Dynamic function service scheduling method for intelligent ultrasonic detection and quality control

Abstract

This application relates to a dynamic function service scheduling method for intelligent ultrasound inspection and quality control, applied to a scheduling system comprising an ultrasound equipment layer, an edge service layer, and a cloud service layer. Each layer deploys different function service variants for function services with the same business logic. The ultrasound equipment layer determines the set of function services to be executed corresponding to the current ultrasound inspection task, extracts task features based on the acquired ultrasound video stream, and reports them to the edge service layer. The function-as-a-service scheduler in the edge service layer determines the target execution level corresponding to each function service in the set based on the task features and the current resource status of each layer. The function-as-a-service scheduler sequentially determines the current function services to be executed and issues scheduling instructions to the corresponding target execution level for each current function service, so that the function service variant of the current function service is invoked at the target execution level for corresponding processing. This method can improve the efficiency of ultrasound image processing.

Description

Technical Field

[0001] This application relates to the field of ultrasound image processing technology, specifically to a dynamic function service scheduling method for intelligent ultrasound detection and quality control. Background Technology

[0002] Ultrasound technology plays a very important role in the medical field. Ultrasound images generated by ultrasound technology can serve as important auxiliary information for doctors in diagnosis.

[0003] In traditional methods, doctors need to rely entirely on their own experience to repeatedly observe ultrasound images in order to extract some information to aid in diagnosis. This requires doctors to expend a great deal of time and effort to obtain the necessary information from ultrasound images, making it very inefficient. Summary of the Invention

[0004] Based on this, it is necessary to provide a dynamic function service scheduling method, system, computer equipment, computer-readable storage medium, and computer program product for intelligent ultrasound detection and quality control that can improve the efficiency of ultrasound image analysis and processing, addressing the aforementioned technical problems.

[0005] Firstly, this application provides a dynamic function service scheduling method for intelligent ultrasound detection and quality control, applied to a scheduling system. This scheduling system comprises an ultrasound equipment layer, an edge service layer, and a cloud service layer. Different function service variants are deployed at each layer for the same business logic. The method includes:

[0006] The ultrasonic equipment layer determines the set of function services to be executed corresponding to the current ultrasonic testing task, extracts task features based on the ultrasonic video stream acquired in real time for the ultrasonic testing task, and reports the task features to the edge service layer.

[0007] The function-as-a-service scheduler in the edge service layer performs cross-layer scheduling decision processing based on task characteristics and the current resource status of each layer, in order to determine the target execution layer corresponding to each function service in the set.

[0008] The Function as a Service scheduler determines the current function service to be executed sequentially according to the execution order of the various function services in the set, generates a corresponding scheduling instruction for each current function service, and sends the scheduling instruction to the corresponding target execution level. The scheduling instruction is used to instruct the function service variant that calls the current function service at the target execution level to process the input data corresponding to the current function service.

[0009] Secondly, this application provides a dynamic function service scheduling system for intelligent ultrasound detection and quality control. The scheduling system includes an ultrasound equipment layer, an edge service layer, and a cloud service layer. Different function service variants are deployed at each layer of the scheduling system for function services with the same business logic.

[0010] The ultrasound equipment layer is used to determine the set of function services to be executed corresponding to the current ultrasound detection task. It extracts task features based on the ultrasound video stream acquired in real time for the ultrasound detection task and reports the task features to the edge service layer.

[0011] The edge service layer is used to make cross-layer scheduling decisions based on task characteristics and the current resource status of each layer using the built-in function-as-a-service scheduler, in order to determine the target execution layer corresponding to each function service in the set; the target execution layer is the ultrasound equipment layer, the edge service layer, or the cloud service layer.

[0012] The edge service layer is also used to use the function-as-a-service scheduler to determine the current function service to be executed in sequence according to the execution order of each function service in the set, generate a corresponding scheduling instruction for each current function service, and send the scheduling instruction to the corresponding target execution level. The scheduling instruction is used to instruct the function service variant that calls the current function service at the target execution level to process the input data corresponding to the current function service.

[0013] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described in the first aspect.

[0014] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of the method described in the first aspect above.

[0015] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the steps of the method described in the first aspect above.

[0016] The aforementioned dynamic function service scheduling method, system, computer equipment, computer-readable storage medium, and computer program product for intelligent ultrasound inspection and quality control utilize Function as a Service (FaaS) technology to construct a collaborative architecture scheduling system comprising an ultrasound equipment layer, an edge service layer, and a cloud service layer. Each layer deploys different adapted function service variants for the same business logic. This decouples complex ultrasound inspection processing into function services, using function services as the scheduling granularity. Based on a dynamic function service scheduling mechanism, and considering the task characteristics of the ultrasound inspection task and the resource status of each layer, the system intelligently and dynamically decides on a cloud-edge-device collaborative scheduling scheme for function services. This effectively balances real-time performance with computational costs, significantly improving the processing efficiency of ultrasound images. Attached Figure Description

[0017] Figure 1 This is an application environment diagram of a dynamic function service scheduling method for intelligent ultrasound detection and quality control in one embodiment;

[0018] Figure 2 This is a flowchart illustrating a dynamic function service scheduling method for intelligent ultrasound detection and quality control in one embodiment.

[0019] Figure 3 This is a flowchart illustrating cross-layer scheduling decision processing in one embodiment;

[0020] Figure 4 This is a schematic diagram illustrating the principle of cross-layer scheduling processing in one embodiment;

[0021] Figure 5 This is a simplified flowchart illustrating the processing involved in the ultrasonic equipment layer in one embodiment;

[0022] Figure 6 This is a simplified diagram illustrating the processing flow involved in the edge service layer in one embodiment;

[0023] Figure 7 This is a simplified diagram illustrating the processing flow involved in the cloud service layer in one embodiment;

[0024] Figure 8 This is a schematic diagram illustrating the principle of a dynamic function service scheduling method for intelligent ultrasound detection and quality control in one embodiment.

[0025] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0027] It should be understood that, unless otherwise defined, all technical terms used herein have the same meaning as commonly understood by those skilled in the art. The technical terms used herein are for the purpose of describing specific examples only and are not intended to limit the scope of this application.

