Portable ultrasonic imaging method based on multi-angle plane waves and neural network

Abstract

The invention provides a portable ultrasonic imaging method based on multi-angle plane waves and a neural network, and belongs to the technical field of neural networks and ultrasonic imaging, and the method comprises the steps: obtaining multi-angle plane wave original data collected by an ultrasonic probe; screening out optimized angle combination data from the multi-angle plane wave original data through a first convolutional neural network model; transmitting the optimized angle combination data to terminal equipment; performing beam forming and image processing on the optimized angle combination data through a second convolutional neural network model in the terminal equipment to generate an ultrasonic image; and displaying the ultrasonic image through the terminal equipment. Aiming at the business scene problems that the multi-angle plane wave data processing efficiency is low and the image quality is influenced by noise and angle selection in the traditional ultrasonic imaging, the imaging quality and efficiency are obviously improved by innovatively fusing the intelligent processes of data screening, transmission and image processing.

Description

Technical Field

[0001] This invention relates to the fields of neural networks and ultrasound imaging technology, and in particular to a portable ultrasound imaging method based on multi-angle plane waves and neural networks. Background Technology

[0002] Traditional ultrasound imaging faces a core technical challenge in processing multi-angle plane wave data: how to balance efficient data processing with high-reliability image quality in resource-constrained medical settings. This contradiction directly impacts the accuracy and practicality of real-time diagnosis. Specifically, traditional methods acquire massive amounts of multi-angle plane wave data containing significant redundancy, leading to low processing efficiency, especially in mobile medical devices or remote medical settings where limited computing resources and bandwidth make real-time processing difficult. Simultaneously, the signal-to-noise ratio of the raw data is limited by noise interference and the uncertainty of angle selection, resulting in low contrast and blurred details, affecting the doctor's accurate lesion diagnosis. Furthermore, high bandwidth consumption during data transmission and the computational complexity of decompression and beamforming on the terminal device further exacerbate the conflict between real-time performance and image quality. Several sub-questions arise around this core issue: how to effectively select high-quality angle combination data to reduce redundancy; how to optimize data compression and transmission efficiency under bandwidth constraints; and how to improve denoising and image resolution on the terminal device using adaptive algorithms. These seemingly minor issues all stem from the core contradiction: balancing the optimization needs of data processing efficiency and image quality with limited resources. This is especially true in dynamically changing medical scenarios, where it is necessary to ensure the adaptability of the imaging process to complex lesions and the reliability of diagnosis. This highlights the unique technical challenges of intelligent optimization of the entire ultrasound imaging process. Summary of the Invention

[0003] This invention provides a portable ultrasound imaging method based on multi-angle plane waves and neural networks, mainly comprising:

[0004] The process involves acquiring raw multi-angle plane wave data from an ultrasound probe; filtering optimized angle combination data from the raw multi-angle plane wave data using a first convolutional neural network model; transmitting the optimized angle combination data to a terminal device; performing beamforming and image processing on the optimized angle combination data using a second convolutional neural network model in the terminal device to generate an ultrasound image; and displaying the ultrasound image on the terminal device.

[0005] Furthermore, the first convolutional neural network model is used to filter out optimized angle combination data from the original multi-angle plane wave data, including: obtaining the signal-to-noise ratio information of the original multi-angle plane wave data; determining the imaging correlation of each angle data based on the signal-to-noise ratio information; judging the angle data that meets the preset correlation threshold based on the imaging correlation; and generating optimized angle combination data based on the angle data that meets the preset correlation threshold.

[0006] Furthermore, transmitting the optimized angle combination data to the terminal device includes: obtaining the compressed format of the optimized angle combination data; transmitting the compressed optimized angle combination data to the terminal device via wired or wireless communication; and decompressing the compressed optimized angle combination data in the terminal device to obtain the transmitted optimized angle combination data.

[0007] Furthermore, the optimized angle combination data is subjected to beamforming and image processing by the second convolutional neural network model in the terminal device to generate an ultrasound image, including: acquiring optimized angle combination data; executing an adaptive beamforming algorithm through the second convolutional neural network model to generate an initial ultrasound image; performing signal denoising processing on the initial ultrasound image to obtain a denoised image; and performing image enhancement processing on the denoised image to generate an ultrasound image.

[0008] Furthermore, the initial ultrasound image is subjected to signal denoising processing to obtain a denoised image, including: obtaining the pixel distribution features of the initial ultrasound image; processing the initial ultrasound image using a deep learning-based filtering algorithm to determine the noise distribution; and removing noise interference from the initial ultrasound image based on the noise distribution to generate a denoised image.

[0009] Furthermore, image enhancement processing is performed on the denoised image to generate an ultrasound image, including: acquiring the resolution information of the denoised image; processing the denoised image using a deep learning-based super-resolution algorithm to determine enhancement features; adjusting the contrast and sharpness of the denoised image according to the enhancement features; and generating an ultrasound image.

[0010] Furthermore, displaying ultrasound images via a terminal device includes: acquiring display parameters of the ultrasound images; adjusting the display ratio of the ultrasound images according to the display parameters; and presenting the adjusted ultrasound images on the display interface of the terminal device.

[0011] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0012] This invention discloses a portable ultrasound imaging method based on multi-angle plane waves and neural networks. Addressing the problems of low data processing efficiency and image quality degradation due to noise and angle selection in traditional ultrasound imaging, this invention significantly improves imaging quality and efficiency by innovatively integrating intelligent processes for data filtering, transmission, and image processing. First, a first convolutional neural network model analyzes the signal-to-noise ratio of the raw multi-angle plane wave data, filtering out optimized angle combinations that meet a preset correlation threshold, thus reducing redundant data. Subsequently, data transmission is optimized through compression formats to reduce bandwidth consumption and ensure efficient decompression by the terminal device. A second convolutional neural network model in the terminal device performs adaptive beamforming, combining deep learning filtering and super-resolution algorithms for denoising and image enhancement, effectively improving image contrast and clarity. Finally, display parameters are adjusted to present high-quality ultrasound images on the terminal device. This invention, through the integration of deep learning and data optimization, achieves intelligent processing from data acquisition to image display, significantly improving the real-time performance, image quality, and diagnostic reliability of ultrasound imaging, making it suitable for resource-constrained medical scenarios. Attached Figure Description

[0013] Figure 1 This is a flowchart of a portable ultrasound imaging method based on multi-angle plane waves and neural networks according to the present invention. Detailed Implementation

[0014] To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0015] like Figure 1 As shown, this embodiment of a portable ultrasound imaging method based on multi-angle plane waves and neural networks may specifically include:

[0016] S101, acquire the raw data of multi-angle plane waves collected by the ultrasonic probe.

[0017] In one embodiment, the ultrasound probe emits plane waves at multiple angles and receives echo signals. The acquired raw data of the multi-angle plane waves includes radio frequency signal sequences at different emission angles. This data is directly output from the probe's sensor array for subsequent processing.

[0018] S102, the first convolutional neural network model is used to filter out the optimized angle combination data from the original data of multi-angle plane waves.

[0019] In one embodiment, the first convolutional neural network model is a lightweight convolutional neural network designed to contain multiple convolutional and pooling layers for extracting features from raw multi-angle plane wave data. This model is deployed at the probe end to accommodate size and power consumption constraints. The model input is the acquired multi-angle data, and the output is a filtered, optimized combination of angles. The filtering rules are learned during training using a labeled dataset of valid angles.

[0020] First, the signal-to-noise ratio (SNR) value for each angle is calculated based on the original data of the multi-angle plane wave.

[0021] For example, the first convolutional neural network model first performs a Fourier transform on the raw data for each angle to obtain the spectral distribution, and then calculates the ratio of the signal peak to the noise level as the signal-to-noise ratio (SNR) value. Angles with high SNR are given priority for inclusion in the optimization combination.

[0022] Secondly, the signal-to-noise ratio is input into the input layer of the first convolutional neural network model for feature extraction and classification, and the optimized angle combination data is output.

