Deployment method, device and system for specific detection model in medical scene

Abstract

The invention provides a deployment method, device and system for a specific detection model in a medical scene, is used for carrying out localized deployment on a mature medical examination model in a hospital, specifically combines a container deployment technology and a container arrangement technology, provides corresponding model calling services for medical personnel more conveniently, efficiently and safely, and improves the medical examination efficiency. Dual requirements of data security compliance and model real-time calling can be effectively balanced, a safe, extensible and high-reliability operation platform is provided for actual landing of a digital algorithm, external dependence is greatly reduced, and the real-time performance of the digital algorithm is improved. The method provides possibility for rapid verification and iterative algorithm of hospital informatization departments and researchers in a clinical environment, achieves intelligent medical assistance of closed-loop, safe, controllable and efficient optimization in a hospital, has large-scale popularization value, and provides robust AI technical support for various clinical diagnosis, inspection analysis and scientific research work.

Description

Technical Field

[0001] The present application relates to the field of medical technology, and specifically to a deployment method, device, and system for a specific detection model in a medical scenario. Background Art

[0002] In recent years, the application of machine learning in the field of medical testing is in a rapid development stage. A large number of studies have shown that algorithm models based on deep learning and traditional machine learning can effectively assist in medical image recognition, test data analysis and clinical diagnostic decision-making.

[0003] However, existing technologies still have limitations in clinical practice. From the perspective of medical staff, the current practice requires manual transcription of test or diagnosis results and other data to a specific website, which is then processed and analyzed by the server. Although this working model allows medical staff to quickly access and use new algorithm models to a certain extent, it faces the following problems:

[0004] 1. Medical data needs to be transmitted through external networks, increasing the risk of data leakage and privacy compliance;

[0005] 2. Data formats and interfaces often lack unified standards, requiring repeated manual entry to complete the analysis process;

[0006] 3. The hospital information system has not yet formed an organic connection with the external algorithm platform, making it difficult to ensure the real-time and reliability of algorithm calls.

[0007] It can be seen that the existing use of medical testing models still has safety risks and inconvenience to a certain extent, which affects the application value. Summary of the Invention

[0008] This application provides a deployment method, device and system for specific detection models in medical scenarios, which are used to locally deploy mature medical test models in the hospital, and specifically combines container deployment technology and container orchestration technology to provide medical staff with corresponding model calling services more conveniently, efficiently and securely. It can effectively balance the dual needs of data security compliance and real-time model calling, and provide a safe, scalable and highly reliable operating platform for the actual implementation of digital algorithms. While greatly reducing external dependence, it also provides the possibility for hospital information departments and scientific researchers to quickly verify and iterate algorithms in clinical environments, and achieve "in-hospital closed-loop, safe and controllable, and efficient optimization" intelligent medical assistance, which has large-scale promotion value and can be widely used in hospital groups, regional medical alliances and larger-scale medical big data platforms, providing robust artificial intelligence (AI) technical support for various clinical diagnosis, test analysis and scientific research work.

[0009] In a first aspect, the present application provides a method for deploying a specific detection model in a medical scenario, the method comprising:

[0010] Determine the target detection model that needs to be called on the hospital's external platform during the processing of the hospital's medical services. Specifically, the target detection model is a machine learning model for medical testing.

[0011] Obtain the corresponding digital algorithms and dependency libraries of the target detection model;

[0012] The digital algorithm and dependent libraries are packaged uniformly into the target container image corresponding to the target detection model, where the target container image includes the execution script, runtime environment, and algorithm weight file;

[0013] In an internal hospital online platform pre-built for deploying different detection models, multiple container copies of the container image are deployed through a container orchestration tool. The internal hospital online platform is built on a private cloud platform or a physical server cluster. The internal hospital online platform is deployed with at least one of a GPU cluster and a CPU cluster. The model service running node and the core database of the internal hospital online platform are isolated. The internal hospital online platform performs model configuration through at least one of a hospital-specific VPN and a physical dedicated line without opening the patient information port to the outside world.

[0014] When an inference request about a target detection model is received, the inference request is distributed to an adapted target container replica among multiple container replicas through a load balancer, so as to provide a call service of the target detection model to the requester of the inference request through the target container replica, where the requester is an authorized user within the hospital.

[0015] In a second aspect, the present application provides a deployment device for a specific detection model in a medical scenario, the device comprising:

[0016] A determination unit is used to determine a target detection model that needs to be called on an external platform of the hospital during the processing of medical services involved in the hospital, wherein the target detection model is specifically a machine learning model for medical testing;

[0017] An acquisition unit, used to obtain the corresponding digital algorithms and dependency libraries of the target detection model;

[0018] A packaging unit is used to package the digital algorithm and dependent libraries into a target container image corresponding to the target detection model, where the target container image includes an execution script, a runtime environment, and an algorithm weight file;

[0019] A deployment unit is used to deploy multiple container copies of a container image through a container orchestration tool in an internal hospital online platform that is pre-built for deploying different detection models, wherein the internal hospital online platform is built on a private cloud platform or a physical server cluster, the internal hospital online platform is deployed with at least one of a GPU cluster and a CPU cluster, the model service running node and the core database of the internal hospital online platform are isolated, and the internal hospital online platform performs model configuration through at least one of a hospital-specific VPN and a physical dedicated line without opening the patient information port to the outside world;

[0020] The response unit is used to distribute the inference request about the target detection model to the adapted target container copy among multiple container copies through the load balancer when receiving the inference request, so as to provide the target detection model calling service to the requester of the inference request through the target container copy, wherein the requester is an authorized user within the hospital.

[0021] In a third aspect, the present application provides a deployment system for a specific detection model in a medical scenario, which is used to execute the method provided in the first aspect of the present application.

[0022] In a fourth aspect, the present application provides a computer-readable storage medium, which stores a plurality of instructions suitable for loading by a processor to execute the method provided in the first aspect of the present application.

