Portable 4D ultrasonic distributed beam and AI joint reconstruction method, device and program product

Abstract

The invention relates to the field of intelligent medical treatment, in particular to a portable 4D ultrasonic distributed beam and AI joint reconstruction method and device and a program product. Comprising the steps that N ultrasonic echo signals are obtained, and N is a natural number larger than 1; carrying out delay correction calculation on the N paths of echo signals and then carrying out weighted stacking to obtain a distributed beam; layering the distributed beam to obtain a display data stream and a reconstruction data stream; carrying out image reconstruction on the reconstructed data stream and then carrying out image target region segmentation to obtain a segmentation mask; performing geometric parameter calculation on the segmentation mask to obtain ultrasonic target area data; obtaining an ultrasonic display image through the display data stream; and superposing and fusing the ultrasonic display image and the ultrasonic target area data to obtain a final ultrasonic image. The application has good clinical value.

Description

Technical Field

[0001] This application relates to the field of intelligent healthcare, specifically to a portable 4D ultrasound distributed beam and AI-based joint reconstruction method, device, and program product. Background Technology

[0002] Existing portable ultrasound devices are mostly based on two-dimensional imaging. Limited by the number of channels, power consumption, and bandwidth, they struggle to achieve stable real-time reconstruction of 4D volumetric data and lack structured 4D dynamic analysis capabilities. Furthermore, ultrasound devices commonly employ Delay-And-Sum (DAS) algorithms for beamforming multi-channel echo signals. This approach only calculates, corrects, and superimposes the delays of multiple channels to obtain a single imaging signal or ultrasound image. However, beamforming is a single-node process, unsuitable for resource-constrained portable system architectures, thus limiting the clinical application value of portable 4D ultrasound. In addition, existing portable ultrasound imaging technologies focus solely on the beamforming process itself, lacking intelligent evaluation and feedback mechanisms for image quality. Imaging results are prone to high sidelobes and strong artifacts, and there are no quality quantification standards. The imaging, display, and post-processing workflows are executed synchronously and serially, and AI inference can easily block real-time display, failing to meet the real-time clinical requirements of portable devices. Summary of the Invention

[0003] To address the above problems, this invention provides a portable 4D ultrasound distributed beamforming and AI-based joint reconstruction method, specifically including: Acquire N echo signals from ultrasound, where N is a natural number greater than 1; After performing delay correction calculations on the N echo signals, a weighted superposition is performed to obtain a distributed beam; The distributed beam is layered to obtain a display data stream and a reconstruction data stream; After image reconstruction of the reconstructed data stream, image target region segmentation is performed to obtain a segmentation mask; Geometric parameters of the segmented mask are calculated to obtain ultrasonic target region data; Ultrasonic display imaging is obtained through the displayed data stream; The ultrasound display imaging and the ultrasound target area data are superimposed and fused to obtain the final ultrasound image.

[0004] Optionally, the geometric parameters include any one or more of the following: area, volume, and rate of change over time; ultrasonic target region data is generated by segmenting the edge contour of the mask and the geometric parameters, the ultrasonic display imaging is a real-time ultrasonic image, and the ultrasonic target region data is superimposed and fused into the real-time ultrasonic image to obtain the final ultrasonic image.

[0005] Optionally, the time change rate is the dynamic rate of change of the target quantization parameters of the target area with the ultrasound imaging time. The calculation of the time change rate includes: acquiring continuous segmentation masks, calculating the target quantization parameters of each mask frame, calculating the instantaneous time change rate of two adjacent frames, and then calculating the average rate within a T-frame period using the instantaneous time change rate to obtain the time change rate, where T is a natural number greater than 1.

[0006] Optionally, the delay correction calculation includes delay calculation and delay correction. After performing delay calculation on the N echo signals, each signal is corrected and aligned with the time domain data to obtain N corrected data. The N corrected data are then distributed and superimposed on the probe side and / or FPGA and / or edge nodes to obtain a distributed beam.

[0007] Optionally, the display data stream includes any one or more of the following: envelope, grayscale data, and real-time ultrasound imaging is performed using envelope and / or grayscale data to obtain ultrasound display imaging; the reconstruction data stream includes any one or more of the following: IQ data, Scanline data, volume data, and image reconstruction and image segmentation are performed using IQ data and / or Scanline data and / or volume data to obtain a segmentation mask.

[0008] Optionally, the segmentation is performed by segmenting the target region using a segmentation model to obtain a segmentation mask. The segmentation model can be any one or more of the following: YOLO series, SAM 3, U-Net, CMaP-SAM, RepSegNet; or the segmentation model can be performed by segmenting the target region using a sequentially connected encoding module, temporal attention fusion module, and decoding module to obtain a segmentation mask. The encoding module includes L encoding layers, each encoding layer including a depthwise separable convolutional layer and a channel attention compression layer. The input data of the segmentation model is encoded by the depthwise separable convolutional layer and the channel attention compression layer, the target features are extracted by the temporal attention fusion module, and then the segmentation mask is obtained by decoding using the decoding model.

[0009] Optionally, the segmentation model further includes a quality adaptive module. The data is first filtered by the quality adaptive module before being input into the segmentation model for segmentation. The quality adaptive module judges the quality of the input reconstructed data. When the quality score is less than a first threshold, it prompts a rescan. When the quality score is greater than or equal to the first threshold and less than a second threshold, the input reconstructed data is input into the artifact suppression module, where features are extracted through dilated convolution and then input into the encoding module of the segmentation model. When the quality score is greater than or equal to the second threshold, the input reconstructed data is input into the fast inference branch, and the output features are input into the encoding module of the segmentation model.

[0010] The purpose of this invention is to provide a computer program product that includes a computer program or instructions, which are executed by a processor to implement the above-described portable 4D ultrasound distributed beamforming and AI-based joint reconstruction method.

[0011] The purpose of this invention is to provide a computer device comprising a memory, a processor, and a computer program or instructions stored in the memory, wherein the computer program or instructions are executed by the processor to implement the above-described portable 4D ultrasound distributed beamforming and AI joint reconstruction method.

[0012] The purpose of this invention is to provide a computer-readable storage medium storing a computer program or instructions thereon, which is executed by a processor to implement the above-described portable 4D ultrasound distributed beamforming and AI-based joint reconstruction method.

[0013] Advantages of this invention: 1. This invention is designed for portable ultrasound architectures and employs a distributed beamforming design. The delay superposition algorithm is executed at the probe side / FPGA / edge node, which is adapted to the resource-constrained architecture of portable devices. Unlike the single-node Matlab simulation of existing technologies, this invention provides a solid foundation for portable ultrasound devices to meet the requirements of real-time and high-quality imaging.