[0028] Figure 1 This is an application environment diagram of a dynamic function service scheduling method for intelligent ultrasound detection and quality control in one embodiment. For example... Figure 1 As shown, this method can be applied to a scheduling system. This scheduling system comprises three layers: the ultrasound equipment layer (i.e., the terminal device layer), the edge service layer, and the cloud service layer.

[0029] The ultrasound equipment layer mainly includes at least one device related to ultrasound testing, such as a handheld wireless ultrasound probe, a trolley-type high-end ultrasound testing instrument, or a portable flatbed ultrasound device, which can be referred to as ultrasound equipment. For example, the ultrasound equipment layer may also include a vibration motor or feedback component for tactile cues. Each device in the ultrasound equipment layer has a built-in high-performance FPGA chip for front-end beamforming and communicates with the edge service layer via a high-speed, low-latency wireless network.

[0030] Ultrasonic equipment not only acquires high frame rate (e.g., >30fps) real-time ultrasound video streams (such as raw RF signals and beamformed B-mode ultrasound images), but also performs functions such as human-computer interaction, haptic feedback, and basic computing power, acting as the system's sensory antennae. To overcome the limitations of traditional ultrasound equipment acting solely as a data acquisition device, this application embeds a micro-function-as-a-service (i.e., micro-FaaS agent) within the ultrasound equipment to run lightweight function service variants. Furthermore, it can invoke a data compression function service before uploading data to the edge service layer, dynamically encoding the data (such as ultrasound video streams) based on network conditions to reduce transmission load.

[0031] The edge service layer consists of edge servers deployed in hospital departments, floor-level computer rooms, or mobile medical vehicles. These edge servers act as local computing centers and data gateways, connecting to the cloud service layer via a high-bandwidth intra-hospital LAN or software-defined wide area network (SDW). Even when disconnected from the cloud service layer, the edge service layer can maintain local ultrasound detection processing and temporarily store the generated offline data, asynchronously synchronizing it to the cloud once the network is restored. The edge service layer is a crucial link in this architecture, responsible for both real-time intelligent inference (e.g., <200ms) and ensuring that sensitive data does not leak out of the hospital network before being anonymized.

[0032] The cloud service layer consists of cloud servers, typically deployed in large-scale data centers on public clouds or private healthcare clouds. These cloud servers possess massive storage and computing resources similar to HPC clusters, serving as a global control center responsible for the long-term cold storage of massive amounts of medical data, distributed training and fine-tuning of complex models, and version management of global algorithms. They play three core roles: first, a deep analysis and assisted diagnosis center, providing secondary verification for difficult cases; second, a side-layer report optimization center, optimizing and logically validating the initial reports generated at the edge layer; and third, an AI model management center, responsible for aggregating sample data from difficult cases, continuously learning and updating models, and distributing optimized model parameters downwards via OTA (Over-The-Air) technology to achieve continuous system evolution.

[0033] The cloud-edge-device three-layer architecture of the scheduling system works in deep collaboration through a dynamic scheduling mechanism based on Function as a Service (FaaS). This decouples complex ultrasound AI applications into a series of loosely coupled, stateless, event-driven independent function services, enabling on-demand allocation of computing resources, millisecond-level elastic scaling, and fault self-healing, thereby improving the processing efficiency of ultrasound images. It should be understood that the output of all function services are intermediate data to assist doctors in diagnosis, automatically providing some information for the doctor's final diagnosis, and are not direct disease diagnosis results.

[0034] In this application, the cloud service layer pre-trains models (i.e., AI models) corresponding to each function service. Function services serve as the runtime environment and logical encapsulation container for these models. The relationship between models and function services is as follows: models exist in the form of weight files (also called model files), each model is used to implement different functions, and different ultrasound detection tasks are completed based on these functions. Function services are executable units containing inference engine code, used to load models, transform input data into tensors that the model can process for computation, and obtain output results. To fully utilize the heterogeneous computing resources of the cloud-edge-device three-layer architecture and achieve optimal scheduling at the function granularity, this application embodiment constructs a function service variant mechanism based on "homogeneous but different effects." Specifically, for function services with the same business logic, the scheduling system pre-sets multiple heterogeneous versions adapted to different hardware environments, i.e., multiple function service variants. The different function service variants corresponding to each function service at each level are generated by performing different transformation processes on the model files corresponding to the function service.

[0035] Different function service variants differ in algorithm model structure, computational complexity, and output accuracy. Specifically, a lightweight function service variant adapted to terminal computing power is deployed at the ultrasonic equipment layer; a standard function service variant adapted to edge computing power is deployed at the edge service layer; and a high-precision function service variant is deployed at the cloud service layer. The lightweight function service variant has the lowest computational complexity, while the high-precision function service variant has the highest computational complexity; the lightweight function service variant has the lowest output accuracy, while the high-precision function service variant has the highest output accuracy. Based on the above variant design, the cross-layer scheduling decision process in this application is essentially a multi-objective optimization decision among the above three function service variants.

[0036] Lightweight function service variants have extremely small model parameter sizes (typically in the MB range), enabling millisecond-level inference using embedded chips on the terminal (such as FPGAs or mobile GPUs). Although their absolute detection accuracy is relatively limited, they possess extremely low latency characteristics, making them suitable for real-time sectional navigation, image artifact cues, and basic quality control during ultrasound scanning, with a focus on meeting the real-time interactive needs of physicians during operation.

[0037] The standard version of the function service variant strikes a balance between inference accuracy and response speed, and can perform tasks such as automatic capture of standard sections, preliminary screening of common diseases and structural measurement, focusing on providing auxiliary test results that meet clinical standards while ensuring near real-time performance.

[0038] The high-precision function service variant boasts the highest detection sensitivity and specificity, capable of identifying minute lesions, analyzing complex anatomical structures, and handling difficult cases. However, it involves enormous computational demands and is limited by network bandwidth. Applicable scenarios include secondary verification of difficult cases, mining of research data, and logical validation of reports generated at the edge, focusing on providing the highest accuracy in detection conclusions.

[0039] For example, the cloud service layer can pre-train a full-precision benchmark model, and then, based on the model file of this full-precision benchmark model, generate a standard function service variant adapted to the edge service layer through half-precision conversion, and a lightweight function service variant adapted to the ultrasound equipment layer through model distillation and quantization techniques. The high-precision function service variants in the cloud service layer can be directly encapsulated from the model file of the full-precision benchmark model.