[0023] In one embodiment, the convolutional layer of the first convolutional neural network model uses a 3x3 kernel to perform convolution operations on the signal-to-noise ratio matrix to extract local features. Then, the most effective angle combination is classified through a fully connected layer. For example, in an abdominal scanning scenario, the model selects 5-7 angles from the original 15, reducing the amount of data by about 60%, thereby improving the efficiency of subsequent transmission.

[0024] For example, in cardiac imaging applications, the model adjusts its selection based on dynamic tissue motion, prioritizing the combination of angles that capture clear images of the ventricular wall. The training data comes from simulated cardiac echoes, ensuring that the signal-to-noise ratio of the optimized combination is improved by more than 20%, which is beneficial to the real-time performance of portable devices.

[0025] The above-mentioned optimized angle combination data specifically includes:

[0026] First, an angle correlation matrix is ​​constructed to assess the complementarity between angles.

[0027] In one embodiment, the angle correlation matrix is ​​generated by calculating the Pearson correlation coefficient between different angle data. Angles with a coefficient higher than 0.8 are considered complementary and included in the matrix for further optimization and screening of the model.

[0028] Secondly, the angle correlation matrix is ​​used to guide the pooling layer operation of the first convolutional neural network model, refining and optimizing the angle combination data.

[0029] For example, the pooling layer uses the max pooling method to reduce the dimensionality of the correlation matrix and output refined combined data, which reduces bandwidth usage and improves imaging stability when transmitted on a tablet.

[0030] In one embodiment, for different probe types, such as linear array probes, the model parameters are adjusted to process higher frequency data, and the optimized combination is selected to cover a wider depth range, which is beneficial to the imaging effect of superficial tissues.

[0031] For example, in a wireless handheld ultrasound system, the first convolutional neural network model is deployed on the probe chip. After filtering, the amount of data is reduced from several GB to several MB. The data is then transmitted to the mobile phone for subsequent synthesis, which significantly reduces latency and improves image quality. Compared with conventional handheld ultrasound architecture, it improves the signal-to-noise ratio and reduces computational complexity, demonstrating the advantages of portable ultrasound imaging architecture design based on multi-angle plane waves and neural networks.

[0032] S103 transmits the optimized angle combination data to the terminal device.

[0033] First, a wireless communication connection is established between the probe and the terminal device, using WiFi or Bluetooth protocols to establish a data transmission channel. After the first convolutional neural network model at the probe end completes angle selection, it encapsulates the selected optimized angle combination data into a data packet, reducing the data packet size by 60% to 80% compared to the original full-angle data.

[0034] Secondly, the optimized angle combination data is compressed and encoded. A lossless compression algorithm is used to encode the filtered ultrasonic echo data to ensure the integrity and real-time requirements of the data during transmission.

[0035] S104 uses the second convolutional neural network model in the terminal device to perform beamforming and image processing on the optimized angle combination data to generate an ultrasound image.

[0036] First, a pre-trained second convolutional neural network model is loaded into the terminal device. This model includes an encoder-decoder architecture. The encoder consists of five convolutional layers, each containing 64 to 512 convolutional kernels of 3×3 size, using the ReLU activation function. The encoder is responsible for extracting deep features from the optimized angle combination data, converting the input multi-angle ultrasound echo signals into a high-dimensional feature representation. The decoder uses transposed convolutional layers for upsampling, progressively reconstructing the spatial resolution of the ultrasound image.

[0037] Secondly, the received optimized angle combination data undergoes preprocessing, including data decompression, normalization, and noise filtering. Normalization adjusts the amplitude range of the ultrasonic echo signal to between 0 and 1, while noise filtering uses a Gaussian filter to remove noise interference introduced during transmission. The preprocessed data forms a three-dimensional tensor with dimensions of angles × depth × lateral position.

[0038] Next, the preprocessed data is focused using a beamforming module within a second convolutional neural network model. This module employs an attention mechanism, assigning weight coefficients to ultrasound data from different angles. These weight coefficients are dynamically adjusted based on signal quality and imaging depth. This module simulates the focusing effect of traditional time-delay superposition beamforming, but learns optimal focusing parameters through neural network learning, avoiding complex geometric calculations and time delay compensation.

[0039] Then, the final ultrasound image is generated using the image reconstruction module of the second convolutional neural network model. The image reconstruction module contains residual connection structures, which preserve shallow feature information through skip connections, preventing the gradient vanishing problem during deep network training. This module outputs a 512×512 pixel grayscale ultrasound image, with pixel values ​​ranging from 0 to 255, corresponding to the intensity distribution of the ultrasound echo.

[0040] Finally, the generated ultrasound images undergo post-processing optimization, including contrast enhancement, edge sharpening, and artifact suppression. Contrast enhancement employs histogram equalization to improve the dynamic range and visual quality of the image. Edge sharpening enhances the clarity of tissue boundaries using the Laplacian operator. The artifact suppression module identifies and reduces imaging artifacts caused by incomplete angular data or noise.

[0041] In one embodiment, the second convolutional neural network model employs the U-Net architecture for beamforming and image processing. The U-Net encoding path contains four downsampling blocks, each consisting of two 3×3 convolutional layers and a 2×2 max-pooling layer. The decoding path contains four upsampling blocks, each restoring the feature map size through transposed convolution and then concatenating it with the corresponding encoding layer feature map. This symmetrical encoder-decoder structure effectively preserves multi-scale feature information, improving the reconstruction quality of ultrasound images.

[0042] For example, when processing cardiac ultrasound imaging, the optimized angle combination data includes echo data from eight optimal angles selected within the range of -30 degrees to +30 degrees. A second convolutional neural network model first extracts features from these eight angles, extracting feature representations of myocardial tissue, cardiac chamber structures, and valves. The beamforming module assigns appropriate focusing parameters to tissues at different depths based on the anatomical characteristics of the heart, ensuring good imaging results in both the near and far fields. The image reconstruction module combines multi-angle information to generate cardiac ultrasound images with high contrast and clear boundaries, clearly displaying ventricular wall thickness, cardiac chamber size, and valve motion.

[0043] In one possible implementation, the second convolutional neural network model is trained using a supervised learning method, with traditional focused beam imaging results serving as the standard reference image. The training dataset includes ultrasound data from different anatomical locations, imaging depths, and patient groups to ensure the model's generalization ability. The loss function employs a weighted combination of mean squared error and structural similarity index, guaranteeing both pixel-level accuracy and maintaining image structural integrity. After training, the model achieves a peak signal-to-noise ratio exceeding 35 dB and a structural similarity index exceeding 0.9 on the test data.

[0044] It should be noted that the second convolutional neural network model in the terminal device has approximately 20 million parameters, a model size of 80MB, and an inference time of approximately 50 milliseconds on a GPU-equipped mobile device, meeting the requirements for real-time ultrasound imaging. During model deployment, quantization techniques are used to convert floating-point parameters into 8-bit integers, reducing memory usage and computational complexity while maintaining image quality.

[0045] Specifically, the attention weight calculation in beamforming is based on the coherence and amplitude characteristics of the signal. For each pixel location, the model calculates the correlation coefficient of data from different angles; angles with high correlation coefficients receive greater weight, while those with low correlation coefficients are suppressed. This adaptive weight allocation mechanism effectively suppresses noise and artifacts, improving the signal-to-noise ratio and contrast resolution of the image. Furthermore, the weight distribution learned by the model is similar to the window function in traditional beamforming, but exhibits stronger adaptability and robustness.

[0046] S105 displays ultrasound images via a terminal device.

[0047] First, the system receives processed ultrasound image data transmitted from the probe. The terminal device then acquires the processed ultrasound image data via a second convolutional neural network, either wirelessly or via a wired connection. Terminal devices include mobile devices such as smartphones and tablets, which possess high-resolution displays and powerful image processing capabilities.

