Heterogeneous collaborative analysis method and system of plane wave ultrasonic intelligent agent

Abstract

The invention discloses a heterogeneous collaborative analysis method and system for a plane wave ultrasonic intelligent agent. The system adopts a heterogeneous double-GPU architecture, the first graphics processing unit is specially used for generating an original radio frequency signal to a multi-modal image, the second graphics processing unit is specially used for AI reasoning, and the first graphics processing unit and the second graphics processing unit realize data non-cache direct transmission through a direct memory access technology. The core of the system lies in that a computing power dynamic distribution and load control algorithm is introduced, and a computing power buffer area is constructed by actively controlling the peak load rate of a graphics processor to be below a preset threshold value. The design not only reduces the architecture delay, but also ensures that the system can immediately respond to the closed-loop feedback adjustment of the imaging parameters by using the reserved computing power when facing the complex focus analysis. In combination with multi-modal data fusion and standard section quality control, the real-time performance, high accuracy and system stability of ultrasonic diagnosis are realized.

Description

Technical Field

[0001] This invention relates to the field of medical ultrasound imaging and artificial intelligence, and in particular to a heterogeneous collaborative analysis method and system for plane wave ultrasound intelligent agents. Background Technology

[0002] Currently, medical ultrasound diagnosis is an indispensable imaging tool in clinical practice, but its diagnostic quality is highly dependent on the operator's experience. While medical ultrasound equipment integrating artificial intelligence is becoming a development trend, existing technological solutions have the following inherent limitations: (1) Architecture delay leads to loss of real-time performance: Existing systems mostly adopt a separate architecture of imaging hardware + external AI server. The raw radio frequency signal data generated by the ultrasound front end is extremely large and must be relayed through the system CPU before being transmitted to the external AI unit for inference via network or bus. This relay process introduces an additional delay of up to 20-30 milliseconds (ms), making the total system delay exceed 50ms. This is difficult to match the real-time diagnostic requirements for rapid physiological processes such as cardiac pulsation and hemodynamics under ultra-high-speed plane wave imaging (frame rate can reach 25000Hz).

[0003] (2) Limited accuracy due to single diagnostic dimensions: Most ultrasound AI models only analyze the grayscale and texture features of B-mode images, failing to effectively integrate key modal data reflecting tissue biomechanical properties, such as real-time shear wave elastography, quantitative sound velocity tissue characterization, and ultrasound attenuation coefficient imaging. This results in an accuracy rate of less than 85% in differentiating between benign and malignant diseases such as thyroid micronodules and early liver fibrosis.

[0004] (3) Lack of intelligent closed loop leads to low operation efficiency: The system lacks a reverse feedback pathway from intelligent diagnostic results to imaging parameter control. Physicians still need to manually and repeatedly adjust the probe to obtain standard sections and observe lesions based on experience. The standard section acquisition rate is only about 60%, and a single examination takes as long as 15-20 minutes.

[0005] (4) Rigid resource scheduling leads to insufficient system stability: The existing system does not perform forward-looking management of computing resources, and the hardware is often running under high load. This makes it impossible for the system to effectively absorb the computing peaks generated when the scanning scene changes suddenly or complex lesions appear, which can easily cause image stuttering or inference delay. At the same time, the fully loaded system does not have the computing power available in time to execute the instructions for optimized scanning, making it difficult to realize the intelligent feedback loop in practice.

[0006] Therefore, there is an urgent need for an integrated intelligent ultrasound diagnostic solution that can achieve deep coupling of imaging and inference, low latency, multimodal fusion, and ensure system stability. Summary of the Invention

[0007] One of the objectives of this invention is to provide a heterogeneous collaborative analysis method and system for plane wave ultrasonic intelligent agents to solve the problems pointed out in the background art.

