Ultrasonic scanning strategy optimization method based on digital twinning and related apparatus

Abstract

The application provides an ultrasonic scanning strategy optimization method based on digital twinning and related devices, belonging to the field of artificial intelligence technology. The method comprises: acquiring multi-modal clinical data of a patient and ultrasonic scanning operation data of a doctor; constructing a digital twin model of the patient according to a human voxel model and the multi-modal clinical data; and determining a plurality of scanning working conditions according to the ultrasonic scanning operation data; constructing a multi-scenario scanning working condition library according to the plurality of scanning working conditions and the ultrasonic scanning operation data; generating a multi-scenario scanning data set according to the digital twin model and the multi-scenario scanning working condition library; training a first scanning strategy model according to the multi-scenario scanning data set to obtain a target scanning strategy model; and controlling an ultrasonic scanning robot to perform an ultrasonic scanning task according to the target scanning strategy model. In this way, the robustness, stability and accuracy of the ultrasonic scanning strategy can be improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and apparatus for optimizing ultrasonic scanning strategies based on digital twins. Background Technology

[0002] Simulating ultrasound scans using digital twin models to generate training data for ultrasound scan control models is a current hot topic in ultrasound scan model research. Training schemes based on digital twin models, and the acquisition of ultrasound scan training data, typically involve generating simulation training data using a general static human anatomical model or a pre-defined standardized scan scenario. However, due to individual differences in anatomical structure, dynamic respiratory and cardiac physiological changes among different patients, simulated scanning data generated using a general static human body model cannot accurately match the anatomical features and dynamic physiological state of real patients. This results in a significant difference in feature domain between simulated training scenarios and real clinical scenarios. At the same time, existing training schemes lack effective coverage of operational error scenarios, extreme anatomical variation scenarios, and sudden abnormal scenarios. This makes it difficult for the trained ultrasound scanning control model to fully learn and master the ability to correct slicing deviations and handle abnormal emergencies. Consequently, the model has poor accuracy, low scanning stability, and insufficient ability to handle abnormal scenarios in real-world scenarios.

[0003] Therefore, improving the robustness, accuracy, and stability of ultrasonic scanning control models is an urgent issue that needs to be addressed. Summary of the Invention

[0004] The purpose of this application is to provide a method and related device for optimizing ultrasonic scanning strategies based on digital twins, which solves the problems of poor accuracy, insufficient robustness, and poor stability of existing ultrasonic scanning control models in real-world application scenarios due to insufficient ability to correct deviations in cross-sections and handle abnormal emergencies during ultrasonic scanning.

[0005] To achieve the objectives of this application, the following technical solution is provided: In a first aspect, embodiments of this application provide a digital twin-based method for optimizing ultrasound scanning strategies, applied to a server in an ultrasound scanning system. The method includes: Acquire multimodal clinical data of patients and ultrasound scan operation data of physicians; the multimodal clinical data includes patient body surface data, ultrasound imaging data and patient physiological data; A digital twin model of the patient is constructed based on a pre-defined human voxel model and the multimodal clinical data; and multiple scanning conditions are determined based on the ultrasound scanning operation data; the multiple scanning conditions include standard scanning conditions, error scanning conditions, extreme scanning conditions, and mutation scanning conditions; a multi-scenario scanning condition library is constructed based on the multiple scanning conditions and the ultrasound scanning operation data; and a multi-scenario scanning dataset is generated based on the digital twin model and the multi-scenario scanning condition library. The first scanning strategy model is trained based on the multi-scenario scanning dataset to obtain the target scanning strategy model; The ultrasonic scanning robot is controlled to perform ultrasonic scanning tasks based on the target scanning strategy model.

[0006] Secondly, this application provides a digital twin-based ultrasonic scanning strategy optimization device, applied to a server in an ultrasonic scanning system, the device comprising: The acquisition unit is used to acquire the patient's multimodal clinical data and the doctor's ultrasound scan operation data; the multimodal clinical data includes the patient's body surface data, ultrasound image data and patient physiological data; The processing unit is configured to construct a digital twin model of the patient based on a preset human voxel model and the multimodal clinical data; and to determine multiple scanning conditions based on the ultrasound scanning operation data; the multiple scanning conditions include standard scanning conditions, error scanning conditions, extreme scanning conditions, and mutation scanning conditions; to construct a multi-scenario scanning condition library based on the multiple scanning conditions and the ultrasound scanning operation data; and to generate a multi-scenario scanning dataset based on the digital twin model and the multi-scenario scanning condition library. The computing unit is used to train the first scanning strategy model based on the multi-scenario scanning dataset to obtain the target scanning strategy model; The control unit is used to control the ultrasonic scanning robot to perform ultrasonic scanning tasks according to the target scanning strategy model.

[0007] Thirdly, embodiments of this application provide a digital twin-based ultrasonic scanning strategy optimization system, wherein the digital twin-based ultrasonic scanning strategy optimization system includes instructions for performing the steps in any of the methods of the first aspect of this application.

[0008] Fourthly, embodiments of this application provide a server, including a processor, a memory, a communication interface, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the processor, and the programs include instructions for performing steps in any of the methods of the first aspect of this application.

[0009] Fifthly, embodiments of this application provide a computer-readable storage medium storing a computer program for electronic data interchange, wherein the computer program causes a computer to perform some or all of the steps described in any method of the first aspect of this application.

[0010] In a sixth aspect, embodiments of this application provide a computer program product, wherein the computer program product includes a non-transitory computer-readable storage medium storing a computer program, the computer program being operable to cause a computer to perform some or all of the steps described in any method of the first aspect of this application, and the computer program product may be a software installation package.

[0011] By implementing the ultrasound scanning strategy optimization method and related apparatus based on digital twins provided in this application embodiment, applied to a server in an ultrasound scanning system, the following steps are taken: First, multimodal clinical data of the patient and ultrasound scanning operation data of the doctor are acquired; a digital twin model of the patient is constructed based on a preset human voxel model and multimodal clinical data; and multiple scanning conditions are determined based on the ultrasound scanning operation data, including standard scanning conditions, erroneous scanning conditions, extreme scanning conditions, and abrupt change scanning conditions; a multi-scenario scanning condition library is constructed based on the multiple scanning conditions and ultrasound scanning operation data; a multi-scenario scanning dataset is generated based on the digital twin model and the multi-scenario scanning condition library; a first scanning strategy model is trained based on the multi-scenario scanning dataset to obtain a target scanning strategy model; and an ultrasound scanning robot is controlled to perform ultrasound scanning tasks based on the target scanning strategy model. Thus, on the one hand, by constructing a digital twin model of the patient based on a preset human voxel model and multimodal clinical data, this digital twin model matches the patient's individual tissue structure and dynamic individual physiological changes, thereby improving the robustness of the ultrasound scan control model compared to simulated scan data generated using a general static human body model. On the other hand, by training the first scan strategy model based on a multi-scenario scan dataset to obtain a target scan strategy model, the digital twin model is combined with a multi-scenario scan condition library to generate a multi-scenario scan dataset corresponding to the time step, and the first scan strategy model is pre-trained under supervision based on the simulated multi-scenario scan dataset to obtain a trained target scan strategy model. This model can learn the ability to correct scission deviations and handle abnormal situations in a large number of error feedbacks and abnormal scenarios, thereby improving the accuracy and stability of the ultrasound scan control model. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is an architectural diagram of an ultrasonic scanning system provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a server provided in an embodiment of this application; Figure 3 This is a flowchart illustrating an ultrasonic scanning strategy optimization method based on digital twins provided in an embodiment of this application. Figure 4 This is a schematic diagram of the architecture of an ultrasound scanning strategy optimization method based on digital twin provided in an embodiment of this application; Figure 5 This is a flowchart illustrating another method for optimizing ultrasonic scanning strategies based on digital twins provided in an embodiment of this application. Figure 6 This is a schematic diagram of an ultrasonic scanning application scenario based on a digital twin model provided in an embodiment of this application; Figure 7 This is a functional module block diagram of an ultrasonic scanning strategy optimization device based on digital twin provided in an embodiment of this application. Detailed Implementation

[0014] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0015] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0016] It should be understood that the term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this document indicates that the preceding and following related objects are in an "or" relationship. In the embodiments of this application, "multiple" refers to two or more.

[0017] In the embodiments of this application, "at least one item" or its similar expression refers to any combination of these items, including any combination of a single item or a plurality of items. "One or more" means one or more, while "multiple" means two or more. For example, "at least one item" of a, b, or c can represent the following seven cases: a, b, c; a and b; a and c; b and c; a, b, and c. Each of a, b, and c can be an element or a set containing one or more elements.

[0018] In this application, the term "connection" refers to various connection methods, such as direct connection or indirect connection, to achieve communication between devices. This application does not impose any limitations on this.

[0019] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] The following is an explanation of the relevant terms used in this application: Digital twin: A digital twin is a technology that uses data such as physical models, sensors, and operational history to create an information model in virtual space that is completely equivalent to a physical entity.

[0021] The following is combined with Figure 1 The hardware architecture of an ultrasound scanning strategy optimization method based on digital twins, as described in this application embodiment, is explained. Figure 1 This is an architecture diagram of an ultrasonic scanning system provided in an embodiment of this application. The ultrasonic scanning system 100 includes a terminal execution layer 110, a local control layer 120, and a cloud computing layer 130.

[0022] The terminal execution layer 110 is used for ultrasound scanning action execution and multimodal raw data acquisition. It serves as the data source and action execution terminal for ultrasound scanning strategy optimization. The terminal execution layer 110 includes an ultrasound probe 111, a depth vision acquisition module 112, and a physiological signal acquisition module 113. The ultrasound probe 111 is the direct execution component for scanning actions. It can be rigidly connected to the collaborative robotic arm through an end effector. It receives instructions from the local control layer to complete ultrasound signal transmission and reception and scan trajectory execution. It supports adaptation to multiple types of probes, such as linear array, convex array, and phased array, to meet the scanning needs of different anatomical sites. The depth vision acquisition module 112 is used to acquire three-dimensional point cloud data and body posture information of the patient's body surface, providing surface feature input for the construction of personalized digital twin models. The physiological signal acquisition module 113 acquires physiological signals such as patient breathing and heartbeat through medical breathing belts, ECG electrodes, and other devices, providing time-series driving data for the physiological motion simulation of the dynamic twin. In an optional embodiment, the terminal execution layer 110 may also integrate a high-precision force control sensor, which is installed at the connection position between the ultrasound probe 111 and the end of the robotic arm, to collect contact pressure data between the probe and the patient's body surface in real time, providing contact mechanics simulation and real-time feedback for scanning action safety control.

[0023] The local control layer 120, used for local real-time control, data preprocessing, and edge-cloud collaborative interaction, serves as a hub connecting the terminal execution layer and the cloud computing layer. The local control layer 120 includes an ultrasonic control host 121, a control device 122, and an interactive terminal 123. The ultrasonic control host 121 is physically connected to the ultrasonic probe 111, completing ultrasonic signal preprocessing and imaging output. It can receive control commands to adjust parameters such as imaging depth and gain, providing real-time image support for the scanning process. The control device 122 is a local control unit that parses and issues scanning motion commands, synchronously collects real-time data such as robotic arm pose and force control sensors, and performs preliminary cleaning and caching of the raw data from the depth vision and physiological signal acquisition modules. The interactive terminal 123 consists of an industrial touchscreen, allowing operators to initiate scanning tasks, configure parameters, and trigger emergency stop commands, achieving human-machine interaction and safety control. In an optional embodiment, the control device 122 can integrate a local edge computing unit, equipped with a lightweight AI model, to achieve low-latency local inference and emergency response during the scanning process, reducing reliance on cloud computing power.

[0024] The cloud computing layer 130 provides high-performance simulation, digital twin modeling, data generation, and AI model training capabilities, offering computing power support for scanning strategy optimization. The cloud computing layer 130 includes a computing server cluster 131, a distributed storage array 132, a network switching device 133, and a management server 134. The computing server cluster 131 consists of heterogeneous servers equipped with GPUs / NPUs, supporting millions of concurrent simulations, individualized digital twin modeling, and AI model training, providing computing power guarantees for massive simulation data generation and model iterative optimization. The distributed storage array 132 stores the digital twin library, the full-scenario working condition library, training datasets, and model training logs, enabling secure storage and efficient retrieval of massive amounts of data. The network switching device 133 is configured with low-latency network equipment to ensure low-latency and high-reliability data transmission between the local control layer and the cloud computing layer, supporting edge-cloud collaborative interaction. The management server 134 is responsible for computing power scheduling, task queue management, and user permission configuration, achieving efficient management and task allocation of cloud computing resources. In an optional embodiment, the management server 134 is deployed redundantly to improve system stability and fault tolerance, and to ensure the continuous operation of the scanning strategy optimization task.

