A medical intelligent teaching model construction method based on ultrasonic AI technology

Abstract

The application relates to the technical field of intelligent medical teaching, and discloses a medical intelligent teaching model construction method based on ultrasonic AI technology. The method collects ultrasonic image sequence data of a target object and corresponding operation records, generates an original teaching data set, carries out multi-dimensional teaching feature analysis on the original teaching data set, extracts a dynamic teaching feature parameter set, constructs an anatomical structure evolution feature tensor according to the parameter set and the time evolution law, calculates the knowledge density distribution of historical typical cases, sets a teaching anomaly discrimination boundary and generates an abnormal feature index set according to the knowledge density distribution, captures ultrasonic image flow and operation behavior data in real time, maps the ultrasonic image flow and the operation behavior data to a multi-scale teaching knowledge space, calculates the spatial distribution similarity of the abnormal feature index set, obtains a real-time teaching deviation coefficient, constructs a teaching risk prediction network model, and generates a teaching operation quality evaluation result and a teaching strategy adjustment scheme.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical teaching technology, specifically a method for constructing a medical intelligent teaching model based on ultrasound AI technology. Background Technology

[0002] In the medical field, ultrasound technology, due to its ease of operation and non-invasive nature, has become an important component of clinical diagnosis and teaching. The standardization of ultrasound operation and the ability to interpret images directly affect diagnostic accuracy; therefore, related teaching has always been a focus of medical education. Traditional ultrasound teaching largely relies on on-site guidance from senior physicians, with students gradually mastering skills through observation, demonstrations, simulations, and case discussions. However, this model is limited by the uneven distribution of teaching resources, resulting in significant differences in teaching quality across different regions and institutions, and some grassroots teaching units struggling to provide a sufficient number of experienced instructors.

[0003] In practical operations, the correspondence between the movement trajectory and angle adjustment of the ultrasound probe and anatomical structures is a significant learning challenge for students. In traditional teaching, instructors often struggle to capture subtle changes in students' movements in real time, and explanations of the dynamic relationship between the movement trajectory and anatomical structures rely heavily on verbal descriptions, making it difficult for students to intuitively understand. Furthermore, improper actions by students during operations often require correction afterward through video playback or physician recollection, hindering timely feedback and potentially leading to repeated errors.

[0004] Current teaching methods for analyzing ultrasound images primarily focus on static images, neglecting the dynamic evolution of the image sequence during the procedure and failing to demonstrate the real-time correlation between operative actions and changes in anatomical structures. Furthermore, the pathological characteristics of different cases vary, making it difficult for students to integrate operative procedures with specific pathological conditions, resulting in insufficient adaptability when facing complex cases. The experience gained from historical teaching cases is difficult to systematically integrate, preventing new students from effectively drawing upon typical problems and solutions from past teaching. This lack of a unified knowledge system undermines the speed and stability of students' skill acquisition. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing a medical intelligent teaching model based on ultrasound AI technology, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a method for constructing a medical intelligent teaching model based on ultrasound AI technology, the method comprising:

[0007] Collect ultrasound image sequence data and corresponding operation record information of the target medical teaching subjects to generate the original teaching dataset;

[0008] Perform multi-dimensional teaching feature parsing operations on the original teaching dataset to extract a dynamic teaching feature parameter set containing anatomical structure features, operation trajectory features, and pathological correlation features;

[0009] Based on the evolution of the dynamic teaching feature parameter set over time, an anatomical structure evolution feature tensor is constructed.

[0010] Based on the tensor of the anatomical structure evolution characteristics, a multi-scale teaching knowledge space is constructed, and the knowledge density distribution of typical historical teaching cases in the multi-scale teaching knowledge space is calculated.

[0011] Based on the knowledge density distribution, a boundary for judging teaching anomalies is set and a set of teaching anomaly characteristic indicators is generated.

[0012] Real-time capture of ultrasound image stream data and operational behavior data during medical teaching operations, and analysis to generate real-time teaching feature vectors;

[0013] The real-time teaching feature vector is mapped to the multi-scale teaching knowledge space, and its spatial distribution similarity with the set of teaching anomaly feature indicators is calculated to generate the real-time teaching deviation coefficient.

[0014] By combining the real-time teaching deviation coefficient, the dynamic teaching feature parameter set, and the anatomical structure evolution feature tensor, a teaching risk prediction network model is constructed to generate teaching operation quality assessment results and teaching strategy adjustment plans.

[0015] Preferably, the process of collecting ultrasound image sequence data and corresponding operation record information of the target medical teaching subjects to generate the original teaching dataset specifically includes:

[0016] Simultaneously acquire the image slice sequences output by the ultrasound equipment and the operation logs recorded by the teaching operation platform;

[0017] The image slice sequence is subjected to temporal domain calibration and invalid frame data that exceeds the preset imaging quality range is removed;

[0018] The time-domain calibrated image slice sequences and operation behavior logs are aligned across modal data based on timestamps to generate a structured teaching dataset containing image slice feature matrices, probe motion trajectory parameters, and operation force change curves.

[0019] Preferably, the step of performing multi-dimensional teaching feature parsing on the original teaching dataset to extract a dynamic teaching feature parameter set containing anatomical structure features, operational trajectory features, and pathological correlation features specifically includes:

[0020] Perform hierarchical feature extraction on the structured teaching dataset;

[0021] Establish a set of teaching feature parameters according to the dimensions of organ identification, operational procedures, and pathological discrimination;

[0022] Calculate the gradient value of the feature evolution of each parameter in the set of teaching feature parameters within the preset teaching cycle;

[0023] The gradient values ​​of the feature evolution are arranged in time sequence to form a three-dimensional anatomical structure evolution feature tensor.

[0024] The first dimension of the three-dimensional anatomical structure evolution feature tensor represents different teaching feature parameter categories, the second dimension represents the feature evolution gradient value sequence, and the third dimension represents the spatial location coordinates.

[0025] Preferably, the construction of a multi-scale teaching knowledge space based on the tensor of the anatomical structure evolution features specifically includes:

[0026] The dimensional architecture of the multi-scale teaching knowledge space is determined based on the number of categories of teaching feature parameters contained in the tensor of anatomical structure evolution features.

[0027] The coordinate axes of the multi-scale teaching knowledge space are set as different teaching feature parameter categories of the anatomical structure evolution feature tensor, and the coordinate points in the space are feature evolution gradient values.

[0028] Calculate the degree of spatial coordinate clustering of typical historical teaching cases within the multi-scale teaching knowledge space to generate a knowledge density distribution value.

[0029] Based on the knowledge density distribution value, a teaching anomaly detection boundary threshold is set, and the teaching feature parameters corresponding to the feature coordinate points that exceed the teaching anomaly detection boundary threshold are selected as the teaching anomaly feature index set.

[0030] Preferably, the step of parsing and generating real-time teaching feature vectors specifically includes:

[0031] It receives continuous image frame data streams transmitted by ultrasound image acquisition equipment and probe pose data captured by motion sensors in real time.

[0032] Real-time organ structure feature parameters are extracted from the continuous image frame data stream, and real-time operation trajectory parameters are parsed from the probe pose data;

[0033] Real-time teaching feature vectors are generated according to the parameter structure order of the dynamic teaching feature parameter set.