[0014] 2. This invention proposes a dual-link parallel scheduling mechanism for asynchronous queues, which decouples the display link from the inference link (the display data stream is used for display imaging, and the reconstruction data stream is used for AI inference to obtain the target contour and geometric parameters). The display link has the highest priority, avoiding the blocking of real-time display by AI inference, and realizing the parallel execution of real-time 4D imaging and intelligent analysis under portable conditions.

[0015] 3. This invention introduces an AI quality gating and segmentation measurement network to achieve quantitative evaluation of imaging quality and automatic 4D analysis of target structures, solving the problems of existing technologies lacking intelligent quality control and quantitative measurement capabilities, and improving imaging quality stability and measurement consistency.

[0016] 4. This invention enables the overlay of inference results with the displayed image on a UI layer, enriching the information dimensions of the imaging results without modifying the original reconstruction process, and providing a more comprehensive reference for clinical diagnosis.

[0017] 5. This invention treats traditional DAS beamforming as only a preliminary basic step. The core innovation is distributed beamforming + asynchronous link + AI quality gating + AI segmentation measurement + result overlay display, forming a closed-loop, decision-making, and feedback-enabled 4D intelligent ultrasound method system. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the distributed beamforming and AI-based joint reconstruction method for portable 4D ultrasound provided in an embodiment of the present invention. Figure 2 A schematic diagram of a portable 4D ultrasound distributed beam and AI joint reconstruction system provided in an embodiment of the present invention; Figure 3 A schematic diagram of a computer device provided in an embodiment of the present invention; Figure 4 This is a diagram illustrating the architecture of a portable 4D ultrasound system provided in an embodiment of the present invention. Figure 5 This is a diagram of a distributed beamforming module provided in an embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0021] In some of the processes described in the specification, claims, and accompanying drawings of this invention, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as S101, S102, etc., are merely used to distinguish different operations and do not represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0022] Figure 1 A schematic diagram of the distributed beamforming and AI-based joint reconstruction method for portable 4D ultrasound provided in this embodiment of the invention is shown, specifically including: S1: Acquire N echo signals from ultrasound, where N is a natural number greater than 1; In one specific embodiment, a 32-channel phased array ultrasound probe is used to acquire raw radio frequency echo signals from the human heart region. Each channel has a sampling frequency of 10MHz, and the number of time-series sampling points is 1024, resulting in 32 raw signals. .

[0023] In one embodiment, the number of echo signal data channels is not less than 16, and preferably 32.

[0024] S2: After performing delay correction calculations on the N echo signals, weighted superposition is performed to obtain a distributed beam; In one embodiment, the delay correction calculation includes delay calculation and delay correction. After performing delay calculation on N echo signals, each signal is corrected and aligned with time-domain data to obtain N corrected data. The N corrected data are then distributed and superimposed on the probe side and / or FPGA and / or edge nodes to obtain a distributed beam.

[0025] In one specific embodiment, a dynamic delay superposition algorithm is executed on the FPGA to calculate the propagation delay of each array element based on the pixel spatial coordinates of the cardiac imaging region. The delay is quickly obtained by pre-calculating a delay lookup table. After delay correction is performed on the 32-channel signals, the Hanning window is used as the weighting coefficient. Perform distributed overlay to obtain focused intermediate data.

[0026] In one embodiment, the distributed delay processing step includes calculating the propagation delay based on the imaging pixel spatial coordinates, using any one or a combination of geometric propagation path resolution, pre-calculated delay lookup table, and dynamic aperture adaptive correction; the formula for calculating the propagation delay is as follows: , For the spatial position of the array element, The location of the target imaging point. This represents the speed of ultrasonic propagation.

[0027] The beamforming step is performed using a weighted delay superposition method, with the weighting coefficients used for sidelobe suppression.

[0028] In one specific embodiment, a dynamic delay superposition algorithm based on a phased array geometric model is used to process the acquired raw signal: (1) Delay calculation: For each imaging pixel, the channel delay is calculated based on its spatial coordinates using any one or a combination of analytical delay calculation based on geometric propagation path, pre-computed delay lookup table (LUT), and adaptive delay correction based on dynamic aperture. (2) Delay correction: Perform delay compensation on each original signal. To achieve data time domain alignment; (3) Distributed weighted superposition: The delayed-corrected signal is weighted and superimposed at the probe side / FPGA / edge node to obtain the focused intermediate data and complete the distributed beamforming.

[0029] Distributed delay superposition: The delayed-corrected signals are partially superimposed at the probe side / FPGA / edge node to obtain... in These are the element weighting coefficients used for sidelobe suppression.

[0030] S3: The distributed beam is layered to obtain a display data stream and a reconstruction data stream; In one embodiment, the display data stream includes any one or more of the following: envelope and grayscale data, and real-time ultrasound imaging is performed using envelope and / or grayscale data to obtain ultrasound display imaging; the reconstruction data stream includes any one or more of the following: IQ data, Scanline data, and volume data, and image reconstruction and image segmentation are performed using IQ data and / or Scanline data and / or volume data to obtain a segmentation mask.

[0031] In one specific embodiment, the intermediate data is used to generate a display data stream (grayscale data) and a reconstruction data stream (IQ data) with a resolution of 512×512.

[0032] In one embodiment, the beamforming data is divided into a display data stream and a reconstruction data stream. The display data stream is envelope / grayscale data, and the reconstruction data stream is IQ / Scanline / volumetric data. The display data stream is used for real-time B / C / volumetric rendering and display, while the reconstruction data stream is used for subsequent AI quality assessment and segmentation measurement. The display data stream and the reconstruction data stream enter different asynchronous queues, with the display queue having a higher scheduling priority than the inference queue. The scanline data is the raw sequence data of IQ data along the ultrasound scanning direction (without grayscale mapping), while the voxel data is a three-dimensional data volume formed by stacking consecutive B-mode frames or IQ data along the time dimension. The B-mode frame is generated by taking the beamforming IQ data as input and performing three steps of calculation: Hilbert transform envelope detection, logarithmic compression, and grayscale mapping. The multi-form design of the reconstruction data stream provides a flexible choice for quality gating: rapid quality control (B-mode frames), high-precision quality control (IQ data), and 4D dynamic quality control (voxel data), adapting to different clinical scenario needs while avoiding data redundancy and fitting the computing power constraints of portable devices.

[0033] The display queue supports real-time refresh, while the inference queue allows for latency and supports batch processing.

[0034] In one specific embodiment, the intermediate data after beamforming is divided into two types of data streams to achieve layered processing of different streams originating from the same source: (1) Display data stream: generated by the low-latency, low-computational-complexity reconstruction path, which is envelope / grayscale data and is directly used for real-time B / C / volume drawing and display; (2) Reconstruct the data stream: retain higher spatial and temporal consistency for IQ / Scanline / volume data, which will be used for subsequent AI quality assessment and segmentation measurement.

