Wearable ultrasonic image quality enhancement method and system based on deep learning

Abstract

The invention belongs to the technical field of image processing. The invention provides a wearable ultrasonic image quality enhancement method and system based on deep learning, and the method comprises the steps: carrying out the IQ demodulation of single-angle echo data, and obtaining complex IQ data containing amplitude and phase information; performing time delay compensation on the complex IQ data to obtain IQ data after time delay; inputting the IQ data after time delay into an encoder to obtain low-dimensional feature representation; and inputting the low-dimensional feature representation into a time-frequency feature fusion module to obtain a reconstructed and enhanced ultrasonic image. According to the method, the echo data of the single-angle plane waves are modeled through the deep learning model, the high frame rate characteristic of single-angle imaging is kept while the resolution, the contrast ratio and the image quality of the ultrasonic image are improved, and it is ensured that the method is suitable for portable and low-power-consumption wearable ultrasonic equipment; the requirements of actual monitoring and clinical application are met.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly to a method and system for enhancing the quality of wearable ultrasonic images based on deep learning. Background Art

[0002] The statements in this section merely provide background art related to the present invention and do not necessarily constitute prior art.

[0003] Wearable ultrasonic devices have gradually attracted attention due to their portability and real-time monitoring capabilities. However, these devices still face many challenges in actual medical applications. Limited by the volume and power consumption of the devices, the number and performance of sensors are often inferior to those of traditional ultrasonic devices, resulting in insufficient clarity and contrast in the acquired ultrasonic images. In addition, factors such as patient movement, environmental noise, and poor probe contact may cause noise, blurring, and artifacts in the ultrasonic images, thereby affecting the accuracy of diagnosis.

[0004] Traditional multi-angle compound imaging (Coherent Plane Wave Compounding, CPWC) significantly improves image quality through multi-angle emission and coherent summation. However, this method requires multiple plane wave emissions, resulting in a significant decrease in the frame rate and is not suitable for real-time wearable ultrasonic devices. Summary of the Invention

[0005] To solve the deficiencies of the prior art, the present invention provides a method and system for enhancing the quality of wearable ultrasonic images based on deep learning. By modeling the echo data of a single-angle plane wave through a deep learning model, while improving the resolution, contrast, and image quality of the ultrasonic image, the high frame rate characteristic of single-angle imaging is retained, ensuring its applicability to portable and low-power wearable ultrasonic devices and meeting the requirements of actual monitoring and clinical applications.

[0006] To achieve the above objective, the present invention adopts the following technical solutions: In the first aspect, the present invention provides a method for enhancing the quality of wearable ultrasonic images based on deep learning.

[0007] A method for enhancing the quality of wearable ultrasonic images based on deep learning includes the following processes: Perform IQ demodulation on the echo data of a single angle to obtain complex IQ data containing amplitude and phase information; Perform time delay compensation on the complex IQ data to obtain time-delayed IQ data; Input the time-delayed IQ data into an encoder to obtain a low-dimensional feature representation; Input the low-dimensional feature representation into a time-frequency feature fusion module to obtain a reconstructed and enhanced ultrasonic image.

[0008] As a further limitation of the first aspect of the present invention, time delay compensation is performed on the complex IQ data to obtain the IQ data after time delay, including: , where is the IQ data after time delay, , where , , and represent the coordinates of the imaging point, and represent the position coordinates of the th probe, represents a single angle, is the complex IQ data, represents the wave speed, represents the propagation time of the transmitted signal at the angle , represents the propagation time of the received signal at the angle , represents the total flight time.

[0009] As a further limitation of the first aspect of the present invention, the time-frequency feature fusion module includes: a time-domain branch and a frequency-domain branch. The time-domain branch obtains a time-domain feature representation through time-domain feature extraction; The frequency-domain branch includes a parallel low-frequency branch and high-frequency branch. The low-frequency branch obtains a low-frequency feature representation through a low-frequency feature extraction module, and the high-frequency branch obtains a high-frequency feature representation through a high-frequency feature extraction module; The feature fusion module performs weighted fusion on the time-domain feature representation, the low-frequency feature representation, and the high-frequency feature representation, and outputs a reconstructed and enhanced ultrasonic image.

