Ultrasound automatic scan control method and robotic system for swallowing function assessment

Abstract

The application discloses an ultrasonic automatic scanning control method and a robot system for swallowing function evaluation, and the method comprises the following steps: acquiring original heterogeneous data sources, including a contact force vector of a probe and skin, an ultrasonic image stream, an electromyogram envelope signal, a current real-time pose of a mechanical arm, performing time sequence synchronization, extracting time domain feature vectors, frequency domain feature vectors, electromyogram-force feature maps and ultrasonic motion features, and constructing a composite feature vector to predict a current swallowing stage label and a state transition probability of a subject; adjusting a physical interaction stiffness according to the current swallowing stage label, calculating a driving instruction acting on a joint of the mechanical arm according to a real-time generated expected target pose; modeling a process of ultrasonic robot scanning and evaluation cooperation as a hierarchical Markov decision process for iterative optimization, outputting a scanning scheme, and adjusting an execution pose vector according to a real-time target pose deviation. The application realizes significant improvement of swallowing dynamic perception depth and self-adaptive tracking capability.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic scanning control technology, specifically to an automated ultrasonic scanning control method and robotic system for swallowing function assessment. Background Technology

[0002] While existing imaging assessment methods for swallowing function have achieved visualization to some extent, the following technical bottlenecks still exist:

[0003] (1) Insufficient multimodal information fusion and real-time analysis capabilities: Existing image assessments such as video fluoroscopy swallowing imaging (VFSS) involve ionizing radiation and cannot be used as a routine bedside screening method; while fiberoptic endoscopy swallowing assessment (FEES) can directly visualize the mucosa, it is invasive and has a white screen period. Although conventional handheld ultrasound is non-invasive and can be performed at the bedside, its scanning section is highly dependent on the physician's manual skills. A single two-dimensional section cannot fully capture the three-dimensional rapid movement trajectory of the hyoid-larynx complex at the moment of swallowing, resulting in low keyframe capture rate and poor repeatability;

[0004] (2) Lack of dynamic standardization and consistency assurance in the scanning process: Traditional ultrasound assessment of swallowing function is often based on static or semi-static observation by the physician holding the probe. Since the swallowing action is completed in a very short time, it is difficult for the manual holding of the probe to maintain a constant contact force between the probe and the skin surface and a precise sound beam incident angle within this time window. Once the probe slips, rotates or the pressure changes drastically, it will cause image artifacts or loss of target structures, and it will be impossible to automatically recover the tracking of anatomical landmarks, which seriously affects the consistency and reliability of the measurement data;

[0005] (3) Multi-joint robotic arms cannot achieve compliant and safe interaction with human subjects: Automatic scanning requires the coordination of multiple tasks, such as path planning of a six-axis robotic arm, probe posture adjustment, contact force sensing feedback, and respiratory motion compensation. Existing research schemes generally implement path planning, visual servoing, and force control separately, lacking force-position hybrid control logic for the deformation of soft tissues in the neck; especially at the moment of larynx elevation, if the robotic arm cannot retreat in real time according to the biomechanical characteristics of the tissue, it is very easy to trigger the patient's cough reflex or soft tissue compression discomfort, which limits the application of robots in bedside assessment of conscious patients;

[0006] (4) It is difficult to balance edge inference and scanning efficiency: Many experimental automated scanning systems rely excessively on external optical tracking equipment or high-performance workstations for offline image analysis. Once they leave the specific experimental environment or the computing power is limited, complex soft tissue deformation prediction and probe repositioning cannot be performed. For community screening or primary medical institutions, this model has the problem of high equipment deployment costs and also seriously restricts the availability and popularization of swallowing assessment technology in non-ideal environments.

[0007] To fully automate the complex scanning-identification-quantification process and enable robots to execute autonomously, precise collaboration is required in soft tissue biomechanical modeling, real-time ultrasound image semantic segmentation, human-machine compliant control, and cross-modal data synchronization of vision and touch. Traditional ultrasound equipment or auxiliary positioning supports can only provide passive support and cannot support dynamic structural tracking and adaptive scanning tasks for live subjects in a state of spontaneous breathing.

[0008] Furthermore, existing research on ultrasound scanning robots mostly focuses on relatively static or periodically moving organs such as the abdomen and heart, using open-loop control through pre-scanning planning or static models. It fails to address automated detection and closed-loop control optimization for explosive, non-periodic, high-speed movements like swallowing. It also lacks a technical path for achieving precise target localization and dynamic tracking under the coupling interference of the subject's spontaneous breathing and swallowing reflexes, through real-time image feedback, contact force sensing, and robotic arm motion calculation. Current methods generally rely on high-precision 3D cameras or prior knowledge of magnetic positioning, lacking real-time compensation and control for objective interferences such as elastic deformation of neck soft tissue, changes in probe-skin coupling pressure, and involuntary patient movement. The scanning process is easily interrupted due to tracking loss and is difficult to recover autonomously, resulting in a lack of temporal integrity and mechanical interpretability. Meanwhile, traditional robotic arm control algorithms cannot support dynamic and compliant following of deep, rapidly moving targets such as the hyoid bone and thyroid cartilage in the narrow anterior cervical triangle area. They lack the ability to autonomously initiate scanning, relocate lost targets, and intelligently capture keyframes. Furthermore, they cannot dynamically adjust scanning sensitivity and spatial range based on the different properties of swallowed substances (such as water, porridge, and solids).

[0009] Therefore, there is an urgent need for an automated ultrasound scanning control technology that can achieve adaptive probe posture control under dynamic respiratory and swallowing coupling interference, has real-time target loss recovery capability, and can provide standardized scanning sections and quantitative indicators for swallowing function assessment. Summary of the Invention

[0010] To overcome the defects and shortcomings of existing technologies, this invention provides an automated ultrasound scanning control method and robotic system for swallowing function assessment. This invention addresses the needs for automated, standardized, and safe ultrasound scanning and dynamic tracking of key anatomical structures in the pharynx during swallowing function assessment. Through multimodal perception fusion, dynamic anatomical atlas decision-making, and hierarchical reinforcement learning, it achieves a complete closed loop from data acquisition to collaborative scanning assessment.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] This invention provides an automated ultrasound scanning control method for swallowing function assessment, comprising the following steps:

[0013] Acquire raw heterogeneous data sources, including the contact force vector between the probe and the skin, ultrasound image stream, electromyographic envelope signal, and the current real-time pose of the robotic arm;

[0014] The original heterogeneous data source is synchronized in time, and data preprocessing and feature extraction are performed to extract time domain feature vectors, frequency domain feature vectors, electromyography-mechanical feature maps, and ultrasound motion features.

[0015] A composite feature vector is constructed based on time-domain feature vectors, frequency-domain feature vectors, electromyographic-mechanical feature maps, and ultrasound motion features. The composite feature vector is used to predict the subject's current swallowing stage label and state transition probability.

[0016] The physical interaction stiffness is adjusted according to the current swallowing stage label, and the driving commands acting on the robotic arm joints are calculated based on the real-time generated desired target pose.