[0035] The display data stream and the reconstruction data stream are sent to separate asynchronous queues to achieve parallel scheduling of the links and ensure clinical real-time performance. (1) Display queue: The scheduling priority is the highest, and the display data stream is refreshed in real time to achieve low-latency output of 4D ultrasound images without waiting for AI inference results; (2) Inference queue: Allows for delays, supports batch processing, caches the reconstructed data stream, and provides data support for subsequent AI inference.

[0036] S4: After image reconstruction of the reconstructed data stream, image target region segmentation is performed to obtain a segmentation mask; geometric parameters of the segmentation mask are calculated to obtain ultrasonic target region data; In one embodiment, the geometric parameters include any one or more of the following: area, volume, and rate of change over time; ultrasonic target region data is generated by segmenting the edge contour of the mask and the geometric parameters, the ultrasonic display imaging is a real-time ultrasonic image, and the ultrasonic target region data is superimposed and fused into the real-time ultrasonic image to obtain the final ultrasonic image.

[0037] The final ultrasound image includes a real-time ultrasound image, as well as a layer showing the edge contour of the target area and the area, volume, or time rate of change parameters of the target area (lesion) superimposed on the real-time ultrasound image.

[0038] In one embodiment, the time change rate is the dynamic rate of change of the target quantization parameters of the target region with the ultrasound imaging time. The calculation of the time change rate includes: acquiring continuous segmentation masks, calculating the target quantization parameters of each mask frame, calculating the instantaneous time change rate of two adjacent frames, and then calculating the average rate within a T-frame period using the instantaneous time change rate to obtain the time change rate, where T is a natural number greater than 1.

[0039] The targets include any one or more of the following: heart ventricles, liver, lesions, and other targets of interest; The target quantization parameters include any one or more of the following: volume, area, and length.

[0040] In one embodiment, the segmentation is performed by segmenting the target region using a segmentation model to obtain a segmentation mask. The segmentation model can be any one or more of the following: YOLO series, SAM 3, U-Net, CMaP-SAM, RepSegNet.

[0041] In another embodiment, the segmentation model segments the target region to obtain a segmentation mask by sequentially connecting an encoding module, a temporal attention fusion module, and a decoding module. The encoding module includes L encoding layers, each of which includes a depthwise separable convolutional layer and a channel attention compression layer. The input data of the segmentation model is encoded by the depthwise separable convolutional layer and the channel attention compression layer, and then the target features are extracted by the temporal attention fusion module and decoded by the decoding model to obtain the segmentation mask. L is a natural number greater than 1.

[0042] Optionally, the encoding module and the decoding module each include L encoding layers and L decoding layers. Each encoding layer includes a depthwise separable convolutional layer and a channel attention layer in sequence, and each decoding layer includes a depthwise separable convolutional layer and an upsampling layer in sequence. The input data of the encoding module is output through the L encoding layers and L decoding layers to obtain the segmentation result.

[0043] Optionally, the temporal attention fusion module is connected in a skip connection between the encoding module and the decoding module, and each encoding layer is connected to the decoding layer after passing through the temporal attention fusion module via a skip connection; or the temporal attention fusion module is connected to the last encoding layer, and the last encoding layer inputs features to the temporal attention fusion module for temporal feature extraction, and then the temporal features are input to the first decoding layer.

[0044] In one embodiment, the segmentation model further includes a quality adaptive module. The data is first filtered by the quality adaptive module before being input into the segmentation model for segmentation. The quality adaptive module judges the quality of the input reconstructed data. When the quality score is less than a first threshold, it prompts a rescan. When the quality score is greater than or equal to the first threshold and less than a second threshold, the input reconstructed data is input into the artifact suppression module, where features are extracted through dilated convolution and then input into the encoding module of the segmentation model. When the quality score is greater than or equal to the second threshold, the input reconstructed data is input into the fast inference branch, and the output features are input into the encoding module of the segmentation model.

[0045] In one specific embodiment, the fast inference branch is a lightweight inference path designed for high-quality 4D ultrasound data (S≥T2) in the segmentation model. Its core is to improve inference speed while ensuring segmentation accuracy by simplifying feature processing steps and skipping redundant artifact suppression steps, thus adapting to the real-time requirements of portable devices. The specific processing flow is lightweight feature mapping without additional feature enhancement + channel dimension adaptation, with no complex convolution operations throughout. The single-frame processing time is reduced by ≥50% compared to the artifact suppression branch. The specific steps are as follows: Input reception: Receive high-quality two-dimensional sequence / three-dimensional volume data (standardized and size-uniform feature maps, denoted as F, with dimensions H×W×C) after quality gating and screening. Input reception: Receive high-quality two-dimensional sequence / three-dimensional volume data (standardized and size-uniform feature maps, denoted as F, with dimensions H×W×C) after quality gating and screening. Batch Normalization (BN layer): Performs batch normalization on the feature map after 1×1 convolution to eliminate data distribution differences and improve inference stability without additional regularization operations; Direct output to the encoding module: The processed feature map (whose size perfectly matches the input of the first layer of the encoding module) is directly fed into the first depthwise separable convolutional layer of the encoding module of the segmentation model, and then enters the subsequent standard encoding-decoding segmentation process.

[0046] The difference between the fast inference branch and the artifact suppression branch is that the fast inference branch has no dilated convolutions, no edge feature enhancements, and no artifact suppression-related operations. It skips all feature repair steps for low / medium quality data, retains the core steps of channel dimension adaptation, ensures inference efficiency for high-quality data, and has a single-frame inference time of ≤30ms, which matches the inference speed of AI quality gating networks. The convolutional layers have no activation functions, avoiding the computational consumption caused by nonlinear transformations, while ensuring the feature integrity of high-quality data.

[0047] In one embodiment, the method further includes quality assessment, wherein the reconstructed data stream first obtains a quality score through quality assessment, performs a first screening based on the quality score, and then inputs the first-screened reconstructed data into a segmentation model, where a quality adaptive module performs a second screening, and the second-screened data is then input into an encoding module.

[0048] In one specific embodiment, the quality assessment model is as follows: This invention adopts a lightweight encoder-regressor architecture, which is adapted to the hardware requirements of portable devices with limited resources, while also taking into account the accurate extraction of ultrasound image quality features. The overall architecture is divided into three core modules, from top to bottom: Input adaptation module: Receives the B-mode frame / IQ data / voxel data I (reconstructed data stream, B-mode frame is obtained by calculation or reconstruction of IQ data) after beamforming, performs standardized preprocessing on the input data (normalized to the [0,1] interval), and uniform size (scaled to 224×224 pixels, voxel data is taken from continuous frame slices and stitched together), without the need for complex data augmentation, reducing computational overhead and adapting to the low power consumption requirements of portable devices; Feature Extraction Encoder: Built on a lightweight convolutional neural network (CNN), it uses depthwise separable convolutions instead of traditional convolutions to reduce the total number of parameters and computational cost. The encoder contains 4-6 convolutional blocks, each consisting of a depthwise separable convolutional layer, a batch normalization layer (BN layer), and a ReLU activation function. Finally, a global average pooling layer outputs a 256-dimensional feature vector. This module can accurately extract the core quality-related features of ultrasound images, including key features such as sharpness, contrast, artifact intensity, and structural integrity. Unlike general image quality assessment networks, it is specifically optimized for feature extraction of quality defects such as artifacts and side lobes unique to ultrasound images. The quality regression output unit consists of two fully connected layers. The first layer compresses the 256-dimensional feature vector output by the encoder to 64 dimensions, and the second layer maps the 64-dimensional features to a single normalized score value. Simultaneously, a softmax layer is connected in parallel to output three quality level labels: low, medium, and high. The regressor uses an L2 loss function constraint to ensure the stability and continuity of the score output.