[0034] Preferably, the step of mapping the real-time teaching feature vector to the multi-scale teaching knowledge space and calculating its spatial distribution similarity with the set of teaching anomaly feature indicators specifically includes:

[0035] The real-time teaching feature vectors are projected onto a multi-scale teaching knowledge space;

[0036] Calculate the spatial distance distribution between the coordinate points of the real-time teaching feature vector and each feature coordinate point in the set of teaching anomaly feature indicators;

[0037] Calculate multidimensional spatial similarity metrics based on spatial distance distribution;

[0038] Real-time teaching deviation coefficients are generated based on the degree of deviation between the multidimensional spatial similarity metric and the preset similarity threshold.

[0039] Preferably, the construction of the teaching risk prediction network model specifically includes:

[0040] Input the real-time teaching deviation coefficient and the dynamic teaching characteristic parameter set into the teaching risk prediction network model;

[0041] The anatomical structure evolution feature tensor is used as the spatiotemporal correlation constraint condition of the teaching risk prediction network model;

[0042] The spatiotemporal convolution module is used to calculate the teaching operation quality assessment parameters.

[0043] The output includes the teaching operation quality assessment results, including the operation standard score, the accuracy of anatomical identification, and the pathological discrimination error.

[0044] Preferably, the generation of teaching operation quality assessment results and teaching strategy adjustment plans specifically includes:

[0045] Based on the difference between the operation standard score in the teaching operation quality assessment results and the preset teaching standard, the probe operation trajectory correction parameters are generated.

[0046] Based on abnormalities in the accuracy of anatomical recognition, adjust the sampling frequency of the anatomical structure feature extraction module;

[0047] Optimize the weight parameters of the pathological feature discrimination network based on the distribution characteristics of pathological discrimination error;

[0048] The aforementioned modified parameters are integrated to generate a dynamic teaching strategy adjustment plan.

[0049] Preferably, it also includes the steps for updating the teaching knowledge graph:

[0050] Obtain the correlation features between typical operational data in the historical teaching case library and real-time teaching feature vectors;

[0051] Calculate the new knowledge density distribution of the aforementioned associated features within the multi-scale teaching knowledge space;

[0052] When the difference between the new knowledge density distribution and the original knowledge density distribution exceeds the update threshold, the dimensional structure of the anatomical structure evolution feature tensor is reconstructed.

[0053] The set of teaching anomaly indicator values ​​is updated based on the tensor of the reconstructed anatomical structure evolution characteristics.

[0054] Preferably, the method further includes a dynamic adaptation step for teaching resources: receiving the teaching operation quality assessment results generated by the teaching risk prediction network model;

[0055] Analyze the abnormal areas in the operational standardization scores in the teaching operation quality assessment results;

[0056] Extract spatiotemporal feature fragments associated with the abnormal region from the tensor of anatomical structure evolution features;

[0057] Based on the aforementioned spatiotemporal feature fragments, a database of typical historical teaching cases is retrieved to generate a targeted teaching demonstration dataset;

[0058] The targeted teaching demonstration dataset is fused with real-time ultrasound image stream data at the feature layer to generate a dynamically enhanced teaching view.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] This method generates a raw teaching dataset by collecting ultrasound image sequence data and corresponding operation record information. It integrates the correlation information between images and operations during the teaching process, breaking the limitation of the separation of images and operation records in traditional teaching. This makes the teaching data more complete and can comprehensively reflect the key information in the teaching scenario. Multi-dimensional teaching feature analysis of operations extracts anatomical structural features, operation trajectory features, and pathological correlation features, systematically integrating the originally scattered teaching elements. This allows the presentation of anatomical structures, the trajectory of operational behaviors, and the correlation of pathological features during the teaching process to form an organic whole, avoiding the one-sidedness of teaching information caused by single feature analysis.

[0061] Based on the evolutionary pattern of dynamic teaching characteristic parameters over time, an anatomical structure evolution characteristic tensor is constructed. This tensor captures the dynamic changes of anatomical structures during manipulation, presenting the state of anatomical structures at different stages of manipulation. Students can intuitively perceive the temporal relationship between manipulation and changes in anatomical structures, and understand the impact of manipulation on image presentation. The multi-scale teaching knowledge space constructed based on this tensor provides a unified storage and analysis framework for typical historical teaching cases. Teaching experiences from different cases can form a quantifiable knowledge density distribution within this space, allowing for the systematic integration of effective past teaching experiences, rather than their fragmented existence.

[0062] By setting boundaries for identifying teaching anomalies and generating a set of characteristic indicators, a clear reference is provided for identifying non-standard operations and abnormal behaviors in the teaching process. This shifts the judgment of deviations in teaching from relying on subjective experience to analysis based on objective characteristics. Real-time capture and parsing of teaching operation data to generate real-time teaching feature vectors enables synchronous tracking of student behavior and video changes, achieving dynamic monitoring of the teaching process and avoiding the problem of delayed feedback in traditional teaching.

[0063] By mapping real-time teaching feature vectors to a multi-scale teaching knowledge space and calculating their spatial distribution similarity with the set of teaching anomaly feature indicators to generate a real-time teaching deviation coefficient, the degree of difference between student operations and standard norms can be quantified, making teaching evaluation more objective and accurate. A teaching risk prediction network model constructed by combining the real-time teaching deviation coefficient, dynamic teaching feature parameter set, and anatomical structure evolution feature tensor can comprehensively generate teaching operation quality evaluation results and teaching strategy adjustment plans based on multi-faceted information. This makes the evaluation results more closely reflect students' actual operation, and the adjustment plans can be flexibly adapted to students' real-time performance, accommodating different students' learning paces and weaknesses, reducing over-reliance on manual guidance, and making the teaching process more targeted and adaptable, covering a wider range of teaching scenarios and student needs. Attached Figure Description

[0064] Figure 1 This is a schematic diagram illustrating the working principle of the medical intelligent teaching model construction method based on ultrasound AI technology described in this invention.

[0065] Figure 2 A flowchart of the method for generating the original teaching dataset;

[0066] Figure 3 A flowchart for a method to generate feature vectors for real-time teaching;

[0067] Figure 4 A flowchart illustrating the method for constructing a network model for predicting teaching risks. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Please see Figure 1 This invention provides a method for constructing a medical intelligent teaching model based on ultrasound AI technology, the method comprising:

[0070] An original teaching dataset is generated by collecting ultrasound image sequence data and corresponding operation records of the target medical teaching subjects. Multi-dimensional teaching feature analysis is performed on this dataset to extract a dynamic teaching feature parameter set, constructing an anatomical structure evolution feature tensor and a multi-scale teaching knowledge space. The knowledge density distribution of historical typical teaching cases is calculated to set the boundary for identifying teaching anomalies. Real-time teaching operation data is captured and analyzed to generate real-time teaching feature vectors, which are then mapped to the multi-scale teaching knowledge space to calculate spatial distribution similarity and generate a real-time teaching deviation coefficient. A teaching risk prediction network model is constructed by combining the dynamic teaching feature parameter set and the anatomical structure evolution feature tensor, ultimately outputting teaching operation quality assessment results and teaching strategy adjustment plans.