[0017] The process of coordinated ultrasound robot scanning and assessment is modeled as a hierarchical Markov decision process and iteratively optimized, including:

[0018] Based on the latent state image features corresponding to the ultrasound image stream, the current contact force vector, the electromyographic envelope features, and the current robotic arm end pose, a multi-source input for the current control cycle is constructed, a reliability score index for real-time scanning is constructed, and the adjustment intensity of the current control feedback strategy is calculated based on the difference between the reliability score index and the standard reference threshold.

[0019] Anatomical landmarks are mapped to nodes in a graph neural network structure. The anatomical co-motor relationships between nodes are learned, and the swallowing function is scored. The scores are converted into node attributes and injected into the scanning knowledge graph. The scanning plan is output, and the execution posture vector is adjusted according to the real-time target pose deviation.

[0020] As a preferred technical solution, data preprocessing and feature extraction specifically include:

[0021] The contact force vector and electromyographic envelope signal are normalized according to their respective preset sensor ranges. A sliding window with a fixed window length is used to divide the normalized signal into frames. The mean, variance, and peak value time-domain statistics are extracted from the signal in the current window to form a time-domain feature vector. Simultaneously, a short-time Fourier transform is performed on the window signal to extract the frequency domain feature vector. ;

[0022] The framed signals are processed by a convolutional layer to extract electromyographic precursor features before swallowing initiation, and an electromyographic-mechanical feature map is output by a fully connected layer. ;

[0023] The ultrasound image flows through a pre-trained keypoint detection network, which outputs the two-dimensional pixel coordinates of each anatomical landmark in the current frame and the corresponding confidence heatmap. This forms a keypoint pixel displacement sequence in temporal order. This sequence is then encoded temporally by a pre-trained long short-term memory network, and its hidden state outputs the ultrasound motion characteristics of the anatomical structure. .

[0024] As a preferred technical solution, a composite feature vector is constructed based on time-domain feature vectors, frequency-domain feature vectors, electromyography-mechanical feature maps, and ultrasound motion features, specifically represented as follows:

[0025] ;

[0026] in, Represents a composite feature vector. This indicates a splicing operation. This indicates a time-averaged pooling operation. This represents the electromyographic-mechanical characteristic map. Represents the time-domain eigenvector. Represents the frequency domain eigenvector. Indicates ultrasonic motion characteristics;

[0027] Based on composite feature vectors, the current swallowing stage label and state transition probability of the subject are predicted, specifically including:

[0028] Using a composite feature vector and a sequence of historical states from the previous few time steps as input observations, a pre-trained Hidden Markov Model is used to probabilistically model the swallowing physiological time sequence, inferring and outputting the label of the subject's current swallowing stage. and state transition probability.

[0029] As a preferred technical solution, the physical interaction stiffness is adjusted according to the current swallowing stage label, and the final driving commands acting on the robotic arm joints are calculated based on the real-time generated desired target pose, specifically including:

[0030] The mapping rules between preset stages and parameters are used to adjust the stiffness and damping matrices of the robotic arm end-effector admittance controller based on the current swallowing stage label.

[0031] The desired joint driving torque vector for the current control cycle is obtained using a quadratic programming algorithm, and is expressed as follows:

[0032] ;

[0033] ;

[0034] ;

[0035] in, This represents the joint driving torque vector. For the Jacobian matrix of the robotic arm, Let be the deviation vector between the desired target pose and the current actual end-effector pose. Indicates the desired target pose. Let this be the vector of deviation between the desired end velocity and the current actual end velocity. For dynamic feedforward compensation torque, Zero spatial torque, It is the identity matrix. Jacobian matrix The false rebellion, Based on the current joint angle Singularity avoidance and joint constraint potential energy functions constructed with respect to physical safety boundaries. The gradient of the potential energy function is given.

[0036] As a preferred technical solution, a trajectory prediction step is also included when the visual target is lost, specifically including:

[0037] When rapid elevation of the larynx causes temporary loss or large-area obstruction of anatomical targets in ultrasound images, a long short-term memory network is used to predict the estimated pose sequence of the hyoid-larynx complex based on historical pixel displacement sequences and composite feature vectors.

[0038] By embedding kinematic constraints based on cervical anatomy maps into probability distribution sampling, a safe target pose is generated, represented as:

[0039] ;

[0040] in, Represents the historical pixel displacement sequence. This represents the estimated pose sequence. Represents a composite feature vector. Indicates kinematic constraints, Indicates the safe target pose.

[0041] As a preferred technical solution, it also includes a soft tissue elastic deformation compensation step and a safety monitoring step. The soft tissue elastic deformation compensation step specifically includes:

[0042] A physical-grade spring-damped mass model is constructed based on the biomechanical parameters of the neck. The contact force vector between the probe and the skin is input into the physical-grade spring-damped mass model to obtain the skin compression deformation depth in the current depth direction.

[0043] The skin compression deformation depth is superimposed and corrected onto the depth component of the desired target pose to achieve soft tissue elastic deformation compensation.

[0044] The specific steps of security monitoring include:

[0045] The observation residual is obtained by calculating the difference between the kinetic energy of the robotic arm and the work done by the external force in real time. The interaction force between the probe and the throat is decomposed into quasi-static pressure component and high-frequency impact component by using low-pass and high-frequency separation filters.

[0046] When the residual or high-frequency impact component exceeds the preset safety force threshold, switch to the safety retraction mode: calculate the safety retraction vector weighted by the joint limit margin in the opposite direction of the current contact force, and guide the robotic arm to retract and unload the force along the direction of the vector.

[0047] As a preferred technical solution, the steps for generating the target pose specifically include:

[0048] Construct the state vector at the current moment, represented as:

[0049] ;

[0050] in, This represents the state vector at the current moment. This refers to the anatomical structure segmentation mask and key point two-dimensional coordinates output in real time from the current ultrasound image by an online ultrasound image segmentation network. The optical flow estimation results are based on two adjacent ultrasound images. This represents the current contact force vector. Indicates the quality score of the ultrasound image;

[0051] The state vector is input into the feedforward dynamic graph modulation network to predict the evolutionary correlation strength between each anatomical landmark node and the current probe observation plane, and the scan correlation matrix within the topology is output. , is represented as:

[0052] ;

[0053] in, and These are the trainable weight matrix and bias term of the modulation network, respectively. Use the Sigmoid activation function;

[0054] The temporal graph convolution operator aggregates the features of each node across the topological dimension, as follows:

[0055] ;

[0056] in, To add a spatial incidence matrix with self-loops, It is the identity matrix. for The degree matrix, For the first Layer node feature embedding, For the first The trainable transformation weight matrix of the layer, It is a non-linear activation function;

[0057] The inference results of the last layer of the graph convolution are regressed by a fully connected layer, outputting a continuous candidate scanning path with confidence scores;

[0058] Multiple semantic metrics are calculated from the dynamic atlas subgraph. Using a multilayer perceptron as the mapping function, these semantic metrics are converted into six-dimensional pose increments of the probe, represented as:

[0059] ;

[0060] ;

[0061] in, This represents the six-dimensional pose increment of the probe. Represents a mapping function. For the decision weight matrix, This is a respiratory movement compensation item. Indicates structural visibility. Indicates the degree of motion coordination. Indicates the force-position coupling degree. Indicates the standard degree of the section;

[0062] Electromyographic precursor features before swallowing initiation are extracted. During the pharyngeal phase, the decision weight for laryngeal tracking is increased through a gating mechanism. During the resting breathing phase, the sectional calibration weight is strengthened. The final generated pose increments are accumulated and used as the desired target pose.