[0008] In a first aspect, the heterogeneous collaborative analysis method for plane wave ultrasonic intelligent agents provided in this invention is executed based on a heterogeneous computing architecture including a first graphics processor and a second graphics processor. The first graphics processor is used for ultrasonic front-end signal processing, and the second graphics processor has a larger memory capacity and computing power than the first graphics processor and is used for artificial intelligence inference. The method includes the following steps: Step S1: The ultrafast imaging unit responds to the current imaging parameters and performs synchronous processing on the raw radio frequency signal acquired by the ultrasound probe based on the first graphics processor to generate multimodal image data aligned in the spatial domain. Step S2: The collaborative scheduling unit monitors the resource load status of the first graphics processor and the second graphics processor in real time, allocates computing tasks according to the dynamic load control algorithm, and transmits the multimodal image data from the first graphics processor to the second graphics processor through direct memory access technology; wherein, before receiving the imaging parameter adjustment instruction, the collaborative scheduling unit controls the peak load rate of the first graphics processor and the second graphics processor to be below a preset threshold in order to reserve a computing power buffer. Step S3: The intelligent inference unit loads the diagnostic model based on the second graphics processor, performs feature fusion analysis on the multimodal image data, and outputs a structured diagnostic report containing diagnostic confidence. Step S4: The feedback adjustment unit parses the diagnostic confidence level. When the diagnostic confidence level meets the preset feedback triggering condition, it uses the computing power buffer to generate an imaging parameter adjustment instruction and sends it to the ultra-fast imaging unit to guide the ultra-fast imaging unit to optimize the subsequent data acquisition process and trigger secondary inference.

[0009] Optionally, in step S1, generating spatially aligned multimodal image data specifically includes: Acquisition layer time synchronization: A unified timestamp is automatically added to the synchronously acquired raw radio frequency signal. The timestamp format is a combination of the acquisition trigger time and the plane wave frame number. Processing synchronization: Based on the imaging parameters, calculate the mapping relationship between pixels and physical space, and convert each modal image into a unified resolution and pixel pitch; Spatial coordinate alignment: A three-dimensional physical coordinate system is established with the center of the ultrasound probe as the origin, and the pixel coordinates of each modal image are converted into coordinate values ​​in the three-dimensional physical coordinate system.

[0010] Optionally, in step S2, the dynamic load control algorithm is based on a comprehensive load score. The formula for task allocation is as follows: ; in, Indicates the graphics processor index. For core load rate, For video memory load rate, Assign weights to the core load.

[0011] Optionally, the core load rate and the aforementioned memory load rate The calculation methods are as follows: ; .

[0012] Optionally, in step S2, the resource allocation strategy of the collaborative scheduling unit includes: During the wide-area scanning phase, 40%-50% of the computing power resources are allocated to the first graphics processor, while 50%-60% of the computing power resources are allocated to the second graphics processor. During the lesion localization assessment phase, 60%-70% of the computing power resources are allocated to the first graphics processor, while 30%-40% of the computing power resources are allocated to the second graphics processor.

[0013] Optionally, in step S3, the diagnostic model integrates a lightweight inference module and a dynamic weight allocation network; The lightweight inference module uses multi-layer feature fusion technology to capture features at different levels and uses quantization technology to compress the model size. The dynamic weight allocation network is configured to automatically optimize the feature contribution weights of different modalities in the multimodal image data based on the target diagnostic disease.

[0014] Optionally, in step S4, the preset feedback triggering condition and execution logic are as follows: Determine whether the diagnostic confidence level is below a high threshold and above a low threshold; If so, it is determined that supplementary information is needed. The collaborative scheduling unit activates the computing power buffer and triggers the ultra-fast imaging unit to execute the imaging parameter adjustment instruction. The imaging parameter adjustment instruction includes adjusting the frame rate, depth, sensitivity, or controlling the probe driver to perform angle fine-tuning.

[0015] Optionally, the ultrafast imaging unit also integrates a standard section quality control module, which is used to perform high-frequency template matching analysis on the real-time image stream using a scale-invariant feature transformation algorithm; The scale-invariant feature transformation algorithm includes Gaussian scale space construction, key point detection, key point localization and orientation assignment, feature descriptor generation and template matching steps; When the matching degree is lower than the qualified threshold, the standard section quality control module guides the operator to adjust the ultrasound probe through a multimodal human-machine interface.