[0048] Secondly, the received ultrasound image data undergoes display format conversion, transforming it into a format suitable for the terminal device's display screen. Display format conversion includes color space conversion, resolution adaptation, and brightness / contrast adjustment. Color space conversion converts the ultrasound image from grayscale to RGB format for correct display on a color screen. Resolution adaptation scales the ultrasound image according to the terminal device's screen size and pixel density, ensuring clear display on devices of different sizes.

[0049] Next, image display optimization processing is performed to enhance the image based on the display characteristics of mobile devices. Image display optimization includes dynamic range compression, edge sharpening, and noise suppression. Dynamic range compression maps the grayscale range of the ultrasound image to the dynamic range of the display device, ensuring that image details are not lost during display through logarithmic compression or piecewise linear mapping. Edge sharpening enhances image edges using the Laplacian or Sobel operators, improving image clarity and contrast. Noise suppression uses median filtering or Gaussian filtering to reduce random noise in the image and improve display quality.

[0050] Then, a real-time display buffer management mechanism is established to ensure smooth display of ultrasound images. The buffer management mechanism employs double or triple buffering to avoid tearing or stuttering during image display. When new ultrasound image data arrives, the system writes it to the background buffer, while the foreground buffer continues to provide image data for display. Buffer switching is triggered by a vertical synchronization signal to ensure smooth image switching.

[0051] Finally, the ultrasound image is rendered on the terminal device's display screen, with the graphics processing unit accelerating the image rendering process. Image rendering includes texture mapping, pixel shading, and frame buffer output. Texture mapping applies the processed ultrasound image data as a texture map onto the display plane, the pixel shader performs color calculations and lighting processing for each pixel, and the frame buffer output transmits the final image data to the display controller.

[0052] In one embodiment, the above-described display format conversion uses different conversion parameters for different types of terminal devices.

[0053] For example, for a 5.5-inch smartphone, the image resolution is adapted to 1920×1080 pixels with a pixel density of 401ppi, and the color space conversion uses the sRGB standard. For a 10.1-inch tablet, the image resolution is adapted to 2560×1600 pixels with a pixel density of 299ppi, and the color space conversion uses the Adobe RGB standard to obtain a wider color gamut. Brightness and contrast are dynamically adjusted based on feedback from the ambient light sensor, increasing image brightness and contrast in bright light environments and reducing brightness in low light environments to protect the user's eyesight.

[0054] In one possible implementation, the image display optimization process described above employs an adaptive parameter adjustment mechanism. The compression ratio of dynamic range compression is automatically adjusted based on the image's histogram distribution; the compression ratio is increased when the image contrast is low and decreased when the image contrast is high. The edge sharpening intensity is adjusted based on the image's edge density; a smaller sharpening intensity is used for areas with abundant edges, and a larger sharpening intensity is used for areas with sparse edges. The noise suppression filtering parameters are dynamically adjusted based on the image's signal-to-noise ratio (SNR); a smaller filtering kernel is used when the SNR is high, and a larger filtering kernel is used when the SNR is low.

[0055] Specifically, the aforementioned real-time display buffer management is optimized for the real-time requirements of handheld ultrasound. The buffer size is dynamically adjusted based on the amount of image data and the transmission rate. When the data transmission rate is high, the buffer size is increased to avoid data overflow; when the data transmission rate is low, the buffer size is decreased to reduce display latency. Buffer switching is triggered by the arrival of new image data, the completion of the display refresh cycle, and user interaction events. The system monitors the buffer's fill status; when the buffer fill rate exceeds 80%, a fast switching mode is triggered; when the buffer fill rate is below 20%, an energy-saving mode is triggered.

[0056] For example, in practical applications, when a doctor uses a handheld ultrasound device to perform an abdominal examination on a patient, the multi-angle plane wave data acquired by the probe is filtered by a first convolutional neural network to select the five most effective angle combinations. This reduces the data volume from the original 15 angles to 5 angles, a reduction of approximately 67% in data transmission. The filtered angle data is then wirelessly transmitted to the doctor's handheld tablet. Upon receiving the data, the tablet immediately performs display format conversion, transforming the 512×512 pixel grayscale ultrasound image into a color image suitable for a 10.1-inch display. Image display optimization processing compresses the dynamic range of the original image from 12 bits to 8 bits while maintaining the clarity of image details. Edge sharpening processing makes organ boundaries more clearly visible. Real-time display buffer management ensures an image refresh rate of 30 frames per second, allowing doctors to observe dynamic images of the patient's internal organs in real time, promptly detect abnormalities, and make diagnoses.

[0057] Optionally, in some embodiments, the above-mentioned filtering of optimized angle combination data from the original multi-angle plane wave data using a first convolutional neural network model includes the following sub-steps S21-S24:

[0058] S21, obtain the signal-to-noise ratio information of the raw data of multi-angle plane waves.

[0059] In one embodiment, step S21 specifically includes: S211, performing time-domain analysis on the original data of multi-angle plane waves to calculate the signal power and noise power of each angle data; S212, obtaining the signal-to-noise ratio information based on the ratio of signal power to noise power.

[0060] For example, in a handheld ultrasound abdominal scan scenario, the raw data might contain plane wave signals from 15 angles. By calculating the average signal intensity at each angle and dividing it by the background noise level, the signal-to-noise ratio (SNR) is obtained, providing the basic data input for subsequent steps. This approach helps to quickly identify data quality, avoid processing low-quality angles, and improve overall screening efficiency.

[0061] S22, determine the imaging correlation of data from each angle based on the signal-to-noise ratio information.

[0062] In one embodiment, step S22 specifically includes: S221, constructing a correlation matrix using signal-to-noise ratio information, wherein the matrix elements represent the cross-correlation coefficients between data from different angles; S222, applying a weighted average to the matrix to quantify the contribution of each angle to the overall imaging, forming an imaging correlation value.

[0063] Specifically, in the portable ultrasound architecture, the lightweight design of the first CNN model allows these values ​​to be calculated in real time at the probe end.

[0064] For example, in cardiac imaging, if the signal-to-noise ratio (SNR) at a certain angle is higher than 20 dB, its correlation value can be calculated by cross-correlation with adjacent angles, increasing it to above 0.8, thereby filtering out angles that contribute significantly to ventricular wall motion. This extension ensures that the SNR output from S21 is directly converted into quantifiable correlation, reducing invalid data uploads, optimizing bandwidth usage, and bringing higher SNR advantages. For example, in technical demonstrations, the model can be used to select effective angles for scanning scenarios, improving imaging results.

[0065] For example, for different tissue depths, such as superficial skin scans, the correlation weight can be adjusted to prioritize angles with high signal-to-noise ratios. In deep liver scans, the correlation of low-frequency angles is emphasized. Through reasoning, if the correlation of an angle with a signal-to-noise ratio of less than 10dB is set to below 0.3, this can support multiple application scenarios and form a consistent optimization argument, namely, reducing the amount of data while improving stability.

[0066] In one embodiment, in another scenario, for vascular imaging, S22 can be extended to calculate dynamic correlation, generating time-varying correlation values ​​based on the signal-to-noise ratio changes of consecutive frames. This is more accurate than static calculation, bringing real-time performance improvement and benefiting the portability of handheld ultrasound.

[0067] S23, for imaging correlation, determine the angle data that meets the preset correlation threshold.

[0068] In one embodiment, step S23 specifically includes: S231, comparing the imaging correlation value with a preset threshold such as 0.6; S232, marking and extracting angle data that meets the threshold.

[0069] For example, in the mobile-connected PalmSuper system, this judgment can quickly filter out the 5 angles with low relevance, retaining only 10, and connecting with the relevance output of S22 to form a logical chain.

[0070] S24. Generate optimized angle combination data based on angle data that meets the preset correlation threshold.