[0049] The total number of network parameters is controlled within 1 million, and the single-frame processing time during inference is ≤20ms, which is fully compatible with the computing power of FPGA / SoC hardware platform and meets the real-time requirements of portable 4D ultrasound equipment.

[0050] Quality assessment model training process: A two-stage pre-training method of "self-supervised pre-training + supervised fine-tuning" is adopted. Combining the characteristics of ultrasound medical imaging, it makes full use of unlabeled ultrasound data to reduce annotation costs, while supervised fine-tuning improves the accuracy of quality assessment and aligns with the quality judgment standards of clinical ultrasound imaging. The specific steps are as follows: Phase 1: Self-supervised pre-training (pre-training on unlabeled data); Pre-training dataset: Collect 10,000+ unlabeled ultrasound images (covering ultrasound images of different parts of the human body, different imaging conditions, and different quality levels, including normal imaging, artifact interference, insufficient clarity and other scenarios, covering common imaging conditions of portable ultrasound equipment), and supplement unlabeled samples from public ultrasound datasets to improve the generalization ability of the pre-trained model and avoid overfitting problems caused by a single dataset. Pre-training task: Masked Image Modeling (MIM) is used as a self-supervised pre-training task. 10%-20% of the input ultrasound image is randomly masked (the masked area is a rectangle or irregular block to simulate artifacts, signal loss and other scenarios in ultrasound images). The network is then forced to recover the image information of the masked area through the features of the unmasked area. This forces the encoder to learn the structural features and quality-related features of the ultrasound image, thus strengthening its feature extraction capabilities. Pre-training parameters: 50 pre-training rounds, batch size set to 32, Adam optimizer used, initial learning rate 1e-4, learning rate decayed to 1 / 10 of the original every 10 rounds, mean squared error (MSE) loss function used to constrain the difference between the restored image and the original image in the mask region; Phase 2: Supervised fine-tuning (fine-tuning with labeled data); Fine-tuning the dataset: Collected 1000+ manually annotated ultrasound images. Each image was annotated by two or more senior ultrasound physicians according to clinical diagnostic criteria with a normalized quality score (0-1 range) and quality level label (low / medium / high). Samples with inconsistent annotations were reviewed and determined by three physicians to ensure the accuracy and authority of the annotated data. Fine-tuning process: Freeze 80% of the encoder parameters obtained from the first stage of self-supervised pre-training, and only fine-tune the top 20% of the convolutional layers of the encoder and the fully connected layers of the regressor to avoid overfitting of the model due to excessive fine-tuning, while retaining the general features learned in the pre-training stage. Fine-tuning parameters: The number of fine-tuning rounds is 20, the batch size is set to 16, the optimizer is Adam optimizer, the initial learning rate is 1e-5, and the loss function is a joint loss of "MSE loss + cross-entropy loss". MSE loss constrains the difference between regression score and manually labeled score, and cross-entropy loss constrains the classification accuracy of quality level labels. Pre-trained model saving: After fine-tuning, the model performance is verified using a validation set (accounting for 20% of the fine-tuning dataset). The model with the smallest MSE loss and the highest classification accuracy on the validation set is selected as the final pre-trained model. The model parameters are saved to the FPGA / SoC hardware platform for subsequent real-time quality assessment inference.

[0051] Additional notes: The pre-training process can be completed on a cloud server or PC. The trained lightweight model parameters can be ported to portable devices without performing the pre-training process on the portable devices, further reducing the hardware burden on the portable devices. At the same time, the pre-training method can be tailored to specific clinical ultrasound application scenarios (such as ultrasound imaging of different parts of the heart, abdomen, etc.), and the pre-training dataset and fine-tuning strategy can be adjusted to improve the scenario adaptability of quality assessment, which meets the needs of the portable 4D ultrasound multi-scenario clinical application of this invention. Furthermore, it references the mature application experience of self-supervised pre-training in the field of medical imaging to ensure the feasibility and effectiveness of the pre-training method.

[0052] The core data processing continues: Input data I is input into the pre-trained ultrasound image quality assessment network, and feature extraction and quality quantification are completed through three steps: input adaptation, feature extraction, and regression output.

[0053] Output data: Normalized quality score Quality grade labels (low / medium / high); Decision processing: Based on preset thresholds , Execute gated logic — S< The system will output a message prompting you to rescan. ≤S< Automatically adjusts imaging parameters (gain, focus, etc.); S≥ This allows for subsequent 4D reconstruction and segmentation measurement steps; Results data: Quality assessment decision instructions, and a high-quality reconstructed data stream V that can be used for subsequent analysis.

[0054] In one embodiment, training the quality assessment model includes: Obtain an unlabeled ultrasound image dataset; After random region masking, the unlabeled ultrasound image dataset is fed into the encoder for self-supervised training. The difference between the restored image generated by the decoder and the original image is compared to calculate the loss function until the loss function remains unchanged, thus obtaining the first quality assessment model. A dataset of ultrasound images labeled with quality levels is obtained and fed into the first quality assessment model for training. The predicted results are compared with the actual results to calculate the loss function until the loss function remains unchanged, thus obtaining the quality assessment model.

[0055] The random region masking refers to masking regions that are rectangular or irregular blocks, simulating artifacts or signal loss; the unlabeled ultrasound images include ultrasound images of different parts of the human body, under different imaging conditions, and of different quality levels.

[0056] In one specific embodiment, conventional ultrasound segmentation models (such as general U-Net and V-Net) have three major adaptation defects and cannot be directly applied to the portable 4D ultrasound scenario of this invention: ① The model has a large number of parameters (usually over a million), slow inference speed, and is not suitable for the low power consumption and real-time requirements of FPGA / SoC hardware platforms; ② It does not adapt to the characteristics of ultrasound image quality fluctuations, is sensitive to artifacts, low contrast, and sidelobe interference, and has unstable segmentation accuracy; ③ It is only designed for static 2D ultrasound images, lacks the ability to learn the temporal correlation of 4D dynamic sequences, and cannot adapt to the dynamic structural analysis requirements of 4D ultrasound.