[0010] As a further limitation of the first aspect of the present invention, the time-domain branch obtains a time-domain feature representation through time-domain feature extraction, including: The input features first pass through a normalization layer, and the normalized features pass through a convolutional layer to capture the change patterns in the time dimension through a local receptive field. After the convolutional operation of the convolutional layer, the features are non-linearly transformed through an activation function, and a time-domain feature representation is output.

[0011] As a further limitation of the first aspect of the present invention, the low-frequency branch obtains a low-frequency feature representation through a low-frequency feature extraction module, including: The low-frequency branch converts the input features from the time domain to the frequency domain through a two-dimensional Fourier transform, passes through a low-pass filter with learnable parameters to extract the low-frequency components of the signal, and the low-frequency components obtain a low-frequency feature representation after passing through the low-frequency feature extraction module and a two-dimensional inverse Fourier transform; The high-frequency branch extracts a high-frequency feature representation through a low-frequency feature extraction module, including: The high-frequency branch converts the input feature from the time domain to the frequency domain through a two-dimensional Fourier transform, passes through a high-pass filter with learnable parameters to extract the high-frequency component of the signal, and obtains the high-frequency feature representation after the high-frequency component passes through the high-frequency feature extraction module and the two-dimensional inverse Fourier transform.

[0012] As a further limitation of the first aspect of the present invention, the feature fusion module performs weighted fusion on the time-domain feature representation, the low-frequency feature representation, and the high-frequency feature representation, including: Feature integration is performed on the time-domain feature representation, the low-frequency feature representation, and the high-frequency feature representation through an adaptive weight learning module to generate three feature weights, and the three feature weights are normalized through a normalization function to ensure that the sum of each weight is 1; The normalized weights are respectively multiplied by the time-domain feature, the low-frequency feature, and the high-frequency feature to perform weighted processing on the features. Finally, the weighted features are linearly superimposed, converted to the frequency domain through a two-dimensional Fourier transform, filtered through learnable parameters, and then transformed back to the spatio-temporal domain to obtain the reconstructed and enhanced ultrasonic image.

[0013] As a further limitation of the first aspect of the present invention, the loss function is a weighted sum of a pixel-level loss, a structural similarity loss, and a wavelet loss.

[0014] In a second aspect, the present invention provides a wearable ultrasonic image quality enhancement system based on deep learning.

[0015] A wearable ultrasonic image quality enhancement system based on deep learning includes: A data demodulation unit, configured to: perform IQ demodulation on the echo data of a single angle to obtain complex IQ data containing amplitude and phase information; A time delay compensation unit, configured to: perform time delay compensation on the complex IQ data to obtain time-delayed IQ data; A data encoding unit, configured to: input the time-delayed IQ data into an encoder to obtain a low-dimensional feature representation; A time-frequency feature fusion unit, configured to: input the low-dimensional feature representation into a time-frequency feature fusion module to obtain the reconstructed and enhanced ultrasonic image.

[0016] In a third aspect, the present invention provides a computer device, including: a processor and a computer-readable storage medium; The processor is adapted to execute a computer program; A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, it implements the method for enhancing the quality of wearable ultrasonic images based on deep learning as described in the first aspect of the present invention.

[0017] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program, which is adapted to be loaded and executed by a processor to implement the method for enhancing the quality of wearable ultrasonic images based on deep learning as described in the first aspect of the present invention.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Through the image reconstruction model based on deep learning, the present invention significantly improves the imaging quality of wearable ultrasonic devices. Under the limitations of low power consumption and single-angle plane wave emission, the generated images approach the effect of multi-angle compound imaging (CPWC), while retaining the high frame rate characteristics of single-angle plane wave imaging.