[0071] In one embodiment, step S24 specifically includes: S241, sorting the angle data that meet the threshold and arranging them from high to low according to the correlation value; S242, applying a combination algorithm such as greedy selection to generate an optimized combination containing at least 5 to 8 angles; S243, verifying the overall signal-to-noise ratio improvement of the combination to ensure that it exceeds the average value of the original data.

[0072] Specifically, in the neural network ultrasonic beam synthesis algorithm based on multi-angle plane waves, this step is deployed in a lightweight CNN at the probe end, connecting to the judgment result of S23.

[0073] For example, in abdominal imaging, 7 angles are selected from 15 angles, reducing the amount of combined data by 60%, and then uploaded to the mobile device for second CNN synthesis. This reduces the transmission burden, improves real-time performance, and fits the focused wave imaging effect, reducing manual processing and aligning with the advantages of the architecture design.

[0074] For example, in one possible implementation, for parameter variation, such as when the threshold is set to 0.7, the generated combination may only have 4 angles, which is suitable for high signal-to-noise ratio scenarios; while in low signal-to-noise ratio cardiac scans, the threshold is reduced to 0.5, and the combination is expanded to 10. This is a progressive logic from core screening to extended variation, ensuring the integrity of the scheme, and enriching the diversity through optional implementations, such as emphasizing the beneficial effects: reduced data volume and improved algorithm stability.

[0075] In one embodiment, in another direction, the combined generation can incorporate weight optimization, based on the correlation weighted summation of S22, to form an enhanced combination, which is consistent with the first CNN screening of the most effective angle combination in the technical briefing, supporting the portable ultrasound architecture, and from multiple perspectives, the common argument is to improve the signal-to-noise ratio and reduce processing complexity.

[0076] Optionally, in some embodiments, transmitting the optimized angle combination data to the terminal device includes the following sub-steps S31-S33:

[0077] S31, Obtain the compression format of the optimized angle combination data.

[0078] First, the optimized angle combination data is preprocessed by rearranging the multi-angle ultrasonic echo signals according to the angle sequence to form an angle-ordered data matrix. During preprocessing, the amplitude range and frequency distribution characteristics of the ultrasonic echo data corresponding to each angle are first identified. Then, based on the time-domain and frequency-domain characteristics of the ultrasonic signals, correlation analysis is performed on the data from adjacent angles to construct a correlation matrix between angles.

[0079] Secondly, based on the sparsity characteristics of ultrasound signals, wavelet transform is used to sparsely represent the angle-ordered data matrix. During the wavelet transform process, the Daubechies wavelet is selected as the basis function, and the ultrasound echo signal at each angle is decomposed into wavelet coefficients for different frequency sub-bands. By setting an adaptive threshold, wavelet coefficients with larger amplitudes are retained, while redundant coefficients with smaller amplitudes are filtered out, thus achieving a sparse representation of the data.

[0080] Next, the sparsified wavelet coefficients are quantized and encoded, and a non-uniform quantization method is used to differentiate the coefficients of different frequency sub-bands. Higher quantization precision is used for low-frequency sub-bands to preserve the basic structural information of the image, while lower quantization precision is used for high-frequency sub-bands to reduce the amount of data. During quantization, appropriate quantization step sizes are assigned to different frequency components based on the characteristics of human visual perception and the diagnostic requirements of ultrasound imaging.

[0081] Finally, entropy coding is used to losslessly compress the quantized coefficients, generating optimized angle combination data in a compressed format. During entropy coding, the probability distribution of the quantized coefficients is statistically analyzed, and a Huffman coding table is constructed. Coefficients with high frequency are assigned shorter codes, while coefficients with low frequency are assigned longer codes, achieving further data compression.

[0082] S32, transmit the optimized angle combination data of the compressed format to the terminal device via wired or wireless communication.

[0083] First, select the appropriate data transmission protocol based on the connection method between the probe and the terminal device. For wired connections, use a USB 3.0 or Type-C interface for high-speed data transmission; for wireless connections, use WiFi 6 or Bluetooth 5.0 protocols for data transmission.

[0084] Secondly, the optimized angle combination data in compressed format is divided according to the preset data packet size, and data packet header information and checksum are added. The data is then sent to the terminal device through the selected communication method.

[0085] S33, the optimized angle combination data in the compressed format is decompressed in the terminal device to obtain the transmitted optimized angle combination data.

[0086] First, compressed data packets are received from the probe, and the packets are reassembled and their integrity verified based on the header information. During the verification process, the correctness of each data packet's transmission is verified using a Cyclic Redundancy Check (CRC) code. For data packets that fail verification, a retransmission request is sent to the probe to ensure complete data reception.

[0087] Secondly, entropy decoding is performed on the recombined compressed data. Based on the pre-stored Huffman decoding table, the compressed binary code stream is restored to the quantized wavelet coefficients. During the decoding process, the coefficient information in the compressed data is parsed step by step in reverse order of the encoding process to reconstruct the complete wavelet coefficient matrix.

[0088] Next, the decoded quantized wavelet coefficients are dequantized. Based on the quantization step size used during encoding, the quantized coefficient values ​​are restored to the original wavelet coefficient amplitudes. During the dequantization process, appropriate dequantization parameters are used for different frequency subbands to ensure accurate recovery of each frequency component.

[0089] Finally, inverse wavelet transform is used to reconstruct the time-domain ultrasonic echo signal from the dequantized wavelet coefficients, obtaining the optimized angle combination data after transmission. During the inverse transform, the coefficients of each frequency sub-band are synthesized according to the inverse process of wavelet decomposition to reconstruct the complete ultrasonic echo data corresponding to each angle.

[0090] In one embodiment, the wavelet transform described above is implemented using a fast algorithm with a lifting scheme. The lifting scheme wavelet transform converts traditional convolution operations into a series of prediction and update operations, significantly reducing computational complexity. In the prediction step, the value of the current sample point is predicted using a linear combination of adjacent sample points, and the prediction error is calculated as a high-frequency coefficient. In the update step, the low-frequency coefficients are corrected using the high-frequency coefficients to maintain the statistical characteristics of the signal. This implementation is particularly suitable for running on low-power processors at the probe end, reducing computational resource consumption while ensuring compression performance.

[0091] In one embodiment, the quantization process described above employs a perceptual quantization method. Perceptual quantization differentiates image information for different regions based on the clinical diagnostic needs of ultrasound imaging. For tissue boundaries and lesion areas containing important diagnostic information, a smaller quantization step size is used to maintain high accuracy. For uniform background areas and noisy areas, a larger quantization step size is used to achieve a higher compression ratio. The selection of the quantization step size is based on the gradient information and texture complexity of the local image, achieving a balance between compression efficiency and image quality through adaptive adjustment.

[0092] For example, when the probe acquires plane wave data from 15 angles, each angle containing echo signals from 128 array elements and 1024 sampling points, the original data size is approximately 30MB. After preprocessing in step S311, the data is arranged in an angle sequence to form a 15×128×1024 three-dimensional data matrix. In step S312, a 4-level wavelet decomposition is used to decompose the signal into one low-frequency sub-band and 12 high-frequency sub-bands. Through adaptive thresholding, approximately 20% of the important wavelet coefficients are retained. In step S313, 8-bit quantization is used for the low-frequency sub-band and 4-bit quantization is used for the high-frequency sub-band, reducing the quantized data size to approximately 6MB. Entropy coding in step S314 further compresses the data to approximately 3MB, achieving a compression ratio of 10:1.

[0093] It should be noted that the choice of the above communication protocol needs to consider the real-time requirements of data transmission. For real-time ultrasound imaging, a frame rate of 30fps or higher is required, and each frame of 3MB compressed data needs to be transmitted within 33 milliseconds. The theoretical transmission speed of the WiFi 6 protocol can reach 1Gbps, and the actual usable bandwidth is about 600Mbps, which can meet the needs of real-time transmission. Bluetooth 5.0 has a lower transmission speed and is suitable for static image transmission scenarios where real-time requirements are not high.