[0040] It should be noted that this application does not set up variations of function services across all three levels for all function services. For example, the data acquisition function service can only be executed using hardware at the ultrasound equipment layer, so there is no need to set up variations for the other two levels. Similarly, deep analysis and assisted diagnosis function services, and edge report analysis and optimization function services can only be executed using hardware at the cloud service layer, so there is no need to set up variations for the ultrasound service layer and the edge service layer. This application only adapts variations of function services at different levels for function services that can execute at multiple levels (such as quality control function services, aspect selection functions, etc., which are not exhaustively listed here).

[0041] like Figure 2 As shown, in one embodiment, a dynamic function service scheduling method for intelligent ultrasound detection and quality control is provided, applied to... Figure 1 In the scheduling system shown, the method specifically includes the following steps:

[0042] S21, the ultrasound equipment layer determines the set of function services to be executed corresponding to the current ultrasound detection task, extracts task features based on the ultrasound video stream acquired in real time for the ultrasound detection task, and reports the task features to the edge service layer.

[0043] Specifically, the interface of the ultrasound equipment in the ultrasound equipment layer can display examination modes for the user to select, such as "routine abdominal examination," "cardiac function assessment," or "emergency trauma scan." When the user selects any examination mode, a corresponding ultrasound examination task is generated. In response to the ultrasound examination task, the ultrasound equipment layer activates its built-in business process parser, retrieving the corresponding function orchestration template from a pre-set business process library to determine the set of function services to be executed (i.e., the function services required to complete the ultrasound examination task).

[0044] For example, a directed acyclic graph task chain containing multiple function services to be executed can be constructed. For instance, if the selected detection mode is "cardiac function assessment", the generated task chain may include: data acquisition function service → preprocessing function service → ventricular wall motion tracking function service → ejection fraction calculation function service → report generation function service → side-side report analysis and optimization function service.

[0045] The ultrasound equipment in the ultrasound equipment layer continuously acquires ultrasound video streams through underlying drivers, for example, by triggering data acquisition function services in the ultrasound equipment layer to acquire ultrasound video streams in real time. The ultrasound equipment layer extracts task features from the real-time acquired ultrasound video streams and reports these task features to the edge service layer. For example, task features include at least one of the following: the data size of the ultrasound video stream, the real-time requirement threshold for the task (e.g., less than 150ms in an emergency scenario), and the privacy level corresponding to the current ultrasound detection task. The privacy level can be divided into "low," "medium," and "high" levels.

[0046] For example, in addition to reporting task characteristics to the edge service layer, the ultrasound device layer can also perform first-level scheduling locally. Specifically, the ultrasound device in the ultrasound device layer has a built-in micro-function-as-a-service agent (i.e., a micro-FaaS agent). This micro-function-as-a-service agent is an extremely lightweight runtime environment based on WebAssembly technology, designed specifically for resource-constrained embedded devices, and has a very low memory footprint. The ultrasound device layer performs first-level scheduling based on the micro-function-as-a-service agent to call lightweight function service variants corresponding to some function services on demand to perform corresponding processing on the real-time acquired video stream, based on the local resource status (such as the local power and computing power of the ultrasound device). In this way, first-level scheduling processing is performed on some function services in the set of function services to be executed. For ease of description, the function services that have been executed (or scheduled) in the first-level scheduling of the ultrasound device layer can be referred to as the first function service. The first function service may include at least one of data acquisition function, preprocessing function service, quality control function service, or feedback function service. In the first-level scheduling, it does not rely on network transmission and is completed directly on the local computing power of the ultrasound device, with latency controlled in the millisecond range.

[0047] Taking the preprocessing function service as an example, a lightweight function service variant corresponding to the preprocessing function service is deployed in the ultrasound equipment layer. The ultrasound equipment layer performs first-level scheduling based on a micro FaaS agent to call the lightweight function service variant corresponding to the preprocessing function service according to the local resource status. This preprocesses each ultrasound image in the ultrasound video stream, resulting in a preprocessed ultrasound video stream. The preprocessed ultrasound video stream is then used as input data for the next function service of the preprocessing function service. For example, if the next function service of the preprocessing function service is a quality control function service, the preprocessed ultrasound video stream can be used as its input data, and the quality control function service can be used to perform quality control detection processing on the ultrasound images in the preprocessed ultrasound video stream.

[0048] For example, the lightweight function service variant corresponding to the preprocessing function service can obtain the ultrasound image in the original ultrasound video stream with zero copy through direct memory access technology, and perform at least one preprocessing such as image denoising, image enhancement, normalized cropping and scaling.

[0049] The process includes: Image denoising: Anisotropic diffusion filtering is used to remove speckle noise specific to ultrasound images while preserving the edge and texture details of key structures to the greatest extent possible, avoiding over-smoothing that could lead to misdiagnosis. Image enhancement: Contrast-limited adaptive histogram equalization is used to dynamically adjust the local contrast of ultrasound images, making grayscale differences more obvious and textures clearer. Standardized cropping and scaling: Contour detection visual algorithms are used to automatically identify valid fan-shaped or rectangular scanning areas, cropping out surrounding useless areas, scaling the cropped ultrasound image to a preset standard size, and normalizing it to generate standard data suitable for subsequent input.

[0050] S22, the function-as-a-service scheduler in the edge service layer performs cross-layer scheduling decision processing based on task characteristics and the current resource status of each layer, in order to determine the target execution layer corresponding to each function service in the set from each layer.

[0051] In addition to acquiring task characteristics reported by the ultrasound equipment layer, the edge service layer can also obtain the resource status of each layer through distributed probes. This resource status includes the computing load of each layer and the network link status between layers (e.g., network latency from the ultrasound equipment layer to the edge service layer, network bandwidth from the edge service layer to the cloud service layer, etc.). A FaaS scheduler (Function as a Service scheduler) is configured in the edge service layer. The FaaS scheduler can be used to perform cross-layer scheduling decisions based on task characteristics and the current resource status of each layer, in order to determine the target execution layer corresponding to each function service in the set.

[0052] In some examples, after determining the target execution level for each function service, a function-level mapping set is formed. Subsequently, based on this function-level mapping set, the corresponding scheduling is performed on each function service in the set of function services to be executed, according to step S23.