[0019] 2. The present invention introduces a time-frequency feature fusion module (TFF). Through a multi-branch design and a dynamic weight fusion mechanism, it effectively extracts and utilizes time-domain and frequency-domain information. Compared with traditional models, it further improves the resolution and contrast of images. The generated high-quality images are suitable for dynamic monitoring and clinical diagnosis, providing reliable support for the practical application of wearable ultrasonic devices.

[0020] Advantages of additional aspects of the present invention will be partially given in the following description, partially become apparent from the following description, or be understood through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings forming a part of this specification are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0022] Figure 1 It is a schematic diagram of the model training process provided in Embodiment 1 of the present invention; Figure 2 It is a schematic diagram of the model application process provided in Embodiment 1 of the present invention; Figure 3 It is a schematic diagram of the effects before and after enhancement provided in Embodiment 1 of the present invention. Among them, (A) is the ultrasonic image before enhancement, and (B) is the ultrasonic image after enhancement; Figure 4 It is a schematic diagram of a system for enhancing the quality of wearable ultrasonic images based on deep learning provided in Embodiment 2 of the present invention; Figure 5 It is a schematic diagram of a computer device provided in Embodiment 3 of the present invention. Detailed Implementation Modes

[0023] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0024] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0025] In the case of no conflict, the embodiments in the present invention and the features in the embodiments may be combined with each other.

[0026] Embodiment 1: This implementation mode proposes a method for enhancing the quality of wearable ultrasonic images based on deep learning, effectively improving the imaging quality of wearable ultrasonic devices and reducing the influence brought by the limitations of device volume, power consumption, and sensor performance. At the same time, this method uses the echo data emitted from a single angle as input to generate high-quality images, and minimizes the number of transmissions of the wearable ultrasonic probe as much as possible without sacrificing image quality to meet the requirements of wearability and low power consumption.

[0027] The present invention obtains echo signal data by using software simulation and the method of collecting on a real human body respectively. Specifically, for the simulated data, first, a scatterer pattern with different acoustic impedance rate distributions is generated, and then the probe echo data is calculated through software simulation. For the data of a real human body, first, the probe is attached to the skin, and the method of plane wave imaging is used to sequentially emit plane waves from multiple angles and receive the echo data, and then after ADC sampling, the sampled echo data is obtained, denoted as , where is the emission angle of the plane wave. The echo data of each angle contains the echo signals from each element, denoted as , where represents the th element. It should be noted that in subsequent calculations, if is not specifically marked and only is marked, it means that in the echo data with the emission angle of , the echo data of each element is calculated sequentially.

[0028] More specifically, if it is multi-angle, assuming there are 75 angles, and each angle contains the echo signals of 128 array elements. The echo signals of each array element are discretely sampled, and the number of sampling points is assumed to be 4,096. Then the shape of the final data is of size RF(75, 128, 4,096), corresponding to RF(alpha, i, t); if it is single-angle, then it is RF(1, 128, 4,096). The number of imaging points is obtained according to the image size. Assuming the imaging size is , then the number of imaging points to be calculated is 65,536, that is, 65,536 (x, y). The meaning of calculating each array element one by one in this implementation method is that first, all angles are traversed (if the angle is 1 here, there is no need to traverse), and then all array elements in each angle are traversed. For each array element [coordinates (x i , y i ), all imaging points [65,536 (x, y)] are traversed, and each imaging point is calculated according to formulas (2) to (5). The final shape of the obtained TDIQ is (1, 128, ), and this TDIQ can be directly input into the model.

[0029] For the echo data RF, any single angle is selected , that is , which is the echo data of the plane wave of a single angle as the input. For all angles , the CPWC method is used for imaging to obtain a high-quality image as the target, denoted as HQI, and HQI form an example in the dataset.

[0030] The training framework of the model is as Figure 1 shown, divided into three parts, including echo data re-representation, image reconstruction model, and loss function constraint.

[0031] In this implementation method, in the echo data re-representation stage, specifically, it includes: First, the single-angle echo data is demodulated by IQ to obtain complex IQ data containing amplitude and phase information, denoted as: (1); Among them, represents the echo signal received by the i-th array element with the transmission angle of , is the carrier frequency, represents time.