[0016] Optionally, the method further includes: using a localized incremental training unit, during system idle periods, selecting images that are determined by the standard section quality control module to have a matching degree higher than the qualified threshold as qualified desensitization samples, and using a transfer learning strategy of parameter freezing and partial fine-tuning to periodically update the diagnostic model.

[0017] Secondly, embodiments of the present invention provide a heterogeneous collaborative analysis system for a plane wave ultrasonic intelligent agent, used to execute the heterogeneous collaborative analysis method described in any one of the first aspects, the system comprising: The ultra-fast imaging unit integrates a high-channel-number plane-wave beamforming circuit and is configured to perform signal acquisition and multimodal image generation based on the first graphics processor; The intelligent inference unit is configured to run the diagnostic model based on the second graphics processor and output the structured diagnostic report; The collaborative scheduling unit, which connects the ultra-fast imaging unit and the intelligent inference unit, is configured to establish a point-to-point data transmission link through a high-speed peripheral component interconnection channel, and execute the dynamic load control algorithm to limit the peak load rate and reserve a computing power buffer. The feedback adjustment unit, which is communicatively coupled to the ultrafast imaging unit and the intelligent inference unit, is configured to generate the imaging parameter adjustment instruction based on the diagnostic confidence using the computing power buffer.

[0018] The present invention has achieved the following beneficial effects: By designing a heterogeneous division of labor architecture for signal translation and intelligent inference, and using DMA direct transfer technology, CPU transfer latency was reduced. An innovative load control strategy was introduced to build a computing power buffer by actively limiting peak load, providing resource guarantees for the system to cope with sudden tasks and provide real-time feedback. By combining multimodal fusion and intelligent closed-loop feedback, high diagnostic accuracy and high system stability were finally achieved.

[0019] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall architecture of the heterogeneous collaborative analysis system for ultrafast plane wave ultrasonic intelligent agents in an embodiment of the present invention; Figure 2 This is a flowchart of the heterogeneous collaborative diagnostic method in this invention; Figure 3 Flowchart of the ultrafast plane wave ultrasound intelligent body heterogeneous collaborative diagnostic method; Figure 4 This is the logical decision diagram of the dynamic computing power allocation and load control algorithm in this invention; Figure 5 This is a schematic diagram illustrating the working principle of the dynamic weight allocation network in this invention; Figure 6 This is a timeline diagram of thyroid nodule diagnosis in an embodiment of the present invention. Detailed Implementation

[0022] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0023] Traditional medical ultrasound equipment architecture has long followed a serial pipeline model of FPGA front-end beamforming and CPU back-end image processing. This model is effective for processing low frame rate data generated by linear array focusing imaging. However, with the increasing clinical demand for observing transient physiological processes (such as shear wave propagation and micro-blood flow perfusion), plane wave imaging technology has emerged. Its single transmission can cover the entire field, and frame rates can easily exceed 10,000 fps or even higher. This has led to an exponential increase in the throughput of raw radio frequency (RF) data, often reaching the GB / s or even TB / s level. At the same time, AI-assisted diagnosis is shifting from simple post-processing to real-time intervention, requiring systems to reverse-engineer imaging parameters based on image content within milliseconds. Existing discrete architectures (i.e., imaging devices and AI workstations connected via Ethernet or USB) fall short in addressing the aforementioned dual challenges: bandwidth limitations of the physical transmission link introduce latency of up to tens of milliseconds, causing AI inference results to lag behind the real-time scanning of the probe; and general operating system scheduling mechanisms cannot guarantee real-time performance under high loads, resulting in system stuttering or sluggish responses during periods requiring bursts of computing power (such as complex lesion analysis). The heterogeneous collaborative analysis method and system for ultrafast plane wave ultrasound intelligent agents proposed in this invention are designed to systematically resolve these deep-seated architectural contradictions. It aims to construct a low-latency, high-throughput ultrasound intelligent agent with autonomous evolution capabilities through underlying hardware reconstruction and upper-level software definition.