[0025] As can be seen, the terminal execution layer, local control layer, and cloud computing layer interact with each other through standardized interfaces and communication protocols, forming an architecture of "data acquisition - local control - cloud optimization - strategy iteration". Furthermore, the multimodal data acquired by the terminal execution layer is preprocessed by the local control layer and then uploaded to the cloud computing layer. The cloud generates massive amounts of simulated data based on digital twin simulation and completes AI model training. The optimized scanning strategy model is then distributed back to the local control layer, achieving continuous optimization through edge-cloud collaboration. This solves the problems of traditional AI model training relying on high-cost real clinical data and having weak generalization ability, while ensuring the real-time performance, safety, and clinical adaptability of the scanning process, thereby improving the robustness, accuracy, and stability of the ultrasound scanning control model.

[0026] The following is combined with Figure 2 The server in the embodiments of this application will be described. Figure 2 This is a schematic diagram of the structure of a server provided in an embodiment of this application, such as... Figure 2 As shown, the server 200 includes a processor 210, a memory 220, a communication interface 230, and one or more programs 221. The processor 210 is communicatively connected to the memory 220 and the communication interface 230 via an internal communication bus.

[0027] The processor 210 is mainly used for: Acquire patients' multimodal clinical data, as well as doctors' ultrasound scanning operation data; A digital twin model of the patient is constructed based on a pre-set human voxel model and multimodal clinical data; and multiple scanning conditions are determined based on ultrasound scanning operation data; among which, multiple scanning conditions include standard scanning conditions, error scanning conditions, extreme scanning conditions and mutation scanning conditions; A multi-scenario scanning condition library is constructed based on multiple scanning conditions and ultrasonic scanning operation data; a multi-scenario scanning dataset is generated based on the digital twin model and the multi-scenario scanning condition library; The first scanning strategy model is trained based on a multi-scenario scanning dataset to obtain the target scanning strategy model; and the ultrasonic scanning robot is controlled to perform ultrasonic scanning tasks based on the target scanning strategy model.

[0028] The one or more programs 221 are stored in the memory 220 and configured to be executed by the processor 210. The one or more programs 221 include instructions for performing any step in the above method embodiments.

[0029] The processor 210 can be a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, units, and circuits described in conjunction with the disclosure of this application. The processor can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The communication unit can be a communication interface, transceiver, transceiver circuit, etc., and the storage unit can be a memory.

[0030] The memory 220 can be volatile memory or non-volatile memory, or it can include both. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0031] It is understood that server 200 may include more or fewer structural elements than those shown in the block diagram above, such as a power module, physical buttons, a Wi-Fi module, a speaker, a Bluetooth module, sensors, a display module, etc., without limitation. It is understood that server 200 may be equipped with... Figure 1 The architecture of the ultrasonic scanning system described above.

[0032] After understanding the software and hardware architecture of this application, the following will be combined with... Figure 3 This application describes an ultrasound scanning strategy optimization method based on digital twins. Figure 3 This is a flowchart illustrating an ultrasound scanning strategy optimization method based on digital twins, provided in an embodiment of this application. The digital twin-based ultrasound scanning strategy optimization method is applied to a server in an ultrasound scanning system. The method specifically includes the following steps: Step S310: Acquire the patient's multimodal clinical data and the doctor's ultrasound scan operation data; the multimodal clinical data includes the patient's body surface data, ultrasound image data and patient physiological data.

[0033] Multimodal clinical data refers to a collection of heterogeneous medical data from multiple sources used to characterize the individual anatomical features, physiological states, and ultrasound imaging characteristics of patients. Patient physiological data characterizes the dynamic physiological changes of patients in real clinical situations and may include respiratory rate data, respiratory waveform data, electrocardiogram signal data, heart rate data, and soft tissue movement cycle data. Physician ultrasound scanning operation data characterizes the physician's operational behavior with the ultrasound probe during actual clinical scanning, including ultrasound probe pose changes, probe motion trajectory data, probe contact pressure data, scanning area switching data, and ultrasound equipment parameter adjustment data.

[0034] Specifically, the visual acquisition device deployed in the ultrasound robot system first performs a 3D scan of the patient's body surface structure to obtain the patient's surface contour information and spatial point cloud data in the current position. Simultaneously, the ultrasound host controls the ultrasound probe to perform an initial scan of the target area to acquire raw ultrasound image data of the corresponding region. During ultrasound image acquisition, the spatial pose parameters, motion trajectory parameters, and contact pressure parameters of the ultrasound probe at each time step are continuously recorded. A time synchronization mechanism is used to associate the ultrasound image data with the probe operation data using a unified timestamp, ensuring temporal consistency among various data types during subsequent digital twin modeling. Simultaneously, the physiological signal acquisition device synchronously acquires the patient's current respiratory and heart rate status to obtain dynamic physiological changes in the patient within a real clinical environment. Then, the acquired surface point cloud data, ultrasound image data, physiological signal data, and probe operation data undergo preprocessing. After preprocessing, the multimodal clinical data and ultrasound scan operation data are associated and stored according to a unified time-series index and uploaded to a cloud training platform for use in subsequent dynamic digital twin model construction, multi-scenario scan condition generation, and simulated scan dataset generation.

[0035] It should be noted that multimodal clinical data can also include CT volume data, MRI volume data, tissue elasticity parameter data, and patients' historical ultrasound diagnostic data, etc., without limitation. In addition, the physician's ultrasound scanning operation data can come not only from the operation records of the actual scanning process, but also from standard scanning cases, abnormal scanning cases, and extreme anatomical variation scanning cases in the historical clinical scanning database.

[0036] Step S320: Construct a digital twin model of the patient based on a preset human voxel model and the multimodal clinical data; and determine multiple scanning conditions based on the ultrasound scanning operation data; the multiple scanning conditions include standard scanning conditions, error scanning conditions, extreme scanning conditions, and mutation scanning conditions.

[0037] Among them, the human voxel model refers to a three-dimensional discretized human tissue model based on the human anatomical structure. It uses voxel units to spatially divide different tissue regions within the human body to represent the spatial structural distribution, tissue type, and acoustic propagation characteristics of human tissues. The digital twin model refers to a dynamically changing virtual human body model constructed based on the patient's real anatomical structure and dynamic physiological state information. It can not only reflect the patient's current static anatomical structure but also simulate the dynamic tissue deformation characteristics during respiration, heartbeat, and tissue compression. Scanning conditions refer to the set of scanning states used to represent different scanning behaviors and clinical scenarios during ultrasound scanning. Standard scanning conditions represent normal scanning processes that conform to clinical standard scanning specifications; error scanning conditions represent non-standard scanning processes caused by probe pose deviation, abnormal pressure, or sectional deviation; extreme scanning conditions represent complex scanning scenarios when patients have special body structures, organ deviations, or anatomical variations; and abrupt change scanning conditions represent dynamic abnormal scenarios such as sudden changes in patient movement, respiration, or contact instability.

[0038] Specifically, the process begins by using a pre-defined human voxel model as the basic anatomical template. Then, based on the patient's surface point cloud data, a global spatial registration process is performed on the voxel model to ensure its external contour matches the patient's current body structure. Next, using the patient's ultrasound and tomographic images, non-rigid deformation registration is applied to local tissue regions within the voxel model. This dynamically corrects the position, size, and boundaries of target tissue structures such as the liver, heart, kidneys, and blood vessels, generating a static digital twin model that corresponds to the patient's actual anatomical structure. After completing the static structural modeling, dynamic physiological cycle parameters are extracted from the patient's respiratory and electrocardiogram data. These parameters then drive periodic dynamic deformation of the target tissues in the static digital twin model. Finally, the spatial relationships and ultrasound propagation paths of each tissue region are updated in real-time based on the dynamically deformed tissue structure, generating a digital twin model that reflects the patient's true dynamic physiological state. Simultaneously, historical ultrasound scan data and real-time scan operation data from doctors are analyzed, and corresponding probe motion trajectory features, posture change features, contact pressure features, and slice quality features are extracted. First, scan data meeting standard slice requirements are screened according to clinical standard scan specifications, and the corresponding scan processes are defined as standard scan conditions. Next, non-standard scan behaviors are identified based on the degree of probe posture deviation, pressure abnormality, and slice offset, and the corresponding scan processes are defined as erroneous scan conditions. Furthermore, complex anatomical scenarios are identified based on the patient's specific body shape characteristics, organ anatomical variations, and abnormal lesion distribution characteristics, and corresponding extreme scan conditions are generated. Then, dynamic abnormal behaviors are identified by combining patient body movement data, respiratory change data, and contact pressure mutation data, and corresponding mutation scan conditions are generated. This allows the constructed digital twin model to simultaneously reflect the patient's true anatomical structure features and dynamic physiological changes, thereby improving the subsequent scan strategy model's generalization and adaptation capabilities to complex scenarios, slice correction capabilities, and emergency handling capabilities for abnormal scenarios.

[0039] In an optional embodiment, the multimodal clinical data includes patient surface point cloud data, ultrasound data, medical tomographic volume data, and patient physiological signals; the step of constructing a digital twin model of the patient based on a preset human voxel model and the multimodal clinical data specifically includes the following steps: 321. Determine the first tissue structure parameter, the skeletal structure parameter, and the first acoustic impedance parameter corresponding to the first tissue structure parameter in the human voxel model; 322. Perform non-rigid registration processing on the human voxel model based on the body surface point cloud data, the ultrasound data, and the medical tomographic volume data to obtain a first digital twin model corresponding to the patient; 323. Generate dynamic physiological deformation data of the patient based on the patient's physiological signals; the dynamic physiological deformation data includes respiratory deformation data and heart rate deformation data; 324. Perform dynamic deformation registration processing on the first digital twin model based on the dynamic physiological deformation data to obtain the second digital twin model; 325. Determine the second tissue structure parameter, the second dynamic deformation parameter, and the second acoustic impedance parameter in the second digital twin model; 326. Determine the range of dynamic tissue structure parameters of the patient based on the first tissue structure parameter and the second tissue structure parameter; and determine the range of dynamic acoustic impedance parameters of the patient based on the first acoustic impedance parameter and the second acoustic impedance parameter; 327. Construct a digital twin model of the patient based on the range of dynamic tissue structure parameters, the range of dynamic acoustic impedance parameters, the bone structure parameters, and the medical tomographic volume data.

[0040] Non-rigid registration specifically refers to a voxel-level deformation mapping algorithm based on Free Form Deformation (FFD). This algorithm calculates a three-dimensional non-rigid deformation field and adapts each voxel of the general model (i.e., the human voxel model) to the actual anatomical contour of the individual. Dynamic physiological deformation data is based on a mass-spring model of computational biomechanics. It is a time-varying deformation field generated by respiratory and electrocardiographic signals, describing the periodic displacement of thoracic and abdominal organs caused by respiratory movements and the impact deformation of adjacent tissues caused by heartbeats. The dynamic tissue structure parameter range refers to the interval between the minimum and maximum values ​​of geometric parameters such as organ boundaries and soft tissue thickness within a complete respiratory-cardiac cycle. The dynamic acoustic impedance parameter range corresponds to the acoustic impedance fluctuation boundaries caused by tissue compression and relaxation. The second tissue structure parameter is the position of each organ boundary; the second dynamic deformation parameter is the amplitude and frequency characteristics of the deformation field; and the second acoustic impedance parameter is the instantaneous fluctuation value of acoustic impedance caused by tissue compression / relaxation.