[0032] IQ demodulation is a demodulation technique for recovering the original baseband signal from the received modulated signal, which is widely used in fields such as wireless communication, radar, and ultrasonic detection. Its core idea is to separate the in-phase (I) and quadrature (Q) components of the signal using orthogonal carriers, thereby extracting the amplitude and phase information of the signal.

[0033] Next, according to the coordinates of the imaging point and the positions of the transmitting array elements, the time of flight (TOF) is calculated, and then time delay compensation is applied to the IQ data to obtain the time-delayed IQ data (TDIQ): (2) ； (3); (4); (5); Among them, and represent the coordinates of the imaging point, and represent the position coordinates of the i-th probe, y is defaulted to 0, c represents the wave speed, represents the propagation time of the transmitted signal at the angle , represents the propagation time of the received signal at the angle , represents the total time of flight.

[0034] Finally, is input into the autoencoder structure for data compression and dimensionality reduction. Specifically, the data first undergoes feature extraction and compression operations through an encoder. The encoder is designed to maximize the retention of the key features of the imaging point while significantly reducing the data volume and redundant information. The output of the encoder is a compact low-dimensional feature representation, which will be used as the input for subsequent image reconstruction. At the same time, to ensure the effectiveness of the compressed features, the autoencoder structure also includes a decoder for restoring the low-dimensional feature representation generated by the encoder to the original data. The purpose of this process is to evaluate the retention of information during the compression and restoration processes and ensure that the low-dimensional features extracted by the encoder can cover the key information of the original data.

[0035] In this implementation method, the image reconstruction model specifically includes: In order to fully extract the Based on the information in the data, the present invention proposes a Time-Frequency Feature Fusion Module (TFF), which is a basic module in the network structure and applicable to network structures including but not limited to Unet and ResNet. Through a multi-branch design, this module extracts and fuses time-domain and frequency-domain features of the input features to generate high-quality feature representations, providing key support for image reconstruction. The image reconstruction model of the present invention takes the time-frequency feature fusion module as the core, and the time-frequency feature fusion module includes two main parts: a feature extraction module (TFE) and a feature fusion module (FFM). The feature extraction module consists of a time-domain branch and a frequency-domain branch, which are used to extract comprehensive time-domain and frequency-domain information from the input features, while the feature fusion module weights and fuses the features output by multiple branches as the output of the time-frequency feature fusion module.

[0036] The feature extraction module first processes the input features which enter the time-domain branch and the frequency-domain branch respectively for processing. The time-domain branch obtains a time-domain feature representation through time-domain feature extraction . Specifically, the input features first pass through a normalization layer to adjust the numerical range of the features and reduce the deviation of the data distribution. The normalized features then pass through a convolutional layer to capture the change patterns in the time dimension through the local receptive field. After the convolutional operation, a non-linear transformation is performed on the features through an activation function to further enhance the expression ability of the features. Finally, the time-domain branch outputs the time-domain feature representation .

[0037] The frequency-domain branch is further divided into a low-frequency branch and a high-frequency branch to process the low-frequency and high-frequency components of the signal respectively. The low-frequency branch first converts the input features from the time domain to the frequency domain through a two-dimensional Fourier transform (2DFFT). The result of the Fourier transform contains the amplitude and phase information of the signal, and then passes through a low-pass filter with learnable parameters to extract the low-frequency components of the signal. Subsequently, the low-frequency components pass through the low-frequency feature extraction module to obtain the low-frequency feature representation . Specifically, the low-frequency components enter a complex-domain normalization layer to adjust the amplitude distribution of the frequency-domain data to make it more suitable for subsequent convolutional operations. The normalized features pass through a complex convolutional layer to further capture the local feature information in the low-frequency components. Subsequently, the convolutional features undergo a non-linear transformation through a complex activation function to enhance the expression ability of the features. Finally, the low-frequency features are restored to the spatio-temporal domain through an inverse two-dimensional Fourier transform (2DIFFT) to generate the low-frequency feature representation .

