: When measuring objects with surfaces of varying curvature, the curvature of the array sensor is treated as 0, and the DAS of each RF is calculated using a predefined sensor model, calculated by the following formula: in,( x f , y f , z f ) is the location of the focal point, ( x c , y c , z c ) is the reference center point of the focal point on the aperture of the sensor. x i , y i , z i () is an arbitrary chip of the sensor i Spatial center coordinates, c It's the speed of sound. t i It is the calculated DAS; The default curvature of the calculated DAS is 0. The DAS obtained at this time is erroneous data. Using this DAS for imaging will produce phenomena such as defocus and distortion. This dataset is regarded as an erroneous sample.

3. The method according to claim 2, characterized in that: Step 2 specifically involves: For each ultrasound image in the DAS imaging dataset that generates an error, there must be a corresponding correct DAS image in this dataset. For objects under test with different surface curvatures, the surface curvature of the object under test is measured by strain sensors. A flexible ultrasound sensor array model is established in the Field II simulation tool, and the curvature of each sensor element is set to the curvature of the object under test measured by the strain sensors. The spatial center coordinates of each sensor element can be obtained through the established sensor array model. The correct DAS is calculated, and this DAS is used for imaging. This imaging will not produce phenomena such as defocusing or distortion, and the generated dataset is used as the correct sample.

4. The method according to claim 3, characterized in that: when T When the value is infinite, the distribution is complete Gaussian noise.

5. The method according to claim 4, characterized in that: During back-diffusion, the distribution of the distorted image after adding noise is processed by Self-Attention and then fed into the U-Net network to predict the noise required to reconstruct the image. The noise predicted by the neural network and the noise added to the distorted image are used to calculate the loss function to update the neural network parameters. The distorted image is reconstructed by predicting the noise required at each step through the neural network.

6. A flexible ultrasonic sensor array imaging distortion correction system, said system operating based on the flexible ultrasonic sensor array imaging distortion correction method of claim 1, characterized in that: The system includes: An error dataset generation module generates an error DAS imaging dataset; A correct dataset generation module generates a correct DAS imaging dataset; The preprocessing module converts the dataset samples into low-dimensional Latent features using a variational autoencoder (VAE) and gradually adds Gaussian noise. The training module adds noise to the Latent of the erroneous samples and feeds them into the Self-Attention and U-Net networks for training to predict the noise required to generate undistorted images. The reconstruction module successively denoises the feature distribution using noise predicted by the neural network, and after complete denoising, reconstructs the undistorted image through the Decoder module of the Variational Autoencoder (VAE).

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the method as claimed in any one of claims 1-5.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the method of any one of claims 1-5.

Images