[0041] Specifically, firstly, three types of baseline parameters are extracted from the pre-defined human voxel model: first tissue structure parameters (i.e., the static spatial distribution of organs and soft tissues), skeletal structure parameters (as a rigid, non-deformable skeleton constraint), and first acoustic impedance parameters corresponding to each tissue. Next, the collected patient surface point cloud data, ultrasound tomography data, and CT / MRI three-dimensional volume data are used as the target basis for individualized deformation. A non-rigid registration algorithm based on FFD is employed to calculate the non-rigid deformation field from the general voxel model to the individual anatomical structure. This deformation field is described by voxel displacement functions (u, v, w) along three axes. The deformation field is applied to each voxel of the general model to complete voxel-level displacement mapping, causing the model to produce deformation consistent with the patient's actual skeletal contours, organ sizes, and relative heights, thus obtaining the first digital twin model. Then, the patient's respiratory zone signal and electrocardiogram signal were simultaneously acquired. Based on a mass-spring model, respiratory deformation fields and cardiac deformation fields were generated respectively. These two types of dynamic deformation fields were superimposed onto the first digital twin model along the time axis, and dynamic deformation registration processing was performed to obtain the second digital twin model. This model not only incorporates individualized anatomical features but also possesses the ability to evolve physiologically over time. It can realistically simulate clinical phenomena such as periodic slippage of the liver section caused by the patient's respiratory fluctuations during scanning and pulsation artifacts in the carotid artery wall caused by cardiac conduction. Based on this, the dynamic change boundaries of key parameters were extracted from the second digital twin model and recorded within a complete respiratory-cardiac cycle. The first and second tissue structure parameters were then compared, and their minimum and maximum values ​​were taken to obtain the range of dynamic tissue structure parameters. Similarly, the range of dynamic acoustic impedance parameters was defined by the first and second acoustic impedance parameters, characterizing the upper and lower limits of acoustic impedance fluctuations caused by tissue deformation. Finally, using skeletal structural parameters as rigid, immobile boundaries and medical tomographic volume data to provide high-resolution internal detail constraints, along with the ranges of dynamic tissue structural parameters and dynamic acoustic impedance parameters as parameterizable sampleable variable domains, a complete and computable digital twin model of the patient was constructed. This digital twin model simultaneously ensures static anatomical accuracy, dynamic physiological realism, and the feasibility of quantifying computational parameters at the physical simulation level. This provides a high-fidelity simulation foundation for subsequent full-scene scanning data generation that conforms to individual anatomical differences and encompasses physiological time-varying patterns, improving the clinical adaptability and generalization ability of the trained scanning strategy.

[0042] Step S330: Construct a multi-scenario scanning condition library based on the multiple scanning conditions and the ultrasonic scanning operation data; generate a multi-scenario scanning dataset based on the digital twin model and the multi-scenario scanning condition library.

[0043] The multi-scenario scanning condition library refers to a set of structured scanning scenarios used to describe different ultrasound scanning behaviors, dynamic physiological states, and abnormal clinical scenarios. It serves to uniformly constrain and control probe movement behavior, tissue dynamic changes, and abnormal event triggering behavior during subsequent simulated scanning processes. The multi-scenario scanning condition library includes a set of scenario parameters corresponding to standard scanning conditions, erroneous scanning conditions, extreme scanning conditions, and abrupt change scanning conditions. Standard scanning conditions correspond to normal scanning processes that conform to clinical standard sectional specifications; erroneous scanning conditions correspond to non-standard scanning processes under conditions of probe pose deviation, sectional deviation, or abnormal pressure; extreme scanning conditions correspond to complex scanning processes under conditions of patients with special body shapes, tissue anatomical variations, or abnormal lesion locations; and abrupt change scanning conditions correspond to dynamic abnormal scenarios such as sudden changes in patient movement, respiratory changes, contact instability, or probe slippage. Each scanning condition includes corresponding anatomical constraint parameters, physiological state constraint parameters, probe pose boundary parameters, contact pressure boundary parameters, and abnormal event triggering parameters. Anatomical constraint parameters are used to define the target scanning area and tissue visibility range of the ultrasound probe; physiological state constraint parameters are used to control the patient's respiratory cycle, heart rate cycle, and dynamic tissue deformation; probe pose boundary parameters are used to define the probe's direction of movement, rotation angle, and range of movement during the scanning process; contact pressure boundary parameters are used to define the safe contact pressure range between the probe and human tissue; and abnormal event triggering parameters are used to trigger abnormal events such as body motion disturbances, sudden changes in contact pressure, or sudden changes in tissue position during the simulated scanning process. The multi-scenario scanning dataset refers to a collection of simulated scanning data generated based on a digital twin model and a multi-scenario scanning condition library. It includes simulated ultrasound image data, simulated probe pose data, simulated contact pressure data, and corresponding condition label data corresponding to multiple time steps, used for supervised training and reinforcement learning training of the subsequent scanning strategy model.

[0044] Specifically, the historical ultrasound scan data is first analyzed to extract corresponding probe motion trajectory parameters, pose change parameters, contact pressure parameters, and section quality parameters. Next, based on the scene type corresponding to different scan conditions, the probe motion trajectory parameters, pose change parameters, contact pressure parameters, and section quality parameters are structurally correlated to establish a mapping relationship between probe control behavior and ultrasound image changes under different conditions. Then, the control rules, abnormal triggering rules, and dynamic tissue change rules under different conditions are encapsulated to construct a multi-scene scan condition library. After completing the construction of the multi-scene scan condition library, a digital twin model is further invoked, and corresponding simulated scan constraints are generated based on the condition parameters in the multi-scene scan condition library. Next, virtual scan control instructions for multiple time steps are generated based on the constraints. Each time step's virtual scan control instruction includes the virtual ultrasound probe's spatial pose parameters, motion velocity parameters, and contact pressure parameters. Finally, the virtual ultrasound probe is controlled to perform dynamic simulated scans in the digital twin model according to the virtual scan control instructions for each time step. During the simulated scan, the digital twin model updates the dynamic deformation results of the tissue in real time based on the patient's respiratory status, heart rate, and tissue pressure status, and determines the ultrasound propagation path according to the probe pose and tissue spatial distribution at the current time step. Next, based on the tissue's acoustic impedance parameters, scattering coefficient parameters, and attenuation coefficient parameters, the propagation, reflection, scattering, and attenuation processes of ultrasound waves in human tissue are simulated to generate simulated ultrasound echo signals for the corresponding time step. The simulated ultrasound images for the corresponding time step are reconstructed based on the generated ultrasound echo signals, and the probe pose data and contact pressure data for the current time step are recorded simultaneously. Finally, the simulated ultrasound image data, simulated pose data, simulated pressure data, and corresponding operational condition label data generated at each time step are uniformly associated and stored to generate a multi-scenario scan dataset.

[0045] In an optional embodiment, generating a multi-scenario scanning dataset based on the digital twin model and the multi-scenario scanning condition database specifically includes the following steps: 331. Determine the first scanning constraint parameters and operation constraint parameters of the ultrasonic probe in the multi-scenario scanning condition database; and determine the second scanning constraint parameters for the digital twin model; 332. Determine the scanning constraint set of the ultrasonic probe based on the first scanning constraint parameter and the second scanning constraint parameter; and determine the safe operation constraint set of the ultrasonic probe based on the operation constraint parameter; 333. Generate simulated scanning control commands for multiple time steps of the ultrasonic probe based on the scanning constraint set and the safe operation constraint set; 334. Control the ultrasound probe to perform a simulated scan in the digital twin model according to the simulated scan control command, and obtain the simulated scan results; 335. Extract the simulated ultrasound image data, simulated pose data, and simulated pressure data corresponding to the multiple time steps from the simulated scanning results; and generate the multi-scene scanning dataset based on the simulated ultrasound image data, the simulated pose data, and the simulated pressure data.

[0046] The first scan constraint parameter refers to the boundary conditions related to the spatial motion of the ultrasound probe extracted from a multi-scenario scan condition database, including the probe's six-degree-of-freedom pose range, upper limit of motion speed, and geometric constraints on the scan trajectory. Operational constraint parameters include contact pressure safety thresholds, abnormal event triggering rules, and emergency operation boundaries. The second scan constraint parameter addresses the anatomical and physical limitations of the current digital twin model, such as extreme values ​​of surface curvature, mechanical interference areas caused by bone protrusions, and accessible organ windows. The scan constraint set is the fusion result of the first and second scan constraint parameters, defining all permissible poses and motion paths of the probe within the simulation space. The safe operation constraint set quantifies the pressure safety range and abnormal response rules during operation. A time step refers to the smallest discrete unit of time in the simulated scan, typically set to the frame interval of clinical ultrasound imaging (10ms). The simulated scan control command is the specific action vector issued to the simulated probe at each time step, including pose adjustment and pressure adjustment.

[0047] Specifically, two types of constraints directly related to probe operation are first extracted from a multi-scenario scanning condition database: the first scanning constraint parameter defines the kinematic boundaries that the probe should follow under ideal conditions. For example, in standardized scanning conditions, the probe pitch angle should be limited to within ±15° and the scanning speed should not exceed 20 mm / s. The operation constraint parameter specifies the safety rules that must be followed regardless of the operating condition, such as an upper limit of 8N for contact pressure and a maximum probe retraction speed of 100 mm / s when an abnormal event is detected. Simultaneously, for the digital twin model, a second scanning constraint parameter is extracted. This second scanning constraint parameter reflects the physical limitations imposed by individual anatomical structures on probe movement, such as allowing only a 10° tilt between ribs and prohibiting large-scale probe roll in areas of abrupt curvature change on the body surface. The first and second scan constraint parameters are logically intersected to obtain the set of scan constraints that the probe can execute in the actual simulation, which is the union of all allowed pose combinations, velocity ranges, and feasible trajectory regions. Simultaneously, the operational constraint parameters are directly mapped to a set of safe operational constraints, which includes not only pressure safety thresholds but also records the mandatory response actions (such as immediately raising the probe or stopping movement) after an anomaly is triggered. Based on this, for each independent task in the million-level concurrent simulation, simulated scan control commands are generated step-by-step according to the time sequence requirements in the working condition descriptor. Using the current probe pose and pressure state as input, the next pose adjustment (Δx, Δy, Δz, Δroll, Δpitch, Δyaw) and pressure adjustment ΔF are generated based on random sampling of the scan constraint set or according to a preset scan strategy. Here, Δx, Δy, and Δz represent the displacement adjustment in the x, y, and z axes, respectively; Δroll, Δpitch, and Δyaw represent the roll angle adjustment, pitch angle adjustment, and yaw angle adjustment, respectively. Next, the safety operation constraint set is verified. If the adjusted pressure value exceeds the safety threshold, it is automatically cropped to the threshold boundary; if the adjusted pose exceeds the scanning constraint set, the command is rejected and regenerated. The verified control command is sent to the digital twin simulation engine, which simultaneously calls the contact mechanics simulation module and the ultrasound image physical rendering module. The contact mechanics simulation module calculates the contact deformation and force feedback between the probe and soft tissue based on finite element analysis, updating the real-time deformation state of the twin; the ultrasound image physical rendering module simulates the propagation, scattering, and attenuation of ultrasound waves in dynamic tissue based on the ultrasonic motion equation and acoustic impedance matching principle, generating a simulated ultrasound image under the current pose. This process is repeated until the entire scanning process is completed. At each time step, three types of data are automatically collected: simulated ultrasound image, simulated pose data (the probe's six-degree-of-freedom coordinate sequence), and simulated pressure data (the instantaneous value of the probe-skin contact force). Finally, the data from all time steps are arranged by timestamp and paired with the corresponding task condition label, automatically generated image quality label, and control parameter label to form a complete simulated scanning sample.

[0048] In an optional embodiment, controlling the ultrasound probe to perform a simulated scan in the digital twin model according to the simulated scan control command to obtain the simulated scan result specifically includes the following steps: 3341. Obtain the simulated scanning control command corresponding to the target time step; the simulated scanning control command includes the pose control parameters and contact pressure control parameters of the simulated ultrasound probe; the target time step is any one of the plurality of time steps; 3342. The ultrasound probe is controlled to move in the digital twin model according to the posture control parameters and the contact pressure control parameters to obtain the simulated tissue structure parameters, simulated dynamic physiological parameters and simulated contact pressure in the digital twin model; 3343. Determine the target simulation scan result corresponding to the target time step based on the simulated tissue structure parameters, the simulated dynamic physiological parameters, and the simulated contact pressure.

[0049] The target time step refers to any sampling moment on the discretized simulation time axis; the simulated scanning control command is the action vector issued to the simulated probe at each time step, which is composed of pose control parameters (six-degree-of-freedom coordinates and Euler angles) and contact pressure control parameters (the normal pressure setpoint of the probe on the skin). The simulated tissue structure parameters refer to the instantaneous state values ​​of the spatial position and geometric shape of the soft tissue and organ boundaries in the digital twin model under the current probe pose and pressure; the simulated dynamic physiological parameters describe the physiological response quantities such as the phase of tissue deformation and movement velocity driven by the respiratory deformation field and the heartbeat deformation field as they change over time; the simulated contact pressure is the actual reaction force between the probe and the body surface calculated by the contact mechanics simulation module, which may differ from the pressure setpoint in the control command. The simulated tissue structure parameters are jointly determined by the superposition of the static anatomical base and the dynamic deformation field of the digital twin model. When the probe moves according to the pose control parameters, the tissue type, organ boundary, and local deformation caused by probe pressure in the corresponding voxel region of the twin model are queried based on the spatial coordinates of the current probe tip. For example, the instantaneous thickness of the left thyroid lobe after compression and the displacement of the carotid artery wall caused by probe compression are quenched. The simulated dynamic physiological parameters reflect the phase state of respiration and heartbeat at the current time step. Since the digital twin model is loaded with respiratory deformation field and heartbeat deformation field driven by the patient's real physiological signals, the instantaneous phase of the respiratory cycle (end of inspiration, end of expiration, or transition phase) and the phase of the cardiac cycle (systole or diastole) are automatically calculated according to the simulation clock at each time step, thereby obtaining dynamic parameters such as diaphragm position, liver lower edge movement speed, and vascular pulsation amplitude. The simulated contact pressure is the real physical reaction force calculated by the finite element analysis algorithm, which comprehensively considers the probe pose, soft tissue stiffness, current physiological deformation phase, and historical pressure cumulative effect. For example, when the control command requires pressing the abdomen with a pressure of 5N, the actual measured simulated contact pressure may fluctuate between 4.7N and 5.3N due to the different respiratory phase (the diaphragm descends during inspiration, causing an increase in abdominal wall tension). If the edge of the probe happens to touch the ribs, the contact mechanics module will detect local stress concentration and feed back a higher reaction force value.