[0024] Please see Figure 1 , Figure 1 This is a schematic diagram of the overall architecture of the heterogeneous collaborative analysis system for ultrafast plane wave ultrasonic intelligent agents in an embodiment of the present invention. The system constructs an embedded dual-GPU heterogeneous parallel computing architecture based on a PCIe 4.0 / 5.0 high-speed switching matrix at the physical level. For example... Figure 1As shown, this system, as a highly integrated logical and physical entity, comprises core components including an ultrafast imaging unit, a collaborative scheduling unit, an intelligent inference unit, a feedback adjustment unit, a standard section quality control module integrated within the imaging unit, and an independent localized incremental training unit. Through hardware-level collaboration, dynamic load control, and intelligent feedback loops, the system achieves multi-site, real-time, and accurate ultrasound-assisted diagnosis. Specifically, the ultrafast imaging unit, based on a first graphics processor (GPU0), integrates a high-channel-count plane wave beamforming circuit and is configured to perform synchronous acquisition and front-end processing of raw ultrasound radio frequency signals. The intelligent inference unit, based on a second graphics processor (GPU1), boasts superior computing power and memory capacity compared to the first graphics processor and is configured to load and run large diagnostic models. The collaborative scheduling unit connects the two units, establishing a point-to-point direct data transmission link through a high-speed peripheral component interconnection channel and direct memory access technology. The feedback adjustment unit is responsible for parsing diagnostic confidence and generating closed-loop control commands.

[0025] Please also refer to Figure 2 and Figure 3 ,in Figure 2 This is a flowchart of the heterogeneous collaborative diagnostic method in this invention. Figure 3 This is a flowchart of an ultrafast plane wave ultrasound intelligent agent heterogeneous collaborative diagnostic method. The specific implementation process of this invention mainly includes the following rigorous logical steps: Step S1: The ultrafast imaging unit responds to the current imaging parameters and synchronously processes the raw radio frequency signal acquired by the ultrasound probe based on the first graphics processor to generate multimodal image data aligned in the spatial domain.

[0026] The ultrafast imaging unit, acting as the system's signal translator, has the core task of acquiring raw data containing rich physical information in an extremely short time and transforming it into spatially aligned multimodal image tensors that AI can understand. In step S1, the system responds to operation commands, loads preset imaging parameters, and simultaneously acquires and processes raw radio frequency signals in real time. After receiving the RF data, GPU0 initiates a massively parallel CUDA Kernel to perform delayed superposition beamforming. The data generated by this invention contains multimodal tensors with rich physical dimensions, specifically including four modes: B-mode image (reflecting anatomical structure), real-time shear wave elasticity map (reflecting tissue stiffness), quantitative sound velocity tissue characterization map (reflecting tissue composition), and ultrasound attenuation coefficient map (reflecting energy loss).

[0027] Because the physical generation mechanisms, temporal resolution, and spatial resolution of the four modalities mentioned above differ, directly inputting them into the AI ​​model would lead to severe feature misalignment. Therefore, the ultrafast imaging unit performs strict alignment operations, specifically including synchronization in the following three dimensions: First, there is time synchronization at the acquisition layer. The system automatically adds a uniform microsecond-level timestamp to the raw radio frequency signal acquired synchronously. The timestamp format is defined as the acquisition trigger time + plane wave frame number, which serves as the unique time index for subsequent image alignment.

[0028] Secondly, synchronization is handled by converting the four modal images to the same resolution (e.g., 1024×768 pixels) and the same pixel pitch (e.g., 0.1mm / pixel).

[0029] The system calculates the mapping relationship between pixels and physical space based on the current imaging parameters of the ultrafast imaging unit (such as probe frequency and imaging depth) to ensure that all images are strictly consistent in size specifications.