[0038] The processing of the high-frequency branch is similar to that of the low-frequency branch. First, the input features are transformed into the frequency domain through a two-dimensional Fourier transform; then, a high-pass filter with learnable parameters is used to extract the high-frequency components of the signal. Subsequently, the high-frequency components pass through a high-frequency feature extraction module to obtain the high-frequency feature representation . Specifically, the high-frequency components pass through a complex domain normalization layer to adjust the distribution of the features; the normalized features capture the detailed characteristics in the high-frequency information through a complex convolution layer. Subsequently, the convolutional features pass through a complex activation function to further enhance the expression ability of the high-frequency features. Finally, the high-frequency features are restored to the spatio-temporal domain through an inverse two-dimensional Fourier transform to generate the final high-frequency feature representation .

[0039] The feature fusion module receives the feature representations output by the time domain branch, the low-frequency branch, and the high-frequency branch, and performs weighted fusion on them. First, an adaptive weight learning module integrates the three types of features to generate three feature weights. These weights are normalized through a normalization function to ensure that the sum of the weights is 1. The normalized weights are multiplied by the time domain features, the low-frequency features, and the high-frequency features respectively to perform weighted processing on the features. Finally, the weighted features are linearly superimposed, then transformed into the frequency domain through a two-dimensional Fourier transform, filtered through learnable parameters, and then transformed back to the spatio-temporal domain to obtain the output of the time-frequency feature fusion module .

[0040] It should be noted that the feature extraction structure of the time domain branch, the feature extraction structure after filtering of the frequency domain branch, and the adaptive weight learning module are not limited to the combination forms of the normalization layer, the convolution layer, and the non-linear layer mentioned above

[0041] Through the above structural design, the image reconstruction model combines and builds TFF as the basic module, and finally reconstructs a high-quality image to ensure that the output of the model has good clarity and contrast

[0042] In this implementation, the loss function constraints specifically include: For the autoencoder, the MSE loss is used to measure the difference between the input data (i.e., ) and the output of the decoder to evaluate the ability of the autoencoder to retain information during the compression process: (6); where, represents the total number of elements in and represents the

[0043] For the reconstruction model, in order to ensure that the model can generate high-quality reconstructed images, the present invention designs a multi-objective loss function constraint mechanism to constrain and optimize the model from multiple levels of pixel level, structure level and frequency domain level. By combining different loss functions, while improving the image resolution and contrast, the model takes into account the preservation of global information and local details.

[0044] The model uses pixel-level loss for the reconstructed image and the target image to constrain the pixel difference between them. Specifically, the pixel-level loss calculates the mean square error (MSE) between the predicted image and the real image to ensure the reconstruction accuracy at the pixel level: (7); where N represents the total number of elements in Y.

[0045] The present invention introduces the structural similarity (SSIM) loss to evaluate the structural consistency between the reconstructed image and the target image. The SSIM loss quantifies the similarity of image brightness, contrast and local structure, effectively improving the model's ability to retain image details and visual quality: (8); represents the mean of the target image Y, represents the reconstructed image 's mean; represents 's square, represents 's square; represents (the square of the standard deviation of the target image Y), that is, the variance of the target image Y, represents (the square of the standard deviation of the reconstructed image ), that is, the variance of the reconstructed image ; represents the covariance between the target image Y and the reconstructed image ; and are both constants. To prevent the denominator from being 0, they can be set by yourself according to needs, or there are default calculation methods, such as , , L is the dynamic range of the image (such as L = 255 for an 8-bit grayscale image), is default to 0.01, is default to 0.03.

[0046] In addition, to further optimize the frequency-domain information, the present invention designs a loss function based on wavelet transform. Specifically, wavelet decomposition is performed on the target image and the predicted image respectively to obtain approximation coefficients and detail coefficients. In the wavelet loss, the mean square error of the approximation coefficients and the detail coefficients is weighted and summed to balance the importance of global information and local details, thereby enhancing the sensitivity of the model to different frequency information.