[0050] Specifically, in each independent simulation thread of the million-level concurrent simulated scans, the scanning process is discretized into a continuous sequence of time steps. For any target time step, the simulated scan control instructions, pre-generated from the scan constraint set and safety operation constraint set, are first retrieved from the task scheduling queue. These instructions include pose control parameters, i.e., the desired six-degree-of-freedom spatial pose of the probe, and contact pressure control parameters, i.e., the desired normal pressure value applied by the probe to the patient's skin surface. Next, the simulation engine calls the contact mechanics simulation module and the ultrasound image physical rendering module to drive the virtual ultrasound probe to move within the digital twin model according to the received control instructions. During this process, simulated tissue structure parameters, simulated dynamic physiological parameters, and simulated contact pressure are acquired in real time. Based on this, the ultrasound image physical rendering module is called to render and generate the simulated ultrasound image at the current time step by solving the ultrasound motion equation or an approximate model based on ray tracing, according to the current simulated tissue structure parameters (which determine the acoustic impedance distribution) and simulated dynamic physiological parameters (which determine the Doppler frequency shift or echo speckle changes caused by tissue movement), combined with the simulated contact pressure (which affects the probe-skin coupling quality and thus determines the transmission efficiency of ultrasound energy). The simulated ultrasound images not only closely match the output of real ultrasound equipment in terms of grayscale distribution and texture features, but also incorporate dynamic effects such as slicing drift caused by respiratory movements and vascular pulsation artifacts caused by heartbeats. Finally, the simulated ultrasound image at the current time step, the actual pose and pressure values, the automatically calculated image quality labels (structural similarity index, cross-section over union, signal-to-noise ratio), and the control parameter labels are packaged and encapsulated to form the simulated scan result for that target time step. This process is repeated for all time steps, and the results are stitched together chronologically to constitute the output of a complete simulated scan task.

[0051] It is evident that by using a time-step instruction parsing-physical simulation-parameter feedback closed loop, the complex physical interaction process between the ultrasound probe and dynamic human tissue can be accurately reproduced in a digital twin environment. Each frame of simulated image is simultaneously constrained by static anatomy, dynamic physiology, and contact mechanics, thereby ensuring the physical authenticity of the generated data and its fit with the clinical scenario. This makes the simulated scanning results highly aligned with real clinical operations in terms of both imaging and mechanical characteristics, providing time-series data with complete physical consistency for model training.

[0052] Step S340: Train the first scanning strategy model based on the multi-scenario scanning dataset to obtain the target scanning strategy model.

[0053] The first scanning strategy model is an intelligent control model used to learn the relationship between probe control behavior and ultrasound image changes during ultrasound scanning. It outputs corresponding ultrasound probe control strategies based on the current ultrasound image state, probe pose state, and the patient's dynamic physiological state. The target scanning strategy model is an ultrasound scanning control model that has undergone supervised learning and reinforcement learning training. It can learn not only the section control rules in standard scanning processes but also section correction strategies and emergency response strategies for erroneous scanning scenarios, complex anatomical scenarios, and dynamic abnormal scenarios. The multi-scenario scanning dataset includes simulated ultrasound image data, simulated pose data, simulated pressure data, and condition label data corresponding to standard scanning conditions, erroneous scanning conditions, extreme scanning conditions, and abrupt change scanning conditions. The condition label data is used to characterize the scenario category, section quality level, and abnormal state type corresponding to the current scanning process.

[0054] Specifically, the simulated ultrasound image data, simulated pose data, and simulated pressure data in the multi-scenario scanning dataset are first subjected to unified preprocessing operations. Specifically, grayscale normalization, artifact removal, and temporal frame alignment are performed on the simulated ultrasound image data to enhance the consistency of image features across different scanning scenarios. Spatial coordinate normalization and trajectory smoothing are performed on the simulated pose data to reduce the impact of probe motion jitter on model training stability. Outlier removal and pressure range normalization are performed on the simulated pressure data to improve the stability of pressure control features. Next, the multi-scenario scanning dataset is divided into scenarios based on different operational condition labels to generate standard scanning training subsets, error scanning training subsets, extreme scanning training subsets, and mutation scanning training subsets. Then, a first scanning strategy model is constructed, and the network parameters in the model are initialized. In the supervised learning training phase, the first scanning strategy model is first pre-trained using the standard scanning training subset. In this process, simulated ultrasound image data under standard scanning conditions and corresponding standard probe control parameters are input into the first scanning strategy model, and a supervised loss function is constructed based on the difference between the model output and the standard control labels. The supervised loss function can be expressed as: ; Where L represents the monitoring loss value; Labels representing the actual control parameters of the ultrasonic probe; This represents the control parameters of the ultrasound probe predicted by the first scanning model; Indicates the number of samples.

[0055] Then, the network parameters in the first scanning strategy model are iteratively updated based on the supervised loss function to enable the model to learn the probe control rules and standard section acquisition rules in the standard scanning process. After completing supervised pre-training, the first scanning strategy model is further trained through trial-and-error reinforcement learning based on erroneous scanning training subsets, extreme scanning training subsets, and mutation scanning training subsets. During reinforcement learning training, an environmental state vector is first constructed based on the simulated ultrasound image state, probe pose state, and pressure state corresponding to the current time step. Subsequently, the first scanning strategy model outputs the corresponding probe action control parameters based on the current environmental state and controls the virtual ultrasound probe to perform the corresponding simulated scanning actions in the digital twin model. Then, corresponding reward values ​​are generated based on the section quality, tissue display integrity, probe pressure safety, and abnormal scene response results corresponding to the current time step. Positive rewards are generated when the probe control actions output by the model can improve the clarity of the standard section, reduce the degree of section offset, and maintain the pressure safety range; negative rewards are generated when the actions output by the model result in section blurring, probe pressure exceeding limits, or abnormal response failure. The action decision parameters in the first scanning strategy model are continuously optimized based on reward feedback, enabling the model to gradually learn cross-section correction strategies for complex scenarios and emergency control strategies for abnormal scenarios. Finally, after the model reward value meets the preset convergence condition, the target scanning strategy model is generated.

[0056] It can be seen that by supervising and reinforcing the first scanning strategy model based on a multi-scenario scanning dataset, the target scanning strategy model can not only learn the pattern acquisition and probe control rules in standard scanning scenarios, but also learn the ability to correct deviations and handle emergencies through trial and error in erroneous scanning scenarios, complex anatomical scenarios, and dynamic abnormal scenarios. This significantly improves the generalization and adaptation ability of the target scanning strategy model to different patient anatomical structures, dynamic physiological changes, and complex clinical scanning scenarios, and enhances the scanning stability, pattern accuracy, and abnormal scenario handling ability of the ultrasound robot in real clinical environments.

[0057] In an optional embodiment, training the first scanning strategy model based on the multi-scene scanning dataset to obtain the target scanning strategy model specifically includes the following steps: 341. Divide the multi-scenario scanning dataset according to the multiple scanning conditions to obtain a standard scanning dataset, an error scanning dataset, an extreme scanning dataset, and a mutation scanning dataset; 342. Supervised training of the first scanning strategy model is performed on the standard scanning dataset to obtain a second scanning strategy model; the second scanning strategy model is used to learn the pose control strategy of the ultrasound probe and the ultrasound image quality optimization strategy under the standard scanning conditions. 343. The second scanning strategy model is trained by reinforcement learning based on the error scanning dataset, the extreme scanning dataset, and the mutation scanning dataset to obtain the target scanning strategy model; the target scanning strategy model is used to learn the scanning processing strategies of the second scanning strategy model for the error scanning conditions, the extreme scanning conditions, and the mutation scanning conditions.

[0058] In supervised training, the input is ultrasound images under standard operating conditions, and the output is the desired probe pose and contact pressure. The real labels are the generated labeled data. Reinforcement learning training, on the other hand, is a training paradigm that optimizes decision-making strategies based on reward signals through agent-environment interaction. In reinforcement learning training, the agent is a scanning strategy model, and the environment is a digital twin simulation environment. The model attempts different actions under three types of conditions: error, extreme, and mutation. It learns correction and emergency response strategies based on the rewards (or penalties) from the simulation feedback. The standard scanning dataset corresponds to standardized scanning procedures in clinical guidelines (e.g., standard thyroid section, maximum oblique diameter section of the right lobe of the liver), characterized by regular probe trajectory, stable pressure, and satisfactory image quality. The error scanning dataset contains routine error sequences such as section deviation, abnormal pressure, and blurred images. The extreme scanning dataset covers scanning data under rare anatomical variations (e.g., dextrocardia, absence of the left lobe of the liver) or special body types (e.g., obesity, scoliosis). The mutation scanning dataset simulates emergency events such as sudden postural changes, respiratory arrest, and abnormal equipment parameters.

[0059] Specifically, firstly, using the condition category labels retained during the data generation phase, the entire dataset is divided into four non-overlapping subsets: the standard scan dataset, the error scan dataset, the extreme scan dataset, and the mutation scan dataset. Next, supervised pre-training is performed. In the first stage of supervised pre-training, the first scan strategy model employs an encoder-decoder architecture. The encoder is a convolutional neural network used to extract high-dimensional depth features from the input ultrasound images; the decoder is a fully connected neural network that maps the depth features to the probe's six-degree-of-freedom pose adjustment and contact pressure adjustment for the next time step. During training, a batch of samples is randomly sampled from the standard scan dataset. Each sample contains the current frame ultrasound image and the corresponding automatically labeled control parameters (desired pose and pressure). The first scan strategy model predicts the control parameters based on the current frame image, calculates the mean squared error loss between the predicted values ​​and the labeled ground truth values, and iteratively updates the model weights using the backpropagation algorithm. After sufficient iterations until the loss function converges, the model learns the basic operational rules in standardized scanning, such as keeping the probe perpendicular to the body surface to obtain the optimal acoustic window, sliding at a constant speed along anatomical landmarks to obtain continuous cross-sections, and fine-tuning the probe angle based on image grayscale feedback to optimize image clarity. At this point, the output second scanning strategy model is capable of completing standard scanning tasks, but it lacks robust strategies to cope with operational deviations, anatomical variations, or sudden anomalies. Therefore, in the second stage, trial-and-error reinforcement learning training is performed, using the second scanning strategy model as the initial policy network, and loading three types of work condition datasets—error, extreme, and mutation—into a digital twin simulation environment. In each training session, the model starts from the current state (including multimodal features such as ultrasound images, probe pose, contact pressure, and force feedback), outputs actions (pose adjustment and pressure adjustment), and the simulation environment executes these actions and returns a new state and an immediate reward value. The reward function is designed using a "positive incentive + negative penalty" approach. Positive rewards are given when the standard cross-sectional structure similarity index generated by the second scanning strategy model is high, the anatomical structure integrity is good, and the image signal-to-noise ratio is excellent. Significant negative penalties are imposed when cross-sectional deviation exceeds a threshold, contact pressure exceeds safety limits, or correct emergency operations (such as emergency probe elevation) are not performed in sudden abnormal scenarios. The second scanning strategy model aims to maximize the cumulative discount reward and continuously optimizes the strategy network parameters using a deep deterministic strategy gradient algorithm. Through repeated trial and error in millions of simulation rounds involving errors and extreme scenarios such as cross-sectional loss, rib obstruction, pressure overload, respiratory motion artifacts, and sudden patient coughing, the second scanning strategy model gradually learns to automatically revert to the standard cross-section from deviation states, adjust the search strategy in anatomically variable regions, and execute safety protection actions when abnormalities occur. When the cumulative reward value converges to a preset threshold, reinforcement learning training terminates, yielding the final target scanning strategy model.This model not only inherits the standard operating capabilities of the supervised pre-training phase, but also learns highly generalizable error correction logic and emergency response strategies from massive error feedback, thus enabling it to cope with complex and ever-changing scanning scenarios in real clinical settings.