[0063] As a preferred technical solution, a reliability scoring index for real-time scanning is constructed, expressed as:

[0064] ;

[0065] in, This represents the cumulative number of valid standard ultrasonic sections actually captured in the current test that exceed a set threshold in terms of matching with the standard section template. The target total number of sections is preset for clinical use. This represents the duration of consecutive successful tracking since the last time the target was lost. This represents the cumulative frequency of lost targets. These are the corresponding preset positive weighting coefficients;

[0066] The current control feedback strategy adjustment strength is calculated based on the difference between the reliability score index and the standard reference threshold, and is expressed as:

[0067] ;

[0068] in, This is the integral gain term. Indicates the standard reference threshold;

[0069] By controlling the feedback strategy to adjust the weight of the intensity scaling force control penalty term, the upper limit of movement speed, and the electromyographic feedforward gain, the individual differences arising from the body shape and swallowed matter characteristics of different subjects can be adapted.

[0070] As a preferred technical solution, the scoring results are converted into node attributes and injected into the scanning knowledge graph to output the scanning plan. The execution posture vector is then adjusted based on the real-time target pose deviation. Specifically, this includes:

[0071] Building a knowledge graph for scanning for:

[0072] ;

[0073] in, A collection of medical entities It is a set of clinical medical logical relationships between entities;

[0074] The output scanning plan is as follows:

[0075] ;

[0076] in, For the subject's historical scan sequence features and action tags, This is a scheme generation network based on a self-attention mechanism. The score indicates the swallowing function.

[0077] Adjust the execution attitude vector based on the real-time target pose deviation. , is represented as:

[0078] ;

[0079] in, For real-time target pose deviation, For the corresponding proportional, integral, and derivative control gains.

[0080] The present invention also provides a robot system equipped with a controller, the controller performing the above-described ultrasonic automatic scanning control method for swallowing function assessment.

[0081] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0082] (1) Current ultrasound assessment of swallowing function mainly relies on the clinical physician’s manual operation experience or the fixation of simple mechanical stents. It lacks the ability to perceive and adapt to the dynamic process of swallowing in real time. The standardization of the scanning section is low and the data is fragmented. It cannot effectively cover multi-dimensional physiological and physical indicators such as electromyographic timing, contact force interaction, and anatomical movement trajectory. At the same time, it lacks the adaptability to individual differences of subjects and human-machine safety collaboration mechanism during the execution of automatic scanning tasks. This results in low key frame capture rate at the moment of swallowing, difficulty in recovering lost targets, poor repeatability of assessment results, and difficulty in obtaining complete and objective quantitative analysis support. This forms a typical bottleneck of difficulty in sensory integration, difficulty in continuous tracking, and difficulty in ensuring safety.

[0083] This invention introduces key technologies such as multimodal swallowing timing prediction, dynamic anatomical atlas decision-making, hierarchical safety constraint reinforcement learning, and knowledge graph fusion reasoning. It supports the collaborative acquisition of information from channels such as ultrasound images, contact force, sEMG electromyography, and probe pose. It integrates real-time swallowing physiological characteristics and uses a time-series graph network and a deep strategy model for autonomous scanning path planning and safe and compliant control. This realizes the transformation of swallowing assessment from subjective experience dependence to objective quantitative autonomy and constructs a closed-loop human-machine collaborative assessment paradigm.

[0084] (2) This invention significantly improves the depth of dynamic swallowing perception and adaptive tracking capability: Compared with the traditional ultrasound scanning method that relies on the physician to manually hold the probe or use a fixed mechanical support for passive support, this invention introduces an adaptive probe compliant control mechanism based on multimodal perception and swallowing timing prediction. It can automatically adjust the admittance control parameters according to the sEMG electromyographic precursor signal, the hyoid bone movement trend and contact force change characteristics in the ultrasound image, and achieve high-precision dynamic tracking of explosive swallowing movements. It supports lost target prediction compensation, force-sensory blind tracking and context perception recovery. It no longer relies solely on a single visual servo, but can integrate neuromuscular electrophysiological information for feedforward prediction, significantly reducing the loss rate of swallowing keyframes. Through variable impedance compliant control technology, the system exhibits follow-up characteristics when the larynx is rapidly raised and stable scanning characteristics during the resting period, ensuring adaptive response to different swallowing phases and significantly improving the dynamic tracking accuracy and clinical applicability of ultrasound scanning.

[0085] (3) This invention achieves a breakthrough in autonomous decision-making and interpretability of scanning paths: Traditional ultrasound assessment relies entirely on the physician's personal experience for selecting scanning sections, lacking standardized path planning and logical explanation. The autonomous decision-making method for scanning paths proposed in this invention, based on dynamic anatomical atlases and swallowing temporal reasoning networks, maps the subject's personalized neck anatomy, hyoid-larynx motion priors, real-time ultrasound image quality, and contact force status in an atlas. This invention can automatically generate a decision chain for anatomical positioning, motion prediction, and section optimization according to different stages of swallowing, realizing dual-modal fusion verification of data-driven and anatomical prior knowledge. It can not only intuitively obtain standardized assessment section images, but also trace back the anatomical basis of each scanning decision, significantly reducing individual differences in manual operation and improving the transparency and comparability of assessment results across operators.

[0086] (4) The safety collaborative control based on hierarchical task planning in this invention greatly improves the fault tolerance of the process: Compared with the traditional open-loop control or single force control strategy of robotic arms, this invention establishes a reentrant scanning-evaluation collaborative task management mechanism through a hierarchical reinforcement learning framework. The automatic scanning process is abstracted into a transferable state of resting positioning, swallowing tracking, safety verification, and keyframe capture. It can automatically identify risks and backtrack to the previous effective scanning state in abnormal situations such as image label loss, sudden change in contact force, or subject body movement, ensuring that the evaluation process does not collapse or interrupt. This reentrant mechanism enables the scanning process to have the ability of "pause-resume-continue", which is especially suitable for clinical evaluation scenarios with strong sudden swallowing movements and large individual variations. Under complex interference conditions, it effectively reduces the rate of manual intervention and improves task continuity and data integrity.