[0023] From the above content, it can be concluded that this application has the following beneficial effects:

[0024] Under the program goal of further improving the application value of medical test models, this application locally deploys mature medical test models in the hospital, and specifically combines container deployment technology and container orchestration technology to provide medical staff with corresponding model calling services more conveniently, efficiently and securely. It can effectively balance the dual needs of data security compliance and real-time model calling, and provide a safe, scalable and highly reliable operating platform for the actual implementation of digital algorithms. While greatly reducing external dependence, it also provides the possibility for hospital information departments and scientific researchers to quickly verify and iterate algorithms in clinical environments, and achieve "closed-loop, safe and controllable, and efficient optimization" intelligent medical assistance. It has large-scale promotion value and can be widely used in hospital groups, regional medical alliances and larger-scale medical big data platforms, providing robust AI technology support for various clinical diagnosis, test analysis and scientific research work. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0026] Figure 1 A flowchart of a method for deploying a specific detection model in a medical scenario in this application;

[0027] Figure 2 A schematic diagram of a structure of a deployment device for a specific detection model in a medical scenario of this application;

[0028] Figure 3 This is a structural diagram of the processing equipment involved in the deployment system of a specific detection model in the medical scenario of this application. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0030] The terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or modules is not necessarily limited to those steps or modules clearly listed, but may include other steps or modules that are not clearly listed or that are inherent to these processes, methods, products or devices. The naming or numbering of steps in this application does not mean that the steps in the method flow must be executed in the time / logical sequence indicated by the naming or numbering. The process steps that have been named or numbered can be changed in the execution order according to the technical purpose to be achieved, as long as the same or similar technical effects can be achieved.

[0031] The division of modules in this application is a logical division. In actual application, there may be other division methods. For example, multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection between modules can be electrical or other similar forms, which are not limited in this application. Moreover, the modules or submodules described as separate components may or may not be physically separated, may or may not be physical modules, or may be distributed into multiple circuit modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this application.

[0032] Before introducing the deployment method of the specific detection model in the medical scenario provided by this application, we first introduce the background content involved in this application.

[0033] The deployment method, device and computer-readable storage medium of specific detection models in medical scenarios provided in this application can be applied to processing equipment for local deployment of mature medical test models in the hospital, and specifically combine container deployment technology and container orchestration technology to provide medical staff with corresponding model calling services more conveniently, efficiently and securely. It can effectively balance the dual needs of data security compliance and real-time model calling, and provide a safe, scalable and highly reliable operating platform for the actual implementation of digital algorithms. While greatly reducing external dependence, it also provides the possibility for hospital information departments and scientific researchers to quickly verify and iterate algorithms in clinical environments, and achieve "closed-loop, safe and controllable, and efficient optimization" intelligent medical assistance. It has large-scale promotion value and can be widely used in hospital groups, regional medical alliances and larger-scale medical big data platforms, providing robust AI technology support for various clinical diagnosis, test analysis and scientific research work.

[0034] The deployment method of a specific detection model in a medical scenario mentioned in this application may be executed by a deployment device of a specific detection model in a medical scenario, or a deployment system of a specific detection model in a medical scenario that integrates the deployment device of the specific detection model in the medical scenario. The deployment device of the specific detection model in the medical scenario may be implemented in hardware or software. The deployment system of the specific detection model in the medical scenario may be specifically built based on different types of processing devices such as servers or physical hosts, and may be specifically configured in the form of a device cluster. The device cluster may further involve user equipment (UE), and the UE may specifically be a terminal device such as a smart phone, tablet computer, laptop computer, desktop computer or personal digital assistant (PDA). The specific device types and device deployment forms involved in the deployment system of the specific detection model in the medical scenario can be flexibly configured according to actual needs, and this application does not make specific limitations.

[0035] Next, we will introduce the deployment method of the specific detection model in the medical scenario provided by this application.

[0036] First, see Figure 1 , Figure 1 A flow chart of a method for deploying a specific detection model in a medical scenario of the present application is shown. The method for deploying a specific detection model in a medical scenario provided by the present application may specifically include the following steps S101 to S105:

[0037] Step S101: determining a target detection model that needs to be called on an external platform during the processing of the hospital's medical services, wherein the target detection model is specifically a machine learning model for medical testing;

[0038] It can be understood that this application focuses on relevant medical test models deployed outside the hospital, that is, by a third party. The core of the solution is to locally deploy such medical test models within the hospital. To this end, we can first determine such medical test models as the target detection model mentioned here, and promote the specific local deployment work within the hospital through this application solution.

[0039] Among them, it can be understood that the determination processing of the target detection model can be either manually specified / determined, or it can be autonomously screened / located by the system during the online operation or recording of related medical services in the hospital. For this autonomous processing method, it can be specifically processed in combination with whether it involves a request to access the external network to call a third-party medical inspection model.

[0040] It is worth noting that this type of target detection model (or algorithm) is a machine learning model configured based on AI technology, mainly a deep learning model. These models themselves are generally mature models that can normally complete the corresponding medical examination goals. These models can directly involve clinical auxiliary diagnostic functions (such as disease risk value, disease classification, diagnosis and treatment prompts, etc.), and can also involve image super-resolution processing, abnormal image recognition (no need to diagnose specific diseases, only need to screen out abnormal images) and other different functions, which are linked to the hospital's diversified medical services.

[0041] Taking the liver cancer risk assessment model as an example, the target detection models involved in this application can specifically be two mature models: the GAAD model and the ASAP model.

[0042] Step S102: Obtain the digital algorithm and dependency library corresponding to the target detection model;

[0043] Corresponding to the localized deployment requirements of the target detection model, this application can obtain the corresponding digital algorithms and dependent libraries of the target detection model. Both are files involved in running the target detection model, or in other words, by running the digital algorithms and dependent libraries that carry the corresponding detection logic, the purpose of running the target detection model can be achieved.

[0044] The obtained digital algorithms and dependent libraries can provide effective data support for subsequent container deployment.