[0060] It can be seen that the phased training framework decouples standardized operation learning from trial-and-error reinforcement learning. While ensuring the accuracy of basic operations, it significantly improves the model's generalization ability to anatomical variations, operational biases and sudden anomalies through reinforcement feedback from massive virtual error scenarios.

[0061] In an optional embodiment, the first scanning strategy model includes an encoder network and a decoder network, and the step of supervising the training of the first scanning strategy model based on the standard scanning dataset to obtain the second scanning strategy model includes: 3421. Based on the standard scanning data set, determine the first ultrasound image data and the first control parameters of the ultrasound probe corresponding to the first ultrasound image data; the first control parameters include a first pose parameter and a first pressure parameter; 3422. Input the first ultrasound image data into the encoder network to obtain ultrasound image features; 3423. Input the ultrasound image features into the decoder network to obtain the predicted second control parameters of the ultrasound probe; 3424. Determine the first loss function of the first scanning strategy model based on the first control parameter and the second control parameter; 3425. The first scanning strategy model is iteratively updated and trained according to the first loss function and the standard scanning dataset. When the first scanning strategy model meets the preset convergence condition, the iteration is stopped to obtain the second scanning strategy model.

[0062] The encoder network is a deep convolutional neural network, which maps the raw ultrasound images into low-dimensional, semantically abstract high-dimensional deep feature vectors. The decoder network is a fully connected neural network, responsible for further mapping the deep feature vectors into specific control commands. The first ultrasound image data refers to the simulated ultrasound grayscale image generated at each time step in a standard scanning procedure, with dimensions consistent with the output of the actual ultrasound equipment. The first control parameter is the control parameter label generated synchronously with the image, including the probe's six-DOF pose and contact pressure value. The second control parameter is the control command predicted by the model based on the current image, i.e., the model's output. The loss function uses mean squared error to quantify the difference between the predicted and true values. The convergence condition is that the loss function value no longer decreases significantly after several consecutive iterations, or falls below a preset threshold (e.g., 1e-4), or the number of iterations reaches the maximum preset value (e.g., 200 epochs).

[0063] Specifically, in the first stage of supervised pre-training, training samples are first read batch by batch from the standard scan dataset. Each training sample consists of a pair of data: the first ultrasound image data and the first control parameter corresponding to that image. The forward computation process of the model is as follows: the first ultrasound image data (normalized to the [0,1] interval) is input into the encoder network. The encoder adopts a multi-layer convolution-pooling structure. The shallow convolution kernels capture low-level features such as edges and textures in the image (such as organ capsule lines and blood vessel wall echoes), while the deep convolution kernels combine to generate more semantic high-level features (such as thyroid contours and blood vessel distribution patterns in the liver). After global average pooling, the encoder outputs a fixed-dimensional ultrasound image feature vector. Then, the decoder network receives this feature vector and maps it step by step through three fully connected layers, finally outputting a seven-dimensional prediction vector as the second control parameter, corresponding to three translational coordinates, three rotation angles, and one pressure value. Since this prediction value is the control command "inversely deduced" by the model based on the image, the difference between it and the first control parameter during actual acquisition directly reflects the model's current understanding of the ultrasound image-control mapping relationship. To calculate the mean squared error loss between the first and second control parameters, the scalar loss value is obtained by averaging the squared differences of the seven components. A larger loss value indicates a greater deviation between the model's predicted pose or pressure and the actual operation that generated the current image. The backpropagation algorithm is then used to calculate the gradient of the loss value layer by layer along the decoder and encoder directions, and the Adam optimizer is used to update the network's weight parameters. The parameters are updated after each batch (e.g., 256 samples), and the loss for the next batch is recalculated. During training, the loss value generally shows a decreasing trend. When the model's loss value on the validation set no longer decreases for 10 consecutive iterations, or the decrease is less than 1e-5, or the total number of iterations reaches the preset 200, the convergence condition is met, and iteration stops. At this point, the second scanning strategy model has learned the mapping relationship from standard ultrasound images to standardized operating instructions. That is, faced with a standard thyroid section image, the model can accurately predict the probe's current pose and pressure, thus possessing the basic ability to autonomously complete standardized scanning tasks.

[0064] It can be seen that supervised pre-training adopts an encoder-decoder architecture and a mean squared error loss function, using automatically labeled standard working condition data as the supervision signal. This enables the model to efficiently learn the inverse mapping relationship from image to control under the drive of a large number of samples, providing a stable initial strategy with basic operational capabilities for the subsequent reinforcement learning stage. This avoids the problems of unstable or inefficient training and significantly improves the overall training efficiency and the robustness of the final model.

[0065] In an optional embodiment, the step of training the second scanning strategy model using reinforcement learning based on the erroneous scanning dataset, the extreme scanning dataset, and the mutation scanning dataset to obtain the target scanning strategy model specifically includes the following steps: 3431. Determine the second ultrasound image data in the erroneous scan dataset, the extreme scan dataset, and the mutation scan dataset, and the third control parameters of the ultrasound probe; 3432. Construct a first environment state vector for reinforcement learning based on the second ultrasound image data and the third control parameters; 3433. Input the first environmental state vector into the second scanning strategy model to obtain the control adjustment parameters of the ultrasound probe; the control adjustment parameters include pose adjustment parameters and pressure adjustment parameters; 3434. Perform scanning simulation in a digital twin model according to the control adjustment parameters to obtain simulation results; the simulation results include ultrasound image quality parameters, scanning section deviation parameters, pressure parameters, and abnormal working condition response parameters; 3435. Construct the reward function for reinforcement learning based on the simulation results; and generate reward values ​​corresponding to the ultrasound image quality parameters, the scanning section deviation parameters, the pressure parameters, and the abnormal working condition response parameters based on the reward function; 3436. The strategy parameters of the second scanning strategy model are iteratively optimized based on the reward value. When the cumulative reward value meets the preset cumulative reward threshold, the optimal solution at which the iteration stops is obtained. 3437. Adjust the strategy parameters according to the optimal solution to obtain the target scanning strategy model.

[0066] In this context, the environment state vector is a comprehensive feature description representing the agent's current environment in reinforcement learning. It is composed of the current frame ultrasound image (processed by the encoder) and the current probe control parameters (pose and pressure). The control adjustment parameters are the action commands output by the policy network, including pose increments and pressure increments, used to update the probe state at the next moment. The simulation results are a set of physical and image indicators fed back by the digital twin environment after executing the action, including image quality parameters (such as structural similarity index SSIM, signal-to-noise ratio SNR), cross-sectional deviation parameters (the degree of anatomical overlap between the current cross-section and the standard cross-section), contact pressure parameters (actual reaction force value), and abnormal condition response parameters (whether the correct emergency action was executed in a sudden scenario). The reward function is a rule that maps the simulation results to scalar reward values, using a composite form combining positive incentives and negative penalties. The cumulative reward value is accumulated based on the initial reward value, and the cumulative reward threshold is a quantitative standard for judging whether the policy has been successfully trained, set as the upper limit of the average reward value over several consecutive rounds.

[0067] Specifically, firstly, training samples are extracted from error scan datasets, extreme scan datasets, and mutation scan datasets. These samples include second ultrasound image data (i.e., the simulated ultrasound image at the current time step) and third control parameters (the actual probe pose and pressure value corresponding to this image). It's important to note that in the reinforcement learning framework, the third control parameter is not directly used as a supervision signal, but rather as a component of the environment state vector to reflect the current operational state. Next, the second ultrasound image data is compressed into a deep feature vector using a pre-trained encoder. This feature vector is then concatenated with the third control parameter (a seven-dimensional pose vector) to form the first environment state vector for reinforcement learning. This state vector includes visual perception information (what anatomical structures are currently scanned, and the image quality) and proprioceptive information (where the probe is currently located, and the pressure applied), providing the policy network with the complete input required for decision-making. Next, the first environmental state vector... The input is fed into the second scanning strategy model (as the initial policy network), and the model outputs control adjustment parameters. Unlike the control parameters in supervised learning, reinforcement learning outputs actions that are adjustments relative to the current state. Pose adjustment parameters specify the incremental motion to be applied across the six degrees of freedom (e.g., a 2° increase in pitch or a 1mm translation to the right), while pressure adjustment parameters specify the increase or decrease in contact pressure (e.g., an increase of 0.2N). Then, the actions are... The simulation is deployed to the digital twin simulation environment. Based on pose and pressure adjustments, the environment updates the position and orientation of the virtual probe, calls the contact mechanics module to calculate the new interaction forces between the probe and soft tissue, and then generates a new frame of simulated ultrasound image through the ultrasound image physical rendering module. Simultaneously, the environment automatically calculates and returns ultrasound image quality parameters (including the structural similarity index between the current frame and the standard section template, anatomical cross-union ratio, and image signal-to-noise ratio), scan section deviation parameters (the spatial offset distance between the current section and the target anatomical structure), pressure parameters (the actual contact pressure value between the probe and skin, in N), and abnormal condition response parameters (if the current condition is a sudden event such as sudden patient movement, it determines whether the model has performed the correct retraction or stop action). A reward function is also constructed based on the simulation results. Among them, the reward function The reward function includes incentive and penalty terms. Positive incentive terms are positively correlated with image quality parameters, encouraging the model to pursue standard cross-sections and clear images. Negative penalty terms impose significant penalties on behaviors such as cross-section deviation exceeding a threshold, stress exceeding the safety limit, and incorrect response in abnormal scenarios (stress exceeding the limit penalty weight is as high as 50%, and abnormal response failure penalty weight is as high as 100%). An instantaneous reward value can be calculated for each set of simulation results using this reward function. It employs a deep deterministic policy gradient algorithm, based on the reward value. The parameters of the policy network are updated: if an action receives a positive reward, the probability of that action being selected is increased; if it receives a negative penalty, its probability is decreased. Through continuous interaction with the simulation environment, trial and error are continuously conducted, and reward feedback is collected, gradually increasing the cumulative reward value. When the cumulative reward value consistently exceeds the preset cumulative reward threshold (e.g., an average reward value of +500 or more per round) over multiple consecutive training rounds, it indicates that the model can make correct decisions in most error, extreme, and abrupt scenarios. At this point, iteration stops, and the current policy network parameters are saved as the optimal solution. Finally, this optimal solution is assigned to the policy model to obtain the final target scanning policy model. This model not only retains the standard operational capabilities of the supervised pre-training phase but also learns robust error correction logic and emergency response strategies from trial and error feedback in massive edge scenarios. This gives it the comprehensive ability to independently complete high-quality ultrasound scanning tasks in complex real-world clinical environments, significantly improving the model's generalization robustness, operational safety, and accuracy across all clinical scenarios.

[0068] In an optional embodiment, constructing the reward function for reinforcement learning based on the simulation results specifically includes the following steps: A1. Determine the image structure similarity parameter between the ultrasound image quality parameter and the preset cross-sectional image quality; and generate the image quality reward parameter for reinforcement learning based on the structure similarity parameter. A2. Determine the ultrasound scan image integrity parameter corresponding to the second ultrasound image data based on the scan section deviation parameter; and generate the image integrity reward parameter for reinforcement learning based on the ultrasound scan image integrity parameter. A3. Determine the pressure difference between the pressure parameter and the preset safe contact pressure threshold; and generate the pressure penalty parameter for reinforcement learning based on the pressure difference. A4. Construct the reward function for trial-and-error reinforcement learning training based on the abnormal operating condition response parameters, the image quality reward parameters, the image integrity reward parameters, and the stress penalty parameters.

[0069] The Structure-Similar Image (SSIM) parameter is a full-reference evaluation index that measures the similarity between two images in terms of brightness, contrast, and structure. Its value ranges from [0,1], with values ​​closer to 1 indicating greater structural similarity between the current ultrasound image and the standard cross-sectional image. The preset cross-sectional image quality refers to the pre-stored standard cross-sectional template image and its corresponding ideal anatomical structure layout for a specific scanning area (e.g., thyroid, liver). The image integrity parameter is represented by the Intersection over Union (IoU), which calculates the pixel-level overlap ratio between the segmented target anatomical structure region in the current image and the corresponding region in the standard template, reflecting whether key anatomical structures are displayed completely. The safe contact pressure threshold is the maximum permissible pressure value of the probe on human skin (e.g., 8N) set according to clinical operating procedures. Exceeding this value may cause patient discomfort or tissue damage. The pressure difference is the algebraic difference between the current measured pressure and the safe threshold; a positive value triggers a penalty. The abnormal condition response parameter is a Boolean or continuous index used to assess whether the model executes the preset correct emergency actions (e.g., immediately raising the probe or stopping advancement) in sudden abnormal scenarios (e.g., sudden patient movement or respiratory arrest).