[0087] (5) This invention constructs a closed-loop management system for behavior based on human-machine collaborative tracking and multimodal feedback. Compared with previous scanning systems that only record static image results, this invention calculates the scanning reliability score by comprehensively analyzing the interaction signals between the scanning agent and the evaluation agent, analyzes the standard section capture progress, anatomical structure tracking time and contact force stability in real time, and automatically generates personalized strategy adjustment instructions. The system can determine whether there is tracking deviation or safety risk based on the clinical priority of swallowing assessment, and dynamically adjust the probe movement speed, force control parameters and image acquisition sequence to achieve closed-loop management of perception, decision-making, execution, feedback and correction. This mechanism enables the system to have multimodal behavior tracking and active safety intervention capabilities, shifting from simple mechanical motion control to deep physiological process collaborative management, which significantly improves the safety, continuity and individualized adaptability of swallowing function assessment.

[0088] (6) Knowledge graph and deep strategy integration to achieve adaptive scanning plan generation: Unlike existing systems that rely on fixed scanning protocols, this invention constructs a swallowing ultrasound knowledge graph that includes anatomical structures, standard sections, probe techniques, and the relationship between assessment indicators. Through graph neural networks, intelligent scanning strategy reasoning is achieved. It can automatically identify scanning difficulties based on the subject's neck anatomical features and real-time electromyography ultrasound data, and retrieve corresponding individualized probe adjustment plans from the graph, forming an intelligent cycle of assessment, decision-making, and optimization. It can not only output highly interpretable scanning operation suggestions, but also dynamically adjust the tracking sensitivity and spatial search range according to the characteristics of the swallowed material, so that the scanning strategy can be flexibly adapted to the subject's characteristics and differences in swallowing tasks, realizing the true personalization and intelligence of swallowing function ultrasound assessment. Attached Figure Description

[0089] Figure 1 This is a flowchart illustrating the automated ultrasound scanning control method for swallowing function assessment according to the present invention. Detailed Implementation

[0090] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0091] Example 1

[0092] like Figure 1 As shown, this embodiment provides an automated ultrasound scanning control method for swallowing function assessment, including the following steps:

[0093] S1: Perform multimodal data synchronous acquisition to obtain raw heterogeneous data sources, providing a high-precision time-aligned benchmark for subsequent feature extraction and state estimation, specifically including:

[0094] S11: Heterogeneous Data Acquisition: Four types of raw heterogeneous data are simultaneously acquired via a six-dimensional force / torque sensor, ultrasound probe, laryngeal surface electromyography electrodes, and robotic arm joint encoder integrated into the robot's end effector. Specifically:

[0095] The contact force vector between the probe and the skin is directly measured by a six-dimensional force / torque sensor. ;

[0096] Ultrasound image stream acquired in real time by an ultrasound imaging system ;

[0097] The electromyography envelope signal is obtained by rectifying and low-pass filtering the surface electromyography (sEMG) signal. ;

[0098] The current real-time pose of the end effector is obtained by solving the joint angle vector fed back by the joint encoder of the robotic arm through the forward kinematics model of the robotic arm. ;

[0099] S12: Real-time timing synchronization: The above four types of heterogeneous data are assigned a unified timing timestamp via the real-time control bus and are synchronously triggered by the system master clock to ensure that the contact force vector between the probe and the skin is synchronized during rapid and dynamic swallowing. Ultrasound image stream Electromyographic envelope signal and position Strict time alignment is achieved, and the synchronized data stream is transmitted to step S2 in real time.

[0100] S2: Multimodal signal preprocessing and feature extraction: Receive the synchronous heterogeneous data stream output from step S1, perform deep feature mining based on the data characteristics of different modalities, and transform it into high-dimensional spatial temporal features, specifically including:

[0101] S21: Extraction of Mechanical and Electromyographic Signal Features:

[0102] For contact force vector and electromyographic envelope signal Each sensor is normalized according to its preset sensor range, and the normalized signal is divided into frames using a sliding window with a fixed window length. The mean, variance, and peak value time-domain statistics are extracted from the signal in the current window to form a time-domain feature vector. Simultaneously, a short-time Fourier transform (STFT) is performed on the window signal to extract the frequency domain feature vector. .

[0103] Furthermore, the segmented signal sequence is input into a pre-trained one-dimensional convolutional neural network (1D-CNN). The convolutional layers automatically capture the electromyographic precursors before swallowing initiation, and the fully connected layers output a deep electromyographic-mechanical feature map. .

[0104] S22: Ultrasound image feature extraction:

[0105] Ultrasound image stream Input a pre-trained lightweight keypoint detection network (this network is pre-trained and converged on an ultrasound image dataset with manually labeled key anatomical landmarks such as the hyoid bone and thyroid cartilage). The network outputs the two-dimensional pixel coordinates of each anatomical landmark in the current frame and the corresponding confidence heatmap, forming a keypoint pixel displacement sequence in time sequence. .

[0106] Subsequently, the pixel displacement sequence The data is fed into a pre-trained Long Short-Term Memory (LSTM) network for temporal feature encoding, and its hidden state outputs the ultrasound motion features of the anatomical structure. All features extracted in step S2 ( , , , and displacement sequence Both are sent to steps S3 and S43.

[0107] S3: Multimodal Feature Fusion and Swallowing Phase Recognition: Based on the multi-source features extracted in step S2, a composite feature representing the current human-computer interaction state of the system is constructed, and the real-time swallowing physiological phase of the subject is inferred. This provides a phase state label for the variable impedance regulation in step S4 and the autonomous decision-making in step S6, specifically including:

[0108] S31: Multimodal feature space fusion: The temporal features extracted in step S21 are fused together. Frequency domain characteristics Depth features and the ultrasonic motion features extracted in step S22 Perform time-average pooling separately, and then concatenate the pooled modal vectors horizontally. This forms a composite feature vector that integrates muscle electrophysiology, biomechanical interaction, and anatomical dynamics. :

[0109] ;

[0110] S32: Real-time prediction during the swallowing stage: combining feature vectors Using the historical state sequence of the preceding time steps as the input observations, a pre-trained Hidden Markov Model (HMM) is used to probabilistically model the swallowing physiological time sequence. The observation probability and transition probability of this HMM model are obtained by training with standard labels from simultaneous fluoroscopy of swallowing (VFSS). The model infers and outputs the label of the subject's current swallowing stage in real time. (Including the oral preparation phase, pharyngeal phase, esophageal phase, or resting phase) and the probability of state transition. Stage labels. and composite features The data will be transmitted in real time to steps S4 and S6 respectively.