[0030] Finally, spatial coordinate alignment is performed. The system extracts imaging parameters for the four modalities (including imaging depth, probe angle, focus position, etc.) and establishes a coordinate system based on the probe's mechanical coordinate system. Specifically, with the probe center as the origin, a three-dimensional physical coordinate system of "depth (Z-axis) - lateral (X-axis) - longitudinal (Y-axis)" is established. The pixel coordinates of all modal images are strictly converted to coordinate values ​​in this physical coordinate system, thereby reducing coordinate system offset caused by dynamic adjustments of imaging parameters. Finally, these four modal images are displayed synchronously, simultaneously presented on the display screen and transmitted to the backend.

[0031] To address the challenges of ultrasound diagnosis's high dependence on operator technique and the difficulty in obtaining standard sections, the ultra-fast imaging unit also integrates a standard section quality control module. This module contains a multi-anatomical site standard section template library and employs the Scale Invariant Feature Transform (SIFT) algorithm to perform high-frequency template matching analysis on the real-time stream. When the matching degree falls below a passing threshold, the operator is guided to adjust the probe via a multimodal human-machine interface combining audiovisual and tactile feedback. The specific implementation of the SIFT algorithm includes the following steps: The first step is Gaussian scale space construction. To simulate the characteristics of images at different scales, the system performs a convolution operation between a Gaussian convolution kernel and the real-time ultrasound image. The calculation formula is as follows: .in, This represents the coordinate position of an image pixel in two-dimensional space. This represents the constructed Gaussian-scale spatial image. Real-time ultrasound images; Let Gaussian function be defined as follows: In the formula, For scale parameters, This indicates a convolution operation.

[0032] The second step is key point detection. This involves subtracting Gaussian images at adjacent scales to generate a difference Gaussian image. This is used to detect extreme points in scale space. The formula is: .in, , This is the scale number. The detection logic is: for... For each pixel in the array, compare it with its own position. Neighborhood and adjacent scales The pixel is compared with 26 neighboring pixels. If the pixel value is the maximum or minimum value among these 26 pixels, it is determined to be a key point.

[0033] The third step is keypoint localization and orientation assignment. To ensure the feature descriptor has rotation invariance, it is necessary to calculate the gradient magnitude of the keypoints. and gradient direction The calculation formulas are as follows: as well as .

[0034] The fourth step is feature descriptor generation. Centered on key points, select... The neighborhood of is divided into Each sub-region is further divided into sub-regions. For each sub-region, a gradient histogram in eight directions is calculated, thereby generating a 128-dimensional feature vector. ,Right now ,in Indicates the first Line 1 Sub-region Histogram statistics for each direction.

[0035] Step 5, Template Matching. Calculate the Euclidean distance between the SIFT feature vector of the real-time ultrasound image and the feature vector of the standard section template. The smaller the distance, the higher the matching degree. The matching degree calculation formula is: Euclidean distance In the formula, For matching degree, For real-time image feature vectors, For standard template feature vectors, For feature dimension, This represents the index number of an element in the feature vector.

[0036] Step S2: The collaborative scheduling unit monitors the resource load status of the first graphics processor and the second graphics processor in real time, allocates computing tasks according to the dynamic load control algorithm, and transmits the multimodal image data from the first graphics processor to the second graphics processor through direct memory access technology; wherein, the collaborative scheduling unit controls the peak load rate of the first graphics processor and the second graphics processor to be below a preset threshold in order to reserve a computing power buffer.