[0071] Example 1: See Figure 2 This process involves collecting ultrasound image sequence data and corresponding operation logs from the target medical teaching subjects, as well as generating the original teaching dataset. This is achieved by simultaneously acquiring image slice sequences output by the ultrasound equipment and operation logs recorded by the teaching operation platform. The image slice sequences are generated by the ultrasound equipment during real-time scanning and contain continuous two-dimensional or three-dimensional medical image frames. The operation logs record parameters such as the probe's movement trajectory, applied force, and operation time during the teaching process. This data is transmitted in real-time through a dedicated data acquisition interface to ensure the integrity and timeliness of the information.

[0072] In the image slice sequence processing stage, a temporal calibration algorithm is used to align consecutive frame data. This algorithm marks each frame based on timestamp information and compensates for time deviations caused by device latency or transmission jitter using interpolation methods. The calibrated image sequence is then inspected by a quality assessment module. This module analyzes parameters such as sharpness, contrast, and signal-to-noise ratio of each frame, automatically removing invalid frame data that does not meet preset imaging quality standards. The criteria for determining invalid frames include, but are not limited to, image blurring, artifact interference, and signal loss. The retained valid image slices are arranged in chronological order to form a continuous image sequence dataset.

[0073] The processing of the operation behavior log includes the analysis of the probe motion trajectory and the generation of operation force curves. The probe motion trajectory is acquired through an inertial measurement unit (IMU) mounted on the ultrasonic probe, which includes sensors such as accelerometers, gyroscopes, and magnetometers, capable of recording the probe's position and attitude changes in real time. After filtering and noise reduction, the sensor data is converted into six-degree-of-freedom motion trajectory parameters. Operation force data is acquired through pressure sensors, recording the changes in pressure applied to the probe by the operator during the teaching process. These parameters are correlated with timestamp information to form a structured operation behavior record.

[0074] The cross-modal data alignment process matches time-domain calibrated image slice sequences with operation behavior logs. This process is based on a unified time reference and implemented using a timestamp alignment algorithm. The aligned dataset contains information in three dimensions: image slice feature matrix, probe motion trajectory parameters, and operation force variation curves. The image slice feature matrix stores the pixel data of each frame of the image and its corresponding spatial coordinate information. The probe motion trajectory parameters describe the position and attitude changes of the probe in three-dimensional space. The operation force variation curves reflect the force applied by the operator to the probe during the teaching process. These data together constitute the original teaching dataset, providing the foundation for subsequent feature analysis.

[0075] When performing multi-dimensional teaching feature analysis on the original teaching dataset, a hierarchical feature extraction method is used to mine teaching features from different dimensions. Feature extraction for organ recognition is achieved through a three-dimensional convolutional neural network. This network takes a continuous sequence of image slices as input and extracts anatomical features at different scales through multi-layer convolution and pooling operations. The network output includes key anatomical information such as organ boundaries, tissue texture, and blood flow signals. Feature extraction for operational standardization is based on probe motion trajectory and operational force data. A trajectory analysis algorithm calculates parameters such as probe motion speed, acceleration, and path smoothness. These parameters reflect the operator's technique and proficiency. Feature extraction for pathological discrimination employs a specialized pathological feature classifier. This classifier takes image slice data as input and outputs feature vectors related to specific pathological changes, such as morphological features of the lesion area or hemodynamic parameters.

[0076] The extracted feature parameters are dynamically analyzed within a preset teaching cycle, and their evolution gradient values ​​over time are calculated. These gradient values ​​reflect the dynamic trends in anatomical structure recognition, operational standardization, and pathological discrimination abilities during the teaching process. These gradient values ​​are arranged chronologically to form a three-dimensional anatomical structure evolution feature tensor. The first dimension of the feature tensor represents different categories of teaching feature parameters, including anatomical structure features, operational trajectory features, and pathological correlation features. The second dimension stores the evolution gradient values ​​of the feature parameters over time, reflecting the dynamic changes during the teaching process. The third dimension records the spatial coordinates of the feature parameters, ensuring the spatial correlation between anatomical structures and operational trajectories. The feature tensor is constructed using a tensor decomposition algorithm, compressing multidimensional feature data into a low-dimensional representation while preserving its main change patterns.

[0077] The generation of structured teaching datasets also includes data standardization and normalization. Data from different sources are scaled to eliminate data bias caused by differences in acquisition equipment or units. Standardized data is more suitable for subsequent machine learning and pattern recognition algorithms. The normalization process maps feature parameters to a uniform numerical range, preventing certain features from having an excessive impact on model training due to different units of measurement. The processed dataset is stored in a dedicated teaching database system, supporting efficient querying and real-time updates.

[0078] The dynamic update mechanism for teaching feature parameters ensures that the system can adapt to the needs of different teaching scenarios. When new teaching data is added, the system automatically triggers the feature parsing process to update the anatomical structure evolution feature tensor. The update process uses an incremental learning algorithm, gradually adjusting the feature parameters and evolution patterns without retraining the entire model. This mechanism enables the system to have continuous learning capabilities, allowing it to continuously optimize feature representations as teaching cases accumulate.

[0079] The results of multi-dimensional teaching feature analysis provide input for the subsequent construction of the teaching knowledge space. The combination of anatomical structural features, operational trajectory features, and pathological correlation features forms a comprehensive teaching feature representation, covering multiple key aspects of medical ultrasound teaching. These features not only reflect the current teaching status but also reveal dynamic changes in the teaching process through time series analysis. The automated design of the feature analysis process reduces manual intervention and improves the system's practicality and scalability.

[0080] The generation and feature parsing of the original teaching dataset involves various data processing algorithms and machine learning models. These algorithms and models run on a dedicated computing platform, utilizing parallel computing and distributed storage technologies to improve processing efficiency. The system's modular design allows for independent upgrades or replacements of individual functional components, adapting to technological advancements and changing teaching needs.

[0081] Example 2: See Figure 3 This paper describes the construction of a multi-scale teaching knowledge space and the generation of real-time teaching feature vectors. This process constructs a multi-dimensional teaching knowledge representation system based on the anatomical structure evolution feature tensor and achieves dynamic monitoring of the teaching status through real-time data stream parsing. The anatomical structure evolution feature tensor serves as the core data structure, with its three-dimensional architecture corresponding to teaching feature parameter categories, temporal evolution gradient sequences, and spatial coordinates, forming a complete feature system covering anatomical recognition, operational norms, and pathological discrimination.

[0082] The construction of a multi-scale teaching knowledge space begins with the determination of its dimensional architecture. The coordinate axis dimensions of the knowledge space are defined based on the number of teaching feature parameter categories contained in the anatomical structure evolution feature tensor. Each coordinate axis corresponds to a class of teaching feature parameters, such as organ boundary clarity, probe motion smoothness, or lesion region contrast. Coordinate points within the space are filled with feature evolution gradient values, reflecting the intensity of change of different teaching features over time. The scale partitioning of the knowledge space employs an adaptive clustering algorithm, dynamically adjusting the spatial resolution based on the distribution density of the feature parameters. High-density regions are partitioned using fine-grained scales to capture subtle differences in teaching patterns; low-density regions are partitioned using coarse-grained scales to reduce computational complexity.