[0070] Specifically, firstly, the simulated ultrasound image generated at the current time step is obtained from the simulation results, and a pre-set standard section image template for the corresponding scanning area is called (this template is derived from clinical guidelines and records the typical anatomical structure layout and echo characteristics that should be presented under a standard section). The structural similarity parameters between the two images are then calculated. Unlike pixel-by-pixel error, Image similarity is comprehensively assessed from three dimensions: brightness, contrast, and structure, and is highly correlated with human visual perception. For example, when the probe angle deviates from the standard section, the thyroid capsule line may become blurred or even broken. The value will decrease significantly; when the probe pressure is appropriate and the acoustic window is good, the image contrast is clear, and the SSIM value is close to 0.95. According to... Values ​​are used to generate image quality bonus parameters. ,in The preset positive reward coefficient ( This makes the model more capable of pursuing high performance. During the process, the probe posture is adjusted to obtain images that are closest to the standard section. Secondly, a pre-trained segmentation network is used to extract the contour regions of target organs (such as the liver and thyroid) from the current ultrasound image, while simultaneously extracting the corresponding anatomical structure regions from the standard section template, and calculating the intersection-over-union ratio (IoU) of the two. . The value reflects whether key tissues are completely visible in the scanning field of view. For example, when scanning the liver, if the probe position is too low, the lower corner of the right lobe of the liver may be cut out of the field of view. The value will decrease significantly. According to Generate complete reward parameters for the image. , =0.3, the excitation model maintains the probe position so that the target structure is always within the scanning window. Meanwhile, the pressure parameters in the simulation results represent the actual contact force. The preset safe contact pressure threshold is And calculate the pressure difference. This means that a non-zero difference is generated only when the actual pressure exceeds the safety threshold. Pressure penalty parameter. , For penalty weights ( This means that if the pressure exceeds the limit, the model will immediately receive a strong negative reward. This design forces the model to strictly control the contact pressure within a safe range while meeting image quality requirements, avoiding patient discomfort or tissue damage due to excessive pressure. Finally, the above factors are combined with the abnormal condition response parameters to construct the final composite reward function. Abnormal operating condition response parameters It is a switch quantity: if the current operating condition is normal, If the model performs the correct emergency action (such as withdrawing the probe) in a sudden abnormal scenario, a positive reward is given; if the model fails to respond correctly, a high penalty is imposed. The complete reward function is expressed as follows: ,in , An adjustable coefficient between 0 and 1 is used, and the penalty weight for abnormal response failure is set to 100 to highlight the safety-first principle. Through its comprehensive calculation, the environmental feedback at each time step is quantified into a scalar reward value. When the model outputs an action that makes... , When the pressure is increased but not exceeded, a positive reward is given; conversely, if the cross-section deviates, the pressure is too high, or emergency operations are missed, a severe penalty is imposed. During training, the model continuously adjusts its policy parameters to maximize cumulative rewards, and ultimately learns to autonomously complete scanning tasks while ensuring image quality and operational safety, and to make correct emergency responses in the event of sudden anomalies.

[0071] It can be seen that by constructing a composite reward function that includes image quality, anatomical integrity, stress safety, and abnormal response, the reinforcement learning training process can simultaneously optimize the diagnostic efficacy and operational safety of the scanning strategy, enabling the final model to obtain high-quality standard sections in complex clinical environments while actively avoiding the risks of stress over-limit and emergency errors.

[0072] Step S350: Control the ultrasonic scanning robot to perform ultrasonic scanning tasks according to the target scanning strategy model.

[0073] An ultrasound scanning robot is an intelligent ultrasound device used to perform automated ultrasound scanning tasks. It includes an ultrasound host, a robotic arm control system, an ultrasound probe, a force feedback sensor, and a visual perception module. The robotic arm control system controls the spatial movement and posture adjustment of the ultrasound probe on the patient's body surface; the force feedback sensor detects changes in contact pressure between the ultrasound probe and the patient's body surface in real time; and the visual perception module acquires the patient's current position and changes in the body surface area. An ultrasound scanning task refers to an automated scanning process that performs continuous ultrasound image acquisition and cross-sectional tracking on a target tissue area. This includes standard cross-sectional acquisition, dynamic cross-sectional tracking, abnormal scene correction, and scanning path optimization.

[0074] Specifically, after training the target scanning strategy model, it is first deployed to the main control system of the ultrasound scanning robot. Then, the target scanning strategy model continuously performs dynamic analysis of the real-time ultrasound image status and adjusts the scanning path in real time based on changes in patient respiration, dynamic tissue displacement, and probe contact status. Simultaneously, it continuously records real-time ultrasound image data, probe motion trajectory data, and pressure change data during the scanning process and performs quality assessment of the current scanning results. When the target tissue area is detected to meet the preset scanning coverage requirements and the corresponding section quality meets standard diagnostic requirements, the current ultrasound scanning task is automatically terminated, and the corresponding target ultrasound image results are output.

[0075] For easier understanding, please refer to Figure 4 , Figure 4This is a schematic diagram of the architecture of an ultrasound scanning strategy optimization method based on digital twin provided in an embodiment of this application. As can be seen, the digital twin simulation engine layer 410 includes an individualized twin modeling module 411, a dynamic physiological and physical simulation module 412, an ultrasound image physical rendering module 413, and a contact mechanics simulation module 414. The individualized digital twin modeling module 411 is used to construct a corresponding individualized digital twin model based on the patient's body surface data, ultrasound image data, and medical image data to achieve three-dimensional reconstruction of the patient's real anatomical structure. The dynamic physiological and physical simulation module 412 is used to perform dynamic deformation simulation on the tissue structure in the digital twin model based on the patient's respiratory signals, electrocardiogram signals, and tissue motion parameters, thereby simulating the dynamic physiological changes of the patient in a real clinical environment. The ultrasound image physical rendering module 413 is used to simulate the propagation, reflection, and attenuation process of ultrasound waves in human tissue based on tissue acoustic impedance parameters, ultrasound propagation paths, and tissue scattering characteristics, and generate corresponding simulated ultrasound images. The contact mechanics simulation module 414 is used to simulate the contact force change process between the ultrasound probe and the patient's body surface to dynamically simulate tissue compression deformation, probe pressure changes, and probe slippage behavior. The simulated scan data generation layer 420 includes a full-scene working condition library management module 421, a concurrent simulated scan scheduling module 422, an automated annotation module 423, and a data domain alignment preprocessing module 424. The system includes several modules: a full-scenario scanning condition library management module 421, which constructs a multi-scenario scanning condition library including standard scanning conditions, error scanning conditions, extreme scanning conditions, and mutation scanning conditions, to achieve unified management of complex clinical scanning scenarios; a concurrent simulation scanning scheduling module 422, which controls multiple virtual ultrasound probes in parallel to perform simulated scanning tasks in a digital twin environment according to different scanning conditions, thereby improving the efficiency of simulated scanning data generation; an automated annotation module 423, which automatically generates corresponding condition labels, section labels, and abnormality labels for the generated simulated ultrasound images based on the probe pose, section quality, and tissue display status during the simulated scanning process; and a data domain alignment preprocessing module 424, which aligns the image distribution differences between simulated scanning data and real clinical scanning data to reduce feature domain deviation between simulated and real data, thereby improving the adaptability of simulated data to real clinical scenarios during subsequent model training. After generating the simulated scanning data, the AI ​​model training layer 430 further trains the ultrasound scanning strategy model. The AI ​​model training layer 430 includes a supervised pre-training module 431, a trial-and-error reinforcement learning module 432, a model fine-tuning and verification module 433, and a model lightweighting module 434.The supervised pre-training module 431 is used to train the initial scanning strategy model under supervised learning based on simulated scanning data corresponding to standard scanning conditions, in order to learn the standard section acquisition rules and standard probe control rules. The trial-and-error reinforcement learning module 432 is used to train the model under trial-and-error reinforcement learning based on simulated scanning data corresponding to erroneous scanning conditions, extreme scanning conditions, and sudden scanning conditions, so that the model can learn section correction strategies and abnormal emergency handling strategies in complex scenarios. The model fine-tuning and verification module 433 is used to fine-tune and optimize the trained model and verify its performance by combining real clinical scanning data, so as to improve the model's generalization ability in real clinical environments. The model lightweighting module 434 is used to compress parameters, prune the structure, and optimize inference of the trained model to reduce the consumption of computing resources during the model deployment process, thereby meeting the real-time requirements of edge deployment of ultrasound scanning robots. In addition, the edge-cloud collaborative closed-loop iteration layer 440 includes an edge-cloud data interaction module 441, a twin library iteration update module 442, a condition library iteration update module 443, and a model incremental training module 444. Among them, the edge-cloud data interaction module 441 is used to establish a data synchronization mechanism between the edge ultrasound scanning device and the cloud training platform to realize real-time scanning data upload and model update distribution; the twin library iterative update module 442 is used to continuously update the digital twin model library based on the patient data added during the real clinical scanning process to improve the coverage of different patient anatomical structures of the digital twin environment; the working condition library iterative update module 443 is used to continuously expand the multi-scenario scanning working condition library based on the added abnormal scanning cases and complex clinical scenarios to improve the system's adaptability to complex scenarios; and the model incremental training module 444 is used to perform online incremental training on the target scanning strategy model based on the added scanning data to continuously optimize the scanning performance of the model in the real clinical environment.

[0076] For easier understanding, please refer to Figure 5 , Figure 5 This is a flowchart illustrating another method for optimizing ultrasound scanning strategies based on digital twins provided in this application. As can be seen, the process follows the logic of "simulation modeling - data generation - model training - continuous iteration". This method is supported by digital twin simulation, generates massive amounts of labeled data by constructing a high-fidelity virtual scanning environment, and then combines hierarchical training and edge-cloud collaboration mechanisms to improve the generalization ability and continuously optimize the AI ​​scanning model.

[0077] Specifically, the process begins with step S501, "Construction of a Dynamic Individualized Digital Twin Library." Based on the patient's 3D point cloud, physiological signals, and a general human anatomical model, a non-rigid registration algorithm is used to create an individualized static twin. Then, a dynamic twin is generated based on dynamic physiological constraints such as respiration and heartbeat. This results in a digital twin library covering all ages, body types, and anatomical variations, providing a high-fidelity foundation for individualized anatomical and physiological simulation for virtual scanning. Next, step S502, "Construction of a Full-Scenario Scanning Condition Library," is performed. Based on clinical ultrasound diagnostic and treatment guidelines and historical scanning data, various conditions such as standardized scanning, operational errors, anatomical variations, and sudden abnormalities are quantitatively defined and categorized, constructing a condition library covering all clinical scenarios to provide multi-dimensional scenario constraints for simulated scanning. Finally, step S503, "Generation of Millions of Concurrent Simulated Scans and Automated Annotated Datasets, as well as Data Domain Alignment and Hierarchical Preprocessing," involves a full permutation and combination of the digital twin library and the full-scenario condition library. The process involves concurrent execution of virtual scanning simulations via a cloud computing cluster, generating massive amounts of simulated scanning data and completing automated annotation. Then, an unsupervised adaptive algorithm is used to reduce the difference in feature distribution between the simulated data and real clinical data, ultimately constructing a standardized hierarchical training dataset. Next, step S504, "Supervised Pre-training of the Ultrasound AI Scanning Control Model," is executed. Based on training data from standardized scenarios, supervised learning is used to pre-train the ultrasound scanning control model, enabling the model to master standardized scanning action logic and basic image recognition capabilities. Finally, step S505, "Model Trial-and-Error Reinforcement Learning Training Based on Error Scenarios, and Model Fine-tuning and Edge-Cloud Collaborative Iteration," utilizes error, extreme, and abnormal scenario data to perform deep reinforcement learning training on the pre-trained model. This allows the model to learn correction and emergency response capabilities through massive "trial and error" processes. A small amount of real clinical data is then combined to complete transfer fine-tuning and model lightweighting. Finally, the scanning strategy is continuously iterated and optimized through an edge-cloud collaborative mechanism until the process concludes.

[0078] As can be seen, the ultrasound scanning strategy optimization process driven by digital twin simulation described above can solve the problem of traditional AI model training relying on high-cost real clinical labeled data. Only a small amount of real data is needed to complete model fine-tuning, which greatly reduces the cost and cycle of model training. At the same time, the introduction of a full-scenario working condition library and trial-and-error reinforcement learning enables the model to cover extreme and abnormal working conditions that are difficult to reproduce in clinical practice, which significantly improves the model's generalization ability and emergency response capability. The edge-cloud collaborative iterative mechanism ensures the continuous optimization of the model in clinical applications, effectively solving the technical pain points of poor generalization and low clinical adaptability of existing ultrasound robot scanning AI models, and providing an efficient, low-cost and highly reliable implementation path for the optimization of ultrasound scanning strategies.