[0045] It is understandable that the acquisition and processing of digital algorithms and dependent libraries here can be either manual entry, or obtained from the external platform involved in the original hospital model call process, or obtained from other places. All of these are possible.

[0046] Step S103: Packaging the digital algorithm and dependent libraries into a target container image corresponding to the target detection model, wherein the target container image includes an execution script, a runtime environment, and an algorithm weight file;

[0047] It can be understood that deploying corresponding application services based on containers may involve the configuration of container images. In this regard, this application can uniformly package the digital algorithms and dependent libraries obtained previously to form the corresponding target container images.

[0048] Among them, the target container image may include data / files such as execution scripts, runtime environment (such as Python and frameworks such as TensorFlow or PyTorch) and algorithm weight files, depending on the operating requirements of the container. These files belong to the existing technology of deploying corresponding application services through container technology, so we will not elaborate on them here.

[0049] Step S104: Deploy multiple container copies of the container image using a container orchestration tool in an internal hospital online platform pre-built for deploying different detection models. The internal hospital online platform is built on a private cloud platform or a physical server cluster, and is deployed with at least one of a GPU cluster and a CPU cluster. The model service running node and the core database of the internal hospital online platform are isolated from each other. The internal hospital online platform performs model configuration through at least one of a hospital-specific VPN and a physical dedicated line without opening the patient information port to the outside world.

[0050] It is understandable that in actual situations involving numerous or different detection models, this application does not deploy all models that need to be locally deployed through one container, nor does it deploy each model through its own model. Instead, each model has multiple container copies, so that a certain model can simultaneously meet the model calling requirements involved in different businesses / services.

[0051] Taking the ASAP model as an example, you can deploy n container replicas, namely ASAP-1, ASAP-2, ...ASAP-n, and distinguish them by labels or routing rules. This allows you to run n ASAP models simultaneously to satisfy the ASAP model call service.

[0052] Different types of target detection models and their numerous container copies can be managed by corresponding container orchestration tools. Container orchestration tools can be understood as automated tools used to deploy, scale, run, and maintain a large number of containerized applications.

[0053] As an example, Kubernetes (k8s for short) can be used in actual applications. Of course, other types of container management tools on the market can also be used. Alternatively, further optimization and improvement can be made based on existing container management tools, or even self-developed novel container management tools can be used. This can be adaptively configured according to actual needs.

[0054] These container replicas are deployed on a pre-planned internal online platform within the hospital. Pre-configuration of this platform involves setting up an intranet environment specifically for pancreatic cancer.

[0055] 1) Establish a computing and storage resource pool within the hospital. Preferably, use a private cloud platform or physical server cluster with a GPU or CPU cluster, and install an operating system and orchestration tools that support containerized deployment.

[0056] 2) Isolate the model service running nodes from the core database through network partitioning (demilitarized zone (DMZ) or intranet zone) and firewall configuration. This ensures that data access flow is limited to the authorized scope. At the same time, model updates can be performed through the hospital's dedicated virtual private network (VPN) or physical dedicated line without opening the patient information port to the outside world.

[0057] In this way, through the above settings, while effectively building the hospital's internal online platform, it can effectively provide a good container operating environment, taking into account high performance, data security and system robustness.

[0058] Specifically, through the container management of the container orchestration tool, corresponding load balancing management can be applied between multiple container copies, thereby effectively responding to outpatient peaks and batch screening scenarios and meeting the high-performance model call service requirements.

[0059] Step S105: When an inference request regarding a target detection model is received, the inference request is distributed to an adapted target container replica among multiple container replicas through a load balancer, so as to provide a call service of the target detection model to the requester of the inference request through the target container replica, wherein the requester is an authorized user within the hospital.

[0060] It is understandable that within the hospital, relevant medical staff can conveniently initiate an inference request to the system about the current situation regarding the target detection model deployed on the platform through the relevant equipment deployed on site or their own UE, so as to prompt the system background side to call the target detection model run by the container copy based on the inference request to perform corresponding model processing. After the model processing is completed, the corresponding inference results can be fed back, and medical staff can obtain the required inference result data support in the first time, which is helpful to assist clinical diagnosis or other data applications, so as to promote the development of high-quality related medical services with high-quality data assistance.

[0061] As for the request allocation involved, or the load balancing management imposed by the load balancer, in layman's terms, it is to match the adapted target container copy with the inference request currently required to be allocated based on the quantified working conditions of each container copy corresponding to the target detection model (mainly involving aspects such as available idle computing resources, request queues, and waiting times). In some cases, the request conditions involved in the inference request currently required to be allocated can be further quantified (such as the required computing resources, required data processing volume, required processing time, etc.) to achieve better adaptation on both sides. In this way, the load balancing management can make the load of each container copy tend to be balanced, and the overall request response efficiency and resource utilization can be improved as much as possible.

[0062] At the same time, this application also makes further restrictions on the requester. The requester who initiates the reasoning request should specifically be an authorized user within the hospital. This means that the reasoning request involves the identification of whether the user is an in-hospital user and whether the user is an authorized user during the initiation process, or after receiving the request, it involves judging whether the user is an in-hospital user and whether the user is an authorized user based on the information carried by the request. This eliminates the possibility of outsiders or unauthorized users within the hospital accidentally or improperly initiating reasoning requests, and ensures in detail that the data access flow is limited to the authorized scope, thereby continuing to ensure the security of data / system.

[0063] Furthermore, in terms of data exchange and security compliance, the system of the present application also has corresponding optimization settings. Correspondingly, as an exemplary embodiment, the method of the present application may also include:

[0064] 1) At the data input and output interfaces, API protocols that comply with preset industry standards are used, where preset industry standards include HL7 and FHIR.

[0065] Among them, the FHIR standard is a new standard that was born after the continuous updating and development of successive HL7 standards.