[0083] The distribution analysis of typical historical teaching cases in the knowledge space employs a kernel density estimation algorithm. This algorithm calculates the clustering degree of spatial coordinate points using nonparametric statistical methods, generating a continuous knowledge density distribution surface. Peak regions on the density distribution surface represent common combinations of teaching characteristic parameters, corresponding to standardized operational techniques and accurate anatomical identification results. Density valley regions reflect rare feature combinations, potentially associated with operational errors or abnormal pathological manifestations. The density distribution is parameterized by fitting a Gaussian mixture model, extracting the positions and distribution ranges of the main density peaks. Based on the fitting results, a threshold for identifying teaching anomalies is set; regions outside the threshold are marked as abnormal feature spaces.

[0084] The generation process of the teaching anomaly feature index set traverses all coordinate points in the knowledge space, filtering out anomalies that exceed the discrimination boundary threshold. Each anomaly is associated with a set of teaching feature parameters and their evolution gradient values, such as sudden shifts in probe motion trajectories or blurring trends in organ boundaries. The anomaly index set is hierarchically stored according to feature category and anomaly degree, forming a structured anomaly knowledge base. This knowledge base supports rapid retrieval and matching, providing a reference standard for real-time teaching monitoring. The knowledge space update mechanism periodically recalculates the density distribution and discrimination boundary to adapt to the dynamic accumulation and pattern evolution of teaching data.

[0085] The generation of real-time teaching feature vectors relies on the collaborative operation of ultrasound image acquisition equipment and motion sensors. The continuous image frame data stream output by the ultrasound equipment undergoes real-time preprocessing, including denoising, enhancement, and normalization. The preprocessed frame data is input into an organ structure feature extraction network, which employs a lightweight design to achieve real-time inference with limited computational resources. The network output includes real-time anatomical information such as organ boundary coordinates, tissue texture features, and hemodynamic parameters. The probe pose data captured by the motion sensor undergoes coordinate transformation and filtering, converting it into a six-degree-of-freedom spatial motion trajectory. The trajectory data is further analyzed into operational feature parameters such as velocity, acceleration, and path curvature.

[0086] The assembly of real-time teaching feature vectors follows the structural specifications of dynamic teaching feature parameter sets. Vector elements are arranged in a fixed order, successively containing dimensions such as anatomical structural features, operational trajectory features, and pathological correlation features. The feature values ​​of each dimension are normalized to eliminate the influence of dimensional differences on spatial mapping. The vector generation module adopts a pipelined architecture to achieve low-latency processing from raw data acquisition to feature vector output. A data quality check is included in the processing to discard temporary data points with insufficient signal-to-noise ratio or obvious anomalies, maintaining the reliability of the feature vectors.

[0087] The interaction between the multi-scale teaching knowledge space and real-time feature vectors is achieved through a spatial projection algorithm. The projection process considers the multidimensional characteristics of the knowledge space, employing an orthogonal decomposition method to map real-time feature vectors onto each coordinate axis. Spatial relationship analysis is performed between the projected coordinate points and the historical case distribution, calculating their distance to the density peak region. An improved Mahalanobis distance algorithm is used for distance measurement, which considers the correlation between different feature dimensions and avoids the limitations of simple Euclidean distance. Spatial similarity analysis not only focuses on absolute distance values ​​but also examines the relative positional relationships of real-time points on the density distribution surface.

[0088] The calculation of the real-time teaching deviation coefficient integrates two factors: spatial distance and distribution density. The coefficient value reflects the overall degree of deviation between the current teaching status and historical typical patterns, with the numerical range normalized to the interval of 0 to 1. The coefficient calculation module adopts a sliding window mechanism to analyze the deviation trends of multiple recent feature vectors, avoiding misjudgments based on single-point fluctuations. The dynamic change curve of the deviation coefficient serves as an important indicator of teaching quality; a sharp increase or a sustained high level triggers subsequent risk prediction processes.

[0089] The visualization function of the knowledge space transforms multidimensional feature distributions into two-dimensional or three-dimensional projections for intuitive understanding by teaching supervisors. The visualization system employs color coding and dynamic marking technology to highlight the location of real-time feature vectors and their relative relationship to abnormal areas. Interactive exploration tools allow users to rotate and zoom the knowledge space view, analyzing teaching feature distribution patterns from different angles. Visualized data is presented synchronously with real-time ultrasound images, forming a multimodal teaching status monitoring interface.

[0090] The implementation of a multi-scale teaching knowledge space requires the support of a high-performance computing platform. The computation and updating tasks of knowledge density distribution are allocated to distributed computing nodes, leveraging parallel processing to accelerate the statistical analysis of large-scale data. A real-time feature vector generation module is deployed on edge computing devices, close to the data acquisition source to reduce transmission latency. The system adopts a microservice architecture, with each functional module deployed and expanded independently to adapt to the needs of teaching scenarios of different scales. The data storage solution combines relational databases and time-series databases to optimize the management efficiency of structured feature data and streaming teaching records.

[0091] The dynamic maintenance mechanism of the teaching knowledge space periodically evaluates the completeness of the feature representation. When new teaching data causes a significant change in the distribution of knowledge density, a space reconstruction process is triggered. The reconstruction process adjusts the coordinate axis direction and scale division through methods such as principal component analysis, ensuring that the knowledge space always maintains a sensitive representation of the teaching model. The version control mechanism records the iteration history of the knowledge space, supporting rollback to a previous stable state when necessary. The entire system design emphasizes robustness and adaptability, enabling continuous optimization of knowledge representation capabilities in ongoing teaching practice.

[0092] The quality of real-time teaching feature vector generation and the accuracy of knowledge space construction directly affect the accuracy of subsequent teaching evaluation. The system employs a multi-level verification mechanism to ensure the reliability of the data processing flow, including input data range checks, feature extraction reproducibility verification, and inverse transformation testing of spatial mapping. Quality control logs record the operational status and abnormal events of each stage, providing a basis for system optimization. During implementation, emphasis is placed on matching computational efficiency with the actual teaching pace to ensure that real-time feedback does not introduce significant operational delays.

[0093] The multi-scale teaching knowledge space, serving as a core medium connecting historical experience with real-time teaching, is constructed in a way that balances mathematical rigor with the professional characteristics of medical teaching. The selection of spatial dimensions covers standard teaching assessment elements while retaining the ability to expand to adapt to new teaching features. Dynamic updates to the density distribution allow the system to adjust assessment criteria as teaching levels improve, avoiding insufficient adaptability caused by fixed thresholds. The hierarchical organization of the abnormal feature indicator set supports differentiated teaching intervention strategies, providing corresponding guidance for abnormal patterns of different degrees and types.

[0094] The real-time data processing pipeline is designed to balance computational accuracy and response speed. The feature extraction algorithm is appropriately simplified while ensuring no key information is lost, meeting real-time constraints. The motion sensor data fusion algorithm compensates for the time asynchrony caused by different sampling rates, ensuring the spatiotemporal consistency between the operation trajectory and the ultrasound image. The feature vector standardization process uses a dynamic benchmark to adapt to the parameter variations of different teaching subjects and scanning sites. The system's runtime resource monitoring module dynamically adjusts the priority of computational tasks to ensure the processing capacity of the critical path.