[0111] S4: Adaptive Compliant Control and Trajectory Replanning: This step, as the underlying motion-driven execution unit, receives the stage label from step S32. Dynamically adjust the physical interaction stiffness and adjust it according to the desired target pose generated in real time. The final drive commands applied to the robotic arm joints are calculated, including:

[0112] S41: Adaptive adjustment of variable impedance parameters: based on the current swallowing stage label predicted in step S32. Based on the preset stage-parameter mapping rules, the stiffness matrix of the robotic arm end-effector admittance controller is dynamically adjusted. With damping matrix Specifically:

[0113] when During the pharyngeal phase (when the larynx rises rapidly), the mapping rule will... Set to low stiffness value Set to high damping value to make the probe exhibit zero-force compliant characteristics that follow the undulations of the skin surface;

[0114] when During the resting period, the mapping rule will Set to a high stiffness value to maintain a stable clamping force to complete the standard scan;

[0115] S42: Quadratic Programming (QP) Solution for Compliant Control Law: Real-time contact force constraints, joint physical limits, and ultrasound image quality scores are incorporated into the optimization objective and constraint set. The quadratic programming (QP) algorithm is used to solve for the expected joint driving torque vector in real time during the current control cycle. :

[0116] ;

[0117] in, The Jacobian matrix of the robotic arm is calculated in real time from the robotic arm's geometric parameters and the currently sampled joint angles. Let be the deviation vector between the desired target pose and the current actual end-effector pose. Indicates the desired target pose. This is the deviation vector between the desired end velocity and the current actual end velocity; The dynamic feedforward compensation torque is calculated online based on the established rigid body dynamics model of the robotic arm, using pre-identified inertial parameters, Coriolis force, gravity, and friction terms. The zero spatial torque is derived based on the gradient projection method: , It is the identity matrix. Jacobian matrix The false rebellion, Based on the current joint angle Singularity avoidance and joint constraint potential energy functions constructed with respect to physical safety boundaries. This is the gradient of the potential energy function;

[0118] S43: Trajectory prediction when visual target is lost (blind tracking catch-up mechanism): When rapid elevation of the larynx causes temporary loss or large-area occlusion of the anatomical target in the ultrasound image, the Long Short-Term Memory (LSTM) network is used to predict the trajectory based on the historical pixel displacement sequence output in step S22. The composite feature vector of step S31 Predicting the estimated pose sequence of the hyoid-larynx complex in future time periods ;

[0119] Simultaneously, kinematic constraints based on cervical anatomical atlases are introduced. (This constraint is obtained from prior neck CT / MRI data through statistical shape modeling to limit the physiological displacement range of the hyoid bone relative to the cervical vertebrae.) By embedding the constraint into probability distribution sampling, a safe target pose conforming to anatomical features is generated:

[0120] ;

[0121] The distribution is decoded using a beam search to determine the target pose for autonomous blind following during target loss. The desired target pose is sent to step S42. One of the sources;

[0122] S5: Soft tissue deformation compensation and safety monitoring. This step works in parallel with the control law calculation in step S4, specifically to offset feedback errors caused by skin pressure deformation and provide physical-dimensional emergency safety protection, including:

[0123] S51: Soft tissue elastic deformation compensation: Using the neck biomechanical parameters (including elastic modulus and damping coefficient) obtained in advance through in vitro tissue physical experiments, a physical-grade spring-damped mass model is constructed. The contact force vector collected in step S11 is then used to... Substituting into the model, the skin compression deformation depth in the current depth direction is calculated. ;

[0124] This embodiment will Real-time overlay correction to the desired target pose in step S42 This compensates for the overshoot caused by tissue deformation in pure visual position feedback by focusing on the depth component.

[0125] S52: Collision Detection and Safe Active Retreat: Establish a state monitoring mechanism based on an energy observer, obtain the observation residual by calculating the difference between the kinetic energy of the robotic arm and the work done by the external force in real time, and decompose the probe-throat interaction force into quasi-static pressure component and high-frequency impact component using low-pass and high-frequency separation filters.

[0126] When the residual or high-frequency impact component exceeds the preset safety force threshold (determined to be due to subject coughing, large-scale body movement or mechanical collision), the conventional control flow is immediately blocked and switched to the safety retreat mode: calculate the safety retreat vector weighted by the joint limit margin along the opposite direction of the current contact force, and guide the robotic arm to quickly retract and unload the force along the direction of the vector.

[0127] S53: Digital Twin Visualization Synchronization: The motion pose, force data and ultrasound image stream of the entire scanning process are mapped to the digital twin interface in real time via the control bus for presentation and full-time recording;

[0128] S6: Autonomous Scanning Path Decision Based on Dynamic Anatomical Atlas: This step integrates prior sensory features to construct a high-dimensional spatiotemporal atlas and uses real-time reasoning to generate the desired scan target pose with standard clinical cross-sectional significance. And then it is sent to step S4, the specific steps of which include:

[0129] S61: State Representation of Four-Dimensional Dynamic Decision Atlas: Using key anatomical landmarks such as the resting hyoid bone, thyroid cartilage, and glottis as spatial references, a four-dimensional dynamic atlas is generated, integrating anatomical structure, motion sequence, mechanical interaction, and image quality. Its state vector at the current moment... Defined as:

[0130] ;

[0131] in, This refers to the anatomical structure segmentation mask and key point two-dimensional coordinates output in real time from the current ultrasound image by an online ultrasound image segmentation network. The optical flow estimation results are based on two adjacent ultrasound images, combined with the electromyographic envelope signal from step S11. The instantaneous motion vector obtained through joint calculation (used to infer the direction of muscle activation); The current six-dimensional contact force / torque vector is directly fed back from step S11; The quality score of ultrasound images is obtained by comprehensively weighting the gray-level co-occurrence matrix of the region of interest (ROI) (to obtain contrast and energy indicators) and the mean value of edge gradient magnitude.

[0132] S62: Scanning the correlation matrix generation: This involves generating the state vector... The input is fed into a feedforward dynamic graph modulation network to predict the evolutionary correlation strength between each anatomical landmark node and the current probe observation plane, and the output is a scan correlation matrix within the topology. :

[0133] ;

[0134] in, and These are the trainable weight matrix and bias term of the modulation network, respectively. This is the Sigmoid activation function. This matrix represents the observation priority of different structures at the current swallowing phase.

[0135] S63: Temporal Graph Convolutional Inference and Candidate Path Generation: Aggregating Node Features Across Topological Dimensions Using Temporal Graph Convolutional Operators (GCNs):

[0136] ;

[0137] in, To add a spatial incidence matrix with self-loops, It is the identity matrix; for The degree matrix; For the first Layer node feature embedding (initial input) From the state vector (obtained through encoding) For the first The trainable transformation weight matrix of the layer; is a non-linear activation function. The inference result of the last layer of the graph convolution is regressed by a fully connected layer, outputting a continuous candidate scan path with confidence scores.

[0138] S64: Semantic-pose Nonlinear Mapping: Real-time computation of four semantic metrics from dynamic atlas subgraphs: structural visibility Motion coordination Force-position coupling degree and cross-sectional standard (The mutual information between the current slice and the preset standard slice template) constitutes the high-level decision feature vector:

[0139] ;

[0140] Using a multilayer perceptron (MLP) as the mapping function The semantic metrics are directly converted into the underlying probe six-dimensional pose increment. :

[0141] ;

[0142] in, This is the decision weight matrix; This is the respiratory motion compensation term (obtained in real-time from the respiratory time sequence collected by the chest surface motion sensor). Based on the electromyographic precursor features extracted in step S21, the decision weights related to laryngeal tracking are automatically increased through a gating mechanism during the pharyngeal phase, while the standard section calibration weights are strengthened during the resting respiratory phase. The resulting pose increments are then accumulated and used as the desired target pose. Enter step S42;

[0143] S65: Adaptive adjustment of swallowing context: fusing the current swallowing task type vector input from the human-computer interface through a graph attention network (GAT). (e.g., one-hot coding of water swallowing or porridge swallowing) and baseline kinematic prior matrix of healthy individuals (Based on the features extracted by principal component analysis of the swallowing imaging trajectory compression of healthy individuals), a personalized decision bias vector is generated to dynamically scale the search space range of the scanning path and the extreme movement speed of the probe to adapt to different clinical subject situations.