[0053] In some embodiments, such as Figure 3 As shown, step S22 includes:

[0054] S221, For each function service, the function-as-a-service scheduler is used to predict the time cost, resource cost, and privacy risk penalty value required for the function service to be executed at each corresponding candidate level based on task characteristics and resource status.

[0055] In some embodiments, the algorithm follows the dual hard constraints of "physical affiliation" and "privacy priority" to determine the set of candidate levels (i.e., executable levels) for each function service.

[0056] Physical constraints refer to the forced locking of candidate layers for function services that rely on ultrasound equipment hardware (such as data acquisition function services) to the ultrasound equipment layer. Examples include data acquisition function services and feedback function services locked to the ultrasound equipment layer. Data acquisition function services are used to acquire ultrasound data (such as ultrasound video streams) in real time. Feedback function services are used to provide real-time feedback. If excessive ultrasound probe pressure, excessive movement speed, or motion blur in the ultrasound image is detected, the ultrasound equipment immediately triggers the local execution of the feedback function service to generate haptic feedback commands, driving the handle to vibrate or playing a prompt sound. Additionally, simple navigation icons are overlaid on the screen edge. If the ultrasound image quality is acceptable, keyframes are automatically cached and silently uploaded, ensuring uninterrupted doctor operation and achieving a low-latency interactive experience.

[0057] Privacy constraints refer to the requirement that for ultrasound inspection tasks with a high privacy level, the candidate layer for function services involving sensitive data for such tasks does not include the cloud service layer. Thus, no matter how strong the cloud computing power is, the function service involving sensitive data is only allowed to be executed on the ultrasound equipment layer or the edge service layer, physically eliminating the risk of privacy leakage.

[0058] For example, the current ultrasound examination task is "VIP ward bedside emergency examination" (privacy level marked as "high"), and the set of function services to be executed includes "automatic lesion identification function service". Although the cloud server has the ability to execute this function service, due to the high privacy attributes of the task, the cloud service layer will be forcibly removed, and only the ultrasound equipment layer and the edge service layer will be retained as candidate layers for this function service.

[0059] For each function service to be executed, predicting the time cost, resource cost, and privacy risk penalty value required to execute the function service at each corresponding candidate level refers to predicting the time cost, resource cost, and privacy risk penalty value that will be generated when calling the function service variant at each candidate level to perform the corresponding processing.

[0060] S222, the time cost, resource cost and privacy risk penalty value are weighted and summed to obtain the comprehensive cost of each function service executed at each corresponding candidate level.

[0061] In some embodiments, each function service is calculated using the following formula. At the candidate level Overall cost of execution :

[0062]

[0063] in: The time cost of prediction; For the predicted resource cost; Privacy risk penalty value; These are the weighting coefficients for the weighted summation. For example, It can be dynamically adjusted according to different scenarios (such as emergency vs. physical examination) or different ultrasound examination tasks.

[0064] For example, the predicted time cost may include at least one of the following: data transmission time, queuing computation time, or function cold start time. The predicted resource cost may include at least one of the following: computational resource cost or expense cost.

[0065] S223, under the premise of satisfying the real-time constraint and with the goal of minimizing the total task cost, determines the target execution level corresponding to each function service; the real-time constraint includes that the total time of the ultrasound detection task is less than or equal to the preset task real-time requirement threshold; the total task cost is obtained by summing the comprehensive costs corresponding to each function service.

[0066] In some examples, the FaaS scheduler at the edge service layer can determine multiple candidate scheduling combinations based on the candidate levels corresponding to each function service to be executed for the ultrasound detection task. For example, after determining multiple preliminary scheduling combinations for the functions to be executed in step S21, for each preliminary scheduling combination, a preset interactive processing function service can be added between the two function services across levels in that preliminary scheduling combination to obtain the final candidate scheduling combination. The interactive processing function service refers to the function service that processes the interactive data (i.e., the input data of the latter function service among the two function services across levels) before cross-level interaction. For example, the interactive processing function service could be a data compression function service, a privacy desensitization function service, etc. It should be understood that each candidate scheduling combination is an updated task chain.

[0067] Each candidate scheduling combination includes all the function services to be executed to complete the ultrasound inspection task, including function services that are uniquely locked to a specific level due to physical constraints. The same function service may correspond to different levels in different candidate scheduling combinations.

[0068] Taking the ultrasound examination task as "cardiac function assessment" as an example, assume the generated original task chain includes: data acquisition function service → preprocessing function service → ventricular wall motion tracking function service → ejection fraction calculation function service → report generation function service → sideline report analysis and optimization function service. Assume two scheduling combinations are obtained, as follows:

[0069] The initial scheduling combination a is: Data acquisition function service (ultrasound equipment layer) → Preprocessing function service (ultrasound equipment layer) → Ventricular wall motion tracking function service (edge ​​service layer) → Ejection fraction calculation function service (edge ​​service layer) → Report generation function service (edge ​​service layer) → Side-side report analysis and optimization function service (cloud service layer). Then, by adding corresponding data compression function services and privacy desensitization function services between the cross-layer interaction function services, a candidate scheduling combination A is obtained: Data acquisition function service (ultrasound equipment layer) → Preprocessing function service (ultrasound equipment layer) → Data compression function service (ultrasound equipment layer) → Ventricular wall motion tracking function service (edge ​​service layer) → Ejection fraction calculation function service (edge ​​service layer) → Report generation function service (edge ​​service layer) → Privacy desensitization function service (edge ​​service layer) → Side-side report analysis and optimization function service (cloud service layer).

[0070] Initial scheduling combination b: Data acquisition function service (ultrasound equipment layer) → Preprocessing function service (ultrasound equipment layer) → Ventricular wall motion tracking function service (ultrasound equipment layer) → Ejection fraction calculation function service (edge ​​service layer) → Report generation function service (edge ​​service layer) → Side-side report analysis and optimization function service (cloud service layer). Adding corresponding data compression and privacy desensitization function services between the cross-layer interaction function services yields candidate scheduling combination B: Data acquisition function service (ultrasound equipment layer) → Preprocessing function service (ultrasound equipment layer) → Ventricular wall motion tracking function service (ultrasound equipment layer) → Data compression function service (ultrasound equipment layer) → Ejection fraction calculation function service (edge ​​service layer) → Report generation function service (edge ​​service layer) → Privacy desensitization function service (edge ​​service layer) → Side-side report analysis and optimization function service (cloud service layer).