[0066] The API protocol settings here that comply with HL7 standards, FHIR standards, and even other industry standards will help achieve seamless communication between the application system and the hospital's existing hospital information system (HIS), laboratory information management system (LIS), or physical examination system, and other different types of business systems, helping to more smoothly integrate the application solution into hospital application scenarios.

[0067] 2) Encrypt transmission and storage of sensitive data, and set differentiated permissions for users of different granularities;

[0068] It is understandable that for sensitive data such as test results, medical information, or patient personal information that is confirmed to be sensitive, encrypted transmission and storage can help improve the security of data and systems.

[0069] Access control based on different roles can provide more flexible differentiated management for users of different granularities (such as different departments, different sections, and different individual users), and open up the calling functions of different models with different permissions, which facilitates flexible and controllable management.

[0070] 3) The data written into the model container for processing is first desensitized to remove privacy content, where privacy content includes name and ID number;

[0071] The private content may overlap with the sensitive data mentioned above, or it may be completely related to the personal privacy of the user or patient. Therefore, the focus is on desensitizing the private content related to personal privacy (which can also be understood as privacy minimization) to protect personal privacy. Only the necessary feature field content required for model reasoning is retained. This can also further simplify the data flow on the data transmission link, which is also beneficial for network transmission.

[0072] 4) Retain the model call log, where the retained content includes the request time, requesting user ID and processing results.

[0073] The retention or evidence processing based on the content of the work log is easy to understand and helps to trace back related situations afterwards. Typically, it can facilitate post-audit and compliance inspections.

[0074] Furthermore, the system of the present application has corresponding optimization settings for containerized deployment and load management. Correspondingly, as an exemplary embodiment, the method of the present application may further include:

[0075] 1) Adaptively scale the number of container replicas corresponding to the target detection model based on past, current, and predicted inference request volumes;

[0076] It is understandable that for a single model, when multiple container copies are involved to provide elastic model call services, in terms of details, the number of container copies can be reduced or expanded by adapting to the corresponding situation through elastic scaling management for container copies, thereby taking into account both resource utilization and model inference performance.

[0077] Among them, the elastic scaling processing of the number of container copies can be carried out based on the amount of inference requests involved in multiple container copies, and specifically can involve past, current and predicted inference request amounts. It can be seen that this application can also involve predictive processing of future inference request amounts based on past and current inference request amounts (as a kind of time series data), so as to better prepare for the elastic scaling work of the number of container copies in an advance planning manner.

[0078] 2) Remove the failed container instance identified by the health check and pull up a new container instance to replace it.

[0079] In addition, it can also involve health checks on container replicas to identify container replicas with abnormal working conditions, automatically remove them, and pull up new container instances to avoid service interruptions and ensure service continuity.

[0080] Furthermore, the system of the present application has corresponding optimization settings for model reasoning services and interface integration. Correspondingly, as an exemplary embodiment, the method of the present application may also include:

[0081] 1) Using RESTful or gRPC methods, a unified inference service application programming interface (API) is exposed to receive inference requests.

[0082] Understandably, by providing a unified API through methods / forms such as RESTful or gRPC, medical staff can easily initiate structured data requests including reasoning requests without having to involve the original manual transcription operations, and the data will automatically flow between hospital systems. Medical staff only need to initiate reasoning requests and view results in a regular workstation interface, eliminating the tedious steps of switching websites or manually copying data, further improving clinical efficiency and reducing the risk of input errors.

[0083] 2) For inference requests whose number of requests or data volume exceeds the corresponding threshold, block processing or queue processing is performed;

[0084] Understandably, when faced with a large number or even massive inference requests, block processing or queue processing can be performed based on the batch mode supported by the system, which can achieve a more orderly request processing effect and ensure the normal operation and efficiency of the system as a whole.

[0085] Among them, whether to trigger block processing or queue processing can also be judged in combination with the corresponding threshold of the number of requests or the corresponding threshold of the data volume. Obviously, if the situation exceeds the normal range or meets the trigger range, block processing or queue processing can be triggered.

[0086] 3) Perform security control operations corresponding to the inference request through the gateway or proxy layer, where security control operations include identity authentication, access token distribution, and logging.

[0087] It can be understood that security control operations such as authentication, access token distribution and logging can achieve good security control effects at the gateway or proxy layer in the network architecture and ensure good network security performance in the network architecture.

[0088] Furthermore, the system of the present application has corresponding optimization settings for model iteration and AB testing mechanisms. Correspondingly, as an exemplary embodiment, the method of the present application may further include:

[0089] 1) For the target detection model, perform multi-version coexistence processing on container replicas, dynamically distribute traffic based on the multi-version coexistence processing, and run new version container replicas in parallel under the test branch, wherein clinical services are not interrupted during the model version update process;

[0090] It is understandable that even the same target detection model, such as the ASAP model, can involve multiple different versions. This corresponds to the possibility that the model algorithm may have different versions in actual situations due to factors such as technical optimization or commercial promotion. When there are significant differences in model performance, it is also possible to use different segmented application scenarios.

[0091] In this regard, different versions of target detection models can be deployed at the same time, that is, multi-version coexistence processing of container copies can be carried out to meet the diverse application needs in actual situations. On this basis, dynamic traffic distribution processing can be carried out under load balancing management to achieve good model call effects between different versions of target detection models.

[0092] In addition, multi-version coexistence processing maintains the operation of old and new version models at the same time (new version container copies run in parallel under the test branch), which is also suitable for the trial of new version models, reducing the operational costs involved when the new version model has anomalies or requires version rollback. At the same time, it also ensures that the model call service can be provided normally during this period, will not interrupt clinical services, and ensure the continuity and robustness of business / systems.

[0093] 2) Using AB testing (or A / B testing), some inference requests are routed to the new version container replica for response processing, and the model processing effect is monitored in real time. The model processing effect includes accuracy, sensitivity, specificity, response latency, and error rate.

[0094] It is understandable that introducing AB testing during the normal operation of the model call service will help better test whether the new version model / container copy can normally meet local business needs and whether it has better performance under the multi-version coexistence processing mechanism.