[0094] Specifically, the entropy decoding process described above needs to handle the complexity of variable-length encoding. Huffman decoding is implemented using a lookup table approach, pre-constructing a decoding tree structure and determining the corresponding original symbol by reading the compressed bitstream bit by bit and searching the path in the decoding tree. To improve decoding speed, a multi-level lookup table method is used, storing frequently used short codewords in a fast lookup table and processing longer codewords using a tree search approach. This hybrid decoding method significantly improves decoding efficiency while ensuring decoding accuracy.

[0095] In one possible implementation, the aforementioned wavelet inverse transform process employs a parallel processing architecture. The multi-core processor of the terminal device can simultaneously process the inverse transform operations of multiple frequency subbands, distributing the data from 15 angles to different processing cores for parallel execution. Each processing core is responsible for the complete reconstruction process of one angle's data, including coefficient synthesis, boundary processing, and data format conversion. Parallel processing reduces the original 150-millisecond serial reconstruction time to approximately 20 milliseconds, meeting the time requirements of real-time imaging. After reconstruction, the correlation between the obtained post-transmission optimized angle combination data and the original data reaches over 0.95, ensuring the accuracy of subsequent neural network processing.

[0096] Optionally, in some embodiments, the above-mentioned beamforming and image processing of the optimized angle combination data using a second convolutional neural network model in the terminal device to generate an ultrasound image includes the following sub-steps S41-S44:

[0097] S41, obtain optimized angle combination data.

[0098] First, the data packets, after angle filtering, are received from the first convolutional neural network model at the probe end.

[0099] Next, the angle index information and corresponding ultrasonic echo data in the data packet are parsed. Step S413: The optimized angle combination data is loaded into the memory buffer of the terminal device to prepare for subsequent beamforming processing.

[0100] S42, an adaptive beamforming algorithm is executed through a second convolutional neural network model to generate an initial ultrasound image.

[0101] First, a multi-layer convolutional neural network architecture is constructed, including an input layer, multiple convolutional layers, pooling layers, and fully connected layers. The input layer receives ultrasonic echo data from different emission angles, with each angle corresponding to a data channel. The convolutional layers use three-dimensional convolutional kernels to extract features from the ultrasonic signals in the temporal and spatial domains. The first convolutional kernel has a size of 3×3×16 and is used to extract local spatiotemporal features. The second convolutional kernel has a size of 5×5×32 and is used to capture a wider range of signal correlations. The pooling layers use max pooling with a pooling window size of 2×2 to reduce data dimensionality while retaining important feature information.

[0102] Secondly, an adaptive weight calculation process is executed. Based on the signal-to-noise ratio and coherence indices of data from different angles, the weight coefficients of each angle data point in beamforming are dynamically adjusted. The coherence factor of each angle data point is calculated, with angle data having higher coherence factors assigned larger weight values. Simultaneously, the cross-correlation between angles is considered, and adjacent angles with excessively high correlation undergo weight normalization to avoid over-amplification of information from certain angles.

[0103] Next, delay-superimposed beamforming is performed. Based on the propagation time of ultrasound at different depths, the delay parameters corresponding to each angle are calculated. The multi-angle data, after delay correction, are then weighted and superimposed according to the weighting coefficients calculated in step S422. During the superposition process, an interpolation algorithm is used to accurately compensate for non-integer delays. Cubic spline interpolation is used to ensure signal continuity and smoothness.

[0104] Finally, the beamforming effect is optimized through the nonlinear mapping capability of the neural network. Fully connected layers map the features extracted by the convolutional layers into pixel-level image data, and the ReLU activation function enhances the network's nonlinear expressive power. The output layer generates the grayscale matrix of the initial ultrasound image, with the image size determined based on the scanning depth and lateral range.

[0105] S43, perform signal denoising processing on the initial ultrasound image to obtain a denoised image.

[0106] First, the types and distribution characteristics of noise in the initial ultrasound images are identified. By analyzing the frequency domain characteristics of the images, different noise components such as speckle noise, electronic noise, and motion artifacts are distinguished. Speckle noise manifests as granular texture in the image, while electronic noise presents as randomly distributed high-frequency components.

[0107] Secondly, an adaptive filtering algorithm is applied for noise suppression. A Wiener filter is used to process electronic noise, with filter parameters adaptively adjusted based on the local signal-to-noise ratio. For speckle noise, an anisotropic diffusion filtering method is used to suppress noise while preserving image edge and detail information. The filtering process is set to iterate 10 times, and the diffusion coefficient is dynamically adjusted according to the image gradient magnitude.

[0108] Finally, morphological post-processing operations are performed. Opening operations are used to remove isolated noise points in the image, with a 3×3 square kernel as the structuring element. Closing operations are used to fill small holes and breaks in the image, preserving the continuity of the tissue structure.

[0109] S44 performs image enhancement processing on the denoised image to generate an ultrasound image.

[0110] First, contrast enhancement is performed. Histogram equalization is used to expand the dynamic range of the image and improve the contrast between different tissues. Second, edge sharpening is applied. The Laplacian operator is used to enhance image edge information, highlighting the contour features of tissue boundaries and lesion areas. Finally, gamma correction is applied to adjust the image brightness distribution, making the image display consistent with the observation habits of medical diagnosis.

[0111] In one embodiment, the adaptive beamforming algorithm in step S42 is specifically embodied in the neural network employing different processing strategies for different scanning depths. During shallow tissue scanning, the network focuses on high-frequency signal components to obtain better axial resolution. During deep tissue scanning, the network utilizes more low-frequency signal components to improve penetration and signal-to-noise ratio. Through training, the network learns this depth-dependent optimal processing method, achieving superior imaging results at different depths compared to traditional fixed-parameter beamforming methods.

[0112] For example, when scanning deep organs such as the heart, the second convolutional neural network model automatically adjusts the weights of the convolutional kernels in each layer to enhance the response to low-frequency echo signals.

[0113] Specifically, the network increases the weight of signal components in the 2-5MHz frequency band to 0.8, while decreasing the weight of high-frequency components above 5MHz to 0.3. This adaptive adjustment allows for clear visualization of cardiac structures within a depth range of 4-8 cm, while effectively suppressing noise interference in deep tissues.

[0114] In one possible implementation, the signal denoising process in step S43 employs differentiated denoising strategies for different anatomical locations. For solid organs such as the liver, speckle noise is the primary noise source, and the denoising algorithm primarily uses spatial domain filtering methods. The filter window size is set to 5×5 pixels, which effectively smooths the speckle texture while maintaining the clarity of the liver outline. For tubular structures such as blood vessels, the denoising process focuses more on preserving edge information, employing an edge-preserving filter that avoids over-smoothing of the vessel walls by detecting local gradient directions.

[0115] It should be noted that the image enhancement processing in step S44 will be adaptively adjusted according to the display characteristics of the terminal device. When the ultrasound image is displayed on a mobile phone screen, considering the brightness and contrast characteristics of the mobile phone screen, the image enhancement algorithm will appropriately increase the overall brightness and enhance the contrast of medium grayscale areas. When the image is displayed on a tablet computer, due to the larger screen size, the image enhancement focuses more on the clarity of details and will perform moderate sharpening processing on the image to compensate for the blurring of details that may be caused by the large screen display.

[0116] Preferably, the entire beamforming and image processing process is executed in parallel on the GPU of the terminal device, making full use of the mobile device's computing power. The inference process of the second convolutional neural network model is accelerated by the GPU, and the processing time of a single frame image is controlled within 50 milliseconds, ensuring the smoothness of real-time imaging. At the same time, the memory usage during processing is optimized through data pipeline design, avoiding the storage of a large amount of intermediate data, enabling the entire algorithm to run stably on memory-constrained mobile devices.

[0117] Optionally, in some embodiments, the above-described signal denoising processing of the initial ultrasound image to obtain a denoised image includes the following sub-steps S431-S433:

[0118] S431, Obtain the pixel distribution characteristics of the initial ultrasound image.