[0071] For each candidate scheduling combination, the total task cost of the scheduling combination is obtained by summing the comprehensive costs of each function service executed at its corresponding level within the combination. Under the premise of satisfying real-time constraints, the target scheduling combination with the minimum total task cost is selected from all scheduling combinations. The level corresponding to each function service in the target scheduling combination is the target execution level for that function service. The total execution time of the target scheduling combination is less than or equal to the preset task real-time requirement threshold. The total execution time of the target scheduling combination refers to the total time spent executing all function services in the target scheduling combination, which is equivalent to the end-to-end total time of the entire task chain corresponding to the ultrasound detection task, and therefore can represent the total execution time of the ultrasound detection task.

[0072] Taking the aforementioned candidate scheduling combinations A and B as examples, if the total task cost of scheduling combination B is lower, and the total execution time corresponding to scheduling combination B is less than the task's real-time requirement threshold, then the level corresponding to each function service in scheduling combination B is its respective target execution level. Subsequently, scheduling instructions are issued based on the mapping relationship between each function in scheduling combination B and its corresponding target execution level (i.e., the function-level mapping relationship).

[0073] It should be understood that the schemes in steps S221 to S223 are essentially based on a soft strategy that follows the hard constraint of "privacy first" (i.e., high privacy-level data is prohibited from being scheduled to the cloud service layer) and the soft strategy of "latency-cost" multi-objective optimization to make scheduling decisions.

[0074] S23, the function-as-a-service scheduler determines the current function service to be executed in sequence according to the execution order of each function service in the set, generates a corresponding scheduling instruction for each current function service, and sends the scheduling instruction to the corresponding target execution level; the scheduling instruction is used to instruct the function service variant that calls the current function service at the target execution level to process the input data corresponding to the current function service.

[0075] It should be understood that the sequential scheduling of each function service (including added interactive processing function services) in the task chain by the FaaS scheduler is called secondary scheduling.

[0076] like Figure 4 As shown, the Function-as-a-Service scheduler in the edge service layer can determine the target execution layer for each function service based on cross-layer scheduling decision rules, combined with task characteristics and resource status at each level. The Function-as-a-Service scheduler can schedule each function service sequentially, with each function service potentially executing at the ultrasound equipment layer, the edge service layer, or the cloud service layer. After each function service completes execution, the scheduling result can be returned, allowing the Function-as-a-Service scheduler to schedule the next function service. The entire secondary scheduling process involves collaborative processing across the cloud, edge, and endpoint layers. For example, the ultrasound equipment layer might need to compress data before uploading it to the edge service layer, or the edge service layer might need to anonymize data before uploading it to the cloud service layer. In this way, collaborative processing enables the scheduling of all function services to be executed in the entire task chain of the ultrasound detection task.

[0077] In some embodiments, if the target execution level corresponding to the current function service determined by the FaaS scheduler in the edge service layer is the ultrasound device layer, and if the current function service is a first function service (i.e., a function service already executed in the first-level scheduling), since the first function service has already been completed in the first-level scheduling stage of the ultrasound device layer, after the scheduling instruction for the first function service is issued to the corresponding target execution level in the second-level scheduling stage, the ultrasound device layer can ignore the scheduling instruction and not repeat the execution of the first function service. Instead, it can directly return feedback information indicating that the scheduling execution is complete to the FaaS scheduler. This saves processing efficiency. If the current function service is not a first function service (i.e., a function service not executed in the first-level scheduling), then the lightweight function service variant of the current function service deployed on the ultrasound device layer is called for processing.

[0078] In some examples, if the execution result of the first function service (i.e., the function service already executed in the first-level scheduling) in the ultrasound equipment layer does not meet the preset requirements, then in the first-level scheduling, the ultrasound equipment layer triggers the data compression function service deployed in the ultrasound equipment layer based on the micro-function-as-a-service proxy to compress the input data of the first function service whose execution result does not meet the preset requirements, and uploads the compressed data to the edge service layer in advance. Thus, it awaits scheduling the first function service to the edge service layer for execution in the second-level scheduling.

[0079] It should be understood that if the execution result of the first function service in the first-level scheduling of the ultrasound equipment layer does not meet the preset requirements, then the target execution level determined by the FaaS scheduler of the edge service layer for the first function service in the cross-layer scheduling decision processing will necessarily not be the ultrasound equipment layer, but can be the edge service layer. Thus, in the second-level scheduling stage, if the current function service determined by the FaaS scheduler is the first function service and the corresponding target execution level is the edge service layer, the standard version function service variant corresponding to the first function service can be called in the edge service layer to perform corresponding processing on the pre-uploaded compressed data.

[0080] like Figure 5As shown, the ultrasound equipment layer deploys lightweight function service variants corresponding to data acquisition, preprocessing, quality control, and data compression functions. In the first-level scheduling, the built-in micro-function-as-a-service agent in the ultrasound equipment layer can, based on local power levels, data content, real-time requirements (i.e., task real-time requirement thresholds), and privacy levels, call the lightweight function service variants corresponding to some of the function services (e.g., data acquisition, preprocessing, and quality control functions) as needed. Taking the call to the lightweight function service variant of the quality control function service as an example, if its quality control effect does not meet the requirements, the data compression function service is triggered in advance to compress the preprocessed ultrasound video stream and upload the compressed data to the function-as-a-service scheduler in the edge service layer. This allows the function-as-a-service scheduler in the edge service layer to schedule the standard version of the quality control function service variant in the edge service layer to perform more accurate quality control processing based on the uploaded compressed data during the second-level scheduling phase. Furthermore, during the secondary scheduling phase, if other function services need to be executed in the ultrasound equipment layer, the function scheduler of the edge service layer can issue scheduling instructions to the ultrasound equipment layer to call the corresponding lightweight function service variant for processing.

[0081] In some embodiments, when the target execution layer is an edge service layer and the source of the input data corresponding to the current function service is an ultrasound device layer, the ultrasound device layer is triggered to call a data compression function service to compress the input data before uploading it to the edge service layer. After receiving and decompressing the compressed data, the standard version of the function service variant deployed on the edge service layer is called to process the decompressed data. For example, the microfunction-as-a-service proxy can dynamically call a lossless compression function service or a lossy compression function service to compress the input data before uploading it to the edge service layer based on the current network status. For instance, if network congestion is detected, it automatically switches to a high compression ratio algorithm to prioritize real-time transmission.