[0079] For easier understanding, please refer to Figure 6 , Figure 6This is a schematic diagram of an ultrasound scanning application scenario based on a digital twin model, provided in an embodiment of this application. As can be seen, the ultrasound scanning robot includes a robotic arm and an ultrasound probe mounted at the end of the robotic arm; the ultrasound host is used to receive real-time ultrasound image data acquired by the ultrasound probe and perform corresponding ultrasound imaging processing; the digital twin model is used to construct a corresponding digital twin model based on the patient's real-time ultrasound scanning status, and to generate corresponding scanning control commands based on the digital twin model, thereby realizing real-time linkage control between the ultrasound scanning robot and the digital twin model. Figure 6 In this procedure, the patient lies supine while the robotic arm controls the ultrasound probe to establish contact with the patient's body surface and performs an automated ultrasound scan of the target tissue area. Simultaneously, the ultrasound host displays the corresponding ultrasound images in real time and establishes a two-way data interaction relationship with the digital twin model. The digital twin model can receive real-time images and output corresponding control commands to the ultrasound scanning robot to achieve dynamic optimization control of the scanning path and probe pose.

[0080] Specifically, once the ultrasound scan is initiated, the robotic arm first moves the ultrasound probe to the target scanning area of ​​the patient and establishes a stable coupling between the probe and the patient's body surface through a contact pressure control mechanism. Next, the ultrasound probe continuously emits ultrasound signals into the patient's body and receives ultrasound echo signals reflected from the tissue. The ultrasound host then performs beamforming, echo enhancement, and image reconstruction on the ultrasound echo signals to generate corresponding real-time ultrasound images. Simultaneously, the ultrasound scan system continuously collects probe pose data, probe contact pressure data, and patient dynamic physiological state data for the current time step and sends this data synchronously to the digital twin model. The digital twin model can predict the spatial position changes of the target tissue area in subsequent time steps based on changes in patient respiration, dynamic tissue displacement, and the ultrasound probe's contact state, and dynamically optimizes the probe's movement direction, rotation angle, and contact pressure based on the prediction results. Then, the digital twin model interaction system outputs corresponding control commands to the robotic arm to adjust the ultrasound probe's pose, thereby maintaining the target tissue area in the optimal ultrasound imaging region. Furthermore, the digital twin model not only reflects the patient's current static anatomical structure information but also simulates dynamic physiological changes such as respiration, heartbeat, and tissue compression in real time. For example, when a change in the patient's respiratory rate causes displacement of the target tissue area, the digital twin model can update the spatial position of the target tissue area in real time and drive the robotic arm to synchronously adjust the scanning path of the ultrasound probe. When the contact pressure of the ultrasound probe changes abnormally, the digital twin model can also predict the possible tissue deformation results based on the current tissue compression state and control the robotic arm to dynamically correct the probe pressure to avoid abnormal tissue deformation or patient discomfort due to excessive pressure. Simultaneously, in complex scanning scenarios, such as when the patient has a special body structure, tissue anatomical variations, or sudden body movements, the digital twin model can also dynamically correct the current scanning path by combining historical scanning data and scanning strategies obtained through reinforcement learning, thereby improving the stability of automatic scanning in complex scenarios.

[0081] As can be seen, by implementing the above-described embodiment of an ultrasound scanning strategy optimization method based on digital twins, the following steps are taken: acquiring the patient's multimodal clinical data and the doctor's ultrasound scanning operation data; constructing a digital twin model of the patient based on the human voxel model and the multimodal clinical data; determining multiple scanning conditions based on the ultrasound scanning operation data; constructing a multi-scenario scanning condition library based on the multiple scanning conditions and the ultrasound scanning operation data; generating a multi-scenario scanning dataset based on the digital twin model and the multi-scenario scanning condition library; training the first scanning strategy model based on the multi-scenario scanning dataset to obtain the target scanning strategy model; and controlling the ultrasound scanning robot to perform ultrasound scanning tasks based on the target scanning strategy model. Thus, by utilizing a target scanning strategy model trained in multiple scenarios to control the ultrasound scanning robot in real time, the robot can dynamically adjust the scanning path and probe control parameters based on real-time ultrasound image status, patient dynamic physiological status, and probe contact status. This enables stable and automated scanning of target tissue areas in complex clinical scenarios. Simultaneously, by introducing anomaly correction strategies and emergency response strategies for abnormal scenarios learned during reinforcement learning training, the system can perform real-time scanning path correction and pressure adjustment in complex scenarios such as changes in patient respiration, dynamic tissue displacement, abnormal probe contact, and section deviation. This improves the accuracy of section acquisition, scanning stability, and abnormal scenario handling capabilities during ultrasound scanning, further enhancing the reliability and clinical adaptability of the ultrasound robot in real clinical environments.

[0082] The above primarily describes the solutions of the embodiments of this application from the perspective of the method execution process. It is understood that, in order to achieve the above functions, the server includes the corresponding hardware structure and / or software modules for executing each function. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the examples described in the embodiments provided herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0083] This application embodiment can divide the server into functional units according to the above method example. For example, each function can be divided into a separate functional unit, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this application embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0084] When dividing each function into modules according to its corresponding function. Figure 7 This is a functional block diagram of an ultrasound scanning strategy optimization device based on digital twins provided in this application embodiment. The digital twin-based ultrasound scanning strategy optimization device 700 is applied to a server in an ultrasound scanning system. The digital twin-based ultrasound scanning strategy optimization device 700 includes: The acquisition unit 710 is used to acquire the patient's multimodal clinical data and the doctor's ultrasound scan operation data; the multimodal clinical data includes the patient's body surface data, ultrasound image data and patient physiological data; Processing unit 720 is configured to construct a digital twin model of the patient based on a preset human voxel model and the multimodal clinical data; and to determine multiple scanning conditions based on the ultrasound scanning operation data; the multiple scanning conditions include standard scanning conditions, error scanning conditions, extreme scanning conditions, and mutation scanning conditions; to construct a multi-scenario scanning condition library based on the multiple scanning conditions and the ultrasound scanning operation data; and to generate a multi-scenario scanning dataset based on the digital twin model and the multi-scenario scanning condition library. The computing unit 730 is used to train the first scanning strategy model based on the multi-scenario scanning dataset to obtain the target scanning strategy model; The control unit 740 is used to control the ultrasonic scanning robot to perform ultrasonic scanning tasks according to the target scanning strategy model.

[0085] In an optional embodiment, the computing unit 730, in training the first scanning strategy model based on the multi-scene scanning dataset to obtain the target scanning strategy model, is specifically used for: The multi-scenario scanning dataset is divided according to the multiple scanning conditions to obtain a standard scanning dataset, an error scanning dataset, an extreme scanning dataset, and a mutation scanning dataset. The first scanning strategy model is trained under supervision based on the standard scanning dataset to obtain the second scanning strategy model; the second scanning strategy model is used to learn the pose control strategy of the ultrasound probe and the ultrasound image quality optimization strategy under the standard scanning conditions. The second scanning strategy model is trained using reinforcement learning based on the error scan dataset, the extreme scan dataset, and the mutation scan dataset to obtain the target scanning strategy model. The target scanning strategy model is used to learn the scanning processing strategies for the error scan conditions, the extreme scan conditions, and the mutation scan conditions learned by the second scanning strategy model.

[0086] In an optional embodiment, the first scanning strategy model includes an encoder network and a decoder network, and the computing unit 730 is specifically used for: supervising the training of the first scanning strategy model according to the standard scanning dataset to obtain the second scanning strategy model. The first ultrasound image data and the first control parameters of the ultrasound probe corresponding to the first ultrasound image data are determined from the standard scan dataset; the first control parameters include a first pose parameter and a first pressure parameter. The first ultrasound image data is input into the encoder network to obtain ultrasound image features; The ultrasound image features are input into the decoder network to obtain the predicted second control parameters of the ultrasound probe; The first loss function of the first scanning strategy model is determined based on the first control parameter and the second control parameter; The first scanning strategy model is iteratively updated and trained based on the first loss function and the standard scanning dataset. When the first scanning strategy model meets the preset convergence condition, the iteration is notified to obtain the second scanning strategy model.

[0087] In an optional embodiment, the computing unit 730, in the process of training the second scanning strategy model using reinforcement learning based on the erroneous scanning dataset, the extreme scanning dataset, and the mutation scanning dataset to obtain the target scanning strategy model, is specifically configured to: Determine the second ultrasound image data in the erroneous scan dataset, the extreme scan dataset, and the mutation scan dataset, and the third control parameters of the ultrasound probe; A first environment state vector for reinforcement learning is constructed using the second ultrasound image data and the third control parameters; The first environmental state vector is input into the second scanning strategy model to obtain the control adjustment parameters of the ultrasound probe; the control adjustment parameters include pose adjustment parameters and pressure adjustment parameters. Based on the control adjustment parameters, a scanning simulation is performed in the digital twin model to obtain simulation results; the simulation results include ultrasound image quality parameters, scanning section deviation parameters, pressure parameters, and abnormal working condition response parameters; The reward function for reinforcement learning is constructed based on the simulation results; and reward values ​​corresponding to the ultrasound image quality parameters, the scanning section deviation parameters, the pressure parameters, and the abnormal working condition response parameters are generated based on the reward function. The strategy parameters of the second scanning strategy model are iteratively optimized based on the reward value. When the cumulative reward value meets the preset cumulative reward threshold, the optimal solution at which the iteration stops is obtained. The strategy parameters are adjusted based on the optimal solution to obtain the target scanning strategy model.

[0088] In an optional embodiment, the computing unit 730 is specifically configured to: construct the reward function for reinforcement learning based on the simulation results. Determine the image structure similarity parameters between the ultrasound image quality parameters and the preset cross-sectional image quality; and generate the image quality reward parameters for reinforcement learning based on the structure similarity parameters. The ultrasound scan image integrity parameter corresponding to the second ultrasound image data is determined based on the scan section deviation parameter; and the image integrity reward parameter for reinforcement learning is generated based on the ultrasound scan image integrity parameter. Determine the pressure difference between the pressure parameter and the preset safe contact pressure threshold; and generate the pressure penalty parameter for reinforcement learning based on the pressure difference; The reward function for trial-and-error reinforcement learning training is constructed based on the abnormal operating condition response parameters, the image quality reward parameters, the image integrity reward parameters, and the stress penalty parameters.

[0089] In one optional embodiment, the multimodal clinical data includes patient surface point cloud data, ultrasound data, medical tomographic volume data, and patient physiological signals. The processing unit 720, in constructing the digital twin model of the patient based on a preset human voxel model and the multimodal clinical data, is specifically used for: Determine the first tissue structure parameter, the skeletal structure parameter, and the first acoustic impedance parameter corresponding to the first tissue structure parameter in the human voxel model; The human voxel model is non-rigidly registered based on the body surface point cloud data, the ultrasound data, and the medical tomographic volume data to obtain a first digital twin model corresponding to the patient. Dynamic physiological deformation data of the patient is generated based on the patient's physiological signals; the dynamic physiological deformation data includes respiratory deformation data and heart rate deformation data. The first digital twin model is dynamically deformed and registered based on the dynamic physiological deformation data to obtain the second digital twin model. Determine the second tissue structure parameter, the second dynamic deformation parameter, and the second acoustic impedance parameter in the second digital twin model; The dynamic tissue structure parameter range of the patient is determined based on the first tissue structure parameter and the second tissue structure parameter; and the dynamic acoustic impedance parameter range of the patient is determined based on the first acoustic impedance parameter and the second acoustic impedance parameter. A digital twin model of the patient is constructed based on the range of dynamic tissue structure parameters, the range of dynamic acoustic impedance parameters, the bone structure parameters, and the medical tomographic volume data.

[0090] In an optional embodiment, the processing unit 720, in generating the multi-scene scanning dataset based on the digital twin model and the multi-scene scanning condition library, is specifically used for: Determine the first scanning constraint parameters and operation constraint parameters of the ultrasonic probe in the multi-scenario scanning condition database; and determine the second scanning constraint parameters for the digital twin model; The ultrasound probe's scanning constraint set is determined based on the first scanning constraint parameter and the second scanning constraint parameter; and the ultrasound probe's safe operation constraint set is determined based on the operation constraint parameter. Based on the scanning constraint set and the safe operation constraint set, simulated scanning control commands for multiple time steps of the ultrasonic probe are generated; The simulation scanning control command controls the ultrasound probe to perform a simulation scan in the digital twin model to obtain the simulation scan result; The simulated ultrasound image data, simulated pose data, and simulated pressure data corresponding to the multiple time steps are extracted from the simulated scan results; and the multi-scene scan dataset is generated based on the simulated ultrasound image data, the simulated pose data, and the simulated pressure data.