[0095] The model processing effect or model evaluation indicators may specifically involve a series of indicators such as accuracy, sensitivity, specificity, response delay, error rate and ROC curve, which provide important basis for model performance optimization and operation and maintenance management. These indicators can usually be existing indicators. Of course, improved or even novel indicators can also be used.

[0096] 3) If the model processing effect meets the preset excellent performance conditions, the request access volume of the new version container copy will be gradually expanded. If the model processing effect does not meet the preset excellent performance conditions or an abnormality occurs, the new version container copy will be rolled back to the old version container copy.

[0097] It can be understood that under AB testing, if it is determined that the new version model / container copy meets the preset excellent performance conditions and has better performance, the request access volume and traffic processing volume of the new version model / container copy can be gradually expanded, so as to gradually increase the proportion of the new version model / container copy in the model call service involved in the target detection model, and further promote its application; on the contrary, if the performance is poor, a version rollback is required, even if it is rolled back to the old version model / container copy to provide model call services.

[0098] In addition, in the model version upgrade process involved above, the corresponding entire upgrade process can also be limited to being carried out only in the hospital's internal network environment, providing a smooth and robust model call service without downtime or interruption of clinical services.

[0099] As an example, in terms of model testing, this application solution is deployed for verification in an intranet environment. Specifically, the following are performed:

[0100] Experimental environment: Set up a server and Kubernetes orchestration cluster, and deploy GAAD and ASAP model images as multiple copies;

[0101] Data source: 600 real case and physical examination data from the hospital HIS system were selected to simulate real-time outpatient call and physical examination batch analysis scenarios;

[0102] Experimental process: Real-time outpatient call: simulate the doctor side to initiate 20 requests at a time, and measure the average response time, success rate and resource consumption;

[0103] Batch analysis: Import 1,000 physical examination data sets and submit them to the inference API at once to observe the total running time and automatic container expansion.

[0104] A / B testing: Deploy the new version of the ASAP model in a test environment, receiving only 10% of the traffic, and compare and evaluate its accuracy and response performance.

[0105] Experimental results show that: 1. Compared with the method of uploading data to the external network for processing, the solution of the present invention can reduce the average response time from several seconds to milliseconds and remain stable under high-concurrency requests, with an order of magnitude advantage; 2. Container automatic scaling can reduce the total processing time by nearly 40% in batch analysis scenarios, avoiding queue blockage or service crashes; 3. During the A / B testing process, the accuracy of the new version model increased by about 5%. If an anomaly is found, it can be instantly rolled back to the old version, and clinical use is not affected.

[0106] Furthermore, in terms of visualization and result writing back, the system of the present application also has corresponding optimization settings. Correspondingly, as an exemplary embodiment, the method of the present application may also include:

[0107] 1) Structuring and visualizing the processing results of the target detection model regarding the reasoning request to output corresponding text descriptions, charts, risk radar charts or risk bar charts, and involving risk warnings or recommended measures;

[0108] It is understandable that the system may also involve a variety of result display methods to meet the result display needs in different situations, which may specifically involve specific display schemes in the form of text descriptions, risk radar charts or risk bar charts.

[0109] In addition, in terms of the result display content, it can also involve risk warnings or recommended measures. In terms of details, different degrees of importance can be further highlighted through different presentation colors to form a more diverse and practical result display effect.

[0110] 2) When the requesting party is an outpatient and emergency physician workstation (outpatient physician workstation and emergency physician workstation) or an inpatient physician workstation, the processing results are written back to the corresponding interface service of the corresponding HIS system;

[0111] It is understandable that automatically pushing the results to doctors in outpatient and emergency departments and inpatient departments will help doctors to view the results immediately or even in real time, and make judgments based on clinical experience, further ensuring high-quality disease diagnosis.

[0112] 3) When there are objections to the processing results, receive data notes, annotation information, and correction information written by medical staff;

[0113] It is understandable that when it comes to some special situations, the processing results given by the target detection model may not be ideal or there may be objections. In this case, authorized medical staff can enter at least one of the data notes, annotation information and correction information to serve as a reminder or correction. At the same time, this once again demonstrates that the model call service provided by the target detection model plays a data-assisted role.

[0114] 4) The data annotation, annotation information and correction information corresponding to the inference request are stored in the database along with the processing results.

[0115] As for the data remarks, annotation information and correction information entered by the medical staff in the above 3), they can be stored in the database together with the corresponding or targeted processing results for retrospective analysis and follow-up.

[0116] Furthermore, in terms of retraining and fine-tuning, the system of the present application also has corresponding optimization settings. Correspondingly, as an exemplary embodiment, the method of the present application may also include:

[0117] The data comments, annotation information, and correction information written by medical staff on the processing results are used as training samples to retrain the target detection model locally or perform parameter fine-tuning to generate a new version container copy of the updated target detection model, which coexists with the old version container copy for multi-version coexistence.

[0118] It can be seen that in addition to directly transplanting the original off-site or third-party target detection model to the hospital for local deployment, and providing convenient, efficient and secure model calling services through container management, this application can also perform adaptive or personalized model optimization work locally to enable the locally deployed target detection model to be further adapted to the specific application scenarios in the hospital in terms of refinement or depth, so as to obtain better detection effects in the specific application scenarios.

[0119] As for the samples of the model optimization work here, the data notes, annotation information and correction information written by the medical staff involved in the previous objections to the processing results of the target detection model can be used (of course, they can also be written for other situations and other application needs). These real follow-up and feedback information provide valuable on-site / actual data for the local optimization work of the model, and can become an important reference for model retraining and parameter fine-tuning.

[0120] Therefore, under this localized model optimization mechanism, it is helpful to improve the model algorithm's precise adaptability to the local disease spectrum and patient characteristics. In the long run, with the continuous accumulation of localized data and repeated optimization of the model, compared with the original simple model algorithm transplantation work, or compared with relying solely on external general data, the introduction of localized data can deeply adapt to the regional disease structure, population differences, testing platforms, and departmental diagnosis and treatment characteristics, and is more sensitive and targeted. The model algorithm will gradually possess and maintain the advantage of high accuracy and high stability in adapting to complex local diagnosis and treatment scenarios.