[0119] First, the initial ultrasound image is statistically analyzed to calculate the grayscale intensity distribution of each pixel. Different tissue structures in ultrasound images exhibit different echo characteristics, and by statistically analyzing the pixel grayscale value distribution, tissue boundaries and internal structural features in the image can be identified.

[0120] Next, the pixel gradient information of the initial ultrasound image is calculated to determine the edge and texture features in the image. The Sobel operator or Canny edge detection operator is used to calculate the gradient magnitude in the horizontal and vertical directions, forming a gradient magnitude matrix and a gradient direction matrix.

[0121] Finally, local texture features of the initial ultrasound image are extracted to generate a pixel distribution feature vector. A local binary mode operator is used to calculate a texture descriptor in the neighborhood of each pixel, combining grayscale distribution, gradient information, and texture features to form a multidimensional feature vector.

[0122] S432, The initial ultrasound image is processed using a deep learning-based filtering algorithm to determine the noise distribution.

[0123] First, a convolutional neural network noise detection model is constructed, comprising a feature extraction layer, a noise recognition layer, and a distribution prediction layer. The feature extraction layer employs multi-scale convolutional kernels to extract multi-level features from the image, including 3×3, 5×5, and 7×7 kernels to capture detailed texture, medium-scale structure, and large-scale background information, respectively. The noise recognition layer learns the difference features between noise and normal tissue signals through residual connection structures and enhances the model's non-linear expressive power using batch normalization and ReLU activation functions. The distribution prediction layer uses a fully convolutional structure to output a noise probability distribution map of the same size as the input image.

[0124] Next, the pixel distribution feature vector is input into the convolutional neural network noise detection model to generate a preliminary noise probability distribution map. The model first performs feature fusion on the input feature vector, adaptively allocating weights to different feature channels through an attention mechanism to highlight feature dimensions effective for noise detection. During feature extraction, dilated convolution is used to expand the receptive field, capturing a wider range of contextual information without increasing the number of parameters.

[0125] Furthermore, based on the speckle noise characteristics of ultrasound images, the preliminary noise probability distribution map is post-processed and optimized. Speckle noise in ultrasound images exhibits multiplicative characteristics and spatial correlation. A statistical model of speckle noise is introduced to correct the output of the deep learning model. The Rayleigh distribution model is used to describe the amplitude distribution characteristics of speckle noise, and the spatial correlation function of the noise is calculated by combining local statistical features to correct false positives and false negatives in the preliminary noise probability distribution map.

[0126] Finally, the final noise distribution map is generated by considering the physical constraints of ultrasound imaging. Based on the attenuation law of ultrasound waves in human tissue, the noise level is higher in deeper areas; this physical constraint is incorporated as prior knowledge into the noise distribution calculation. A time gain compensation curve is used to weight and adjust the noise distribution at different depths to ensure that the noise distribution map conforms to the physical characteristics of ultrasound imaging.

[0127] In one embodiment, the convolutional neural network noise detection model employs an improved version of the U-Net architecture. The encoder is responsible for extracting multi-scale features, while the decoder is responsible for restoring spatial resolution and generating a noise distribution map. The encoder includes four downsampling stages, each containing two convolutional layers and one max-pooling layer, with the feature map size halved and the number of channels doubled sequentially. The decoder uses transposed convolutions for upsampling and fuses the encoder's feature map with the decoder's feature map through skip connections, preserving detailed information while restoring spatial resolution.

[0128] In one possible implementation, a multi-task learning strategy is used to train the noise detection model for different types of ultrasound probes and imaging modes. For the linear scanning mode of a linear array probe, the focus is on learning the noise difference characteristics between the near and far fields. For the fan-shaped scanning mode of a convex array probe, the focus is on learning the geometric distortion noise characteristics at the edge of the fan. For the fan-shaped depth imaging of a phased array probe, the focus is on learning the coherent noise characteristics introduced during the multi-angle plane wave synthesis process.

[0129] S433, Remove noise interference from the initial ultrasound image according to the noise distribution to generate the denoised image.

[0130] First, adaptive filter parameters are constructed based on the noise distribution map. The filter strength is dynamically adjusted according to the noise probability value at each pixel location, with strong filter parameters used in areas with high noise probability and weak filter parameters used in areas with low noise probability, to avoid loss of image details due to over-filtering.

[0131] Secondly, a nonlocal mean filtering algorithm is used to denoise the initial ultrasound image. Nonlocal mean filtering utilizes the correlation between similar structural blocks in the image for denoising. It searches for pixel blocks similar to the current pixel block and calculates a weighted average as the denoising result. The weights of the similarity metric are adjusted based on the noise distribution map, increasing the search window and similarity threshold in areas with strong noise.

[0132] Finally, edge protection processing is performed on the denoised image to restore important tissue boundary information. Gradient information from the original image is used as an edge prior to provide special protection for edge regions during denoising, preventing excessive smoothing of important anatomical structure boundaries. A bilateral filtering approach is employed, considering pixel similarity in both the spatial and grayscale domains.

[0133] Specifically, in cardiac ultrasound imaging applications, there is a significant echo difference between myocardial tissue and blood chambers, requiring noise detection models to accurately identify the noise distribution near the endocardial boundary. By training data containing images of different cardiac cycles, the model learns the noise variation patterns during myocardial contraction and relaxation. After determining the noise distribution, an adaptive filter employs anisotropic filtering in the endocardial boundary region, smoothing along the boundary direction while maintaining sharpness perpendicular to it, thus removing noise while preserving crucial anatomical boundary clarity.

[0134] For example, in abdominal ultrasound imaging, speckle noise in the liver parenchyma exhibits a uniform distribution, while noise at the liver-gallbladder interface shows a boundary enhancement effect. The noise detection model, by learning noise patterns from a large number of abdominal ultrasound images, can accurately distinguish between speckle noise within the tissue and artifact noise generated by interface reflections. In the denoising process, uniform speckle suppression filtering is applied to the liver parenchyma region, while edge protection filtering is applied to high-contrast interfaces such as the gallbladder wall to ensure accurate lesion detection.

[0135] In one embodiment, combining the multi-angle plane wave imaging architecture mentioned in the technical disclosure, the second CNN model integrates the aforementioned denoising process while performing beamforming. When the effective angle data selected by the first CNN at the probe end is processed on the mobile phone / tablet, it first generates an initial ultrasound image through beamforming, and then enters the denoising process. Since the data volume has been significantly reduced through angle selection, the mobile phone / tablet has sufficient computing resources to run complex deep learning denoising algorithms, achieving high-quality real-time ultrasound imaging.

[0136] Optionally, in some embodiments, the above-described image enhancement processing of the denoised image to generate the ultrasound image includes: S441, obtaining the resolution information of the denoised image; S442, processing the denoised image using a deep learning-based super-resolution algorithm to determine enhancement features; S443, adjusting the contrast and sharpness of the denoised image according to the enhancement features; and S444, generating the ultrasound image.

[0137] Step S441: Obtain the resolution information of the denoised image.

[0138] Step S4411: Read the pixel matrix dimensions of the denoised image and determine the width and height parameters of the image. Step S4412: Calculate the pixel density and spatial resolution values ​​of the image.

[0139] Step S442: The denoised image is processed using a deep learning-based super-resolution algorithm to determine the enhancement features.

[0140] Step S4421: Construct a residual dense network as the backbone network for super-resolution reconstruction. This network contains multiple residual dense blocks, each containing five convolutional layers. The outputs of the first four convolutional layers are passed to subsequent layers through dense connections, and the last convolutional layer integrates features through a local feature fusion layer. The residual dense blocks directly pass the input to the output through skip connections, achieving local residual learning.

[0141] Step S4422: Design a multi-scale feature extraction module, which uses convolutional kernels of different sizes to process the denoised image in parallel. This module contains three branches, which use 3×3, 5×5 and 7×7 convolutional kernels to extract feature information of different receptive fields, respectively. The outputs of the three branches are weighted and fused through a channel attention mechanism to generate multi-scale fused features.