[0082] like Figure 6 As shown, the edge service layer can deploy standard variants of function services such as quality control function services (i.e., function services for quality control and inspection of ultrasound data), section selection function services, auxiliary diagnostic analysis function services, report generation function services, and privacy desensitization function services. These are only illustrative examples and should not be considered limiting. After the ultrasound equipment layer uploads the input data, the function scheduler of the edge service layer can determine whether the function service corresponding to the input data is executed in this layer. If so, it calls the corresponding standard variant of the function service in the edge service layer to perform the corresponding processing on the input data. If not, it desensitizes the input data and then uploads the desensitized data to the cloud service layer.

[0083] For example, if the required function is a section selection function, the standard version of this function service variant is invoked to automatically identify and extract the best ultrasound image frame conforming to medical measurement standards from the ultrasound video stream (i.e., the ultrasound video stream after decompressing the input data uploaded from the ultrasound equipment layer), which can be denoted as a standard section image. If the next required function is an auxiliary diagnostic analysis function, the standard version of the auxiliary diagnostic analysis function service variant is invoked to further infer the extracted standard section image, achieving automatic detection of lesions within the section, contour segmentation, and structural measurement, thus obtaining image analysis results. If a report generation function service is further required, its standard version on the edge service layer is invoked to map the above image analysis results into structured text, assemble a preliminary report, and return the report to the ultrasound equipment layer. If the function-as-a-service scheduler determines that the preliminary generated report needs to be uploaded to the cloud service layer, it forcibly triggers the privacy desensitization function service to desensitize the report and upload the desensitized data to the cloud service layer.

[0084] In some embodiments, when the target execution layer is the cloud service layer and the source of the input data corresponding to the current function service is the edge service layer, the edge service layer is triggered to call the privacy-de-identifying function service to de-identify the input data before uploading it to the cloud service layer; after receiving the de-identified data, the high-precision function service variant of the current function service deployed on the cloud service layer is called to process the de-identified data.

[0085] like Figure 7 As shown, the cloud service layer deploys high-precision function service variants of the deep analysis and auxiliary diagnosis function service and the edge-side report analysis and optimization function service. If the current function service to be scheduled is the deep analysis and auxiliary diagnosis function service, after receiving the anonymized difficult case data uploaded from the edge service layer, the cloud service layer calls the high-precision function service variant deployed on the cloud service layer to perform secondary detection and analysis on the difficult case data, generating a deep detection result (this detection result is not a direct disease diagnosis result, but intermediate detection data used to assist doctors in disease diagnosis). Further, after the edge-side report analysis and optimization function service is scheduled, its corresponding high-precision function service variant is called to perform consistency analysis between the report uploaded from the edge service layer and the deep detection result to determine the detection bias, and then optimize the report from the edge service layer based on the detection bias. Regarding... Figure 7 The cloud service layer model shown in the diagram is a full lifecycle function service that is executed periodically, rather than being invoked for a specific current task. Therefore, it will not be discussed here, but will be described in detail later.

[0086] Figure 8 This is a schematic diagram illustrating the principle of a dynamic function service scheduling method for intelligent ultrasound detection and quality control in one embodiment. Figure 8As shown, various function service variants are deployed in the ultrasound equipment layer, edge service layer, and cloud service layer (for illustration, each function service variant is referred to as an abbreviation). The micro-functions, i.e., service proxies, built into the ultrasound equipment layer can perform first-level scheduling and trigger preprocessing, i.e., trigger the preprocessing function service to preprocess the ultrasound video stream acquired by the data acquisition function service. In addition, the ultrasound equipment layer can also extract task features for the current ultrasound detection task and report these task features to the edge service layer (not shown in the figure). Based on the reported task features and the collected resource status at each level (not shown in the figure), the edge service layer determines the final scheduling strategy and performs second-level scheduling. The scheduling path for each function service to be executed in the ultrasound detection task is one of three scheduling paths. If it is scheduling path 1, the function service should be executed locally in the edge service layer, so a scheduling instruction is sent to the standard version function service variant corresponding to the function service in the edge service layer. If it is scheduling path 2, the function service should be executed in the cloud service layer, so a scheduling instruction is sent to the cloud service layer to instruct the invocation of the high-precision function service variant corresponding to the function service to send the scheduling instruction. If it is scheduling path 3, the function service should be executed at the ultrasound equipment layer. Therefore, a scheduling instruction is sent to the ultrasound equipment layer to instruct the invocation of the lightweight function service variant corresponding to the function service.

[0087] In the above method, function services are used as the scheduling granularity. Based on the task characteristics of ultrasound detection tasks and the resource status of each level, the cloud-edge-device collaborative scheduling scheme of function services is intelligently and dynamically determined, which greatly improves the processing efficiency of ultrasound images.

[0088] In some embodiments, a prediction model calibration mechanism is introduced to adapt to dynamic environmental changes. The time cost predicted in step S221 is predicted by the execution latency prediction model, and the resource cost is predicted by the resource consumption prediction model.

[0089] The FaaS scheduler can collect the actual execution time (such as actual data transmission time, actual queuing computation time, or actual function cold start time) and / or actual cost (such as actual computing resource consumption cost or actual expense cost) of each function service at the corresponding target execution level. Based on the deviation between the actual execution time and the predicted time cost, it adjusts the model parameters of the execution latency prediction model. For example, if the actual data transmission time is consistently higher than the predicted value, it automatically increases the weight coefficient of the network bandwidth factor to make the prediction result at the next moment more consistent with the current network conditions. Furthermore, the FaaS scheduler can also adjust the model parameters of the resource consumption prediction model based on the deviation between the actual cost and the predicted resource cost.

[0090] In some embodiments, a cost calculation calibration mechanism is introduced to adapt to dynamic environmental changes. That is, when calculating the overall cost of executing some function services at candidate levels, the penalty weight corresponding to each level (equivalent to a level-based penalty weight) can be considered, and the overall cost of executing the function service at the corresponding level can be dynamically adjusted based on this penalty weight. This penalty weight can be determined based on the historical execution performance of the function service at each level. For example, if a function service 1 has a poor historical execution performance at a certain level, the penalty weight of that level is increased, so that when the function service 1 is rescheduled next time, it is forced to be executed preferentially at other levels to obtain better execution performance.