[0047] (9); Among them, and are weight coefficients, both being 0.5; A and D represent the approximation coefficient and the detail coefficient after wavelet transform of the target image Y; and represent the approximation coefficient and the detail coefficient after wavelet transform of the reconstructed image ; , , , are all tensors, and the subscript means taking the th element; means 's number of elements (the same as ), means 's number of elements (the same as ), and here is also used for traversal in the summation symbol.

[0048] Combining the above loss functions, the total loss function of the model is composed of a weighted combination of multiple loss functions, and the overall quality of image reconstruction is effectively improved through multi-objective constraints. Each loss function acts synergistically during the training process to guide the model to generate high-resolution and high-contrast ultrasound images, meeting the requirements for image quality of wearable ultrasound devices: (10).

[0049] Among them, , and are weight coefficients.

[0050] The model is first trained on a large number of simulation data sets to obtain pre-trained weights, and then the autoencoder structure is fine-tuned on a real human data set to obtain the final model. The above loss function is used for constraint during the training process.

[0051] The image reconstruction model of the present invention can generate high-quality ultrasound images based on the plane wave echo data emitted by the wearable ultrasound device once. Specifically, the application of the model includes the following steps: AsFigure 2 As shown in the figure, first, a wearable ultrasonic probe is used to emit single-angle plane waves and receive corresponding echo signals. The received echo signals are processed through analog-to-digital conversion (ADC sampling) to generate the original sampled data. Subsequently, according to the coordinate information of the imaging area, the time of flight (TOF) of each imaging point is calculated, and the original echo data is compensated for TOF to generate IQ data after time delay ( ); then, the generated is compressed and dimension-reduced by an encoder to extract a compact low-dimensional feature representation; subsequently, the compressed feature representation is input into an image reconstruction model, and finally, an enhanced ultrasonic image is generated. The quality of the output image is close to the effect of multi-angle compound imaging (CPWC), while retaining the high frame rate advantage of single-angle plane wave imaging. The generated high-quality images are suitable for clinical diagnosis or dynamic monitoring tasks, such as cardiovascular health assessment or real-time functional monitoring. As Figure 3 shown, it is a schematic diagram of the effect before and after enhancement. Figure 3 In the figure (A), it is the ultrasonic image before enhancement. Figure 3 In the figure (B), it is the ultrasonic image after enhancement. It can be seen that after adopting the method of this implementation, the image quality has been significantly enhanced.

[0052] Embodiment 2: As Figure 4 shown, this implementation provides a deep learning-based wearable ultrasonic image quality enhancement system, including: A data demodulation unit, configured to: perform IQ demodulation on the single-angle echo data to obtain complex IQ data containing amplitude and phase information; A time delay compensation unit, configured to: perform time delay compensation on the complex IQ data to obtain IQ data after time delay; A data encoding unit, configured to: input the IQ data after time delay into an encoder to obtain a low-dimensional feature representation; A time-frequency feature fusion unit, configured to: input the low-dimensional feature representation into a time-frequency feature fusion module to obtain a reconstructed and enhanced ultrasonic image.

[0053] For the specific working methods of each unit, see the introduction in Embodiment 1, which will not be elaborated here.

[0054] It can be understood that each of the above units can be separately or wholly combined into one or several other units to form, or some of them can be further split into multiple smaller units in terms of function to form, which can achieve the same operation without affecting the realization of the technical effects of the embodiments of the present application. The above units are divided based on logical functions. In practical applications, the function of one unit can also be realized by multiple units, or the functions of multiple units can be realized by one unit. In other embodiments of the present application, the system can also include other units. In practical applications, these functions can also be assisted by other units and can be realized by the cooperation of multiple units.

[0055] According to another embodiment of the present application, the system described in this embodiment can be constructed, and the method of Embodiment 1 of the present application can be implemented by running a computer program (including program code) capable of executing the respective steps involved in the corresponding method described in Embodiment 1 on a general computing device such as a computer including processing elements and storage elements such as a central processing unit (CPU), a random access memory (RAM), and a read-only memory (ROM). The computer program can be recorded on a computer-readable recording medium, for example, and loaded into the above computing device through the computer-readable recording medium and run therein.