[0037] Please see Figure 4 , Figure 4 This is the logical decision diagram of the dynamic computing power allocation and load control algorithm in this invention. In step S2, the collaborative scheduling unit plays a crucial role. It establishes a point-to-point data transmission link through high-speed peripheral component interconnection channels and direct memory access technology, achieving cache-free, low-latency transmission of multimodal image data. More importantly, this unit incorporates a built-in dynamic computing power allocation and load control algorithm to monitor the resource utilization of the dual processors in real time and dynamically allocate computing tasks. This invention employs a specific comprehensive load scoring formula to quantify the GPU status: .in: For graphics processor indexes, Represents the imaging GPU (first graphics processor). Represents inference GPU (second graphics processor); This indicates the overall load score; Core load rate reflects the activity level of the GPU cores, ranging from 0% to 100%, and is calculated using the following formula: ; This is the video memory load rate, used to avoid task interruptions caused by video memory overflow. Its calculation formula is: ; The weight of the core load is set to 0.7 in this embodiment to highlight the importance of the computing core load. Based on the above formula, the collaborative scheduling unit adaptively adjusts the computing power ratio by sampling the memory usage and core utilization of the dual processors at high frequency and combining the current clinical scenario mode. A core strategy of this invention is to force the peak load rate of each processor to be preset to below 70% when performing critical real-time tasks. This design constructs a computing power buffer by actively reserving 30% computing power margin, which provides the necessary resource guarantee for the system to cope with unpredictable peak computing demand (such as tissue interface mutations, in-depth analysis of complex lesions) and closed-loop feedback tasks that are started immediately, fundamentally ensuring the absolute smoothness and responsiveness of the diagnostic process. The specific allocation strategy includes: (1) In the wide-area scanning stage, the physician moves the probe quickly to find lesions, and the system needs to maintain an ultra-high imaging frame rate. At this time, the collaborative scheduling unit allocates 40%-50% of the computing power resources to the imaging unit and 50%-60% of the computing power resources to the inference unit to achieve rapid full-field lesion screening. (2) During the lesion localization assessment stage, the physician fixes the probe to observe details, and the system needs to improve the image resolution and signal-to-noise ratio. At this time, the computing power ratio of the imaging unit is increased to 60%-70%, while the computing power ratio of the inference unit is adjusted to 30%-40% to focus on the refined feature analysis of local areas.

[0038] Step S3: The intelligent inference unit loads the diagnostic model based on the second graphics processor, performs feature fusion analysis on the multimodal image data, and outputs a structured diagnostic report containing diagnostic confidence.

[0039] Please see Figure 5 , Figure 5 This is a schematic diagram illustrating the working principle of the dynamic weight allocation network in this invention. In step S3, the intelligent inference unit, based on the second graphics processor, loads and runs a large diagnostic model, extracts features from the received multimodal image data, performs fusion analysis, and outputs a structured diagnostic report and confidence level. To meet real-time requirements, the intelligent inference unit integrates a lightweight inference module. This module adapts and simplifies the original multimodal large language model, employing multi-layer feature fusion technology to capture features at different levels of the visual encoder. It can recognize fine-grained details such as small characters and minor flaws in images, and accurately align images and text, avoiding recognition errors. Quantization technology compresses the model size to its maximum, without significantly sacrificing image recognition accuracy, while reducing memory usage and computational power consumption. Furthermore, the intelligent inference unit also incorporates a dynamic weight allocation network, trained on a large-scale labeled clinical dataset, which can automatically optimize the feature contribution weights of different modalities based on the target diagnostic disease. Figure 3 As shown, when thyroid nodules are detected, the system automatically increases the weight of elastic modulus and microcalcification features; while when liver lesions are detected, the weight of attenuation coefficient and sound velocity value is increased.

[0040] Step S4: The feedback adjustment unit parses the diagnostic confidence level. When the diagnostic confidence level meets the preset feedback triggering condition, it uses the computing power buffer to generate an imaging parameter adjustment instruction and sends it to the ultra-fast imaging unit to guide the ultra-fast imaging unit to optimize the subsequent data acquisition process and trigger secondary inference.