[0142] Step S4423 introduces a sub-pixel convolutional layer for upsampling, converting the low-resolution feature map into a high-resolution output. The sub-pixel convolutional layer rearranges the pixel positions in the feature map through periodic shuffling operations, achieving resolution improvement while maintaining the integrity of feature information.

[0143] Step S4424: Apply a perceptual loss function to guide the network training process. This loss function combines pixel-level mean squared error loss and perceptual loss based on a pre-trained VGG network. The perceptual loss ensures the visual perception quality of the reconstructed image by comparing the feature representation differences between the reconstructed image and the target image in the intermediate layers of the VGG network.

[0144] Step S4425: Extract the feature maps from the intermediate layers of the network as enhancement features. These features contain texture information, edge information, and structural information of the image. The enhancement features are stored in the form of multidimensional tensors, with each channel corresponding to different types of image enhancement information.

[0145] Step S443: Adjust the contrast and sharpness of the denoised image according to the enhancement features.

[0146] Step S4431: Establish the mapping relationship between enhancement features and contrast adjustment parameters. Calculate the contrast enhancement coefficient by analyzing the brightness distribution and dynamic range features in the enhancement features. This coefficient is determined based on histogram statistics and local contrast measurements in the enhancement features and is used to adjust the brightness and contrast of the image.

[0147] Step S4432: Construct a method for calculating the sharpness adjustment parameters. Edge intensity features and high-frequency detail features are extracted from the enhancement features, and the weight parameters of the sharpening filter are calculated using these features. The sharpening filter employs an improved form of the Laplacian operator, adaptively adjusting the filter intensity based on the enhancement features.

[0148] Step S4433: Design an adaptive gamma correction algorithm. This algorithm dynamically adjusts the gamma value based on the brightness distribution information in the enhancement features. By analyzing the cumulative distribution function of the image and the brightness statistical characteristics in the enhancement features, the algorithm assigns different gamma correction parameters to different brightness areas, achieving fine adjustment of local contrast.

[0149] Step S4434: Apply a bilateral filter for edge-preserving smoothing. The spatial and intensity weights of the bilateral filter are adaptively adjusted based on the edge information and noise level in the enhanced features, maintaining the sharpness of image edges while removing residual noise.

[0150] Step S444: Generate the ultrasound image.

[0151] Step S4441: Convert the adjusted image data into a standard ultrasound image format and set the corresponding image metadata information. Step S4442: Output the final ultrasound image, which has enhanced contrast and clarity.

[0152] In one embodiment, the residual dense network in step S442 specifically adopts an RDN architecture, which fully utilizes the information from all preceding layers through dense connections. The network contains 16 residual dense blocks, and the five convolutional layers within each block output 32 feature channels respectively. After feature dimensionality reduction through 1×1 convolution, the features are passed to the next residual dense block. This design enables the network to achieve good super-resolution reconstruction results with a relatively small number of parameters, making it particularly suitable for deployment on mobile devices such as smartphones and tablets.

[0153] For example, the channel attention mechanism in the multi-scale feature extraction module generates importance weights for each channel through global average pooling and fully connected layers. For the input 256-channel feature map, the channel attention module first compresses each channel's feature map into a single value, then learns the correlation between channels through two fully connected layers, ultimately outputting 256 weight values ​​to weight features at different scales. This mechanism can automatically learn which scales of features are more important to the current ultrasound image.

[0154] In one possible implementation, the subpixel convolutional layer's upsampling factor is set to 4x, essentially magnifying the input low-resolution image by a factor of 4. This layer learns convolutional kernels for 16 output channels and then rearranges these 16 channels into 4×4 pixel blocks, achieving a mapping from 1 pixel to 16 pixels. Compared to traditional bilinear interpolation upsampling, subpixel convolution can learn an upsampling pattern more suitable for the characteristics of ultrasound images.

[0155] Specifically, in the calculation of the perceptual loss function, the pixel-level loss weight is set to 0.1, and the perceptual loss weight is set to 1.0. The perceptual loss uses the features of the conv3_4 and conv4_4 layers of the pre-trained VGG-19 network, which can capture the mid-level semantic features of the image. Guided by this loss function, the reconstructed ultrasound image not only closely approximates the target image in pixel values ​​but also better conforms to the perceptual characteristics of the human eye in terms of visual quality.

[0156] In one embodiment, the contrast adjustment in step S443 employs an adaptive histogram equalization method. This method divides the image into 8×8 sub-blocks based on local contrast information in the enhancement features, and each sub-block undergoes independent histogram equalization processing. The sub-block boundaries are smoothly transitioned using bilinear interpolation to avoid block artifacts. This local processing approach can better enhance the contrast of tissues at different depths in ultrasound images.

[0157] Preferably, the sharpening filter in the sharpness adjustment employs an improved version of the unsharpened mask algorithm. The algorithm first applies a Gaussian blur to the original image to obtain a low-pass filter result, then calculates the difference between the original image and the blurred image to obtain high-frequency details. Finally, it adaptively adjusts the enhancement level of the high-frequency details based on the edge strength information in the enhancement features. For regions with strong edges, a higher sharpening intensity is set, while for smooth regions, a lower sharpening intensity is maintained.

[0158] It should be noted that the gamma value calculation in the adaptive gamma correction algorithm is based on the image's average brightness and standard deviation. When the enhanced features show that the image is generally dark, the gamma value is set to less than 1.0 to brighten the image; when the image is generally bright, the gamma value is set to greater than 1.0 to darken the image. The specific gamma value is determined by looking up a preset mapping table, which is established based on the statistical characteristics of a large number of ultrasound images.

[0159] In one embodiment, the spatial standard deviation of the bilateral filter is set to 5 pixels, while the intensity standard deviation is dynamically adjusted based on noise estimates in the enhancement features. A larger intensity standard deviation is set when noise levels are high to achieve a stronger smoothing effect, and a smaller intensity standard deviation is set when noise levels are low to preserve more image detail. This adaptive adjustment mechanism ensures an optimal balance between filtering effectiveness and image quality.

[0160] Optionally, in some embodiments, the above-mentioned display of the ultrasound image through the terminal device includes: S51, obtaining display parameters of the ultrasound image; S52, adjusting the display ratio of the ultrasound image according to the display parameters; S53, presenting the adjusted ultrasound image on the display interface of the terminal device.

[0161] Step S51, obtaining the display parameters of the ultrasound image, including:

[0162] Step S511: Extract image size information from the ultrasound image processed by the second CNN. Image size information includes the pixel width and pixel height, which are directly derived from the dimensions of the image matrix output after neural network beamforming. In the portable ultrasound imaging architecture, the second CNN model completes beamforming and image processing on a mobile phone or tablet, and the output image data contains complete size attributes.

[0163] Step S512 involves detecting the dynamic range parameter of the ultrasound image. The dynamic range parameter reflects the ratio between the brightest and darkest pixel values ​​in the image. It is obtained by traversing the image pixel matrix to obtain the maximum and minimum grayscale values. Since multi-angle plane wave imaging has a better signal-to-noise ratio after neural network processing, its dynamic range is generally superior to traditional focused wave imaging. The detection process is completed through pixel value statistics, calculating the image histogram distribution characteristics, and determining the effective grayscale value range.

[0164] Step S513: Obtain the screen display capability parameters of the terminal device. These parameters include the screen's physical resolution, pixel density, color depth, and brightness adjustment range. Terminal devices such as mobile phones and tablets have different screen specifications, requiring the acquisition of the device's display interface parameters to ensure image display compatibility. These parameters are obtained by calling the terminal device's hardware interface, providing basic data for subsequent display ratio adjustments.