[0057] To address this, this invention proposes an improved segmentation model. Based on the conventional U-Net, it replaces traditional standard convolutions with depthwise separable convolutions, removes redundant pooling and fully connected layers, and adds a channel attention compression module. This keeps the total number of model parameters below 800,000, with a single-frame inference time ≤30ms, matching the inference speed of AI quality-gated networks and fully adapting to the computing power of FPGA / SoC hardware platforms. This solves the problems of high power consumption and slow inference in conventional models, aligning with the core constraints of portable devices. Secondly, based on the quality score S output by the AI ​​quality-gated step, an adaptive inference branch is designed: when the quality score is 0.6 ≤ S < 0.8 (medium quality), the artifact suppression branch is automatically activated, enhancing edge feature extraction through dilated convolutions and suppressing artifacts and sidelobe interference; when the quality score S ≥ 0.8 (high quality), a fast inference branch is enabled to improve segmentation efficiency; when the quality score S < 0.6, segmentation is directly rejected (triggered by the quality-gated step for rescanning), avoiding invalid segmentation results. This mechanism addresses the shortcomings of conventional segmentation models, such as sensitivity to ultrasound image quality and unstable segmentation accuracy, and forms a collaborative closed loop with the quality gating process of this invention.

[0058] Furthermore, a temporal attention fusion module is added between the encoder and decoder ends of the segmentation network. It takes 3-5 consecutive frames of 4D ultrasound sequence data as input, learns the dynamic motion patterns and temporal correlations of target structures (such as the heart and liver), and captures the structural change features between adjacent frames through temporal convolution. This avoids the inter-frame segmentation misalignment problem caused by conventional static segmentation models, and ensures the continuity and consistency of segmentation results in 4D dynamic analysis, which is in line with the core application scenario of 4D ultrasound in this invention.

[0059] Finally, combining the clinical diagnostic standards of ultrasound medicine, morphological constraints of organ structures (such as the morphological proportions of the heart ventricles and the continuity constraints of blood vessels) are embedded in the loss function of the model. A joint loss function of "cross-entropy loss + Dice loss + morphological constraint loss" is adopted to guide the model to learn organ structural features that conform to clinical cognition. This is different from the general loss function design of conventional segmentation models, improves the clinical adaptability of segmentation results, and reduces segmentation errors (average error ≤3%, consistent with the measurement accuracy in the specific embodiment).

[0060] In one specific embodiment, a temporal attention fusion module is connected in series between the output of the last encoding layer and the input of the first decoding layer of the segmentation model, adding temporal feature extraction capability to the main connection path of encoding-decoding. (Spatial feature map output from the last encoding layer → temporal attention fusion module → extracting inter-frame temporal correlation features → outputting a feature map with fused spatiotemporal features → sent to the first decoding layer for upsampling and decoding). Deployed only in the core connection layer, the number of parameters increases by ≤5%, and the single-frame inference time increases by ≤2ms, fully adapting to the computing power constraints of FPGA / SoC and meeting the real-time requirements of portable devices. Applicable scenarios: routine clinical 4D ultrasound examinations (such as abdominal and obstetric basic screening). In another embodiment, the temporal attention fusion module is embedded in the skip connection path between each coding layer and the corresponding decoding layer, while retaining the temporal feature extraction capability of the main connection layer (spatial feature map output from the nth coding layer → temporal attention fusion module in the corresponding skip connection → extracting inter-frame temporal features at this scale → fusing with the feature map upsampled by the decoding layer → sending to the next decoding layer). Temporal correlation information is incorporated into multi-scale spatial features, reducing inter-frame segmentation misalignment error by ≥40% and improving the accuracy of 4D dynamic analysis. Applicable scenarios: high-precision 4D ultrasound functional assessment (such as heartbeat, fetal cardiac dynamic analysis).

[0061] The temporal attention fusion module is a lightweight spatiotemporal feature extraction structure with an overall parameter count of ≤50,000 and a single-frame processing time of ≤3ms. It consists of a frame feature stacking layer, a lightweight temporal convolutional layer, and a spatiotemporal attention weighting layer. The frame feature stacking layer stacks spatial feature maps of 3-5 consecutive frames along the temporal dimension to form a spatiotemporal feature cube. The lightweight temporal convolutional layer uses a 1×1×3 depthwise separable temporal convolutional kernel to extract inter-frame temporal correlation features. The spatiotemporal attention weighting layer uses dual weighting of temporal attention and channel attention to filter effective spatiotemporal features and pool them along the temporal dimension, outputting a single-frame spatial feature map that fuses temporal correlation information, which is then sent to the decoding module for further processing.

[0062] The frame feature stacking layer stacks the same-scale spatial feature maps of K consecutive frames of 4D ultrasound input along the "time dimension (T dimension)" to obtain a spatiotemporal feature cube, completing the format conversion from "single-frame spatial features" to "multi-frame spatiotemporal features," providing a foundation for subsequent temporal feature extraction. The lightweight temporal convolutional layer uses a 1×1×3 depthwise separable temporal convolutional kernel to convolve the spatiotemporal feature cube. After batch normalization (BN layer) and ReLU6 activation function (adapting to low-precision calculations in embedded devices), overfitting is suppressed and nonlinear expression is enhanced, resulting in a spatiotemporal feature cube that integrates inter-frame temporal correlations (extracting the structural motion patterns of adjacent frames). (e.g., edge motion features of ventricular systole / diastole); the spatiotemporal attention weighted layer combines temporal attention and channel attention to weight and filter spatiotemporal features, highlighting effective temporal features and suppressing noise features. The temporal attention branch performs global temporal pooling on the spatiotemporal feature cube, while the channel attention branch performs global spatial pooling on the time-weighted feature cube. The double-weighted spatiotemporal feature cube is then averaged along the time dimension, completing the transformation from "spatiotemporal features" to "spatial features fused with temporal sequence". This results in a single-frame spatial feature map that fuses the temporal correlation information of K consecutive frames, which is then directly sent to the decoding module for further processing.

[0063] In one embodiment, the lightweight architecture of the segmentation network achieves feature dimension simplification and effective feature selection through a channel attention compression layer. This module is an original lightweight structure that integrates the dual functions of channel attention weighting and feature compression. It is not an existing SE or CBAM module. The total number of parameters is ≤10,000. It includes a global channel pooling layer, a lightweight channel attention weighting layer, and a channel feature fusion and compression layer. It can reduce the number of model parameters and computational load without losing key features, adapt to the computing power constraints of portable devices, and suppress invalid channel features caused by ultrasound artifacts.