[0041] Please see Figure 6 , Figure 6This is a time-series diagram of thyroid nodule diagnosis in an embodiment of the present invention. The feedback adjustment unit analyzes the diagnostic confidence level output by the intelligent inference unit and automatically generates imaging parameter adjustment instructions based on preset rules. Its working mechanism is as follows: when the diagnostic confidence level is not lower than the high threshold (e.g., 0.95), the diagnosis is determined to be clear, and a diagnostic report is directly output. When the confidence level is lower than the high threshold but higher than the low threshold (e.g., between 0.75 and 0.95), it is determined to be "suspicious but lacking in information," requiring supplementary information, and the imaging parameter optimization process is triggered. At this time, the collaborative scheduling unit uses its preset computing power buffer (i.e., the aforementioned reserved 30% computing power) to immediately allocate the computing resources required for the imaging unit to execute the optimization instructions, instructing it to adjust parameters such as frame rate, depth, and sensitivity, and control the probe driver to perform angle fine-tuning, focusing on the region of interest for supplementary data acquisition, and then starting secondary inference. For example: during the system initialization and data acquisition phase, the physician starts the examination, and the ultra-fast imaging unit generates multimodal images. During the collaborative scheduling phase, if the GPU memory usage is low, computing power is allocated according to the load control algorithm, and data is transferred via DMA. During the intelligent inference phase, the initial diagnosis is: "Left inferior lobe nodule, possibly malignant (confidence 0.92)". In the feedback optimization phase, the feedback adjustment unit identifies a confidence level of 0.92 (<0.95, a high threshold threshold; depending on specific settings, if 0.92 still requires confirmation of blood supply, it falls within the optimization range) and decides to initiate optimization. The collaborative scheduling unit responds immediately, utilizing computing power to instantly increase the computing power allocation of the imaging unit, instructing it to reduce the frame rate to 15000Hz to improve resolution and increase blood flow sensitivity by 2 levels, while simultaneously controlling probe fine-tuning by +3°. In the secondary confirmation phase, after focusing and collecting supplementary data for 3 seconds, the secondary inference incorporates blood flow characteristics with a high resistance index (RI=0.75), ultimately outputting a diagnostic report. Finally, to ensure the system's continuous evolution capability, this invention also includes a localized incremental training unit. This unit uses Advanced Encryption Standard (AES) to desensitize clinical data and automatically selects qualified samples. By utilizing the system's computing resources during nighttime or off-peak hours, and employing a transfer learning strategy involving parameter freezing and partial fine-tuning (such as LoRA), the diagnostic model is periodically updated incrementally. This mechanism ensures that the model can continuously adapt to changes in data distribution caused by different regions and varying degrees of equipment aging, achieving an effect of becoming more accurate with use.

[0042] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A heterogeneous collaborative analysis method of plane wave ultrasound agents, characterized in that, Based on the execution of a heterogeneous computing architecture comprising a first graphics processor for ultrasonic front-end signal processing and a second graphics processor with a larger memory capacity and computing power than the first graphics processor and used for artificial intelligence inference, the method comprises the following steps: Step S1, the super-fast imaging unit, in response to the current imaging parameters, performs synchronous processing on the raw radio frequency signals collected by the ultrasonic probe based on the first graphics processor to generate multi-modal image data aligned in the spatial domain; Step S2, the cooperative scheduling unit monitors the resource load state of the first graphics processor and the second graphics processor in real time, allocates computing tasks according to a dynamic load control algorithm, and transmits the multi-modal image data from the first graphics processor to the second graphics processor through direct memory access technology; wherein, before receiving the imaging parameter adjustment instruction, the cooperative scheduling unit controls the peak load rate of the first graphics processor and the second graphics processor below a preset threshold to reserve a computing power buffer; Step S3, the intelligent inference unit loads a diagnostic model based on the second graphics processor, performs feature fusion analysis on the multi-modal image data, and outputs a structured diagnostic report containing a diagnostic confidence; Step S4, the feedback adjustment unit analyzes the diagnostic confidence, and when the diagnostic confidence meets a preset feedback triggering condition, generates an imaging parameter adjustment instruction using the computing power buffer and sends it to the super-fast imaging unit to guide the super-fast imaging unit to optimize the subsequent data acquisition process and trigger secondary inference.