[0165] Step S52, adjusting the display scale of the ultrasound image according to the display parameters, includes:

[0166] Step S521: Calculate the image-to-screen size matching ratio. Compare the pixel size of the ultrasound image with the displayable area of ​​the terminal device screen, and calculate the scaling ratios in the horizontal and vertical directions. When the image size exceeds the screen display area, select a smaller scaling ratio to ensure the image is displayed completely. When the image size is smaller than the screen display area, it can be appropriately enlarged to improve the display effect.

[0167] Step S522: Image scaling is performed based on bilinear interpolation. Bilinear interpolation determines the new pixel value by calculating the weighted average of the four neighboring pixels surrounding the target pixel. For each scaled target pixel location, the corresponding floating-point coordinate position is found in the original image, and then the distance weights of the four integer coordinate pixels surrounding that position are calculated. The grayscale value of the target pixel is obtained by weighted averaging, ensuring the smoothness and continuity of the scaled image.

[0168] Step S523: Adjust the grayscale mapping relationship of the image to adapt to the screen display range. Based on the detected image dynamic range and screen display capabilities, a grayscale value mapping function is established. When the image dynamic range exceeds the screen display range, linear compression mapping is used to compress the image grayscale values ​​to the screen's displayable range. When the image dynamic range is smaller than the screen display range, linear stretching mapping is used to expand the image contrast. The mapping process maintains the relative grayscale relationship of the image, ensuring the integrity of the diagnostic information in the ultrasound image.

[0169] Step S53: The adjusted ultrasound image is presented on the display interface of the terminal device. The ultrasound image data, after size and grayscale adjustment, is transmitted to the display buffer of the terminal device, and the final ultrasound image is displayed on the screen through the graphics rendering interface.

[0170] In one embodiment, the image size information acquisition process in step S511 specifically demonstrates the advantages of a portable ultrasound imaging architecture. Due to the adoption of a dual-CNN architecture, the first lightweight CNN performs angle filtering at the probe end, significantly reducing the amount of data transmitted to the mobile phone / tablet. The second CNN processes the filtered angle data on the mobile phone / tablet, outputting an image with a fixed size format.

[0171] For example, the original multi-angle data contains plane wave information from 75 different angles. After the first CNN filters the data, only the 15 most effective angles are retained, reducing the data volume by 80%. The second CNN processes the data from these 15 angles and outputs a 512×512 pixel ultrasound image, with the image size information obtained directly from the output matrix.

[0172] In one possible implementation, the dynamic range detection in step S512 fully leverages the imaging advantages of neural network beamforming. Traditional manual signal processing algorithms are prone to artifacts and noise, resulting in limited image dynamic range. In contrast, CNN-based beamforming can fit the focused wave imaging effect, achieving a higher signal-to-noise ratio and a wider dynamic range.

[0173] Specifically, the detection process iterates through all pixel values ​​in a 512×512 image matrix, statistically determining a maximum grayscale value of 245, a minimum grayscale value of 8, and a dynamic range ratio of 30.6. Simultaneously, an image histogram is constructed, revealing that effective grayscale values ​​are primarily distributed between 20 and 230, providing a precise parameter basis for subsequent display adjustments.

[0174] For example, the size matching ratio calculation process in step S521 reflects the practical needs of portable ultrasound devices. Mobile phone screen resolution is typically 1920×1080 pixels, while tablet screen resolution may be 2560×1600 pixels. When a 512×512 ultrasound image is displayed on a mobile phone, the available display area is assumed to be 1080×800 pixels. The calculated horizontal scaling ratio is 1080 / 512 = 2.11, and the vertical scaling ratio is 800 / 512 = 1.56. Choosing the smaller 1.56 as the uniform scaling ratio ensures that the image is fully displayed in the center of the screen, resulting in a final display size of 799×799 pixels. This adaptive scaling guarantees a consistent display effect of the ultrasound image across different terminal devices.

[0175] In one embodiment, the bilinear interpolation process in step S522 ensures the quality of image scaling. For a pixel with coordinates (x, y) in the target image, its corresponding floating-point coordinates in the original image are (x / 1.56, y / 1.56). Assuming the calculated coordinates are (15.3, 28.7), interpolation is performed using the grayscale values ​​of four pixels (15, 28), (16, 28), (15, 29), and (16, 29) in the original image. Based on distance weights, the horizontal weights are 0.3 and 0.7, and the vertical weights are 0.7 and 0.3. The grayscale value of the target pixel is obtained by weighted averaging. This interpolation method preserves the edge sharpness and texture details of the ultrasound image, avoiding the jagged effect that may occur with simple scaling.

[0176] It should be noted that the grayscale mapping adjustment process in step S523 fully considers the medical diagnostic needs of ultrasound imaging. The grayscale information of ultrasound images is directly related to the identification of tissue structures and the detection of lesions; therefore, grayscale mapping must maintain the accuracy of diagnostic information. When the detected image dynamic range is 20 to 230, while the screen display range is 0 to 255, a linear stretching mapping formula is used to redistribute the original grayscale values ​​across the entire screen. The mapped image contrast is enhanced, and tissue boundaries are clearer, which is beneficial for doctors to make accurate diagnoses. At the same time, the mapping process maintains the relative grayscale differences between different tissues, ensuring that the medical value of the ultrasound image is not affected.

[0177] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A portable ultrasound imaging method based on multi-angle plane wave and neural network, characterized in that, The method comprises the following steps: acquiring multi-angle plane wave raw data collected by an ultrasonic probe; screening optimized angle combination data from the multi-angle plane wave raw data through a first convolutional neural network model; transmitting the optimized angle combination data to a terminal device; performing beam synthesis and image processing on the optimized angle combination data through a second convolutional neural network model in the terminal device to generate an ultrasonic image; displaying the ultrasonic image through the terminal device.

2. The method of claim 1, wherein, The step of screening the optimized angle combination data from the multi-angle plane wave raw data through the first convolutional neural network model comprises the following steps: acquiring signal-to-noise ratio information of the multi-angle plane wave raw data; determining imaging correlation of angle data according to the signal-to-noise ratio information; judging angle data satisfying a preset correlation threshold according to the imaging correlation; generating the optimized angle combination data according to the angle data satisfying the preset correlation threshold.

3. The method of claim 1, wherein, The step of transmitting the optimized angle combination data to the terminal device comprises the following steps: acquiring a compression format of the optimized angle combination data; transmitting the optimized angle combination data in the compression format to the terminal device through wired or wireless communication; decompressing the optimized angle combination data in the compression format in the terminal device to obtain the transmitted optimized angle combination data.

4. The method of claim 1, wherein, The step of performing beam synthesis and image processing on the optimized angle combination data through the second convolutional neural network model in the terminal device to generate an ultrasonic image comprises the following steps: acquiring the optimized angle combination data; performing an adaptive beam synthesis algorithm through the second convolutional neural network model to generate an initial ultrasonic image; performing signal denoising processing on the initial ultrasonic image to obtain a denoised image; performing image enhancement processing on the denoised image to generate the ultrasonic image.

5. The method of claim 4, wherein, The step of performing signal denoising processing on the initial ultrasonic image to obtain a denoised image comprises the following steps: acquiring pixel distribution characteristics of the initial ultrasonic image; determining noise distribution by processing the initial ultrasonic image through a filtering algorithm based on deep learning; generating the denoised image by removing noise interference in the initial ultrasonic image according to the noise distribution.

6. The method of claim 4, wherein, The step of performing image enhancement processing on the denoised image to generate the ultrasonic image comprises the following steps: acquiring resolution information of the denoised image; determining enhancement characteristics by processing the denoised image through a super-resolution algorithm based on deep learning; adjusting contrast and definition of the denoised image according to the enhancement characteristics; generating the ultrasonic image.

7. The method of claim 1, wherein, The step of displaying the ultrasonic image through the terminal device comprises the following steps: acquiring display parameters of the ultrasonic image; adjusting display proportions of the ultrasonic image according to the display parameters; presenting the adjusted ultrasonic image on a display interface of the terminal device.

Images