[0144] S66: Decision-Execution Timing Synchronization and Multi-Step Compensation: The absolute timestamp of swallowing initiation is jointly determined and latched by the start time when the sEMG envelope signal exceeds the resting baseline threshold and the frame time when the anatomical structure in the ultrasound image first shows a significant displacement. Calculate the moment when the high-level decision-making instruction is issued. Synchronization deviation with the absolute timestamp of the swallowing initiation :

[0145] ;

[0146] in, This describes the subject response delay balance coefficient obtained from preliminary experimental calibration. The control unit evaluates the current timing state deviation every 10 ms and adaptively reduces or increases the inference step size of the HCN. If a prediction error occurs due to processing delay, the timing prediction compensation is activated in real time. :

[0147] ;

[0148] in, This is the preset smoothing factor constant; The future anatomical structure location predicted in step S43; The compensation amount is directly incorporated into the final control position command to ensure that the path converges quickly to the real motion trajectory, as the actual position of the anatomical structure obtained by the key point detection network in step S22 is the actual position of the anatomical structure.

[0149] S7: Scan-Assessment Collaborative Control Based on Hierarchical Reinforcement Learning: This step implements a decision optimization mechanism by abstracting the operational states of steps S1 to S6 into a reinforcement learning environment. Through online policy iteration, it achieves joint optimization of the accuracy of the entire scanning process and clinical assessment. Specifically, this includes:

[0150] S71: Definition of Hierarchical Markov Decision Architecture: The process of ultrasonic robot scanning and evaluation coordination is modeled as a hierarchical Markov decision process (HMDPS) seven-tuple. :

[0151] in, The global perception state space is formed by horizontally concatenating and combining all the high-dimensional features output from steps S2 and S3. The joint action space includes the continuous amount of probe fine-tuning pose at the bottom layer and the discrete triggering action of key standard frame capture at the top layer. The state transition probability; A weighted multi-objective reward function that integrates standard image cross-section matching accuracy, throat tracking accuracy, force stability (torque variance), and task completion time; Discount factor; The strategy is a hierarchical coordination strategy (the higher-level strategy is responsible for making decisions and scanning the macroscopic model, while the lower-level strategy is responsible for outputting continuous pose control commands). It is a set of safety constraints consisting of a six-dimensional force upper limit, a joint motion soft limit, and an image quality baseline.

[0152] State evolution strictly follows:

[0153] ;

[0154] Among them, actions are sampled from the current policy network. .

[0155] S72: Multi-source state adaptive fusion and policy network update: Defining the multi-source input vector for the current control cycle :

[0156] in, The hidden state image features of LSTM in step S22; This is the current contact force data; This is a characteristic of electromyographic envelope; This represents the current pose of the robotic arm's end effector.

[0157] Each modal data is processed by a dedicated fully connected network encoder. After extraction, the current stage probability output from step S32 is activated by Softmax to generate dynamic weight coefficients. The weighted fusion yields a globally unified state vector. :

[0158] ;

[0159] The hierarchical policy network optimization employs the Multi-Agent Proximal Policy Optimization (MAPPO) algorithm, which prunes the objective loss function. Defined as:

[0160] ;

[0161] in, This represents the probability ratio between the old and new strategies; The value of the current dominance function is calculated using the Generalized Dominance Estimation (GAE) algorithm. The preset strategy clipping magnitude hyperparameter is used. The network ultimately outputs the optimal action based on the action value function. This action acts as a fine-tuning instruction and is applied to the controller in step S4 in real time.

[0162] S73: Agent Fault Tolerance and Operation Backtracking: Defining Binary Logic Gating Functions ;

[0163] When image quality score When the set ultrasound image quality threshold is met, the image status is marked as ImageStatus = Valid;

[0164] When the contact force satisfies ( When the preset safe force limit is reached, the force status is marked as ForceStatus=Safe.

[0165] If both conditions are met, then let the logic gate function... Execute the standard vision-force fusion intelligent control process; otherwise, order It automatically switches to the local fault-tolerant blind following mode (based on the electromyographic precursor features and historical motion trajectory of step S21 for feedforward estimation), or directly triggers sudden regression.

[0166] At the same time, maintain a stack of records of executed operations in memory. When image loss continues for more than a set period, a backtracking function is called. The robotic arm is controlled to autonomously drive the probe to restore it to the pose point of the most recent reliable anatomical structure observation.

[0167] S74: Comprehensive Assessment and Personalized Feedback Adjustment of Scan Reliability: Constructing a Real-Time Scan Reliability Scoring Index :

[0168] ;

[0169] in, This represents the cumulative number of valid standard ultrasonic sections actually captured in the current test that exceed a set threshold in terms of matching degree with the standard section template; The target total number of sections is preset for clinical use; The duration of continuous successful tracking since the last time the target was lost; This represents the cumulative frequency of lost targets. These are the corresponding preset positive weighting coefficients.

[0170] Based on this real-time score and the standard reference threshold The difference is used to dynamically calculate the adjustment intensity of the current control feedback strategy. :

[0171] ;

[0172] in, This is the integral gain term. Through this intensity... The dynamic online scaling of the force control penalty weight, upper limit of movement speed, and electromyographic feedforward gain in step S4 automatically adapts to individual differences caused by different subjects' body shapes and swallowed food characteristics.

[0173] S75: Key Anatomical Structure Graph Neural Network (GNN) Modeling: Key anatomical landmarks such as the hyoid bone, thyroid cartilage, and cricoid cartilage are mapped to nodes in a graph topology. The edges between nodes and their weights are established based on the kinematic correlations in a statistical shape model (SSM) constructed from a large amount of neck clinical imaging data. The anatomical co-motion relationships between nodes are learned using graph convolutional layers, and its hierarchical evolution formula is as follows:

[0174] ;

[0175] in, For anatomical nodes In the Graph feature representation of the layer (initial input) (Obtained by splicing the two-dimensional pixel coordinates of the key point with local appearance features); For nodes The set of adjacent nodes; The normalization coefficient used to preserve the feature scale is set to a value of [value missing]. ; This is a trainable feature mapping matrix between layers; Using a non-linear activation function, this model is used to accurately predict the anatomically associated movements of hyoid bone elevation and larynx anterior displacement;

[0176] S76: Medical Knowledge Graph Reasoning and Personalized Scanning Plan Generation: Constructing a scanning knowledge graph that includes standard section definitions, typical anatomical landmarks, clinical scanning techniques (sliding, tilting, rotation), and functional assessment indicators. ,in, A collection of medical entities It is a set of clinical medical logical relationships between entities.