[0064] In one specific embodiment, the channel attention compression layer is a core component of the lightweight architecture of the segmentation network in this invention. It integrates two main functions: "channel attention weighting (filtering effective channel features)" and "feature dimension compression (reducing computational overhead)." Its core objective is to reduce the number of parameters and computational load of the segmentation model by filtering effective channels and compressing redundant channel dimensions without losing key features. This adapts to the low-power and real-time requirements of FPGA / SoC hardware platforms, while simultaneously improving the targeting of ultrasound image feature extraction (suppressing artifacts, sidelobes, and other invalid channel features). The channel attention compression layer includes a global channel pooling layer, a lightweight channel attention weighting layer, and a channel feature fusion and compression layer. This invention's channel attention compression layer is not a simple modification of existing SE and CBAM modules, but rather an original integrated "attention + compression" structure designed based on the computational constraints and image characteristics of portable 4D ultrasound. It solves the shortcomings of existing modules, such as large parameter counts and incompatibility with portable devices, and improves the efficiency and robustness of ultrasound image feature extraction through feature compression and attention filtering.

[0065] The module has a serial structure with no redundant fully connected layers. The number of parameters is controlled to within 10,000, and the processing time per frame is ≤1ms. It is fully adaptable to the computing power constraints of portable devices. It consists of three core sub-layers from top to bottom. The specific structure and execution logic are as follows: Sub-layer 1: Global channel pooling layer (feature compression pre-processing); Input: Spatial feature map output from the coding layer of the segmentation network (size is H×W×C, where H / W is the height / width of the feature map, and C is the number of channels); a. Processing logic: Global average pooling is used to average the spatial features of each channel, compressing the H×W×C feature map into a 1×1×C channel feature vector, which achieves both initial compression of feature dimensions and preservation of global feature information of each channel. b. Output: 1×1×C channel feature vectors, providing a simplified channel feature input for subsequent attention weighting.

[0066] Sub-layer 2: Lightweight channel attention weighting layer (core attention mechanism); Core design: Abandoning the dual fully connected layer structure of existing SE and CBAM modules, only a single 1×1 convolutional layer is used to map channel attention weights, significantly reducing the number of parameters; Processing logic: Input the 1×1×C channel feature vector into a 1×1 convolutional layer (the number of convolutional kernels is C / 2, which further compresses the channel dimension), and after passing through the Sigmoid activation function, output a 1×1×(C / 2) channel attention weight vector, with weight values ​​∈ [0,1], which is used to characterize the feature importance of each channel; Output: A 1×1×(C / 2) channel attention weight vector, which completes the selection and weight allocation of effective channels.

[0067] Sub-layer 3: Channel feature fusion and compression layer (final output adaptation); a. Processing logic: The H×W×C feature map output by the original coding layer is fused with the attention weight vector in a channel-dimensional weighted manner, retaining the effective channel features with higher weights. At the same time, the number of channels in the feature map is compressed from C to C / 2 through channel concatenation and 1×1 convolution, thus completing the final compression of the feature dimension. b. Output: A lightweight feature map of H×W×(C / 2) is directly fed into the next coding layer or the temporal attention fusion module, which reduces the computational load of subsequent processing while retaining key effective features.

[0068] The overall execution flow of the module is as follows: Encoding layer feature map → Global channel pooling (preliminary compression) → Lightweight attention weighting (selecting effective channels) → Channel fusion and compression (final lightweight output).

[0069] In one embodiment, the rate of change over time refers to the rate of change of the core quantitative parameters (such as volume, area, and length) of the target organ / structure (such as the ventricle of the heart, liver lesions, etc.) with the ultrasound imaging time in a 4D ultrasound dynamic sequence. It is used to quantify the dynamic motion law or morphological change trend of the target structure. It is the core analytical indicator that distinguishes 4D ultrasound from 3D / 2D ultrasound and meets the clinical needs for dynamic organ function assessment (such as heartbeat, fetal limb movement, etc.).

[0070] Based on the target structure mask output by the segmentation network (one mask per frame of temporal image), and combined with the 4D ultrasound imaging frame rate, the calculation is completed in three steps, fully adapting to the computing power requirements of portable devices. The calculation process can be efficiently executed on the FPGA / SoC. The specific steps are as follows: Step 1: Obtain basic data (based on segmentation results and 4D temporal data); - The segmentation network outputs a target structure mask consisting of T consecutive frames (T≥3 frames, adapted to 4D dynamic sequences, such as 5 consecutive frames at a frame rate of 10fps in the specific embodiment). The core quantization parameter value corresponding to each frame is calculated through the mask, denoted as... ( The frame number, ); - Core Quantization Parameters The selection can be made according to clinical needs, with priority given to volume (such as ventricular volume) and area (such as lesion cross-sectional area). Length (such as blood vessel diameter) can also be selected. All of these are calculated from the pixel information corresponding to the mask (e.g., volume = number of mask pixels × actual spatial volume corresponding to a single pixel, area = number of mask pixels × actual area corresponding to a single pixel). - Obtain 4D ultrasound imaging frame rate (Unit: fps, frames per second), calculate the time interval between two adjacent frames. (Unit: seconds), such as frame rate in a specific embodiment. ,but .

[0071] Step 2: Calculate the instantaneous rate of change (rate of change between single frames); - For two adjacent frames (the first) Frame and the (frames), instantaneous time change rate calculated using the finite difference method. The calculation formula is:

[0072] - illustrate: This indicates that the target parameter increases over time (e.g., the ventricular volume increases during diastole). This indicates that the target parameter decreases over time (e.g., the volume of the ventricle decreases during systole). This indicates that the target parameters have not changed significantly. This calculation method is simple and efficient, requires no complex computing power, and is suitable for portable devices.

[0073] Step 3: Calculate the average time change rate (average rate over multiple frame periods). - For a continuous T-frame dataset, calculate the average of all instantaneous rates of change over time to obtain the average rate of change over time. The calculation formula is:

[0074] - Note: The average rate of change over time is used to reflect the overall trend of change of the target structure over a dynamic period, such as the average rate of change of ventricular volume over one heartbeat cycle, providing a quantitative basis for clinical functional assessment. For more accurate trend analysis, the instantaneous rate of change can be smoothed (e.g., by moving average) to further reduce the impact of ultrasound image noise on the calculation results.

[0075] Additional explanation: The calculation of the time change rate is based on the accurate mask data output by the segmentation network of this invention, combined with the design of 4D temporal characteristics, which is different from the defects of conventional static segmentation models that lack dynamic analysis capabilities; the calculation process adopts lightweight algorithms, with a single cycle (T=5 frames) calculation time ≤10ms, which matches the speed of AI inference and display links, does not affect the real-time performance of portable devices, and the calculation results can be directly used as the core parameters of 4D analysis results, superimposed on the display image or included in the measurement report, which fits the integrated design of "real-time display + intelligent analysis" of this invention.

[0076] In one embodiment, quality assessment is performed using a quality assessment model to obtain a quality score. The quality assessment model includes an input adaptation module, a feature encoding module, and a quality regression module. The data stream is preprocessed by the input adaptation module and then input to the feature encoding module. The feature encoding module includes N convolutional blocks and a global average pooling layer. After encoding by the N convolutional blocks, the core quality features are output through the global average pooling layer. The core quality features are input to the quality regression module, where feature compression and scoring are performed through K fully connected layers, outputting a quality score and a classification result. N and K are natural numbers greater than 1.