[0056] Embodiment 3: As Figure 5 shown, this implementation provides an electronic device, which includes a processor 1001, a communication interface 1002, and a computer-readable storage medium 1003. Among them, the processor 1001, the communication interface 1002, and the computer-readable storage medium 1003 can be connected through a bus or other means.

[0057] Among them, the communication interface 1002 is used to receive and send data. The computer-readable storage medium 1003 can be stored in the memory of the electronic device. The computer-readable storage medium 1003 is used to store a computer program, and the computer program includes program instructions. The processor 1001 is used to execute the program instructions stored in the computer-readable storage medium 1003.

[0058] The processor 1001 (or CPU (Central Processing Unit, central processor)) is the computing core and control core of the electronic device, which is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to implement the corresponding method flow or corresponding function.

[0059] The processor 1001 is configured to execute the following process: Perform IQ demodulation on the echo data of a single angle to obtain complex IQ data containing amplitude and phase information; Perform time delay compensation on the complex IQ data to obtain IQ data after time delay; Input the IQ data after time delay into an encoder to obtain a low-dimensional feature representation; Input the low-dimensional feature representation into a time-frequency feature fusion module to obtain a reconstructed and enhanced ultrasonic image.

[0060] For the specific working method, please refer to the introduction in Embodiment 1 and will not be elaborated here.

[0061] Embodiment 4: This implementation provides a computer-readable storage medium (Memory). A computer-readable storage medium is a memory device in an electronic device for storing programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the electronic device and, of course, the extended storage medium supported by the electronic device. The computer-readable storage medium provides a storage space, and this storage space stores the processing system of the electronic device.

[0062] Moreover, one or more instructions suitable for being loaded and executed by a processor are also stored in this storage space. These instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory; optionally, it can also be at least one computer-readable storage medium located far from the aforementioned processor.

[0063] In one embodiment, one or more instructions are stored in the computer-readable storage medium; the processor loads and executes one or more instructions stored in the computer-readable storage medium to implement the following process: Perform IQ demodulation on the echo data of a single angle to obtain complex IQ data containing amplitude and phase information; Perform time delay compensation on the complex IQ data to obtain IQ data after time delay; Input the IQ data after time delay into an encoder to obtain a low-dimensional feature representation; Input the low-dimensional feature representation into a time-frequency feature fusion module to obtain a reconstructed and enhanced ultrasonic image.

[0064] For the specific working method, please refer to the introduction in Embodiment 1 and will not be elaborated here.

[0065] Those of ordinary skill in the art can realize that the units and algorithm steps of each example described in combination with the embodiments disclosed in this application can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of this application.

[0066] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions according to the embodiments of this application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center in a wired manner (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or a wireless manner (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that the computer can access or a data processing device such as a server or data center that includes one or more integrated available media. The available medium can be a magnetic medium (such as a floppy disk, hard disk, magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0067] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for enhancing the quality of wearable ultrasound images based on deep learning, characterized in that, It includes the following processes: Perform IQ demodulation on the echo data of a single angle to obtain complex IQ data containing amplitude and phase information; Perform time delay compensation on the complex IQ data to obtain time-delayed IQ data; Input the time-delayed IQ data into an encoder to obtain a low-dimensional feature representation; Input the low-dimensional feature representation into a time-frequency feature fusion module to obtain a reconstructed and enhanced ultrasound image.

2. The method for enhancing the quality of wearable ultrasound images based on deep learning according to claim 1, wherein Performing time delay compensation on the complex IQ data to obtain time-delayed IQ data includes: , where is the IQ data after time delay, , where , , and represent the coordinates of the imaging point, and represent the position coordinates of the -th probe, represents a single angle, is the complex IQ data, represents the wave speed, represents the propagation time of the transmitted signal at angle , represents the propagation time of the received signal at angle , represents the total flight time.