[0177] The swallowing function score of the subjects output by the deep reinforcement learning network The node attributes are converted and injected into the knowledge graph. The association weights between entities are updated through a graph neural network, thereby inferring and recommending personalized initial probe placement pose, global scanning path, and initial force control preset schemes for the current subject, either offline or online. The dynamic generation and update function of the scanning scheme is expressed as follows:

[0178] ;

[0179] in, G represents the historical scan sequence features and action labels recorded by the subject in the system database; G is a scheme generation network based on the self-attention mechanism.

[0180] S77: Continuous Optimization of the Closed-Loop Control Process: After a single complete swallowing assessment task, multi-dimensional feedback data is collected throughout the entire process, and the network parameters of the overall hierarchical reinforcement learning network are optimized using the policy gradient algorithm. Perform a complete set of backpropagation optimizations:

[0181] ;

[0182] in, The learning rate is preset. At the underlying execution control level, a classic incremental PID controller fine-tunes the probe's attitude vector based on the calculated deviation. :

[0183] ;

[0184] in, This refers to the real-time target pose or force deviation. The corresponding proportional, integral, and derivative control gains are tuned using the Ziegler-Nichols method. This creates a closed-loop continuous optimization management system encompassing evaluation, intervention, and feedback, achieved through both macro-level strategies and micro-level drivers.

[0185] S8: Robust recovery under abnormal operating conditions: This step serves as the boundary robustness guarantee unit for the entire process control. It is triggered when the state evaluation function in step S7 or the image feature confidence level in step S3 exceeds the preset safety limit, in order to ensure the reproducibility and safety of the solution under unstable clinical operating conditions.

[0186] S81: Handling Severe Acoustic Window Attenuation: When the subject experiences violent swallowing causing misalignment of the coupling agent between the probe and the skin, or when changes in the tracheal position cause severe attenuation of the ultrasound acoustic window, the image quality score calculated in step S61 may be affected. When the ultrasonic attenuation threshold is lower than preset, the control flow immediately suspends the visual tracking channel and automatically switches to a pure force-driven servo mode: utilizing only the contact force vector fed back by the six-dimensional force sensor in step S11. By analyzing the temporal variation trend, the current upward movement direction of the throat is estimated through a constructed Kalman filter, which controls the robotic arm to drive the probe to maintain physical contact and continuous following.

[0187] S82: Active micro-probe verification under low confidence: When the maximum value of the anatomical landmark confidence heatmap output by the key point detection network is lower than the set confidence threshold (indicating that the ultrasound section is suspected of having serious artifacts or drift), the system activates the active micro-probe mechanism: Under the premise of maintaining basic contact force control, a set of high-frequency, small-amplitude sinusoidal physical micro-vibrations are applied to the probe-skin interaction interface through the bottom force control channel (the vibration frequency and amplitude are adaptively set according to the tissue impedance characteristics in step S51), and the mechanical impedance change feedback of the robotic arm end is monitored in real time.

[0188] If the impedance matching characteristics match the human skin boundary, the image reduction system is determined to be an artifact interference, and the original pose tracking is maintained; if the impedance characteristics are abnormal, the probe is determined to be out of contact, the scan is immediately stopped, and an error is reported.

[0189] S83: State Smooth Convergence and Resting Relocation: After determining the end of a single swallowing action or the recovery from a brief loss of target, the system relocates the current dynamic decision state vector according to the set state convergence index model. A smooth transition resets the robot to its initial resting standard repositioning mode, preventing sudden changes in speed and torque.

[0190] ;

[0191] in, This is the system state vector for the next control cycle during the transition phase; These are the convergence rate coefficients of the positive constants obtained through experimental calibration, used to constrain the acceleration during the repositioning process; This is the duration of the timer after the loss of the target event or the completion of the swallowing action. After convergence and reset, it returns to the resting scan state and resets the control loop, waiting for step S1 to trigger the dynamic following of the next new swallowing cycle.

[0192] The step numbers in the above embodiments are only set for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0193] Example 2

[0194] This embodiment provides a robot system equipped with a controller, which executes the ultrasound automatic scanning control method for swallowing function assessment described in Embodiment 1 above.

[0195] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for automated ultrasound scanning control for swallowing function assessment, characterized in that, Includes the following steps: Acquire raw heterogeneous data sources, including the contact force vector between the probe and the skin, ultrasound image stream, electromyographic envelope signal, and the current real-time pose of the robotic arm; The original heterogeneous data source is synchronized in time, and data preprocessing and feature extraction are performed to extract time domain feature vectors, frequency domain feature vectors, electromyography-mechanical feature maps, and ultrasound motion features. A composite feature vector is constructed based on time-domain feature vectors, frequency-domain feature vectors, electromyographic-mechanical feature maps, and ultrasound motion features. The composite feature vector is used to predict the subject's current swallowing stage label and state transition probability. The physical interaction stiffness is adjusted according to the current swallowing stage label, and the driving commands acting on the robotic arm joints are calculated based on the real-time generated desired target pose. The process of coordinated ultrasound robot scanning and assessment is modeled as a hierarchical Markov decision process and iteratively optimized, including: Based on the latent state image features corresponding to the ultrasound image stream, the current contact force vector, the electromyographic envelope features, and the current robotic arm end pose, a multi-source input for the current control cycle is constructed, a reliability score index for real-time scanning is constructed, and the adjustment intensity of the current control feedback strategy is calculated based on the difference between the reliability score index and the standard reference threshold. Anatomical landmarks are mapped to nodes in a graph neural network structure. The anatomical co-motor relationships between nodes are learned, and the swallowing function is scored. The scores are converted into node attributes and injected into the scanning knowledge graph. The scanning plan is output, and the execution posture vector is adjusted according to the real-time target pose deviation.

2. The automated ultrasound scanning control method for swallowing function assessment according to claim 1, characterized in that, Data preprocessing and feature extraction specifically include: The contact force vector and electromyographic envelope signal are normalized according to their respective preset sensor ranges. A sliding window with a fixed window length is used to divide the normalized signal into frames. The mean, variance, and peak value time-domain statistics are extracted from the signal in the current window to form a time-domain feature vector. Simultaneously, a short-time Fourier transform is performed on the window signal to extract the frequency domain feature vector. ; The framed signals are processed by a convolutional layer to extract electromyographic precursor features before swallowing initiation, and an electromyographic-mechanical feature map is output by a fully connected layer. ; The ultrasound image flows through a pre-trained keypoint detection network, which outputs the two-dimensional pixel coordinates of each anatomical landmark in the current frame and the corresponding confidence heatmap. This forms a keypoint pixel displacement sequence in temporal order. This sequence is then encoded temporally by a pre-trained long short-term memory network, and its hidden state outputs the ultrasound motion characteristics of the anatomical structure. .