[0077] Furthermore, the quality score is compared with a preset threshold. When the score is less than the first preset threshold, a rescan is prompted. When the score is greater than or equal to the first preset threshold and less than the second threshold, the imaging parameters are automatically adjusted. When the score is greater than or equal to the second preset threshold, a pass is indicated.

[0078] Optionally, each convolutional block in the quality assessment model includes a depthwise separable convolutional layer, a batch normalization layer, and an activation function layer. The input data of the convolutional block is output after passing through the depthwise separable convolutional layer, the batch normalization layer, and the activation layer in sequence.

[0079] In one specific embodiment, the IQ data in the inference queue (reconstruction data stream branch) is input into a pre-trained ultrasound image quality assessment network, which outputs a quality score Q. When Q≥0.8, the data is input into a cardiac structure segmentation CNN network, which outputs masks of structures such as the left ventricle, right ventricle, and myocardium. Based on the masks, parameters such as left ventricular ejection fraction and ventricular volume are calculated to generate a 4D dynamic analysis curve of the cardiac structure.

[0080] In one embodiment, the quality gating step includes inputting the reconstructed data stream into a pre-trained ultrasound image quality assessment network and outputting a normalized imaging quality score. The quality score Q is used to control whether to execute subsequent segmentation measurement steps. When Q ≥ a preset threshold, the segmentation measurement step is initiated; when Q < a preset threshold, a rescan prompt is output or the imaging parameters are automatically adjusted.

[0081] The segmentation measurement step includes inputting a quality-gated reconstructed data stream into a segmentation neural network, which outputs a target structure mask. This target structure mask is used to perform geometric measurements, obtaining parameters such as length, area, volume, and rate of change over time. The AI ​​inference results include a quality score, segmentation contour information, numerical measurement parameters, and 4D structural dynamic analysis results.

[0082] AI inference includes using reconstructed data streams for quality assessment, mask segmentation, geometric parameter measurement, and calculation.

[0083] In this invention, the AI quality gating, as the core quality control link, is not a redundant design, but a targeted improvement for the blank in the quality control of ultrasonic imaging in the prior art. The problems it needs to solve, the benefits it brings, and the essential differences from data preprocessing are as follows. Specifically, the necessity of it in the overall technical solution is clarified: 1) There is no quantitative standard for imaging quality, and the quality is uneven: Ultrasonic imaging is easily affected by artifacts, sidelobe interference, imaging parameter deviation, etc. The prior art cannot quantitatively evaluate the imaging quality, and cannot distinguish between "qualified images available for clinical analysis" and "inferior images without analysis value", which easily leads to subsequent segmentation measurement and diagnostic analysis based on inferior data, resulting in serious errors and misleading clinical judgments. 2) Ineffective data occupies hardware resources and affects real-time performance: The computing power and power consumption of portable devices are limited. In the prior art, regardless of the imaging quality, all data streams after beamforming will enter the subsequent AI inference and segmentation measurement links. The ineffective processing of inferior data will occupy a large amount of FPGA / SoC computing power, resulting in an increase in real-time display delay and AI inference lag, which violates the clinical real-time requirements of portable devices.

[0084] 3) There is no quality feedback mechanism and it cannot be actively optimized: The prior art can only passively output imaging results and cannot adjust imaging parameters (such as gain, focus) according to the imaging quality. When inferior imaging occurs, it can only rely on the doctor to manually rescan, which is cumbersome and cannot guarantee the quality after rescan, reducing the clinical examination efficiency.

[0085] 4) The accuracy of 4D dynamic analysis cannot be guaranteed: 4D ultrasound relies on the consistency of continuous frame data. Without quality control in the prior art, inferior frames may be mixed in continuous frames, resulting in frame misalignment and parameter deviation in subsequent calculation of time change rate and dynamic structure analysis, and unable to achieve accurate 4D dynamic evaluation.

[0086] Therefore, this invention quantifies the imaging quality through a quality control evaluation model, screens effective data, saves hardware resources, ensures real-time performance, rejects inferior data (S < T1) from entering subsequent high-computing-power-consuming links, and only performs simple parameter adjustment on medium-quality data (T1 ≤ S < T2) without unnecessary AI inference, greatly reducing the computing power occupancy of FPGA / SoC, ensuring the real-time performance of the display link and effective data inference, adapting to the resource constraints of portable devices, and ensuring the consistency and accuracy of 4D dynamic analysis.

[0087] S5: Obtain ultrasonic display imaging through the display data stream; In a specific embodiment, asynchronous queue scheduling: The display data stream is sent to the display queue to achieve real-time display of cardiac 4D ultrasound images at a frame rate of 30fps; the reconstruction data stream is sent to the inference queue for caching at a batch processing rate of 10fps.

[0088] S6: The ultrasound display imaging and ultrasound target area data are superimposed and fused to obtain the final ultrasound image.

[0089] In one embodiment, the inference results are overlaid onto the real-time 4D ultrasound image output by the display link in the form of layers, annotations, or parameters. This is a UI / system-level overlay, not a signal-level overlay.

[0090] The final output data includes real-time 4D ultrasound images, AI quality score results, segmentation contours and measurement parameters, and an overlay display interface.

[0091] In one specific embodiment, the cardiac structure segmentation contour, ejection fraction measurement value, and quality score Q are superimposed on the real-time displayed 4D cardiac ultrasound image in the form of a red layer, and the superimposed display is realized on the device. At the same time, the measurement parameters generate a standardized report and store it locally on the device, which can be uploaded to a cloud server for remote consultation.

[0092] In one specific embodiment, the architecture of the portable 4D ultrasound system of the present invention is as follows: Figure 4 As shown, the architecture consists of a Phased Array Probe → Analog Front-End + Analog-to-Digital Converter (AFE + ADC) → FPGA / SoC Core Processing Module → Beamforming + AI Joint Reconstruction Module. This architecture is a distributed design, with beamforming completed at the FPGA / SoC level and AI inference executed at edge nodes. It adapts to the hardware architecture of portable devices, achieving low power consumption and high real-time processing requirements. The distributed waveform formation process is as follows: Figure 5 As shown, the execution logic of the dynamic delay superposition algorithm based on the phased array geometric model is as follows: Delay Calculation → Delay Compensation → Weighted Sum Output, which is different from the single superposition process of existing technologies.

[0093] In one embodiment, the method is applicable to battery-powered portable ultrasound devices with a total power consumption of ≤10W. The method supports real-time display of 4D ultrasound imaging with a display latency of ≤50ms. The method supports remote consultation or cloud analysis, and the inference results can be uploaded to a cloud server. The method is implemented through software and stored in a computer-readable storage medium, and can run on an FPGA / SoC hardware platform. The imaging parameters include gain and focus, and the method can automatically adjust the imaging parameters based on quality scores.