[0095] The integration of a teaching knowledge space with real-time monitoring functions creates a closed-loop teaching quality management system. Typical patterns accumulated during the space construction phase provide a benchmark for real-time assessment, while real-time data continuously feeds back to optimize the spatial representation. This two-way interaction enables the system to continuously evolve, gradually establishing an assessment system tailored to specific teaching environments and student groups. During implementation, the professional standards and safety requirements of medical education were fully considered, and all algorithm designs and parameter settings were reviewed and confirmed by clinical teaching experts.

[0096] Example 3: See Figure 4 This paper describes the mapping and analysis of real-time teaching feature vectors in a multi-scale teaching knowledge space, and the construction process of a teaching risk prediction network model. This process achieves dynamic evaluation of teaching quality and risk warning through spatial similarity calculation and spatiotemporal feature fusion. Real-time teaching feature vectors are derived from real-time analysis of ultrasound image streams and operational behavior data, encompassing dimensions such as anatomical structural features, operational trajectory features, and pathological correlation features. The multi-scale teaching knowledge space, as an abstract representation of historical teaching experience, has coordinate axes corresponding to different categories of teaching feature parameters, and contains a set of feature points from typical historical cases.

[0097] The spatial mapping of real-time teaching feature vectors employs an orthogonal projection algorithm, decomposing high-dimensional feature vectors onto various coordinate axes of the knowledge space. The projection process preserves the relative weights of the vectors in each feature dimension, ensuring that the mapped coordinates accurately reflect the comparison between the current teaching state and historical patterns. A dimension weighting factor is introduced into the projection calculation to adjust the degree of influence of different feature categories on the final spatial location. This weighting factor is dynamically configured according to clinical teaching standards; for example, the weight of anatomical structure features is increased during organ recognition training, while the consideration of operational trajectory features is emphasized during operational skills assessment.

[0098] Spatial similarity measurement is achieved by calculating the distribution relationship between real-time projected points and points in the set of teaching anomaly feature indicators. The set of anomaly feature indicators is stored in a specific region of the multi-scale teaching knowledge space, containing feature points historically marked as operational errors or misidentifications. The similarity measurement considers both spatial distance and density distribution, employing an improved distance function to assess the proximity between real-time points and anomaly regions. This distance function is defined as:

[0099]

[0100] in, Represents the regularized distance value. For the real-time feature vector at the th dimensional projection coordinates, The set of abnormal feature indicators represents the set of indicators in the first place. The mean of the dimension, The standard deviation of the corresponding dimension. It is a small constant to prevent division by zero. This distance metric takes into account the distribution differences of different feature dimensions, relaxing the distance requirement on dimensions with large variance and strictly matching on dimensions with concentrated distributions.

[0101] The real-time teaching deviation coefficient is generated based on a non-linear transformation of regularized distance values. The transformation function uses a sigmoid curve to map the distance values ​​to a coefficient range between 0 and 1. A coefficient value close to 1 indicates that the current teaching state is highly similar to historical abnormal patterns, posing a significant operational risk; a coefficient value close to 0 reflects compliant teaching practices. The deviation coefficient is calculated using a sliding time window mechanism, analyzing the average deviation trend of multiple recent feature vectors to avoid misjudgments caused by instantaneous fluctuations. The window size is dynamically adjusted according to the teaching stage; a shorter response window is used for basic skills training, while a longer observation period is used for complex case diagnosis.

[0102] The architecture of the teaching risk prediction network model is a spatiotemporal hybrid neural network. The network input layer receives a combination of real-time teaching deviation coefficients and dynamic teaching feature parameter sets. The dynamic teaching feature parameter set contains all feature parameters extracted from the current teaching stage and their evolution gradients, reflecting the dynamic changes in the teaching process. The network hidden layers contain a hybrid structure of 3D convolutional modules and long short-term memory modules, respectively handling spatial feature associations and temporal series patterns. The kernel function parameters of the 3D convolutional modules are initialized from the anatomical structure evolution feature tensor, maintaining continuity with historical teaching knowledge. The long short-term memory module tracks the temporal evolution trend of teaching features, capturing the formation process of operational habits.

[0103] The anatomical structure evolution feature tensor serves as a constraint for network training, influencing feature selection in the hidden layers through an attention mechanism. Attention weights calculate the degree of matching between current teaching features and historical typical patterns, guiding the network to focus on feature dimensions with significant differences. The network output layer contains three parallel branches, generating predicted values ​​for operational standardization score, anatomical recognition accuracy, and pathological discrimination error, respectively. The operational standardization score is calculated based on the smoothness of the probe's motion trajectory and the preservation of standard sections; the anatomical recognition accuracy assesses the localization precision of the target organ's boundary; and the pathological discrimination error reflects the specificity and sensitivity of abnormal region identification. The three output values ​​are normalized and then combined into a comprehensive teaching quality evaluation result.

[0104] The network training process employs a two-stage strategy: offline pre-training based on historical teaching data, followed by online fine-tuning using real-time feedback. The offline stage optimizes initial network parameters using a large-scale historical case library, while the online stage adapts to specific teaching scenarios and student characteristics through incremental learning. The loss function is designed as a multi-task weighted combination to balance the training signal strength of different evaluation metrics. Regularization constraints prevent the model from overfitting historical data, maintaining its adaptability to new teaching models. The update frequency of network parameters is adjusted according to the teaching progress, decreasing the update frequency during stable phases and increasing the adjustment frequency during periods of rapid change.

[0105] The real-time evaluation result generation process integrates spatial similarity analysis and network prediction output. Teaching segments with deviation coefficients exceeding a threshold trigger detailed evaluations, with the network conducting in-depth analysis of the teaching characteristics of the relevant time periods. Evaluation results comprise both quantitative scores and qualitative descriptions. Quantitative scores are used for longitudinal comparison and progress tracking, while qualitative descriptions identify specific problem types and improvement suggestions. The result output interface supports multiple formats, including visual dashboards, structured reports, and real-time voice prompts, adapting to the feedback needs of different teaching scenarios.

[0106] The teaching risk prediction system is implemented based on a distributed computing architecture. The real-time analysis module is deployed on edge computing nodes, close to the data source to reduce transmission latency; batch training tasks run on a cloud computing platform, utilizing elastic resources to handle large-scale computations. The data pipeline adopts a stream-batch integrated design, with real-time data streams directly used for teaching monitoring, while persistent storage is used for model optimization. The system resource manager dynamically allocates computing tasks, prioritizing the processing capacity of the real-time evaluation link. The status monitoring component tracks the operating metrics of each module and automatically triggers a recovery mechanism when performance degrades.

[0107] The dynamic maintenance of the multi-scale teaching knowledge space and the risk prediction model form a collaborative optimization loop. Regular updates to the spatial structure provide the model with more accurate feature representations, and the model's predictions, in turn, guide the re-division of spatial regions. This collaborative mechanism enables the system to continuously track improvements in teaching quality and constantly adjust evaluation criteria. Strict version management of the space and model ensures the consistency of historical data analysis.

[0108] The real-time feedback mechanism is designed with the psychological impact of the teaching process in mind. The visualization of deviation coefficients employs a progressive warning strategy: minor deviations are displayed as suggestive markers, while only serious anomalies trigger an interruption alarm. The tone and wording of the voice guidance content are specially designed to avoid putting excessive pressure on learners. Feedback timing is chosen in conjunction with the teaching pace, providing guidance during breaks rather than critical steps to minimize distractions.