[0121] Furthermore, in terms of details, for the localized data used for localized model optimization, that is, the data notes, annotation information and correction information written by medical staff for the processing results in different situations such as when there are objections to the processing results, the first weight of the specific hospital department, user title and individual user of the corresponding writer can be determined, as well as the second weight of the disease level, degree of model processing deviation and other aspects. Starting from a deep level, the model is guided to identify the contribution that different localized data can make to the localized model optimization processing in the corresponding aspects, and to perform highly targeted model optimization in more detailed specific hospital application scenarios, to obtain a more delicate localized model optimization effect, so as to continue to strengthen the advantages that can be achieved by the localized model optimization mechanism.

[0122] Finally, for the above solution content, in general, it can be concluded that under the solution goal of further improving the application value of medical test models, this application locally deploys mature medical test models in the hospital, and specifically combines container deployment technology and container orchestration technology to provide medical staff with corresponding model calling services more conveniently, efficiently and safely. It can effectively balance the dual needs of data security compliance and real-time model calling, and provide a safe, scalable and highly reliable operating platform for the actual implementation of digital algorithms. While greatly reducing external dependence, it also provides the possibility for hospital information departments and scientific researchers to quickly verify and iterate algorithms in clinical environments, and achieve "closed-loop, safe and controllable, and efficient optimization" intelligent medical assistance. It has large-scale promotion value and can be widely used in hospital groups, regional medical alliances and larger-scale medical big data platforms, providing robust AI technology support for various clinical diagnosis, test analysis and scientific research work.

[0123] The above is an introduction to the deployment method of a specific detection model in a medical scenario provided by this application. In order to facilitate better implementation of the deployment method of a specific detection model in a medical scenario provided by this application, this application also provides a deployment device for a specific detection model in a medical scenario from the perspective of a functional module.

[0124] See Figure 2 , Figure 2 This is a structural diagram of a deployment device for a specific detection model in a medical scenario of this application. In this application, the deployment device 200 for a specific detection model in a medical scenario may specifically include the following structure:

[0125] A determination unit 201 is configured to determine a target detection model that needs to be invoked on an external platform during the processing of medical services involved in the hospital, wherein the target detection model is specifically a machine learning model for medical testing;

[0126] An acquisition unit 202 is used to acquire a digital algorithm and dependency library corresponding to the target detection model;

[0127] A packaging unit 203 is configured to package the digital algorithm and dependent libraries into a target container image corresponding to the target detection model, wherein the target container image includes an execution script, a runtime environment, and an algorithm weight file;

[0128] A deployment unit 204 is configured to deploy multiple container copies of a container image through a container orchestration tool in an internal online platform of a hospital that is pre-built for deploying different detection models, wherein the internal online platform of the hospital is built on a private cloud platform or a physical server cluster, the internal online platform of the hospital is deployed with at least one of a GPU cluster and a CPU cluster, the model service running node of the internal online platform of the hospital is isolated from the core database, and the internal online platform of the hospital performs model configuration through at least one of a hospital-specific VPN and a physical dedicated line without opening a patient information port to the outside world;

[0129] The response unit 205 is used to distribute the inference request about the target detection model to the adapted target container copy among multiple container copies through the load balancer when receiving the inference request, so as to provide the target detection model calling service to the requester of the inference request through the target container copy, wherein the requester is an authorized user within the hospital.

[0130] In an exemplary embodiment, in terms of data exchange and security compliance, the device further includes a security management unit configured to:

[0131] At the data input and output interfaces, API protocols that comply with preset industry standards are used, including the HL7 standard and the FHIR standard.

[0132] Encrypt transmission and storage of sensitive data, and set differentiated permissions for users of different granularities;

[0133] The data written into the model container for processing is first desensitized to remove private content, including name and ID number.

[0134] The model call log is retained, where the retained content includes the request time, request user ID and processing results.

[0135] In another exemplary embodiment, with respect to containerized deployment and load management, the apparatus further includes a load management unit configured to:

[0136] Adaptively scale the number of container replicas corresponding to the target detection model based on past, current, and predicted inference request volumes.

[0137] Remove failed container instances identified by health checks and start new ones to replace them.

[0138] In another exemplary embodiment, with respect to the integration of model reasoning services and interfaces, the apparatus further includes a service management unit configured to:

[0139] Use RESTful or gRPC methods to expose a unified inference service API to receive inference requests.

[0140] For inference requests whose number of requests or data volume exceeds the corresponding threshold, block processing or queue processing is performed;

[0141] Through the gateway or proxy layer, security control operations corresponding to inference requests are performed, where security control operations include authentication, access token distribution, and logging.

[0142] In another exemplary embodiment, with respect to the model iteration and AB testing mechanism, the apparatus further includes a test management unit configured to:

[0143] For target detection models, multiple versions of container replicas coexist, and based on this coexistence, dynamic traffic allocation is performed. Furthermore, new version container replicas are run in parallel on the test branch, without interrupting clinical services during the model version update process.

[0144] Using A / B testing, some inference requests are routed to the new version container replica for response processing, and the model processing effect is monitored in real time. The model processing effect includes accuracy, sensitivity, specificity, response latency, and error rate.

[0145] If the model processing effect meets the preset excellent performance conditions, the request access volume of the new version container copy will be gradually expanded. If the model processing effect does not meet the preset excellent performance conditions or an abnormality occurs, the new version container copy will be rolled back to the old version container copy.