3. The method for enhancing the quality of wearable ultrasound images based on deep learning according to claim 1, wherein The time-frequency feature fusion module includes a time domain branch and a frequency domain branch. The time domain branch obtains a time domain feature representation through time domain feature extraction; The frequency domain branch includes a parallel low-frequency branch and high-frequency branch. The low-frequency branch obtains a low-frequency feature representation through a low-frequency feature extraction module, and the high-frequency branch obtains a high-frequency feature representation through a high-frequency feature extraction module; The feature fusion module performs weighted fusion on the time domain feature representation, the low-frequency feature representation, and the high-frequency feature representation, and outputs a reconstructed and enhanced ultrasound image.

4. The method for enhancing the quality of wearable ultrasound images based on deep learning according to claim 3, wherein The time domain branch obtains a time domain feature representation through time domain feature extraction, including: The input feature first passes through a normalization layer. After normalization, the feature passes through a convolutional layer to capture the change pattern in the time dimension through a local receptive field. After the convolutional operation of the convolutional layer, the feature is non-linearly transformed through an activation function, and the time domain feature representation is output.

5. The method for enhancing the quality of wearable ultrasound images based on deep learning according to claim 3, wherein The low-frequency branch obtains a low-frequency feature representation through a low-frequency feature extraction module, including: The low-frequency branch converts the input feature from the time domain to the frequency domain through a two-dimensional Fourier transform, passes through a low-pass filter with learnable parameters to extract the low-frequency component of the signal, and the low-frequency component obtains a low-frequency feature representation after passing through the low-frequency feature extraction module and the two-dimensional inverse Fourier transform; The high-frequency branch obtains a high-frequency feature representation through a high-frequency feature extraction module, including: The high-frequency branch converts the input feature from the time domain to the frequency domain through a two-dimensional Fourier transform, passes through a high-pass filter with learnable parameters to extract the high-frequency component of the signal, and the high-frequency component obtains a high-frequency feature representation after passing through the high-frequency feature extraction module and the two-dimensional inverse Fourier transform.

6. The method for enhancing the quality of wearable ultrasound images based on deep learning according to claim 3, wherein The feature fusion module performs weighted fusion on the time domain feature representation, the low-frequency feature representation, and the high-frequency feature representation, including: The time-domain feature representation, the low-frequency feature representation, and the high-frequency feature representation are integrated through an adaptive weight learning module to generate three feature weights. The three feature weights are normalized through a normalization function to ensure that the sum of each weight is 1; The normalized weights are multiplied by the time-domain features, low-frequency features, and high-frequency features respectively to perform weighted processing on the features. Finally, the weighted features are linearly superimposed, transformed to the frequency domain through a two-dimensional Fourier transform, filtered through learnable parameters, and then transformed back to the spatio-temporal domain to obtain the reconstructed and enhanced ultrasound image.

7. The deep learning-based wearable ultrasound image quality enhancement method according to any one of claims 1-6, wherein The loss function is a weighted sum of pixel-level loss, structural similarity loss, and wavelet loss.

8. A wearable ultrasonic image quality enhancement system based on deep learning, characterized in that, It includes: A data demodulation unit configured to: perform IQ demodulation on the single-angle echo data to obtain complex IQ data containing amplitude and phase information; A time delay compensation unit configured to: perform time delay compensation on the complex IQ data to obtain time-delayed IQ data; A data encoding unit configured to: input the time-delayed IQ data into an encoder to obtain a low-dimensional feature representation; A time-frequency feature fusion unit configured to: input the low-dimensional feature representation into a time-frequency feature fusion module to obtain a reconstructed and enhanced ultrasound image.

9. A computer device, characterized in that, It includes: A processor and a computer-readable storage medium; The processor is adapted to execute a computer program; The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the deep learning-based wearable ultrasound image quality enhancement method according to any one of claims 1 to 6 is implemented.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and the computer program is adapted to be loaded and executed by the processor to implement the deep learning-based wearable ultrasound image quality enhancement method according to any one of claims 1 to 6.

Images