[0091] Taking the quality control function service as an example, if the target execution level of the quality control function service in this task is the ultrasound equipment layer, the actual quality control pass rate of the quality control function service executed on the ultrasound equipment layer can be collected. If it is found that the actual quality control pass rate executed on the ultrasound equipment layer is significantly lower than the historical benchmark (indicating that the terminal model has a large number of false alarms, affecting the efficiency of doctors), the penalty weight of the ultrasound equipment layer for the quality control function service can be automatically increased, thereby reducing the scheduling priority of the quality control function service executed on the ultrasound equipment layer. This will cause the overall cost of the quality control function service executed on the ultrasound equipment layer to surge in the next calculation, thereby forcing the quality control function service to be scheduled to the edge service layer or cloud service layer for execution, sacrificing a small amount of latency for higher accuracy.

[0092] like Figure 7 As shown, the cloud service layer also has a model lifecycle management function service, which will be described in more detail below. The cloud service layer can periodically call the model lifecycle management function service to update and train the model based on anonymized difficult case sample data, perform model evaluation and management to verify the performance metrics of the new model, and then optimize the model and function service version. That is, after the performance metrics of the new model reach the optimization target, function service version optimization is performed based on the updated new model to obtain new function service variants adapted to different levels. Then, the new version function service variants are distributed to update the function service variants at various levels in the scheduling system.

[0093] Specifically, the cloud service layer collects anonymized difficult-to-train case sample data uploaded by the edge service layer. For each model, the cloud service layer periodically updates the training performance of the model based on the difficult-to-train case sample data, resulting in a new version of the model file. Afterwards, the business logic code of the function service is not changed; instead, a rebuild or configuration update of the function service image is triggered. That is, the new version of the model file is packaged and encapsulated into the function service image, generating a new version of the function service image; the new version of the function service image includes function service variants adapted to different layers.

[0094] Furthermore, update commands are issued, causing the ultrasound device layer and edge service layer to update the corresponding function service variants from the new version of the function service image to their respective local environments. For example, through the OTA mechanism, differentiated distribution is performed based on the hardware attributes of the devices. That is, a standard function service update command is issued to the edge service layer, and a lightweight function service update command is issued to the ultrasound device layer. This allows the corresponding new version of the function service variants from the new version of the function service image to be selectively distributed to the edge service layer or the ultrasound device layer. The edge service layer or the ultrasound device layer then uses container restart or hot reloading technology to replace the old function service variants, ensuring that the updated function service variants match the local computing power. Through this mechanism, intelligent upgrades of function services can be completed solely through data-driven model iteration without changing the business logic code.

[0095] In some embodiments, to support the aforementioned large-scale concurrency and real-time response, this application has made the following multi-dimensional optimizations in its architecture design.

[0096] 1. FaaS containerization technology and elastic scaling:

[0097] To address the distinct peak and trough characteristics of ultrasound testing requests (such as peak outpatient hours), a container-based lightweight virtualization technology is employed to achieve millisecond-level elastic scaling of function services.

[0098] Specifically, each function service variant (obtained by model encapsulation) is an independent, stateless function service image. These images share the host operating system kernel, avoiding the traditional operating system boot process during virtual machine startup. When a sudden surge in ultrasonic testing requests occurs, the FaaS platform can leverage copy-on-write technology and a lightweight runtime to concurrently launch hundreds or thousands of function container instances within milliseconds, achieving real-time scaling of computing power. When the request volume decreases, the system automatically destroys idle containers to free up memory and GPU resources.

[0099] 2. Instance warm-up and cold start elimination based on time series analysis:

[0100] To address the "cold start" delay issue caused by the long loading time of AI models, this system introduces a proactive instance preheating mechanism. Specifically, this includes:

[0101] 2-1 Traffic Prediction: The system continuously collects function call frequency data from historical moments through distributed probes, and uses time series analysis algorithms to predict and model business traffic for a future period (such as the next 10 minutes), identifying traffic peaks such as outpatient peaks and shift handover periods.

[0102] 2-2 Preheating: Based on the prediction results, the Function as a Service scheduler triggers the creation instruction of function instances in advance before the traffic peak arrives, so as to preload the corresponding function service variants into the GPU memory of edge servers or cloud servers in the edge service layer, so that the function service variants adapted to each level are in a "hot standby" state.

[0103] 2-3 Dynamic Adjustment: When the predicted traffic arrives, the request is directly distributed to the pre-warmed (function service variant) instance for processing, thereby eliminating the several-second delay caused by model loading and container initialization, ensuring that doctors receive smooth real-time feedback during operation.

[0104] 3. Edge Intelligent Caching: In the edge service layer, edge servers are not only computing nodes but also caching centers for hot data. By building a caching proxy for container image repositories on the edge side, the ultrasonic device layer can directly pull the corresponding lightweight variants of the new version from the edge service layer for updates without consuming wide area network bandwidth.

[0105] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0106] Based on the same inventive concept, this application also provides a dynamic function service scheduling system for implementing the aforementioned intelligent ultrasound detection and quality control. The solution provided by this system is similar to the implementation scheme described in the above method. Therefore, the specific limitations of one or more embodiments of the dynamic function service scheduling system for intelligent ultrasound detection and quality control provided below can be found in the limitations of the dynamic function service scheduling method for intelligent ultrasound detection and quality control described above, and will not be repeated here.

[0107] In one embodiment, a dynamic function service scheduling system for intelligent ultrasound detection and quality control is provided. This system is used to execute the dynamic function service scheduling method for intelligent ultrasound detection and quality control described in this embodiment. Each module in this dynamic function service scheduling system for intelligent ultrasound detection and quality control can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0108] In one embodiment, a computer device is provided, which can be any device at any level in a scheduling system. The internal structure diagram of the computer device can be as follows: Figure 9 As shown, the computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a dynamic function service scheduling method for intelligent ultrasound detection and quality control.

[0109] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0110] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the embodiments of this application.

[0111] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the embodiments of this application.

[0112] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the embodiments of this application.

[0113] It should be noted that the user information (including but not limited to user device information, user attribute content, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0114] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory.