[0146] In another exemplary embodiment, in terms of visualization and result writing back, the apparatus further includes a result management unit configured to:

[0147] Perform structured and visual conversion on the processing results of the object detection model regarding the inference request to output corresponding text descriptions, risk radar charts or risk bar charts, and involve risk warnings or recommended measures;

[0148] When the requesting party is an outpatient or emergency physician workstation or an inpatient physician workstation, the processing results will be written back to the corresponding interface service of the corresponding HIS system;

[0149] When there are objections to the processing results, receive data notes, annotation information and correction information written by medical staff;

[0150] The data annotation, annotation information and correction information corresponding to the inference request are stored in the database along with the processing results.

[0151] In yet another exemplary embodiment, in terms of retraining and fine-tuning, the apparatus further includes an optimization management unit configured to:

[0152] The data comments, annotation information, and correction information written by medical staff on the processing results are used as training samples to retrain the target detection model locally or perform parameter fine-tuning to generate a new version container copy of the updated target detection model, which coexists with the old version container copy for multi-version coexistence.

[0153] This application also provides a deployment system for a specific detection model in a medical scenario from the perspective of hardware structure. It can be understood that the deployment system for a specific detection model in a medical scenario can be either a single device or a device cluster. For the convenience of explanation, the devices involved in the deployment system for a specific detection model in a medical scenario are collectively referred to as processing devices. Figure 3 , Figure 3 A schematic diagram of the structure of the processing device involved in the deployment system of the specific detection model in the medical scenario of the present application is shown. Specifically, the processing device involved in the deployment system of the specific detection model in the medical scenario of the present application may include a processor 301, a memory 302 and an input and output device 303. The processor 301 is used to execute the computer program stored in the memory 302 to implement the following Figure 1 The deployment method of the specific detection model in the medical scenario in the corresponding embodiment corresponds to the steps of the execution subject; or, when the processor 301 is used to execute the computer program stored in the memory 302, the following is implemented Figure 2 The memory 302 is used to store the functions of each unit in the embodiment corresponding to the processor 301. Figure 1 The deployment method of the specific detection model in the medical scenario in the corresponding embodiment corresponds to the computer program required by the execution subject.

[0154] For example, the computer program may be divided into one or more modules / units, one or more of which are stored in the memory 302 and executed by the processor 301 to complete the present application. One or more modules / units may be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program in a computer device.

[0155] The processing devices involved in the deployment system of a specific detection model in a medical scenario may include, but are not limited to, a processor 301, a memory 302, and an input / output device 303. Those skilled in the art will understand that the illustration is merely an example of the processing devices involved in the deployment system of a specific detection model in a medical scenario, and does not constitute a limitation on the processing devices involved in the deployment system of a specific detection model in a medical scenario. The illustration may include more or fewer components than shown in the illustration, or a combination of certain components, or different components. For example, the processing devices involved in the deployment system of a specific detection model in a medical scenario may also include a network access device, a bus, etc., and the processor 301, the memory 302, the input / output device 303, etc. are connected via a bus.

[0156] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor. The processor is the control center of the processing equipment involved in the deployment system of a specific detection model in a medical scenario, and uses various interfaces and lines to connect various parts of the entire device.

[0157] The memory 302 can be used to store computer programs and / or modules. The processor 301 implements various functions of the computer device by running or executing the computer programs and / or modules stored in the memory 302 and calling the data stored in the memory 302. The memory 302 can mainly include a program storage area and a data storage area. The program storage area can store an operating system, at least one application required for a function, etc.; the data storage area can store data created by the use of processing equipment involved in the deployment system of a specific detection model in a medical scenario, etc. In addition, the memory can include high-speed random access memory and non-volatile memory, such as a hard disk, internal memory, a plug-in hard disk, a smart memory card (SmartMedia Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0158] When the processor 301 is used to execute the computer program stored in the memory 302, it can specifically implement the following functions of the corresponding execution entities:

[0159] Determine the target detection model that needs to be called on the hospital's external platform during the processing of the hospital's medical services. Specifically, the target detection model is a machine learning model for medical testing.

[0160] Obtain the corresponding digital algorithms and dependency libraries of the target detection model;

[0161] The digital algorithm and dependent libraries are packaged uniformly into the target container image corresponding to the target detection model, where the target container image includes the execution script, runtime environment, and algorithm weight file;

[0162] In an internal hospital online platform pre-built for deploying different detection models, multiple container copies of the container image are deployed through a container orchestration tool. The internal hospital online platform is built on a private cloud platform or a physical server cluster. The internal hospital online platform is deployed with at least one of a GPU cluster and a CPU cluster. The model service running node and the core database of the internal hospital online platform are isolated. The internal hospital online platform performs model configuration through at least one of a hospital-specific VPN and a physical dedicated line without opening the patient information port to the outside world.

[0163] When an inference request about a target detection model is received, the inference request is distributed to an adapted target container replica among multiple container replicas through a load balancer, so as to provide a call service of the target detection model to the requester of the inference request through the target container replica, where the requester is an authorized user within the hospital.

[0164] Those skilled in the art will clearly understand that for the convenience and brevity of description, the specific working process of the deployment device, system and corresponding units of the specific detection model in the medical scenario described above can refer to the following. Figure 1 The description of the deployment method of the specific detection model in the medical scenario in the corresponding embodiment will not be repeated here.

[0165] Those skilled in the art will appreciate that all or part of the steps in the various methods of the above embodiments may be accomplished by instructions, or by controlling related hardware through instructions. The instructions may be stored in a computer-readable storage medium and loaded and executed by a processor.

[0166] To this end, the present application provides a computer-readable storage medium, which stores a plurality of instructions, which can be loaded by a processor to execute the present application as follows: Figure 1 For the steps of the deployment method of a specific detection model in a medical scenario in the corresponding embodiment, the specific operations can be referred to as follows Figure 1 The description of the deployment method of the specific detection model in the medical scenario in the corresponding embodiment will not be repeated here.

[0167] The computer-readable storage medium may include a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0168] Due to the instructions stored in the computer readable storage medium, the present application can be executed as follows: Figure 1 The steps of the deployment method of the specific detection model in the medical scenario in the corresponding embodiment, therefore, the application can be implemented as follows Figure 1 The beneficial effects that can be achieved by the deployment method of a specific detection model in a medical scenario in the corresponding embodiment are detailed in the previous description and will not be repeated here.