3. The automated ultrasound scanning control method for swallowing function assessment according to claim 1, characterized in that, A composite feature vector is constructed based on time-domain feature vectors, frequency-domain feature vectors, electromyography-mechanical feature maps, and ultrasound motion features, specifically represented as follows: ； in, Represents a composite eigenvector. This indicates a splicing operation. This indicates a time-averaged pooling operation. This represents the electromyographic-mechanical characteristic map. Represents the time-domain eigenvector. Represents the frequency domain eigenvector. Indicates ultrasonic motion characteristics; Based on composite feature vectors, the current swallowing stage label and state transition probability of the subject are predicted, specifically including: Using a composite feature vector and a sequence of historical states from the previous few time steps as input observations, a pre-trained Hidden Markov Model is used to probabilistically model the swallowing physiological time sequence, inferring and outputting the label of the subject's current swallowing stage. and state transition probability.

4. The automated ultrasound scanning control method for swallowing function assessment according to claim 1, characterized in that, The physical interaction stiffness is adjusted based on the current swallowing stage label, and the final drive commands acting on the robotic arm joints are calculated based on the real-time generated desired target pose. Specifically, these commands include: The mapping rules between preset stages and parameters are used to adjust the stiffness and damping matrices of the robotic arm end-effector admittance controller based on the current swallowing stage label. The desired joint driving torque vector for the current control cycle is obtained using a quadratic programming algorithm, and is expressed as follows: ； ； ； in, This represents the joint driving torque vector. For the Jacobian matrix of the robotic arm, Let be the deviation vector between the desired target pose and the current actual end-effector pose. Indicates the desired target pose. Let this be the vector of deviation between the desired end velocity and the current actual end velocity. For dynamic feedforward compensation torque, Zero spatial torque, It is the identity matrix. Jacobian matrix The false rebellion, Based on the current joint angle Singularity avoidance and joint constraint potential energy functions constructed with respect to physical safety boundaries. The gradient of the potential energy function is given.

5. The automated ultrasound scanning control method for swallowing function assessment according to claim 1, characterized in that, It also includes a trajectory prediction step when a visual target is lost, specifically including: When rapid elevation of the larynx causes temporary loss or large-area obstruction of anatomical targets in ultrasound images, a long short-term memory network is used to predict the estimated pose sequence of the hyoid-larynx complex based on historical pixel displacement sequences and composite feature vectors. By embedding kinematic constraints based on cervical anatomy maps into probability distribution sampling, a safe target pose is generated, represented as: ； in, Represents the historical pixel displacement sequence. This represents the estimated pose sequence. Represents a composite eigenvector. Indicates kinematic constraints, Indicates the safe target pose.

6. The automated ultrasound scanning control method for swallowing function assessment according to claim 1, characterized in that, It also includes soft tissue elastic deformation compensation steps and safety monitoring steps. The soft tissue elastic deformation compensation steps specifically include: A physical-grade spring-damped mass model is constructed based on the biomechanical parameters of the neck. The contact force vector between the probe and the skin is input into the physical-grade spring-damped mass model to obtain the skin compression deformation depth in the current depth direction. The skin compression deformation depth is superimposed and corrected onto the depth component of the desired target pose to achieve soft tissue elastic deformation compensation. The specific steps of security monitoring include: The observation residual is obtained by calculating the difference between the kinetic energy of the robotic arm and the work done by the external force in real time. The interaction force between the probe and the throat is decomposed into quasi-static pressure component and high-frequency impact component by using low-pass and high-frequency separation filters. When the residual or high-frequency impact component exceeds the preset safety force threshold, switch to the safety retraction mode: calculate the safety retraction vector weighted by the joint limit margin in the opposite direction of the current contact force, and guide the robotic arm to retract and unload the force along the direction of the vector.

7. The automated ultrasound scanning control method for swallowing function assessment according to claim 1, characterized in that, The specific steps for generating the desired target pose include: Construct the state vector at the current moment, represented as: ； in, This represents the state vector at the current moment. This refers to the anatomical structure segmentation mask and key point two-dimensional coordinates output in real time from the current ultrasound image by an online ultrasound image segmentation network. The optical flow estimation results are based on two adjacent ultrasound images. This represents the current contact force vector. Indicates the quality score of the ultrasound image; The state vector is input into the feedforward dynamic graph modulation network to predict the evolutionary correlation strength between each anatomical landmark node and the current probe observation plane, and the scan correlation matrix within the topology is output. , represented as: ； in, and These are the trainable weight matrix and bias term of the modulation network, respectively. Use the Sigmoid activation function; The temporal graph convolution operator aggregates the features of each node across the topological dimension, as follows: ； in, To add a spatial incidence matrix with self-loops, It is the identity matrix. for The degree matrix, For the first Layer node feature embedding, For the first The trainable transformation weight matrix of the layer, It is a non-linear activation function; The inference results of the last layer of the graph convolution are regressed by a fully connected layer, outputting a continuous candidate scanning path with confidence scores; Multiple semantic metrics are calculated from the dynamic atlas subgraph. Using a multilayer perceptron as the mapping function, these semantic metrics are converted into six-dimensional pose increments of the probe, represented as: ； ； in, This represents the six-dimensional pose increment of the probe. Represents a mapping function. For the decision weight matrix, This is a respiratory movement compensation item. Indicates structural visibility. Indicates the degree of motion coordination. Indicates the force-position coupling degree. Indicates the standard degree of the section; Electromyographic precursor features before swallowing are extracted. During the pharyngeal phase, the decision weight for laryngeal tracking is increased through a gating mechanism. During the resting breathing phase, the sectional calibration weight is strengthened. The final generated pose increments are accumulated and used as the desired target pose.

8. The automated ultrasound scanning control method for swallowing function assessment according to claim 1, characterized in that, A reliability score for real-time scanning is constructed, expressed as: ； in, This represents the cumulative number of valid standard ultrasonic sections actually captured in the current test that exceed a set threshold in terms of matching with the standard section template. The target total number of sections is preset for clinical use. This represents the duration of consecutive successful tracking since the last time the target was lost. This represents the cumulative frequency of lost targets. These are the corresponding preset positive weighting coefficients; The current control feedback strategy adjustment strength is calculated based on the difference between the reliability score index and the standard reference threshold, and is expressed as: ； in, This is the integral gain term. Indicates the standard reference threshold; By controlling the feedback strategy to adjust the weight of the intensity scaling force control penalty term, the upper limit of movement speed, and the electromyographic feedforward gain, the individual differences arising from the body shape and swallowed matter characteristics of different subjects can be adapted.

9. The automated ultrasound scanning control method for swallowing function assessment according to claim 1, characterized in that, The scoring results are converted into node attributes and injected into the scanning knowledge graph. A scanning plan is output, and the execution posture vector is adjusted according to the real-time target pose deviation. Specifically, this includes: Building a knowledge graph for scanning for: ； in, A collection of medical entities It is a set of clinical medical logical relationships between entities; The output scanning plan is as follows: ； in, For the subject's historical scan sequence features and action tags, This is a scheme generation network based on a self-attention mechanism. The score indicates the swallowing function. Adjust the execution attitude vector based on the real-time target pose deviation. , represented as: ； in, For real-time target pose deviation, For the corresponding proportional, integral, and derivative control gains.

10. A robot system, comprising a controller, characterized in that, The controller performs the automated ultrasound scanning control method for swallowing function assessment as described in any one of claims 1-9.

Images