[0115] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. The above embodiments only illustrate several implementation methods of this application, but should not be construed as limiting the scope of this application. For those skilled in the art, several modifications and improvements can be made without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A dynamic function service scheduling method for intelligent ultrasound detection and quality control, characterized in that, The method is applied to a scheduling system, wherein the scheduling system comprises an ultrasound equipment layer, an edge service layer, and a cloud service layer; different function service variants are deployed for the same business logic at each layer; the method includes: The ultrasound equipment layer determines the set of function services to be executed corresponding to the current ultrasound detection task, extracts task features based on the ultrasound video stream acquired in real time for the ultrasound detection task, and reports the task features to the edge service layer; the set is a task chain formed by the various function services that need to be executed to complete the ultrasound detection task; the task features include at least one of the following: the data size of the ultrasound video stream, the real-time requirement threshold of the task, and the privacy level corresponding to the ultrasound detection task. The function service scheduler in the edge service layer performs cross-layer scheduling decision processing based on the task characteristics and the current resource status of each layer, so as to determine the target execution layer corresponding to each function service in the set from each layer; The ultrasound equipment layer performs first-level scheduling based on a built-in micro-function service agent to call lightweight function service variants corresponding to some of the function services on demand according to the local resource status to perform corresponding processing; wherein, the function service executed in the first-level scheduling is denoted as the first function service; the first function service includes at least one of data acquisition function service, preprocessing function service, quality control function service, or feedback function service; the micro-function service agent is used to run lightweight function service variants; The function-as-a-service scheduler determines the current function service to be executed sequentially according to the execution order of each function service in the set, generates a corresponding scheduling instruction for each current function service, and sends the scheduling instruction to the corresponding target execution level; the scheduling instruction is used to instruct the function service variant that calls the current function service at the target execution level to process the input data corresponding to the current function service; Specifically, if the target execution level corresponding to the current function service is the ultrasound equipment layer, and if the current function service is the first function service, the ultrasound equipment layer ignores the scheduling instruction and directly returns feedback information indicating that the scheduling execution is complete to the function-as-a-service scheduler; if the current function service is not the first function service, the lightweight function service variant of the current function service deployed on the ultrasound equipment layer is invoked for processing.

2. The method of claim 1, wherein, If the execution result of the first function service in the first-level scheduling of the ultrasound equipment layer does not meet the preset requirements, the target execution level determined for the first function service in the cross-layer scheduling decision processing is not the ultrasound equipment layer. The method further includes: In the first-level scheduling, the ultrasound equipment layer triggers a data compression function service based on the micro-function service agent to compress the input data of the first function service whose execution result does not meet the preset requirements, and uploads the compressed data to the edge service layer in advance. If the current function service determined by the function-as-a-service scheduler is the first function service and the corresponding target execution level is the edge service layer, the standard version function service variant corresponding to the first function service is called in the edge service layer to perform corresponding processing on the pre-uploaded compressed data.

3. The method of claim 1, wherein, The first function service is a preprocessing function service; The ultrasound equipment layer performs first-level scheduling based on a built-in micro-function-as-a-service proxy to call lightweight function service variants corresponding to some of the function services on demand, according to local resource status, to perform corresponding processing, including: The ultrasound device layer performs first-level scheduling based on the micro-function as a service agent to call the lightweight function service variant corresponding to the preprocessing function service according to the local resource status, and preprocesses each ultrasound image in the ultrasound video stream to obtain the preprocessed ultrasound video stream; the preprocessed ultrasound video stream is used as input data for the next function service of the preprocessing function service.

4. The method of claim 1, wherein, After issuing the scheduling instruction to the corresponding target execution level, the method further includes: When the target execution layer is the edge service layer and the source of the input data corresponding to the current function service is the ultrasound equipment layer, the ultrasound equipment layer is triggered to call the data compression function service to compress the input data and then upload it to the edge service layer; after receiving and decompressing the compressed data, the standard version function service variant of the current function service deployed on the edge service layer is called to process the decompressed data; When the target execution layer is the cloud service layer and the source of the input data corresponding to the current function service is the edge service layer, the edge service layer is triggered to call the privacy de-identification function service to de-identify the input data and then upload it to the cloud service layer; after receiving the de-identified data, the high-precision function service variant of the current function service deployed on the cloud service layer is called to process the de-identified data.

5. The method of claim 1, wherein, The resource status includes the computing load of each level and the network link status between each level.

6. The method of claim 5, wherein, The function-service scheduler in the edge service layer performs cross-layer scheduling decision processing based on the task characteristics and the current resource status of each layer, to determine the target execution layer corresponding to each function service in the set, including: For each of the function services, the function-as-a-service scheduler is used to predict the time cost, resource cost, and privacy risk penalty value required for the function service to be executed at each corresponding candidate level, based on the task characteristics and the resource status. The time cost, resource cost, and privacy risk penalty value are weighted and summed to obtain the comprehensive cost of each function service executed at each corresponding candidate level; To meet real-time constraints and minimize the total task cost, the target execution level for each function service is determined. The real-time constraints include that the total time of the ultrasound detection task is less than or equal to a preset real-time requirement threshold. The total task cost is obtained by summing the comprehensive costs corresponding to each function service.

7. The method of claim 6, wherein, The time cost is predicted by an execution latency prediction model; the resource cost is predicted by a resource consumption prediction model; the method further includes: Collect the actual execution time and / or actual cost of each function service at the corresponding target execution level; The model parameters of the execution latency prediction model are adjusted based on the deviation between the actual execution time and the predicted time cost, and / or the model parameters of the resource consumption prediction model are adjusted based on the deviation between the actual consumption cost and the predicted resource cost.

8. The method according to any one of claims 1 to 7, characterized in that, The cloud service layer pre-trains models corresponding to each function service; the different function service variants adapted to each level for each function service are generated by performing different transformation processes on the model files of the model corresponding to the function service. The method further includes: The cloud service layer collects anonymized difficult case sample data uploaded by the edge service layer; For each of the models, the cloud service layer periodically updates and trains the model based on the difficult case sample data to obtain a new version of the model file; The new version of the model file is packaged and encapsulated into the function service image to generate a new version of the function service image; the new version of the function service image includes function service variants adapted to different levels; An update command is issued, causing the ultrasound equipment layer and the edge service layer to update the corresponding function service variants in the new version of the function service image to their respective local locations.

Images