[0091] In an optional embodiment, the processing unit 720, in controlling the ultrasound probe to perform a simulated scan in the digital twin model according to the simulated scan control command and obtaining the simulated scan result, is specifically used for: Obtain the simulated scanning control command corresponding to the target time step; the simulated scanning control command includes the pose control parameters and contact pressure control parameters of the simulated ultrasound probe; the target time step is any one of the plurality of time steps; The ultrasound probe is moved within the digital twin model according to the pose control parameters and the contact pressure control parameters to obtain simulated tissue structure parameters, simulated dynamic physiological parameters, and simulated contact pressure in the digital twin model. The target simulation scan result corresponding to the target time step is determined based on the simulated tissue structure parameters, the simulated dynamic physiological parameters, and the simulated contact pressure.

[0092] As can be seen, this application provides a digital twin-based ultrasound scanning strategy optimization device, applied to a server in an ultrasound scanning system. This device acquires multimodal clinical data of the patient and ultrasound scanning operation data of the doctor; constructs a digital twin model of the patient based on the human voxel model and the multimodal clinical data; determines multiple scanning conditions based on the ultrasound scanning operation data; constructs a multi-scenario scanning condition library based on the multiple scanning conditions and the ultrasound scanning operation data; generates a multi-scenario scanning dataset based on the digital twin model and the multi-scenario scanning condition library; trains a first scanning strategy model based on the multi-scenario scanning dataset to obtain a target scanning strategy model; and controls an ultrasound scanning robot to perform ultrasound scanning tasks based on the target scanning strategy model. This enables the generation of multi-scenario simulated scanning training data that matches real clinical scenarios based on a dynamic and individualized digital twin model environment, thereby improving scanning accuracy. Furthermore, by driving the first scanning strategy model through trial-and-error reinforcement learning training in error and abnormal scenarios, the ultrasound scanning strategy model's adaptability to different patient anatomical structures, dynamic physiological changes, and complex clinical scenarios is significantly improved. At the same time, the model's error correction ability, scanning stability, and emergency response capability in sudden abnormal scenarios are enhanced, further improving the scanning accuracy of the ultrasound robot in real clinical scenarios.

[0093] This application also provides a digital twin-based ultrasonic scanning strategy optimization system, wherein the digital twin-based ultrasonic scanning strategy optimization system includes some or all of the steps for performing any of the methods described in the above method embodiments.

[0094] This application also provides a computer-readable storage medium storing a computer program for electronic data interchange, which causes a computer to perform some or all of the steps of any of the methods described in the above method embodiments, wherein the computer includes a server.

[0095] This application also provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps of any of the methods described in the above method embodiments. The computer program product may be a software installation package, and the computer may include a server.

[0096] It should be noted that, for the sake of simplicity, the above embodiments are all described as a series of actions. Those skilled in the art should understand that this application is not limited to the described order of actions, as some steps in the embodiments of this application can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions, steps, modules, or units involved are not necessarily essential to the embodiments of this application.

[0097] In the above embodiments, the descriptions of each embodiment in this application have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0098] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

[0099] The steps of the methods or algorithms described in the embodiments of this application can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in RAM, flash memory, ROM, EPROM, electrically erasable programmable read-only memory (EEPROM), registers, hard disk, portable hard disk, read-only optical disk (CD-ROM), or any other form of storage medium well known in the art. An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium.

[0100] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the embodiments of this application. It should be understood that the above descriptions are merely specific embodiments of the embodiments of this application and are not intended to limit the protection scope of the embodiments of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solutions of the embodiments of this application should be included within the protection scope of the embodiments of this application.

Claims

1. A method for optimizing ultrasonic scanning strategies based on digital twins, characterized in that, A server applied in an ultrasonic scanning system, the method comprising: Acquire multimodal clinical data of patients and ultrasound scan operation data of physicians; the multimodal clinical data includes patient body surface data, ultrasound imaging data and patient physiological data; A digital twin model of the patient is constructed based on a pre-defined human voxel model and the multimodal clinical data; and multiple scanning conditions are determined based on the ultrasound scanning operation data; the multiple scanning conditions include standard scanning conditions, error scanning conditions, extreme scanning conditions, and mutation scanning conditions; a multi-scenario scanning condition library is constructed based on the multiple scanning conditions and the ultrasound scanning operation data; and a multi-scenario scanning dataset is generated based on the digital twin model and the multi-scenario scanning condition library. The first scanning strategy model is trained based on the multi-scenario scanning dataset to obtain the target scanning strategy model; The ultrasonic scanning robot is controlled to perform ultrasonic scanning tasks based on the target scanning strategy model.

2. The method as described in claim 1, characterized in that, The step of training the first scanning strategy model based on the multi-scenario scanning dataset to obtain the target scanning strategy model includes: The multi-scenario scanning dataset is divided according to the multiple scanning conditions to obtain a standard scanning dataset, an error scanning dataset, an extreme scanning dataset, and a mutation scanning dataset. The first scanning strategy model is trained under supervision based on the standard scanning dataset to obtain the second scanning strategy model; the second scanning strategy model is used to learn the pose control strategy of the ultrasound probe and the ultrasound image quality optimization strategy under the standard scanning conditions. The second scanning strategy model is trained using reinforcement learning based on the error scan dataset, the extreme scan dataset, and the mutation scan dataset to obtain the target scanning strategy model. The target scanning strategy model is used to learn the scanning processing strategies for the error scan conditions, the extreme scan conditions, and the mutation scan conditions learned by the second scanning strategy model.

3. The method as described in claim 2, characterized in that, The first scanning strategy model includes an encoder network and a decoder network. The step of supervising the training of the first scanning strategy model using the standard scanning dataset to obtain the second scanning strategy model includes: The first ultrasound image data and the first control parameters of the ultrasound probe corresponding to the first ultrasound image data are determined from the standard scan dataset; the first control parameters include a first pose parameter and a first pressure parameter. The first ultrasound image data is input into the encoder network to obtain ultrasound image features; The ultrasound image features are input into the decoder network to obtain the predicted second control parameters of the ultrasound probe; The first loss function of the first scanning strategy model is determined based on the first control parameter and the second control parameter; The first scanning strategy model is iteratively updated and trained based on the first loss function and the standard scanning dataset. When the first scanning strategy model meets the preset convergence condition, the iteration stops, and the second scanning strategy model is obtained.

4. The method as described in claim 2, characterized in that, The step of training the second scanning strategy model using reinforcement learning based on the erroneous scanning dataset, the extreme scanning dataset, and the mutation scanning dataset to obtain the target scanning strategy model includes: Determine the second ultrasound image data in the erroneous scan dataset, the extreme scan dataset, and the mutation scan dataset, and the third control parameters of the ultrasound probe; A first environment state vector for reinforcement learning is constructed using the second ultrasound image data and the third control parameters; The first environmental state vector is input into the second scanning strategy model to obtain the control adjustment parameters of the ultrasound probe; the control adjustment parameters include pose adjustment parameters and pressure adjustment parameters. Based on the control adjustment parameters, a scanning simulation is performed in the digital twin model to obtain simulation results; the simulation results include ultrasound image quality parameters, scanning section deviation parameters, pressure parameters, and abnormal working condition response parameters; The reward function for reinforcement learning is constructed based on the simulation results; and reward values ​​corresponding to the ultrasound image quality parameters, the scanning section deviation parameters, the pressure parameters, and the abnormal working condition response parameters are generated based on the reward function. The strategy parameters of the second scanning strategy model are iteratively optimized based on the reward value. When the cumulative reward value meets the preset cumulative reward threshold, the optimal solution at which the iteration stops is obtained. The strategy parameters are adjusted based on the optimal solution to obtain the target scanning strategy model.

5. The method as described in claim 4, characterized in that, The step of constructing the reward function for reinforcement learning based on the simulation results includes: Determine the image structure similarity parameters between the ultrasound image quality parameters and the preset cross-sectional image quality; and generate the image quality reward parameters for reinforcement learning based on the structure similarity parameters. The ultrasound scan image integrity parameter corresponding to the second ultrasound image data is determined based on the scan section deviation parameter; and the image integrity reward parameter for reinforcement learning is generated based on the ultrasound scan image integrity parameter. Determine the pressure difference between the pressure parameter and the preset safe contact pressure threshold; and generate the pressure penalty parameter for reinforcement learning based on the pressure difference; The reward function for trial-and-error reinforcement learning training is constructed based on the abnormal operating condition response parameters, the image quality reward parameters, the image integrity reward parameters, and the stress penalty parameters.

6. The method according to any one of claims 1-5, characterized in that, The multimodal clinical data includes patient surface point cloud data, ultrasound data, medical tomographic volume data, and patient physiological signals; the construction of the patient's digital twin model based on a preset human voxel model and the multimodal clinical data includes: Determine the first tissue structure parameter, the skeletal structure parameter, and the first acoustic impedance parameter corresponding to the first tissue structure parameter in the human voxel model; The human voxel model is non-rigidly registered based on the body surface point cloud data, the ultrasound data, and the medical tomographic volume data to obtain a first digital twin model corresponding to the patient. Dynamic physiological deformation data of the patient is generated based on the patient's physiological signals; the dynamic physiological deformation data includes respiratory deformation data and heart rate deformation data. The first digital twin model is dynamically deformed and registered based on the dynamic physiological deformation data to obtain the second digital twin model. Determine the second tissue structure parameter, the second dynamic deformation parameter, and the second acoustic impedance parameter in the second digital twin model; The dynamic tissue structure parameter range of the patient is determined based on the first tissue structure parameter and the second tissue structure parameter; and the dynamic acoustic impedance parameter range of the patient is determined based on the first acoustic impedance parameter and the second acoustic impedance parameter. A digital twin model of the patient is constructed based on the range of dynamic tissue structure parameters, the range of dynamic acoustic impedance parameters, the bone structure parameters, and the medical tomographic volume data.

7. The method according to any one of claims 1-5, characterized in that, The step of generating a multi-scenario scanning dataset based on the digital twin model and the multi-scenario scanning condition database includes: Determine the first scanning constraint parameters and operation constraint parameters of the ultrasonic probe in the multi-scenario scanning condition database; and determine the second scanning constraint parameters for the digital twin model; The ultrasound probe's scanning constraint set is determined based on the first scanning constraint parameter and the second scanning constraint parameter; and the ultrasound probe's safe operation constraint set is determined based on the operation constraint parameter. Based on the scanning constraint set and the safe operation constraint set, simulated scanning control commands for multiple time steps of the ultrasonic probe are generated; The simulation scanning control command controls the ultrasound probe to perform a simulation scan in the digital twin model to obtain the simulation scan result; The simulated ultrasound image data, simulated pose data, and simulated pressure data corresponding to the multiple time steps are extracted from the simulated scan results; and the multi-scene scan dataset is generated based on the simulated ultrasound image data, the simulated pose data, and the simulated pressure data.

8. The method as described in claim 7, characterized in that, The step of controlling the ultrasound probe to perform a simulated scan in the digital twin model according to the simulated scan control command, and obtaining the simulated scan results, includes: Obtain the simulated scanning control command corresponding to the target time step; the simulated scanning control command includes the pose control parameters and contact pressure control parameters of the simulated ultrasound probe; the target time step is any one of the plurality of time steps; The ultrasound probe is moved within the digital twin model according to the pose control parameters and the contact pressure control parameters to obtain simulated tissue structure parameters, simulated dynamic physiological parameters, and simulated contact pressure in the digital twin model. The target simulation scan result corresponding to the target time step is determined based on the simulated tissue structure parameters, the simulated dynamic physiological parameters, and the simulated contact pressure.

9. A device for optimizing ultrasonic scanning strategies based on digital twins, characterized in that, A server used in an ultrasonic scanning system, the device comprising: The acquisition unit is used to acquire the patient's multimodal clinical data and the doctor's ultrasound scan operation data; the multimodal clinical data includes the patient's body surface data, ultrasound image data and patient physiological data; The processing unit is configured to construct a digital twin model of the patient based on a preset human voxel model and the multimodal clinical data; and to determine multiple scanning conditions based on the ultrasound scanning operation data; the multiple scanning conditions include standard scanning conditions, error scanning conditions, extreme scanning conditions, and mutation scanning conditions; to construct a multi-scenario scanning condition library based on the multiple scanning conditions and the ultrasound scanning operation data; and to generate a multi-scenario scanning dataset based on the digital twin model and the multi-scenario scanning condition library. The computing unit is used to train the first scanning strategy model based on the multi-scenario scanning dataset to obtain the target scanning strategy model; The control unit is used to control the ultrasonic scanning robot to perform ultrasonic scanning tasks according to the target scanning strategy model.

10. A digital twin-based ultrasonic scanning strategy optimization system, characterized in that, The digital twin-based ultrasound scanning strategy optimization system includes instructions for performing the steps of the method as described in any one of claims 1-8.

Images