[0169] The above is a detailed introduction to the deployment method, device, system and computer-readable storage medium of the specific detection model in the medical scenario provided by this application. Specific examples are used in this article to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the core idea of ​​this application; at the same time, for technical personnel in this field, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on this application.

Claims

1. A method for deploying a specific detection model in a medical scenario, characterized in that: The method comprises: Determine the target detection model that needs to be called on the hospital's external platform during the processing of the hospital's medical services, where the target detection model is specifically a machine learning model for medical testing; Obtain the digital algorithm and dependent library corresponding to the target detection model; The digital algorithm and the dependent library are uniformly packaged into a target container image corresponding to the target detection model, wherein the target container image includes an execution script, a runtime environment, and an algorithm weight file; In an internal hospital online platform pre-built for deploying different detection models, multiple container copies of the container image are deployed through a container orchestration tool, wherein the internal hospital online platform is built on a private cloud platform or a physical server cluster, and the internal hospital online platform is deployed with at least one of a GPU cluster and a CPU cluster. The model service running node and the core database of the internal hospital online platform are isolated from each other, and the internal hospital online platform performs model configuration through at least one of a hospital-specific VPN and a physical dedicated line without opening the patient information port to the outside world; When an inference request regarding the target detection model is received, the inference request is distributed to an adapted target container replica among the multiple container replicas through a load balancer, so as to provide a calling service of the target detection model to the requester of the inference request through the target container replica, wherein the requester is an authorized user within the hospital.

2. The method according to claim 1, characterized in that In terms of data exchange and security compliance, the method also includes: At the data input and output interfaces, an API protocol that complies with preset industry standards is adopted, wherein the preset industry standards include the HL7 standard and the FHIR standard; Encrypt transmission and storage of sensitive data, and set differentiated permissions for users of different granularities; The data written into the model container for processing is first desensitized to remove the private content, wherein the private content includes the name and ID number; The model call log is retained, where the retained content includes the request time, request user ID and processing results.

3. The method according to claim 1, characterized in that Regarding containerized deployment and load management, the method further includes: For the multiple container replicas corresponding to the target detection model, adaptively scaling the number of the multiple container replicas according to past, current, and predicted inference request volumes; Remove failed container instances identified by health checks and start new ones to replace them.

4. The method according to claim 1, wherein Regarding the integration of model reasoning services and interfaces, the method further includes: Use RESTful or gRPC methods to expose a unified inference service API to receive the inference request; For the inference requests whose number of requests or data volume exceeds the corresponding threshold, perform block processing or queue processing; The security control operations corresponding to the inference request are performed through the gateway or proxy layer, wherein the security control operations include identity authentication, access token distribution and logging.

5. The method according to claim 1, wherein Regarding the model iteration and AB testing mechanism, the method further includes: For the target detection model, performing multi-version coexistence processing on the container replica, and, based on the multi-version coexistence processing, performing dynamic traffic allocation processing, and running the new version container replica in parallel under the test branch, wherein clinical services are not interrupted during the model version update process; Using AB testing, some of the inference requests are routed to the new version container replica for response processing, and the model processing effect is monitored in real time, where the model processing effect includes accuracy, sensitivity, specificity, response latency, and error rate; If the model processing effect meets the preset excellent performance conditions, the request access volume of the new version container copy will be gradually expanded. If the model processing effect does not meet the preset excellent performance conditions or an abnormality occurs, the new version container copy will be rolled back to the old version container copy.

6. The method according to claim 1, characterized in that In terms of visualization and result writing back, the method further includes: Performing structural and visual conversion on the processing results of the object detection model regarding the reasoning request to output corresponding text descriptions, risk radar charts or risk bar charts, and involving risk warnings or recommended measures; When the requesting party is an outpatient and emergency physician workstation or an inpatient physician workstation, writing the processing result back to the corresponding interface service of the corresponding HIS system; When there is any objection to the processing result, receive the data notes, annotation information and correction information written by the medical staff; The data annotation, the annotation information, and the correction information corresponding to the inference request are stored in a database along with the processing result.

7. The method according to claim 1, characterized in that In terms of retraining and fine-tuning, the method further includes: The data comments, annotation information and correction information written by medical staff on the processing results are used as training samples, and the target detection model is trained again or parameter adjustments are performed locally to generate a new version container copy of the updated target detection model, which coexists with the old version container copy for multi-version coexistence.

8. A deployment device for a specific detection model in a medical scenario, characterized in that: The device comprises: a determination unit, configured to determine a target detection model that needs to be invoked on an external platform of the hospital during the processing of the medical services involved in the hospital, wherein the target detection model is specifically a machine learning model for medical testing; An acquisition unit, configured to acquire a digital algorithm and a dependency library corresponding to the target detection model; A packaging unit, configured to uniformly package the digital algorithm and the dependent library into a target container image corresponding to the target detection model, wherein the target container image includes an execution script, a runtime environment, and an algorithm weight file; A deployment unit is configured to deploy multiple container copies of the container image through a container orchestration tool in an internal hospital online platform pre-built for deploying different detection models, wherein the internal hospital online platform is built on a private cloud platform or a physical server cluster, the internal hospital online platform is deployed with at least one of a GPU cluster and a CPU cluster, the model service running node and the core database of the internal hospital online platform are isolated from each other, and the internal hospital online platform performs model configuration through at least one of a hospital-specific VPN and a physical dedicated line without opening a patient information port to the outside world; A response unit is used to distribute the inference request regarding the target detection model to an adapted target container replica among the multiple container replicas through a load balancer when receiving the inference request, so as to provide the target detection model calling service to the requester of the inference request through the target container replica, wherein the requester is an authorized user within the hospital.

9. A deployment system for a specific detection model in a medical scenario, characterized by: Used to perform the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the method according to any one of claims 1 to 7.

Images