[0109] During system integration, rigorous synchronization of multimodal data is performed. Timestamps for ultrasound image streams, probe motion data, and evaluation results are aligned to the millisecond level to ensure precise correspondence between feedback information and specific operations. The synchronization mechanism employs a hardware-level time source to avoid deviations caused by software clock drift. Data latency monitoring displays the transmission status of each channel in real time, automatically activating compensation strategies when thresholds are exceeded.

[0110] Example 4: This example focuses on generating teaching operation quality assessment results and dynamically adjusting teaching strategies, combined with a continuous updating mechanism for the teaching knowledge graph, to form a closed-loop teaching optimization system. This process uses real-time teaching deviation coefficients and dynamic teaching characteristic parameter sets as inputs, generates targeted teaching improvement plans through multi-dimensional analysis, and maintains the system's adaptability through knowledge graph evolution. The implementation process is illustrated below using a teaching case study of liver ultrasound scanning.

[0111] During the teaching operation quality assessment phase, the system receives raw assessment data output from the risk prediction network model. Taking the long-axis section scan of the portal vein of the liver as an example, the assessment data includes indicators in three dimensions: probe operation trajectory score, accuracy of liver lobe anatomy identification, and potential lesion discrimination error. After standardization, these indicators are compared with preset teaching standard values ​​for difference analysis. The difference analysis adopts a grading strategy, classifying the degree of deviation into three levels: slight, moderate, and severe. The results of the difference analysis for the liver scan case are shown in the table below.

[0112] Table 1: Results of the differential assessment of liver scan cases are shown in the table below.

[0113]

[0114] Based on the evaluation results shown in the table, the system initiates the process of generating a teaching strategy adjustment plan. For moderate deviations in the probe operation trajectory, the trajectory correction module analyzes relevant spatiotemporal feature segments in the tensor of anatomical structure evolution characteristics. These segments show regular fluctuations in the probe rotation angle when the trainee maintains the long axis section of the portal vein. The correction plan generates a two-stage adjustment strategy: initially increasing the display density of virtual guide lines, and later gradually transitioning to a keyframe prompt mode. Specific parameters are set to add a dynamic guide line every 15 seconds, and once the trajectory stability improves to 80 points, the standard section image is prompted every 30 seconds.

[0115] Slight deviations in liver lobe anatomical identification triggered a sampling frequency adjustment mechanism. The system detected that the caudate lobe boundary recognition problem mainly occurred at scanning depths exceeding 8 cm, where image resolution decreased, making feature extraction difficult. The strategy adjustment increased the sampling frequency in areas deeper than 7 cm from 15 Hz to 20 Hz, while simultaneously enhancing contrast processing of deep tissues. The adjusted parameters were sent to the ultrasound simulation equipment in real time via the teaching control interface and took effect immediately in subsequent scans.

[0116] Severe biases in lesion discrimination necessitated weight optimization of the pathological feature discrimination network. The system retrieved 20 similar small cyst identification cases from its historical teaching case library. The feature patterns of these cases closely matched the current learners' misjudgments, exhibiting a common feature confusion pattern. Network optimization employed a transfer learning strategy, adjusting the weight distribution of the last three convolutional kernels while maintaining the core network structure. The optimized network demonstrated improved sensitivity to enhancement features at the edge of small cysts and enhanced suppression of vascular cross-section artifacts on test data.

[0117] The knowledge graph update process and strategy adjustments are carried out simultaneously. The system records the feature data of this liver scan case, including abnormal operation trajectory patterns, deep-dependent anatomical recognition problems, and specific lesion discrimination errors. After anonymization, this data is analyzed for feature association with a historical database of typical cases. The association analysis uses a graph embedding algorithm to map case features to a high-dimensional space to calculate similarity. When the feature combination of a new case is more than a threshold away from the core region of the original knowledge density distribution, knowledge graph reconstruction is triggered.

[0118] The knowledge graph reconstruction process begins with a dimensionality analysis of the anatomical structure evolution feature tensor. In the liver scan case, the system found a significant increase in the correlation between depth parameters and image quality features, which was not fully represented in historical data. The reconstruction algorithm adds a depth compensation dimension to the feature tensor and adjusts the representation of the original spatial coordinates. The updated feature tensor more accurately reflects the changes in anatomical structures at different scanning depths, providing a more refined reference framework for subsequent teaching assessments.

[0119] The update of the teaching anomaly feature index set adopts a gradual strategy. Newly identified anomaly patterns first enter an observation period, and are only formally included in the anomaly index set when they recur in three consecutive teaching cases. The pattern of "blurred caudate lobe boundary at a depth exceeding 8 cm" found in the liver scan case was added as a new anomaly feature index because it appeared in both the current case and two historical cases. The index description includes metadata such as depth threshold, degree of blurring, and associated anatomical structures, supporting precise matching and similarity calculation.

[0120] The effectiveness of the dynamic teaching strategy adjustment scheme was verified through a closed-loop monitoring mechanism. After applying the probe trajectory correction parameters, the system continuously tracked the trajectory stability index for the subsequent five scanning cycles. Data showed that the trajectory score gradually increased from 76.5 to 82.3, proving the effectiveness of the adjustment strategy. Similarly, the anatomical recognition accuracy after the sampling frequency adjustment increased from 88.2 to 89.7 within three teaching cycles, approaching the preset standard value. This monitoring data, in turn, served as feedback signals for strategy optimization, forming a virtuous cycle of continuous improvement.

[0121] The system employs differentiated update strategies tailored to the characteristics of different teaching stages. During the basic skills training phase, knowledge graph updates focus on supplementing operational standardization features; during the advanced diagnostic training phase, the optimization of pathological discrimination features is emphasized. The liver scan teaching case is in the intermediate transition phase, and the system automatically balances the update weights of operational skills and anatomical recognition to avoid one-sided emphasis. This adaptive mechanism ensures that the teaching strategy adjustments align with the learners' current developmental level.

[0122] Example 5: Focusing on a dynamic adaptation mechanism for teaching resources, this method generates targeted teaching demonstration data by analyzing the results of teaching operation quality assessments in real time and intelligently integrates it with real-time ultrasound image streams, ultimately outputting a dynamically enhanced teaching view. This process is based on in-depth analysis of the tensor of anatomical structure evolution characteristics, combined with intelligent retrieval technology from a database of historical typical teaching cases, to achieve precise matching of teaching resources with the current operational status of trainees.

[0123] The dynamic adaptation process for teaching resources begins with the reception and parsing of the quality assessment results output by the teaching risk prediction network model. The system extracts multi-dimensional features from the assessment results, focusing on abnormal areas in the operational procedure scores. Taking thyroid ultrasound scanning instruction as an example, when the assessment results show a significant deviation in the probe pressure control dimension, the system automatically marks the abnormal feature area and initiates the associated feature retrieval process. Feature retrieval is based on the timestamp alignment principle, establishing a mapping relationship between the abnormal operation period and the corresponding spatiotemporal segment in the anatomical structure evolution feature tensor.

[0124] The analysis of the anatomical structure evolution feature tensor employs a hierarchical feature matching algorithm. The system first spatially locates the extent of the anatomical structures affected by abnormal manipulation, identifying the local area of ​​the thyroid capsule as the key region of interest. Subsequently, it analyzes the evolution patterns of feature parameters in this region over time, identifying the curves showing changes in tissue deformation characteristics caused by abnormal pressure. This spatiotemporal correlation analysis precisely defines the target scope for adapting teaching resources, avoiding interference from irrelevant areas.