2. The method of claim 1, wherein, In step S1, generating multi-modal image data aligned in the spatial domain specifically includes: Time synchronization of the acquisition layer: automatically adding a uniform timestamp to the synchronously collected raw radio frequency signals, the timestamp format being a combination of the acquisition trigger time and the plane wave frame number; Processing synchronization: based on the imaging parameters, calculating the mapping relationship between pixels and physical space, and converting each modal image to a uniform resolution and pixel spacing; Spatial coordinate alignment: establishing a three-dimensional physical coordinate system with the center of the ultrasonic probe as the origin, and converting the pixel coordinates of each modal image to coordinate values in the three-dimensional physical coordinate system.

3. The method of claim 1, wherein, In step S2, the dynamic load control algorithm is based on a composite load score The task allocation is performed with the formula: ； wherein, represents a graphics processor index, is a core load rate, is a memory load rate, is a core load distribution weight.

4. The method of claim 3, wherein, The core load rate And the calculation method of the video memory load rate Respectively: ； 。 5. The method of claim 1, wherein, In step S2, the resource allocation strategy of the cooperative scheduling unit includes: In the wide-area scanning stage, allocate 40%-50% of the computing power resources to the first graphics processor, and at the same time, allocate 50%-60% of the computing power resources to the second graphics processor; In the lesion site evaluation stage, allocate 60%-70% of the computing power resources to the first graphics processor, and at the same time, allocate 30%-40% of the computing power resources to the second graphics processor.

6. The method of claim 1, wherein, In step S3, the diagnostic model integrates a lightweight inference module and a dynamic weight distribution network; The lightweight inference module uses multi-layer feature fusion technology to capture features at different levels and compresses the model size through quantization technology; The dynamic weight distribution network is configured to automatically optimize the feature contribution weights of different modal data in the multi-modal image data according to the target diagnostic disease.

7. The method of claim 1, wherein, In step S4, the preset feedback trigger condition and execution logic are: determining whether the diagnostic confidence is lower than a high threshold and higher than a low threshold; if yes, determining that information needs to be supplemented, the cooperative scheduling unit enables the computing power buffer, triggers the super-fast imaging unit to execute the imaging parameter adjustment instruction, and the imaging parameter adjustment instruction includes adjusting the frame frequency, depth, sensitivity, or controlling the probe driver to perform angle fine adjustment.

8. The method of claim 1, wherein, The super-fast imaging unit is also integrated with a standard section quality control module for high-frequency template matching analysis of real-time image streams using a scale-invariant feature transform algorithm. The scale-invariant feature transform algorithm includes the steps of Gaussian scale space construction, key point detection, key point positioning and direction assignment, feature descriptor generation, and template matching. When the matching degree is lower than the qualified threshold, the standard section quality control module guides the operator to adjust the ultrasonic probe through a multi-modal human-machine interface.

9. The method of claim 1, wherein, The method further includes: using a localized incremental training unit to screen images determined by the standard section quality control module as having a matching degree higher than the qualified threshold as desensitized qualified samples during system idle periods, and using a parameter freezing and partial fine-tuning transfer learning strategy to periodically update the diagnostic model incrementally.

10. A heterogeneous collaborative analysis system of planar wave ultrasound agents, characterized in that, A system for performing the heterogeneous cooperative analysis method of any one of claims 1-9, the system comprising: a super-fast imaging unit integrated with a high-channel plane wave beam synthesis circuit configured to perform signal acquisition and multi-modal image generation based on the first graphics processor; an intelligent inference unit configured to run the diagnostic model based on the second graphics processor and output the structured diagnostic report; a cooperative scheduling unit connected to the super-fast imaging unit and the intelligent inference unit, configured to establish a point-to-point data transmission link through a high-speed peripheral component interconnect channel, and execute the dynamic load control algorithm to limit the peak load rate and reserve a computing power buffer; a feedback adjustment unit communicatively coupled to the super-fast imaging unit and the intelligent inference unit, configured to generate the imaging parameter adjustment instruction using the computing power buffer based on the diagnostic confidence.

Images