[0094] This embodiment is implemented on a battery-powered portable 4D ultrasound device with a total power consumption of ≤10W, a real-time display delay of ≤50ms, an AI inference delay of ≤200ms, an imaging quality score Q≥0.8 in more than 95% of cases, and an average error of ≤3% in segmentation measurement, meeting the needs of clinical real-time diagnosis and quantitative analysis.

[0095] The present invention also discloses a computer program product or system, including a computer program that, when executed by a processor, implements the above-described method steps.

[0096] Figure 2 A schematic diagram of the portable 4D ultrasound distributed beamforming and AI-based joint reconstruction system provided in this embodiment of the invention specifically includes: Acquisition Unit: Acquires N echo signals from ultrasound, where N is a natural number greater than 1; Distributed beamforming unit: The distributed beam is obtained by performing delay correction calculations on the N echo signals and then weighting and superimposing them. Layered unit: The distributed beam is layered to obtain the display data stream and the reconstruction data stream; Segmentation unit: After image reconstruction of the reconstructed data stream, segmentation of the target region of the image is performed to obtain a segmentation mask; Geometric parameters of the segmented mask are calculated to obtain ultrasonic target region data; Display unit: Obtains ultrasound display imaging through the display data stream; Imaging unit: The ultrasound display imaging is superimposed and fused with the ultrasound target area data to obtain the final ultrasound image.

[0097] Figure 3 An embodiment of the present invention provides a schematic diagram of a computer device, specifically including: A memory and a processor; the memory is used to store program instructions; the processor is used to invoke the program instructions, when any of the program instructions are executed, according to the above-described portable 4D ultrasound distributed beamforming and AI joint reconstruction method.

[0098] The present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, represents any of the above-described methods for the distributed beamforming and AI-based reconstruction of portable 4D ultrasound.

[0099] The verification results of this verification embodiment show that assigning inherent weights to indications can improve the performance of this method compared to the default settings. Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for example, the division of units is merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces, indirect coupling or communication connection of devices or units, and may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separated; the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of this embodiment. Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. This program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0100] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0101] The computer device provided by the present invention has been described in detail above. For those skilled in the art, there will be changes in the specific implementation and application scope based on the ideas of the embodiments of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A portable 4D ultrasound distributed beamforming and AI-based joint reconstruction method, characterized in that, include: Acquire N echo signals from ultrasound, where N is a natural number greater than 1; After performing delay correction calculations on the N echo signals, a weighted superposition is performed to obtain a distributed beam; The distributed beam is layered to obtain a display data stream and a reconstruction data stream; After image reconstruction of the reconstructed data stream, image target region segmentation is performed to obtain a segmentation mask; Geometric parameters of the segmented mask are calculated to obtain ultrasonic target region data; Ultrasonic display imaging is obtained through the displayed data stream; The ultrasound display imaging and the ultrasound target area data are superimposed and fused to obtain the final ultrasound image.

2. The method for combined distributed beamforming and AI reconstruction of portable 4D ultrasound according to claim 1, characterized in that, The geometric parameters include any one or more of the following: area, volume, and rate of change over time; ultrasonic target region data is generated by segmenting the edge contour of the mask and the geometric parameters, the ultrasonic display imaging is a real-time ultrasonic image, and the ultrasonic target region data is superimposed and fused into the real-time ultrasonic image to obtain the final ultrasonic image.

3. The method for combined distributed beamforming and AI reconstruction of portable 4D ultrasound according to claim 2, characterized in that, The time change rate is the dynamic rate of change of the target quantization parameters of the target area with the ultrasound imaging time. The calculation of the time change rate includes: acquiring continuous segmentation masks, calculating the target quantization parameters of each mask frame, calculating the instantaneous time change rate of two adjacent frames, and then calculating the average rate within a T-frame period using the instantaneous time change rate to obtain the time change rate, where T is a natural number greater than 1.

4. The method for combined distributed beamforming and AI reconstruction of portable 4D ultrasound according to claim 1, characterized in that, The delay correction calculation includes delay calculation and delay correction. After performing delay calculation on N echo signals, each signal is corrected and aligned with the time domain data to obtain N corrected data. The N corrected data are then distributed and superimposed on the probe side and / or FPGA and / or edge nodes to obtain a distributed beam.

5. The method for combined distributed beamforming and AI reconstruction of portable 4D ultrasound according to claim 1, characterized in that, The display data stream includes any one or more of the following: envelope and grayscale data, and real-time ultrasound imaging is performed using envelope and / or grayscale data to obtain ultrasound display imaging; the reconstruction data stream includes any one or more of the following: IQ data, Scanline data, and volume data, and image reconstruction and image segmentation are performed using IQ data and / or Scanline data and / or volume data to obtain a segmentation mask.

6. The method for combined distributed beamforming and AI reconstruction of portable 4D ultrasound according to claim 1, characterized in that, The segmentation is performed by segmenting the target region using a segmentation model to obtain a segmentation mask. The segmentation model can be any one or more of the following: YOLO series, SAM 3, U-Net, CMaP-SAM, RepSegNet; or the segmentation model can be performed by segmenting the target region using a sequentially connected encoding module, temporal attention fusion module, and decoding module to obtain a segmentation mask. The encoding module includes L encoding layers, each of which includes a depthwise separable convolutional layer and a channel attention compression layer. The input data of the segmentation model is encoded by the depthwise separable convolutional layer and the channel attention compression layer, and then the target features are extracted by the temporal attention fusion module and decoded by the decoding model to obtain the segmentation mask.

7. The method for combined distributed beamforming and AI reconstruction of portable 4D ultrasound according to claim 6, characterized in that, The segmentation model also includes a quality adaptive module. The data is first filtered by the quality adaptive module before being input into the segmentation model for segmentation. The quality adaptive module judges the quality of the input reconstructed data. When the quality score is less than a first threshold, it prompts a rescan. When the quality score is greater than or equal to the first threshold and less than a second threshold, the input reconstructed data is input into the artifact suppression module, where features are extracted through dilated convolution and then input into the encoding module of the segmentation model. When the quality score is greater than or equal to the second threshold, the input reconstructed data is input into the fast inference branch, and the output features are input into the encoding module of the segmentation model.

8. A computer program product comprising a computer program or instructions, characterized in that, The computer program or instructions are executed by the processor to implement the distributed beamforming and AI-based joint reconstruction method for portable 4D ultrasound as described in any one of claims 1-7.

9. A computer device comprising a memory, a processor, and a computer program or instructions stored in the memory, characterized in that, The computer program or instructions are executed by the processor to implement the distributed beamforming and AI-based joint reconstruction method for portable 4D ultrasound as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, The computer program or instructions are executed by the processor to implement the distributed beamforming and AI-based joint reconstruction method for portable 4D ultrasound as described in any one of claims 1-7.

Images