[0125] The historical typical teaching case database employs a multi-level screening strategy. The first level of screening is based on anatomical similarity, selecting all thyroid scan cases from the database. The second level incorporates operational characteristic constraints, retaining cases that include standardized pressure control records. The third level considers the trainee's current operating environment parameters, matching cases with similar probe models and ultrasound equipment settings. The candidate case set after these three levels of screening typically comprises 5-8 of the most valuable exemplary cases.

[0126] The generation process of the targeted teaching demonstration dataset integrates the essence of multiple excellent case studies. The system performs feature clustering on the stress control operation segments in the candidate cases, extracting common ideal stress curve patterns. Simultaneously, it retains the unique operational skill details of each case, forming a multi-level demonstration dataset containing basic standards and advanced techniques. The demonstration dataset uses a timeline labeling method, marking the key control points and common error forms of each operation step, providing a time alignment reference for subsequent real-time fusion.

[0127] Feature extraction from real-time ultrasound image stream data is kept synchronized with the demonstration data. The system establishes a dual-channel processing pipeline: the main channel processes the real-time image stream, while the secondary channel parses the demonstration dataset. Both channels share the same feature extraction algorithm and parameter settings, ensuring consistency in feature representation. In the thyroid scan case, features such as capsule clarity, blood flow signal intensity, and glandular echo uniformity are extracted simultaneously through both channels, establishing a basis for comparability between real-time operation and the ideal demonstration.

[0128] The feature layer fusion process employs an adaptive weight allocation mechanism. The system dynamically adjusts the fusion weights based on the degree of difference between real-time operation and demonstration data. When real-time operation approaches the standard, the display intensity of the demonstration data is reduced; as the deviation increases, the visual salience of the demonstration guidance is gradually enhanced. The fusion algorithm considers ergonomic principles to ensure that the enhanced view does not cause visual fatigue or distraction. In the thyroid scan stress control scenario, demonstration data is displayed as a semi-transparent outline, with its transparency inversely proportional to the stress deviation value.

[0129] The dynamically enhanced teaching view rendering combines multiple visualization technologies. Standardized operation trajectories are presented as dynamic guide lines, while abnormal operation areas are marked with pulsed highlighted borders. Anatomical structure comparison uses split-screen slider technology, allowing trainees to manually adjust the contrast between real-time images and the demonstration standard. Pathological feature discrimination auxiliary information is displayed as labeled bubbles, popping up only when the probe remains stationary beyond a set threshold to avoid interfering with the normal scanning rhythm. The color coding of all visual elements conforms to medical image display standards, with important warning information using orange-red tones and general prompts using blue-green tones.

[0130] The interactive view design prioritizes ease of use during the teaching process. Students can switch between different demonstration perspectives and rotate the view to see the 3D operation path using gestures. Key and difficult steps can be replayed in a loop with adjustable playback speed. Questions can be marked in real-time, and the system automatically retrieves relevant micro-lecture video clips. These interactive features maintain the main ultrasound scan workflow without disruption while providing ample learning support.

[0131] The system employs an intelligent load balancing strategy for resource management during runtime. Real-time image processing threads enjoy the highest priority, ensuring a smooth scanning process. Demonstration data retrieval and fusion computing tasks are dynamically scheduled based on CPU load, automatically reducing the accuracy requirements of non-critical tasks when computing resources are strained. Memory management uses a recently used priority retention strategy, maintaining a high-speed cache of frequently used demonstration data. This resource allocation method ensures system responsiveness even under limited hardware conditions.

[0132] The effectiveness of dynamic adaptation of teaching resources is evaluated using an indirect measurement method. The system records the improvement in students' operations before and after using the enhanced view, and statistically analyzes the trends of key indicators. Evaluation indicators include the duration of abnormal operations, the speed of achieving standard cross-sections, and the fluctuation range of operation parameters. This data is used to optimize adaptation strategy parameters, such as adjusting the timing and intensity of demonstration data display. The evaluation process does not interrupt the normal teaching flow, and all data collection is completed silently in the background.

[0133] An anomaly handling mechanism ensures system robustness. When severe artifacts or interruptions occur in the real-time video stream, the system automatically switches to the previous valid frame in the cache to continue analysis. In the event of a timeout in the demonstration data retrieval, a simplified version of the general demonstration mode is activated. All anomalies are logged in the system log, including the time of occurrence, duration, and countermeasures taken, for subsequent analysis and improvement.

[0134] The system's deployment environment adaptability design supports various teaching scenarios. In standard ultrasound teaching laboratories, it fully utilizes hardware features such as large touchscreens and force feedback probes; in mobile teaching scenarios, the interface layout is optimized for tablet displays. When network connectivity is unstable, the system automatically activates locally cached core demonstration datasets. This flexibility expands the system's applicability, providing support from formal classroom teaching to bedside instruction.

[0135] Privacy protection measures are implemented throughout the entire process. All teaching case data is anonymized before being stored in the database, removing any personally identifiable information. Real-time video stream data is encrypted and temporarily stored in memory, and then immediately deleted after processing. System access is strictly controlled, and students can only view demonstration data relevant to the current course. These measures comply with medical data security standards and protect the privacy rights of patients and students.

[0136] A knowledge update mechanism keeps teaching resources up-to-date. The system regularly checks the consistency of the demonstration dataset with the latest clinical guidelines and marks content that needs updating. New teaching cases, after expert review, may have their best operational segments adopted as new demonstration standards. This continuous evolution capability ensures that the system always reflects the forefront of current medical practice.

[0137] The user feedback channel is designed to facilitate continuous system improvement. A convenient problem reporting portal is provided during the teaching process, allowing students to quickly mark unsatisfactory adaptation effects. The teacher's side provides a sample data evaluation tool to rate and annotate the system-recommended cases. Feedback data, after cluster analysis, is used to optimize retrieval algorithms and fusion strategies, forming a user-participatory system improvement cycle.

[0138] The integration of a multimodal prompting system enhances teaching effectiveness. In addition to enhanced visual views, the system can provide voice prompts and tactile feedback, depending on the configuration. In the thyroid scan case, a slight vibration at a specific frequency is emitted when the probe pressure approaches the ideal range; a confirmation sound is played when the model standard is perfectly matched. This multisensory feedback helps trainees build muscle memory and accelerates the internalization of correct operating patterns.

[0139] System performance optimization employs a progressive strategy. The initial version focuses on the reliable implementation of core functions to ensure basic teaching needs are met. Subsequent iterations gradually add advanced features, such as 3D anatomical structure tracking and multi-student collaborative training. Thorough compatibility testing is conducted before each version update to ensure seamless integration with existing teaching progress. This robust development strategy enables the system to continue functioning effectively in actual teaching.

[0140] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0141] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for constructing a medical intelligent teaching model based on ultrasound AI technology, characterized in that, Includes the following steps: Collect ultrasound image sequence data and corresponding operation record information of the target medical teaching subjects to generate the original teaching dataset; Perform multi-dimensional teaching feature parsing operations on the original teaching dataset to extract a dynamic teaching feature parameter set containing anatomical structure features, operation trajectory features, and pathological correlation features; Based on the evolution of the dynamic teaching feature parameter set over time, an anatomical structure evolution feature tensor is constructed. Based on the tensor of the anatomical structure evolution characteristics, a multi-scale teaching knowledge space is constructed, and the knowledge density distribution of typical historical teaching cases in the multi-scale teaching knowledge space is calculated. Based on the knowledge density distribution, a boundary for identifying teaching anomalies is set, and a set of teaching anomaly characteristic indicators is generated. Real-time capture of ultrasound image stream data and operational behavior data during medical teaching operations, and analysis to generate real-time teaching feature vectors; The real-time teaching feature vector is mapped to the multi-scale teaching knowledge space, and its spatial distribution similarity with the set of teaching anomaly feature indicators is calculated to generate the real-time teaching deviation coefficient. By combining the real-time teaching deviation coefficient, the dynamic teaching feature parameter set, and the anatomical structure evolution feature tensor, a teaching risk prediction network model is constructed to generate teaching operation quality assessment results and teaching strategy adjustment schemes. The construction of a multi-scale teaching knowledge space based on the tensor of the evolutionary features of the anatomical structure specifically includes: The dimensional architecture of the multi-scale teaching knowledge space is determined based on the number of categories of teaching feature parameters contained in the tensor of anatomical structure evolution features. The coordinate axes of the multi-scale teaching knowledge space are set as different teaching feature parameter categories of the anatomical structure evolution feature tensor, and the coordinate points in the space are feature evolution gradient values. Calculate the degree of spatial coordinate clustering of typical historical teaching cases within the multi-scale teaching knowledge space to generate a knowledge density distribution value. Based on the knowledge density distribution value, a teaching anomaly detection boundary threshold is set, and the teaching feature parameters corresponding to the feature coordinate points that exceed the teaching anomaly detection boundary threshold are selected as the teaching anomaly feature index set.

2. The method for constructing a medical intelligent teaching model based on ultrasound AI technology according to claim 1, characterized in that, The process of collecting ultrasound image sequence data and corresponding operation record information from the target medical teaching subjects to generate the original teaching dataset specifically includes: Simultaneously acquire the image slice sequences output by the ultrasound equipment and the operation logs recorded by the teaching operation platform; The image slice sequence is subjected to temporal domain calibration and invalid frame data that exceeds the preset imaging quality range is removed; The time-domain calibrated image slice sequences and operation behavior logs are aligned across modal data based on timestamps to generate an original teaching dataset containing image slice feature matrices, probe motion trajectory parameters, and operation force change curves.

3. The method for constructing a medical intelligent teaching model based on ultrasound AI technology according to claim 2, characterized in that, The step of performing multi-dimensional teaching feature parsing on the original teaching dataset to extract a dynamic teaching feature parameter set containing anatomical structure features, operational trajectory features, and pathological correlation features specifically includes: Perform hierarchical feature extraction on the original teaching dataset; Establish a set of teaching characteristic parameters according to the dimensions of organ identification, operational procedures, and pathological discrimination; Calculate the gradient value of the feature evolution of each parameter in the set of teaching feature parameters within the preset teaching cycle; The gradient values ​​of the feature evolution are arranged in time sequence to form a three-dimensional anatomical structure evolution feature tensor. The first dimension of the three-dimensional anatomical structure evolution feature tensor represents different teaching feature parameter categories, the second dimension represents the feature evolution gradient value sequence, and the third dimension represents the spatial location coordinates.

4. The method for constructing a medical intelligent teaching model based on ultrasound AI technology according to claim 1, characterized in that, The process of generating real-time teaching feature vectors through parsing specifically includes: It receives continuous image frame data streams transmitted by ultrasound image acquisition equipment and probe pose data captured by motion sensors in real time. Real-time organ structure feature parameters are extracted from the continuous image frame data stream, and real-time operation trajectory parameters are parsed from the probe pose data; Real-time teaching feature vectors are generated according to the parameter structure order of the dynamic teaching feature parameter set.

5. The method for constructing a medical intelligent teaching model based on ultrasound AI technology according to claim 4, characterized in that, The step of mapping the real-time teaching feature vector to the multi-scale teaching knowledge space and calculating its spatial distribution similarity with the set of teaching anomaly feature indicators specifically includes: The real-time teaching feature vectors are projected onto a multi-scale teaching knowledge space; Calculate the spatial distance distribution between the coordinate points of the real-time teaching feature vector and each feature coordinate point in the set of teaching anomaly feature indicators; Calculate multidimensional spatial similarity metrics based on spatial distance distribution; Real-time teaching deviation coefficients are generated based on the degree of deviation between the multidimensional spatial similarity metric and the preset similarity threshold.

6. The method for constructing a medical intelligent teaching model based on ultrasound AI technology according to claim 5, characterized in that, The construction of the teaching risk prediction network model specifically includes: Input the real-time teaching deviation coefficient and the dynamic teaching characteristic parameter set into the teaching risk prediction network model; The anatomical structure evolution feature tensor is used as the spatiotemporal correlation constraint condition of the teaching risk prediction network model; The spatiotemporal convolution module is used to calculate the teaching operation quality assessment parameters. The output includes the teaching operation quality assessment results, including the operation standard score, the accuracy of anatomical identification, and the pathological discrimination error.

7. The method for constructing a medical intelligent teaching model based on ultrasound AI technology according to claim 6, characterized in that, The generation of teaching operation quality assessment results and teaching strategy adjustment plans specifically include: Based on the difference between the operation standard score in the teaching operation quality assessment results and the preset teaching standard, the probe operation trajectory correction parameters are generated. Based on abnormalities in the accuracy of anatomical recognition, adjust the sampling frequency of the anatomical structure feature extraction module; Optimize the weight parameters of the pathological feature discrimination network based on the distribution characteristics of pathological discrimination error; The aforementioned modified parameters are integrated to generate a dynamic teaching strategy adjustment plan.

8. The method for constructing a medical intelligent teaching model based on ultrasound AI technology according to claim 7, characterized in that, It also includes the steps for updating the teaching knowledge graph: Obtain the correlation features between typical operational data in the historical teaching case library and real-time teaching feature vectors; Calculate the new knowledge density distribution of the aforementioned associated features within the multi-scale teaching knowledge space; When the difference between the new knowledge density distribution and the original knowledge density distribution exceeds the update threshold, the dimensional structure of the anatomical structure evolution feature tensor is reconstructed. The set of teaching anomaly indicator values ​​is updated based on the tensor of the reconstructed anatomical structure evolution characteristics.

9. The method for constructing a medical intelligent teaching model based on ultrasound AI technology according to claim 1, characterized in that, The method also includes a dynamic adaptation step for teaching resources: receiving the teaching operation quality assessment results generated by the teaching risk prediction network model; Analyze the abnormal areas in the operational standardization scores in the teaching operation quality assessment results; Extract spatiotemporal feature fragments associated with the abnormal region from the tensor of anatomical structure evolution features; Based on the aforementioned spatiotemporal feature fragments, a database of typical historical teaching cases is retrieved to generate a targeted teaching demonstration dataset; The targeted teaching demonstration dataset is fused with real-time ultrasound image stream data at the feature layer to generate a dynamically enhanced teaching view.

Images